Datasets:

Saeed

/

TissueGPT_dataset

like 0

10. 1001/jamacardio. 2016. 2750

JAMA cardiology

Potential Strategies to Address the Major Clinical Hurdles Facing Stem Cell Regenerative Therapy for Cardiovascular Disease: A Review

Importance While progress continues to be made in the field of stem cell regenerative medicine for the treatment of cardiovascular disease, significant barriers to clinical translation still exist that have thwarted the delivery of cell therapy to the bedside. Objective The purpose of this review is to summarize the major current hurdles for the clinical implementation of stem cell therapy and discuss potential strategies to overcome them. Evidence Review Information for this review was obtained through a search of PubMed and the Cochrane database for English language studies published between January 1, 2000 and June 15, 2016. Ten randomized clinical trials and eight systematic reviews were included in this review. Findings One of the major clinical hurdles facing the routine implementation of stem cell therapy is the limited and inconsistent benefit observed thus far. Reasons for this are unclear but may be due to poor cell retention and survival, as suggested by numerous preclinical studies and a handful of human studies incorporating cell fate imaging. Additional cell fate imaging studies in humans are needed to determine how these factors contribute to limited efficacy. Treatment strategies to address poor cell retention and survival are under investigation and include the following: 1) co-administering of immunosuppressive and pro-survival agents, 2) delivering cardioprotective factors packaged in exosomes rather than the cells themselves, and 3) using tissue engineering strategies to provide structural support for cells. If larger grafts are achieved using the aforementioned strategies, it will be imperative to carefully monitor the potential risks of tumorigenicity, immunogenicity, and arrhythmogenicity. Conclusions and Relevance Despite important achievements to date, stem cell therapy is not yet ready for routine clinical implementation. Significant research is still needed to address the clinical hurdles outlined herein before the next wave of large clinical trials is underway.

No full text available

10. 1001/jamanetworkopen. 2021. 22607

JAMA Network Open

Effect of a Novel Macrophage-Regulating Drug on Wound Healing in Patients With Diabetic Foot Ulcers

Key Points Question Can the topical application of ON101 cream demonstrate a superior therapeutic benefit in wound healing among patients with diabetic foot ulcers (DFUs) compared with standard care? Findings In this randomized phase 3 clinical trial of 236 patients with DFUs, topical application of ON101 with gauze immediately after debridement demonstrated significant healing efficacy compared with an absorbent dressing in all patients, including those with DFU-related risk factors. Meaning Topical treatment with ON101 resulted in improved healing of DFUs.

Introduction Approximately 80% of lower limb amputations are preceded by chronic diabetic foot ulcers (DFUs), resulting in a heavy burden of medical care and expenditure. 1, 2 The current treatment for DFUs in clinical practice focuses primarily on local wound care, including debridement, off-loading, infection control, and maintaining a moist environment with dressings, 3, 4 whereas adjunctive therapies such as the use of growth factors, tissue engineering products, hyperbaric oxygen, and negative pressure wound therapies are applied if the DFUs worsen. 5 Although current treatments featuring tissue repair or the use of anti-inflammatory agents might help in closing or controlling the progression of DFUs, most of these treatments are not well supported by clinical evidence or are not recommended for routine care by the International Working Group on the Diabetic Foot. 6 In addition, the annual increase in amputations also suggests that treatment improvement is needed. 7 Diabetic foot ulcers are pathologically complex mostly because the ulceration is undermined by the existence of multiple risk factors, such as poor patient adherence to treatment, severity of the ulcer, ulcer location and duration, vascular condition, control of glycated hemoglobin (HbA 1c ) levels, smoking habits, and kidney dysfunction. 8, 9 These factors impose a significant clinical need for novel and effective interventions to tackle this life-debilitating and life-threatening disease. Accumulating scientific evidence has revealed that targeting macrophage phenotypes might be a potentially effective therapy in DFUs because hyperglycemia increases the ratio of proinflammatory M1 to proregenerative M2 macrophages. 9, 10, 11, 12, 13, 14 ON101 (supplied by Oneness Biotech Co, Ltd; previously given the research code WH-1) exerts its therapeutic effect through regulation of the balance between M1 and M2 macrophages. ON101 is composed of 2 active pharmaceutical ingredients: PA-F4 from an extract of Plectranthus amboinicus and S1 from an extract of Centella asiatica, 2 medicinal plants reported to have significant pharmacological activities in wound healing. 10, 11, 12 With 48 in vitro and in vivo studies performed, these 2 ingredients, which contribute to a synergistic effect on regulation of the M1:M2 macrophage ratio, have been defined and formulated in a cream base using a proprietary formula. One of these ingredients, PA-F4, significantly attenuates M1 macrophages by suppressing the NLRP3-mediated inflammasome pathway and the production of downstream inflammatory cytokines such as interleukin 1β and interleukin 6, 13 which arrest the inflammation phase. On the other hand, the extract of C asiatica has been reported to activate M2 macrophages by increasing collagen synthesis and by stimulating fibroblast proliferation and the migration of keratinocytes. 14, 15 ON101 has been further demonstrated to accelerate wound healing efficiently in a db/db mouse model of diabetes, obesity, and dyslipidemia by decreasing inflammatory M1 macrophage activity and enriching M2 macrophage populations through granulocyte colony-stimulating factor–mediated M2 polarization, which changed the ulcer status from the inflammatory phase to the proliferation and remodeling stages (eFigure 1 in Supplement 2 ). A clinical pharmacokinetic study on 12 patients with DFUs showed that topical administration of ON101 twice daily in single and multiple doses yielded very limited systemic exposure (Kai-Min Chu, MD, PhD, oral communication, September 4, 2017). Thus, the maximum body concentrations from days 1 and 14 were similar, demonstrating that topical ON101 has no obvious accumulation in the body. No treatment-related adverse events were observed. In a clinical research trial conducted in 24 patients with chronic DFUs classified as grade 3 according to the Wagner system, 10 treatment with ON101 for 2 weeks resulted in an approximately 20% reduction in wound size, and no serious adverse events were reported. Of the 21 patients with evaluable data, the mean wound size at baseline was 359 (range, 20-2352) mm 2, decreasing to 293 mm 2 after 2 weeks of ON101 treatment. 10 Another clinical trial was performed with 30 patients with Wagner grade 1 chronic DFUs treated with ON101 for as long as 12 weeks (Yu-Yao Huang, MD, PhD, oral communication, August 22, 2011). The final incidence of healing was 50%. The mean wound area at baseline was 577 (range, 303-1225) mm 2, decreasing to 163 mm 2 after 12 weeks of ON101 treatment. The topical use of ON101 is supported with a safety profile from the manufacturer and has clear therapeutic potential in promoting wound healing based on previous studies. 10 This multicenter, phase 3 randomized clinical trial was designed to evaluate whether ON101 could treat chronic DFUs by comparing it with a standard primary wound care absorbent dressing. Methods We followed adequate and well-controlled studies as categorized by the US Food and Drug Administration 16 to design a randomized, controlled, evaluator-blinded phase 3 trial to evaluate the efficacy of ON101 applied topically twice daily for treating chronic DFUs (the trial protocol is available in Supplement 1 ). This treatment was compared with an absorbent dressing (Hydrofiber; ConvaTec Ltd) as a comparator in the control group for treating chronic DFUs. This multicenter study was performed with institutional review board approval from 21 medical/clinical centers (eTable 1 in Supplement 2 ) with wound care specialty across the US, China, and Taiwan, where these investigational new drug programs were initiated; all patients provided written informed consent at enrollment. The study followed the International Council on Harmonization guideline 17 and the Consolidated Standards of Reporting Trials ( CONSORT ) reporting guideline. From November 23, 2012, to May 11, 2020, we enrolled outpatients with type 1 or 2 diabetes (as defined by World Health Organization criteria) aged 20 to 80 years, with a baseline HbA 1c level of less than 12% measured during screening or within 3 months before randomization (to convert to proportion of total hemoglobin, multiply by 0. 01). The target ulcer classified as grade 1 or 2 based on the Wagner system on the foot (below the ankle) needed to measure from 1 to 25 cm 2 after debridement, without active infection, and present for at least 4 weeks despite receiving standard of care (according to the International Working Group on the Diabetic Foot guidelines 18 ) before randomization. To avoid possible premature discontinuation of the patient treatments during the trial, we excluded patients with an ankle-brachial before randomization; those with necrosis, purulence, or sinus tracts in the target ulcer not removable by debridement during the screening visit; or those with acute Charcot neuroarthropathy as defined by the American Diabetes Association and the American Podiatric Medical Association, which indicates perturbations of bone metabolism. 19 In addition, revascularization procedures aimed at increasing blood flow in the target limb must have been performed at least 4 weeks before randomization. Eligible participants judged by the principal investigators (Y. -Y. H. , N. -C. C. , H. -H. C. , K. -F. H. , K. -Y. T. , H. -L. H. , P. -Y. L. , C. -K. P. , B. S. , C. L. , Y. M. , Y. C. , Y. L. , Y. X. , Q. L. , G. N. , and S. -C. C. ) on completion of the screening period (≤7 days) were assigned to receive ON101 or absorbent dressing for as long as 16 weeks in a 1:1 allocation by a computer-generated block randomization scheme (eMethods 1 in Supplement 2 ). 20 Individual investigators and research staff were blinded to the size of the block and remained blinded to the treatment assignment before randomization, eliminating the possibility of predetermining the prospective participant’s treatment assignment. The investigator was informed of the randomized treatment assignment in a sealed envelope containing the individual treatment code at the baseline visit. The end-of-treatment visit (visit 10) was the visit in the 16th week after randomization or the visit in which complete wound closure was confirmed, whichever happened first. The independent evaluator assessed the degree of wound closure. The independent evaluator and the study statistician were blinded to the participants’ treatment throughout the study until the clinical database had been locked. To ensure masking throughout the trial, a standardized procedure was established including camera settings, photographing and image-encoding, image delivery to the independent evaluator, and outcome assessment based on the digitally encoded images to delink the patients’ identification, treatment groups, visits, or site information. The detailed blinding procedure is described in eMethods 2 in Supplement 2. Interventions Demographic data, medical history, disease status, radiography, and eligibility were evaluated during the screening period (before randomization). Participants were scheduled for return visits every 2 weeks to receive wound cleansing and debridement with an assessment of wound status, wound size measurement, physical examination results, and concomitant medication records throughout the 16 weeks of the study period once the interventions were administered. The principal investigators and nurses were trained to use standardized study materials, ON101 or absorbent dressings, camera setting, and off-loading recommendations. The instruction for use of off-loading devices was given to the patients with plantar ulcers as assessed by the clinical investigators. All adverse events were recorded at every visit once the intervention was applied. Blood samples for laboratory tests (including hematologic and biochemical analysis) were collected at the screening visit, then every 4 weeks during the treatment period and at the last visit of the follow-up period to detect the levels of factors such as alanine aminotransferase and aspartate aminotransferase to measure liver status, creatinine and blood urea nitrogen to measure kidney status, and albumin to measure nutritional status. Levels of HbA 1c and blood glucose were measured to monitor diabetes-related safety concerns. ON101, a topical cream composed of PA-F4 and S1, was supplied by Oneness Biotech Co, Ltd, and manufactured in Taiwan in a facility in compliance with Good Manufacturing Practice certified by the Pharmaceutical Inspection Cooperation Scheme. Participants in the ON101 treatment group were shown how to self-administer the cream twice daily in an amount to cover the target ulcer fully without exceeding 2 mm in thickness at each visit. The absorbent dressing containing sodium carboxymethylcellulose (Aquacel; ConvaTec Ltd) needed to be changed daily or 2 to 3 times weekly subject to exudate level following the product’s instructions or the investigators’ discretion. The only secondary dressing allowed was sterile gauze for both groups. The amount of ON101 used or the frequency of absorbent dressing changes for each patient was recorded at every visit during the treatment period. No systemic prescriptions were contraindicated during the treatment period, whereas topical antimicrobials and antiseptic agents were not allowed. In cases where the target ulcer worsened (defined as Wagner grade 3), the investigators could determine whether to terminate treatment. If the ulcer was judged by the blinded evaluator as having undergone complete epithelialization for 2 consecutive visits during the treatment period (at or before visit 10), the intervention (ON101 or absorbent dressing) was stopped, and a visit 10 was scheduled after this judgment. If the patients were confirmed to have an unhealed target ulcer at visit 10, continual standard of care with the absorbent dressings was provided to them regardless of the allocated group during the 12-week follow-up period. Data Collection and Outcome Measures The primary efficacy outcome was to compare the incidence of complete healing between the 2 groups at the end of the 16-week treatment period. Complete healing, defined as complete epithelialization maintained without drainage or requirement of dressings for at least 2 consecutive visits, was determined by an independent evaluator blinded to the patient’s information and treatment allocation. Secondary ulcer-related outcomes included time to complete ulcer healing (from baseline visit to first 100% re-epithelialization visit), percentage of change in ulcer surface area from baseline (to the latest treatment visit or complete wound closure), percentage of patients with a 50% reduction in ulcer surface area, and incidence of infection of the target ulcer. The exploratory, ulcer-related outcome data included any incidence of ulcer recurrence during the 12-week follow-up period. Target wound size was measured by an investigator using digital planimetry at every visit after any necessary debridement. In addition, efficacy variables were further assessed for subgroups for the incidence of complete healing, characterized according to the prior duration of ulcers recorded at the baseline visit (6 months as a cutoff), 21 ulcer size (5 cm 2 as a cutoff), 22 and HbA 1c level (9% as a cutoff regarded as poor glycemic control according to the definition of the American Diabetes Association). Safety outcomes were used to assess adverse events and clinical laboratory values. Statistical Analysis The sample size was calculated based on the results of ON101 in the previous trial by hypothesizing a 20% superiority in the incidence of wound closure compared with the efficacy of the absorbent dressing (Yu-Yao Huang, MD, PhD, oral communication, August 22, 2011). With a 1:1 randomization ratio in the 2 groups, 236 participants were required to be enrolled to ensure that at least 212 had evaluable data for achieving 80% power with a 2-sided α value of 5% nominal significance. All analyses were performed using SAS software, version 9. 4 (SAS Institute Inc). The intention-to-treat (ITT) principle was applied to the full-analysis set (FAS), which included all randomized patients irrespective of the actual receipt of study intervention and adherence to the protocol or the occurrence of adverse events. The FAS was used to analyze all efficacy and safety data. A modified ITT (mITT) protocol was applied to exclude patients in the FAS with ineligible target ulcers at baseline. The mITT was used for supportive analysis of efficacy data as appropriate. For the primary end point, we used a χ 2 test and a logistic regression model with intervention as a fixed factor, with the baseline ulcer size and Wagner grade adjusted as covariates. The results of the logistic regression model are presented in terms of the odds ratio (OR), with P values and associated 95% CIs. Some outcomes are expressed as the hazard ratio (HR). Exploratory post hoc analyses of pertinent variables, such as ulcer duration, ulcer size, and patients’ HbA 1c levels, were also performed. The time to complete ulcer healing was calculated using the Kaplan-Meier method with a log-rank test. The HRs and 95% CIs were estimated using a Cox proportional hazards regression model. The percentile changes in ulcer surface area and ulcer surface area change from baseline were subjected to regression analysis adjusted by baseline ulcer area and Wagner grades. The incidences of infection of target ulcers and of recurrence were evaluated using the Fisher exact test. The adverse events were regarded as treatment emergent if they occurred after the intervention started. Adverse events, treatment-emergent adverse events, and serious adverse events were summarized by frequency and proportion of total patients by system organ class and by preferred terms. All adverse event–related comparisons between the 2 groups were performed using the Fisher exact test. The clinical laboratory test data were used to tabulate the change in values from baseline and were compared between groups using analysis of covariance. All tests were 2 tailed, and P <. 05 was considered statistically significant. For possible early study termination, an independent data monitoring committee was established to monitor data when the patient numbers reached approximately 50% and 90% of the planned enrollment. The futility or superiority of ON101 cream was assessed by the independent data monitoring committee using the Lan-DeMets alpha-spending approach, in which the boundaries were determined by the type of O’Brien-Fleming spending function. 23 The superiority of ON101 was confirmed by the independent data monitoring committee ( P <. 001, much less than the boundary of 0. 03476) on achieving 90% of the planned enrollment (212 participants with evaluable data) so the interim analysis could proceed. The trial was not terminated despite ON101 achieving superiority in the interim analysis because the 236th patient with evaluable data was already enrolled before this point. Results A total of 236 patients were included in the FAS (175 men [74. 2%]; 61 women [25. 8%]; mean [SD] age, 57. 0 [10. 9] years). The mean (SD) HbA 1c level was 8. 1% (1. 6%) at baseline and did not change significantly at the end of treatment (mean [SD] HbA 1c of 8. 0% [1. 8%] in the ON101 group vs 7. 9% [1. 6%] in the comparator group), and 144 patients (61. 0%) were diagnosed as having had diabetes for more than 10 years. Patients in the FAS were randomly allocated to treatment: 114 (48. 3%) to the comparator group and 122 (51. 7%) to the ON101 group. Sixteen patients (13. 1%) in the ON101 group vs 21 (18. 4%) in the comparator group had an early termination (total of 37) ( Figure 1 ). The instructions for using off-loading devices were given to the patients who were assessed by the clinical investigators. Some patients did not follow the suggestion because of the humidity in Taiwan ( Table 1 ). Among the 236 patients in the FAS, 184 (78. 0%) were classified as having Wagner grade 2 ulcers, 117 (49. 6%) had ulcers in the plantar region, and 64 (27. 1%) had a baseline HbA 1c level of at least 9%. The mean (SD) ulcer size was 4. 8 (4. 4) cm 2, and the mean (SD) prior duration of the target ulcer was 7. 2 (13. 4) months at entry ( Table 1 ). Figure 1. CONSORT Diagram of Study Flow A total of 236 patients were randomized. Absorbent dressing was Hydrofiber (ConvaTec Ltd). To convert glycated hemoglobin (HbA 1c ) to proportion of total hemoglobin, multiply by 0. 01. ABI indicates ankle-brachial index; FAS, full-analysis set; and mITT, modified intention to treat. a Judged by the investigator to be unsuitable for the study for any other reason. Table 1. Baseline Patient Characteristics and Intervention During the Study Characteristic Patient group a ON101 (n = 122) Absorbent dressing (n = 114) All (N = 236) Baseline patient characteristics Age, mean (SD), y 57. 4 (10. 6) 56. 6 (11. 3) 57. 0 (10. 9) Sex Male 93 (76. 2) 82 (71. 9) 175 (74. 2) Female 29 (23. 8) 32 (28. 1) 61 (25. 8) Type 2 diabetes 121 (99. 2) 113 (99. 1) 234 (99. 2) Diabetes duration, y ≤10 55 (45. 1) 37 (32. 5) 92 (39. 0) >10 67 (54. 9) 77 (67. 5) 144 (61. 0) HbA 1c level, % Mean (SD) 8. 1 (1. 5) 8. 1 (1. 8) 8. 1 (1. 6) <9 90 (73. 8) 82 (71. 9) 172 (72. 9) ≥9 32 (26. 2) 32 (28. 1) 64 (27. 1) BMI <25 59 (48. 4) 50 (43. 9) 109 (46. 2) ≥25 63 (51. 6) 64 (56. 1) 127 (53. 8) Hypertension 78 (63. 9) 73 (64. 0) 151 (64. 0) CVD history b 25 (20. 5) 23 (20. 2) 48 (20. 3) Kidney status eGFR, mL/min/1. 73 m 2 ≥60 90 (73. 8) 81 (71. 1) 171 (72. 5) <60 32 (26. 2) 33 (28. 9) 65 (27. 5) ABI, mean (SD) 1. 1 (0. 2) 1. 1 (0. 1) 1. 11 (0. 1) Amputation history c 56 (45. 9) 60 (52. 6) 116 (49. 2) Wound conditions, Wagner grade 1 29 (23. 8) 23 (20. 2) 52 (22. 0) 2 93 (76. 2) 91 (79. 8) 184 (78. 0) Ulcer size, cm 2 Mean (SD) 5. 0 (4. 4) 5. 1 (4. 7) 4. 8 (4. 4) 1-5 88 (72. 1) 77 (67. 5) 165 (69. 9) >5 33 (27. 0) 36 (31. 6) 69 (29. 2) Ulcer duration, mo Mean (SD) 7. 2 (13. 0) 7. 3 (13. 9) 7. 15 (13. 4) <6 86 (70. 5) 79 (69. 3) 165 (69. 9) ≥6 36 (29. 5) 35 (30. 7) 71 (30. 1) Plantar ulcers 64 (52. 5) 53 (46. 5) 117 (49. 6) Intervention during the study Off-loading in plantar ulcer d Use 33 (51. 6) 34 (64. 2) 67 (57. 3) No use 15 (23. 4) 9 (17. 0) 24 (20. 5) Not specified 16 (25. 0) 10 (18. 9) 26 (22. 2) Diabetes medication prescribed Metformin 62 (50. 8) 51 (44. 7) 113 (47. 9) Insulin 67 (54. 9) 67 (58. 8) 134 (56. 8) Any oral hypoglycemic agent 84 (68. 9) 81 (71. 1) 165 (69. 9) Use of antibiotics 30 (24. 6) 26 (22. 8) 56 (23. 7) Abbreviations: ABI, ankle-brachial index; BMI, body mass index (calculated as weight in kilograms divided by height in meters squared); CVD, cardiovascular disease; eGFR, estimated glomerular filtration rate; HbA 1c, glycated hemoglobin. SI conversion factor: To convert HbA 1c to proportion of total hemoglobin, multiply by 0. 01. a Unless otherwise indicated, data are expressed as number (%) of patients. Owing to missing data, numbers may not total column headings or percentages may not total 100. Absorbent dressing was Hydrofiber (ConvaTec Ltd). b Includes ischemic heart disease, coronary artery disease, or cerebral vascular accident with embolic, ischemic, or hemorrhagic stroke. c Due to previous diabetic foot ulcers. d Includes only patients with plantar ulcer. Primary Outcome Seventy-four patients (60. 7%) in the ON101 group vs 40 (35. 1%) in the comparator group achieved ulcer closure within 16 weeks (OR, 2. 84; 95% CI, 1. 66-4. 84; P <. 001) ( Table 2 ). Similar results were also noted in the mITT population, where 73 of 118 patients (61. 9%) in the ON101 group and 38 of 112 (33. 9%) in the comparator group had ulcer closure (OR, 3. 15; 95% CI, 1. 82-5. 43; P <. 001) ( Table 2 and eTable 2 in Supplement 2 ). The independent evaluator assessed the degree of wound closure. Table 2. Primary and Secondary Outcomes a Outcome Patient group OR (95% CI) P value ON101 (n = 122) Absorbent dressing (n = 114) Complete healing, No. (%) FAS 74 (60. 7) 40 (35. 1) 2. 84 (1. 66-4. 84) <. 001 b mITT 73 (61. 9) 38 (33. 9) 3, 15 (1. 82-5. 43) <. 001 b Secondary Change in WSA from baseline to visit 10, mean (SD), % −78. 0 (42. 6) −78. 0 (34. 9) NA. 89 Incidence of patients with 50% reduction in WSA on visit 10, No. (%) 101 (82. 8) 98 (86. 0) 0. 80 (0. 39-1. 62). 53 b Incidence of wound infection 6 (4. 9) 7 (6. 1) NA. 78 Ulcer recurrence, No. (%) c 15 (20. 3) 7 (17. 5) NA. 81 Safety Patients with TEAEs, No. (%) 76 (62. 3) 77 (67. 5) NA. 42 No. of TEAEs 207 235 NA Related TEAEs Patients, No. (%) 7 (5. 7) 5 (4. 4) NA. 77 No. of events 11 5 NA NA Serious TEAEs Patients, No. (%) 14 (11. 5) 9 (7. 9) NA. 39 No. of events 24 14 NA NA Related serious TEAEs in events, No. (%) 0 1 (0. 9) NA <. 48 TEAE leading to death, No. 0 0 NA NA Subgroup analysis Wound closure, No. /total No. (%) HbA 1c level <9% 59/90 (65. 6) 33/82 (40. 2) 2. 81 (1. 50-5. 26) <. 001 b ≥9% 15/32 (46. 9) 7/32 (21. 9) 3. 14 (1. 04-9. 50). 04 b Ulcer size, cm 2 1-5 55/88 (62. 5) 31/77 (40. 3) 2. 46 (1. 31-4. 61). 005 b >5 18/33 (54. 5) 8/36 (22. 2) 4. 09 (1. 42-11. 80). 009 b Ulcer duration, mo <6 mo 62/86 (72. 1) 36/79 (45. 6) 3. 07 (1. 59-5. 95) <. 001 b ≥6 mo 12/36 (33. 3) 4/35 (11. 4) 3. 99 (1. 09-14. 63). 04 b Abbreviations: FAS, full-analysis set; HbA 1c, glycated hemoglobin; mITT, modified intent-to-treat; NA, not applicable; OR, odds ratio; TEAEs, treatment-emergent adverse events; WSA, wound (ulcer) surface area. a The absorbent dressing used was Hydrofiber (ConvaTec Ltd). b Calculated using a logistic regression model. Treatment was the main exposure variable; the baseline wound size in cm 2 and Wagner grade were covariates. c Ulcer recurrence was recorded once the ulcer had healed completely and was observed during the follow-up period. Ulcer duration, ulcer size, and HbA 1c levels are known to be associated with poor prognosis of DFUs. 9, 24, 25 Therefore, a subgroup analysis was conducted on baseline ulcer duration (6 months as a cutoff), baseline ulcer area (5 cm 2 as a cutoff size), and baseline HbA 1c level (9% as a cutoff). The subgroup analysis displayed a significant OR in favor of the ON101 group compared with the comparator group (OR, 3. 14 [95% CI, 1. 04-9. 50; P =. 04] for HbA 1c level ≥9%; OR, 3. 99 [95% CI, 1. 09-14. 63; P =. 04] for ulcer duration ≥6 months; OR, 4. 09 [95% CI, 1. 42-11. 80; P =. 009] for ulcer size >5 cm 2 ) ( Table 2 ). In addition, we subgrouped patients with an ulcer reduction of less than 10% during the screening period and analyzed the primary efficacy variable. The result also favored the ON101 treatment (32 of 64 [50. 0%] vs 18 of 66 [27. 3%]; P =. 02) (eTable 6 in Supplement 2 ). Secondary Outcome Patients in the ON101 group had a better healing rate than those in the comparator group (HR, 1. 80 [95% CI, 1. 23-2. 65; P =. 002]) ( Figure 2 ) in the FAS as well as the mITT population (HR, 1. 91 [95% CI, 1. 29-2. 83; P =. 001]) (eFigure 2 in Supplement 2 ). The cumulative incidence of complete healing at each week also reflected the continual higher probability in the ON101 group for reaching complete wound closure from week 4 onward. The time to reach median population healing was 98 days in the ON101 group, whereas it was undeterminable in the comparator group because ulcers of only 40 patients (35. 1%) healed in this group during the treatment period ( Figure 2 ). The mean reduction in ulcer surface area (from the last treatment visit to baseline) was 78. 0% in both groups (SDs, 42. 6% for the ON101 group and 34. 9% for the comparator group; P =. 89), and the incidence of a 50% reduction in ulcer surface area was not significantly different between both groups (101 of 122 [82. 8%] vs 98 of 114 [86. 0%]) ( Table 2 ). Only a few episodes of target ulcer infection occurred in both groups during the treatment period (6 in the ON101 group and 7 in the comparator group; P =. 78) ( Table 2 ). The incidence of recurrence in completely healed wounds during the follow-up phase was 15 of 74 (20. 3%) in the ON101 group and 7 of 40 (17. 5%) in the comparator group without statistical significance ( P =. 81) ( Table 2 ). Figure 2. Kaplan-Meier Plot of Time to Complete Healing in the Full-Analysis Set Population The survival curve indicates the incidence of ulcers healed at each visit in the full-analysis set population. Complete healing was defined as epithelialization without drainage observed at 2 consecutive visits. A full-analysis set cohort randomly assigned to the absorbent dressing (Hydrofiber; ConvaTec Ltd) group (n = 114) or ON101 group (n = 122) was used for Kaplan-Meier analysis. Adverse Events In terms of safety, there were no clinically significant changes or differences between the 2 treatment groups in hematology, biochemistry (including HbA 1c and fasting glucose levels), or vital signs ( Table 2 and eTable 3 in Supplement 2 ). Treatment-emergent adverse events were reported in 76 patients in the ON101 group and 77 in the comparator group, of whom 7 of 122 (5. 7%) in the ON101 group and 5 of 114 (4. 4%) from the comparator group were considered related to the treatments ( Table 2 and eTable 4 in Supplement 2 ). None of the serious adverse events was related to ON101 treatment, whereas there was 1 case of osteomyelitis reported to be linked to the comparator group in which 1 patient (0. 8%) assigned to ON101 died of septic shock, acute kidney injury, and acute respiratory failure, which were not considered to be related to treatment or to ulcer progression (eTable 5 in Supplement 2 ). Discussion To our knowledge, this study is the first international phase 3 randomized clinical trial of an investigational drug able to regulate M1/M2 macrophage activities in patients with DFUs. ON101 exhibited better efficacy in facilitating the complete healing of DFUs. Hyperglycemia is an underlying cause of chronic DFUs in which the M1-to-M2 macrophage transition is delayed and the inflammatory stage is prolonged. 26, 27 ON101 can restore the balance of M1/M2 macrophages caused by hyperglycemia. The robust efficacy in patients with high-risk factors suggests that ON101 might provide multiple and proactive ways to improve wound healing by promoting the M1-to-M2 transition and thereby accelerating wound healing for ulcers not only in terms of early formation but also with high-risk factors including ulcer duration of at least 6 months, ulcer size greater than 5 cm 2, and an HbA 1c level of at least 9%. The design of this study followed US Food and Drug Administration guidelines. 16 The complete healing rate of the comparator group at 16 weeks (35. 1%) was consistent with the 28. 2% shown by ITT analysis at week 12 disclosed in a previous trial by Jeffcoate et al. 28 This finding verifies the suitability of the design and implementation of this study in conforming to other randomized clinical trials. The application of ON101 after debridement—which can be self-administered at home—indicated the same level of convenience of use as for the absorbent dressing. Despite the statistically significant wound closure and healing rates provided by ON101, the ulcer reduction outcomes, including changes in ulcer area from baseline and rate of 50% reduction in the wound area, were not statistically significant between the 2 groups during the treatment period. This discrepancy possibly arose from the use of 2-dimensional measurements on the wound area without considering the wound depth. In this study, 78. 0% of the ulcers were Wagner grade 2, meaning that they extended into tendon, bone, or capsule. Thus, the measurement of wound area instead of volume might not reflect the actual volumetric change. Similar outcomes were also noted in the pivotal study (study 92-22120-K) of becaplermin (Regranex; Smith & Nephew plc). The use of 3-dimensional measurement tools should be considered in future studies. Limitations This study has some limitations. The first was the open-label design, which did not allow us to mask the interventions to patients or clinical investigators; therefore, blinded evaluation was implemented to minimize any possible bias. Second, the inclusion and exclusion criteria ruled out patients requiring dialysis, which, to a certain extent, reflects some types of patients with DFUs. Using the ankle-brachial index as the sole criterion in judging blood perfusion could not exclude patients with ischemia completely. Last, the lack of a 2-week run-in period was a potential flaw in the design, because possible rapid healers might not have been excluded in the study. To assess whether this factor affected the trial results, a separate analysis of the complete ulcer healing rate was performed by excluding those patients with an ulcer reduction of at least 10% during the screening period, the results of which favored ON101 treatment (32 of 64 [50. 0%] vs 18 of 66 [27. 3%]; P =. 02) (eTable 6 in Supplement 2 ). Conclusions The results of this randomized clinical trial demonstrate a clinically and statistically superior therapeutic efficacy of ON101 in the treatment of DFUs in both FAS and mITT populations in terms of complete healing rate and time to complete healing compared with absorbent dressing. For chronic wounds in patients with high-risk factors, the therapeutic efficacy of ON101 could be sustained in ulcers that last for more than 6 months or measure greater than 5 cm 2 or in patients with high HbA 1c levels. The findings of this study suggest that ON101, a macrophage regulator that behaves differently from moisture-retaining dressings, represents an active-healing alternative for home and primary care of patients with chronic DFUs.

10. 1001/jamaophthalmol. 2013. 4319

JAMA ophthalmology

Accelerated

Purpose To design patterned, transparent silk films with fast degradation rates for the purpose of tissue engineering corneal stroma, Methods β-sheet (crystalline) content of silk films was decreased significantly by using a short water annealing time. Additionally, a protocol combining short water annealing time with enzymatic pretreatment of silk films with protease XIV was developed. Results Low β-sheet content (17–18%) and enzymatic pre-treatment provided film stability in aqueous environments and accelerated degradation of the silk films in the presence of human corneal fibroblasts in vitro. The results demonstrate a direct relationship between reduced β-sheet content and enzymatic pre-treatment and overall degradation rate of the protein films. Conclusions The novel protocol developed here provides new approaches to modulate the regeneration rate of silk biomaterials for corneal tissue regeneration needs. Translational relevance Patterned silk protein films possess desirable characteristics for corneal tissue engineering, including optical transparency, biocompatibility, cell alignment and tunable mechanical properties, but current fabrication protocols do not provide adequate degradation rates to match the regeneration properties of the human cornea. This novel processing protocol makes silk films more suitable for the construction of human corneal stroma tissue and a promising way to tune silk film degradation properties to match corneal tissue regeneration.

No full text available

10. 1001/jamaoto. 2013. 5669

JAMA otolaryngology-- head & neck surgery

Defining the critical-sized defect in a rat segmental mandibulectomy model

Importance Advances in tissue engineering offer potential alternatives to current mandibular reconstructive techniques; however, prior to clinical translation of this technology, a relevant animal model must be used to validate possible interventions. Objective This study aims to establish the critical-sized segmental mandibular defect that does not heal spontaneously in the rat mandible. Design Prospective study using an animal model. Setting Animal laboratory. Participants Sprague-Dawley rats. Interventions Twenty-nine Sprague-Dawley rats underwent creation of one of four segmental mandibular defects: 0-mm, 1-mm, 3-mm and 5-mm. All mandibular wounds were internally fixated with 1-mm microplates and screws and allowed to heal for 12-weeks. Main Outcomes and Measures Mandibles were analyzed with micro-computed tomography (microCT) and bony healing was graded on a semi-quantitative scale. Results Seven animals were utilized in each experimental group. No 5-mm segmental defects successfully developed bony union, whereas all 0-mm and 1-mm defects had continuous bony growth across the original defect on micro-CT. Three of the 3-mm defects had bony continuity, and three had no healing of the bony wound. Bony union scores were significantly lower in the 5-mm defects compared to 0-mm, 1-mm and 3-mm defects (all p < 0. 01). Conclusion and Relevance The rat segmental mandible model cannot heal a 5-mm segmental mandibular defect. Successful healing of 0-, 1- and 3-mm defects confirms adequate stabilization of bony wounds with internal fixation with 1-mm microplates. The rat segmental mandibular critical-sized defect provides a clinically relevant testing ground for translatable mandibular tissue engineering efforts.

No full text available

10. 1002/0471143030. cb2309s61

Current protocols in cell biology / editorial board, Juan S. Bonifacino. . . [et al. ]

Expanding Mouse Ventricular Cardiomyocytes through GSK-3 Inhibition

Controlled proliferation of cardiac myocytes remains a major limitation in cell biology and one of the main underlying hurdles for true modern regenerative medicine. Here we provide a technique to robustly expand early fetal-derived mouse ventricular cardiomyocytes on a platform usable for high-throughput molecular screening, tissue engineering or potentially useful for in vivo translational experiments. This method provides a small molecule-based approach to control proliferation or differentiation of early beating cardiac myocytes through modulation of the Wnt/β-catenin signaling pathway. Moreover isolation and expansion of fetal cardiomyocytes takes less than 3 weeks, yields a relatively pure (~70%) functional myogenic population and is highly reproducible.

No full text available

10. 1002/1873-3468. 12285

Febs Letters

Inherited heart disease – what can we expect from the second decade of human i

Induced pluripotent stem cells (i PSC s) were first generated 10 years ago. Their ability to differentiate into any somatic cell type of the body including cardiomyocytes has already made them a valuable resource for modelling cardiac disease and drug screening. Initially human i PSC s were used mostly to model known disease phenotypes; more recently, and despite a number of recognised shortcomings, they have proven valuable in providing fundamental insights into the mechanisms of inherited heart disease with unknown genetic cause using surprisingly small cohorts. In this review, we summarise the progress made with human i PSC s as cardiac disease models with special focus on the latest mechanistic insights and related challenges. Furthermore, we suggest emerging solutions that will likely move the field forward.

Abbreviations ALPK3, alpha kinase 3 ARVC, arrhythmogenic right ventricular cardiomyopathy CMs, cardiomyocytes DCM, dilated cardiomyopathy HCM, familial hypertrophic cardiomyopathy hiPSC, human iPSC hPSC, human pluripotent stem cell iPSCs, induced pluripotent stem cells LQTS, long‐QT syndrome RBM20, RNA‐binding motif protein 20 Since their discovery in 2006 1, induced pluripotent stem cells (iPSCs) have enabled scientists to study the physiological and pathological mechanisms of both development and disease in a new way 2, 3, 4. Inherited cardiovascular disorders and in particular channelopathies have been among the first human diseases studied using iPSCs 5, 6, 7, 8. Indeed, although animal models have been and continue to be essential in advancing the understanding of cardiovascular disease 9, 10, 11, interspecies differences hamper translation of many results directly to humans. Because human iPSCs (hiPSCs) can be derived from virtually any patient of interest and can usually differentiate efficiently into cardiomyocytes (hiPSC‐CMs) and there are well‐established techniques for their functional characterisation in vitro, they have rapidly been exploited for disease modelling and drug screening, and in the future are expected to offer new opportunities for regenerative medicine and personalised medicine. The ambition of the Precision Medicine Initiative is to employ a combination of clinical, genetic or genomic, and molecular data to develop tailored therapies for subgroups of patients 12. The pathogenesis of many inherited cardiac diseases remains insufficiently understood and it has been difficult to account for incomplete penetrance and variable severity. Any methodology that could help deciphering predisposition or causative molecular and cellular mechanisms therefore could be helpful. Because hiPSCs can capture the complex genetic background of a patient, expectations are that they will contribute to this goal. Whether patient‐specific hiPSC‐CMs can provide information that predicts disease penetrance and outcome remains to be determined. However, recent studies using hiPSC‐CMs have lead to optimism with respect to the use of this technology in providing new mechanistic insights into disease pathogenesis 13 and cardiotoxicity 14, 15. In this review we discuss the use of hiPSCs as models of inherited heart diseases, with special focus on underlying disease mechanisms that have not been evident from other approaches, current challenges, and emerging solutions for moving the field forward. Generation of hiPSC‐CMs (for cardiac disease modelling) In using hiPSCs for disease modelling, the first step is reprogramming somatic cells collected from primary tissue samples. Most frequently, hiPSCs are derived from patients with known disease‐causing mutations. An alternative, but increasingly used, approach is to introduce site‐specific genetic changes (including knock‐out and precise nucleotide changes) in wild‐type human pluripotent stem cell (hPSC) lines by gene targeting 4. Even in cases where no causative mutations are yet known, the rationale of using hiPSCs is that they are able to capture any genetic predisposition in patients aside from specific mutations, which are thought to play determinant roles not only in disease manifestation and progression but also in the context of external environmental factors that may precipitate the condition. These could include exercise, cardiac‐ and noncardiac drugs, fever, food supplements and the like 15, 16. Once hiPSCs have been obtained, several methods to induce cardiac differentiation can be used 17, 18. Although these techniques were initially inefficient and not readily transferable across cell lines, there are now a number of more robust protocols available and CMs at > 95% purity can be produced 19, 20. In addition, a number of defined media and commercial kits have become available of late which seem particularly efficient across lines, including several apparently differentiation refractory hiPSC lines. However, it is noteworthy that although the efficiency of differentiation protocols has undergone a multifold increase over recent years as a result of culture condition optimisation, this has not been paralleled by improvements in maturation of the electrophysiological properties of hiPSC‐CMs: resting membrane potential is depolarised, and upstroke velocity and ion channels expression remain low in comparison with adult cardiomyocytes 21, 22, 23. This suggests that optimisation has impacted quantitative rather than qualitative aspects of differentiation. Of note, most of these differentiation protocols result in mixed populations of ventricular‐, atrial‐ and nodal‐like subtypes, with ventricular CMs being the most represented. Some recent studies have succeeded in directing hPSC differentiation towards atrial 24, 25 and pacemaker 26 subtypes, however, their application for studying molecular mechanisms related to disease is still under investigation. Maturation of hiPSC‐CMs Improving maturity in hiPSC‐CMs remains one of the major priorities of the field, since phenotypic immaturity limits their ability to successfully model critical aspects of cardiac disorders including adult‐onset diseases 21. Comparison with human fetal hearts suggests that in vitro ‐derived hPSC‐CMs are similar to first trimester gestational stage CMs with regard to gene expression, structure and function and only in certain culture conditions do they become more similar to second trimester fetal CMs 27, 28. Channelopathies are among the cardiac diseases that suffer least from these limitations, since most (but not all) of the relevant ion channels for the generation of the cardiac action potential are expressed in hiPSC‐CMs. This is the reason why the long‐QT syndrome (LQTS) was one of the first cardiac arrhythmia conditions to be modelled using hiPSC‐CMs 5. Since then, approximately one‐fourth of the publications on cardiac disease modelling have studied LQTS‐causing mutations. Although some key features of other inherited heart diseases, such as catecholaminergic polymorphic ventricular tachycardia (CPVT) 29, arrhythmogenic right ventricular cardiomyopathy (ARVC) 30, familial hypertrophic cardiomyopathy (HCM) 31 and familial dilated cardiomyopathy (DCM) 32 have also been recapitulated, certain molecular mechanisms will only be reproduced when more mature cardiac phenotypes are achieved. Existing hiPSC models of inherited cardiac diseases Human iPSC technology has succeeded in modelling cardiovascular and cardiometabolic diseases with different inheritance patterns: the most common autosomal dominant forms (LQTS 5, 6, 7, 33, CPVT1 29, DCM 32, HCM 31, ARVC 34 ) but also the rarer autosomal recessive forms (CPVT2 35, Jervell and Lange‐Nielsen syndrome (JLNS) 36, Pompe disease 37 ), the X‐linked dominant forms (Danon disease 38, Fabry disease 39 ), the X‐linked recessive forms (Barth syndrome 40, Duchenne muscular dystrophy or DMD 41 ) and also finally the nontypical Mendelian forms (hypoplastic left heart syndrome or HLHS 42 ). All of these examples result from genetic defects with cell‐autonomous mechanisms of action in CMs, which means that the pathological phenotype is evident in the cardiomyoctes expressing the (mutated) gene without the need to interact with other cell types (Fig. 1 ). This may not always be the case and there is an increasing number of examples in which the interaction between two or more cell types is needed to reveal the disease phenotype since the cell expressing the mutation sends defective signals to its neighbours 43. As hiPSC technology advances, the ability to establish heterotypic cultures and complex structures increases, we expect that noncell‐autonomous disease mechanisms will be also recapitulated such as those leading to heart failure due to vascular diseases (thrombosis, atherosclerosis) and myocardial infarction (Fig. 1 ). Figure 1 Cell‐autonomous versus noncell‐aututonomous diseases. hi PSC ‐ CM s have already proven their value in recapitulating cell‐autonomous cardiovascular diseases, such as arrhythmic syndromes ( LQTS, JLNS, CPVT ), cardiomyopathies ( DCM, HCM, ARVC, DMD ), cardiometabolic disorders (Pompe disease, Fabry disease, Danon disease, Barth syndrome). More challenging to be modelled are noncell‐autonomous cardiovascular disorders, such as diabetic cardiomyopathy, heart failure due to vasculature diseases, for example, thrombosis, atherosclerosis or myocardial infarction. Drug screening, toxicology assays and safety pharmacology One of the fundamental applications of hiPSC cardiac disease models is the development of treatments that ideally will eventually be translated into the clinic to cure (reverse) or relieve (delay) disease symptoms, much like that already achieved for some neurodegenerative disorders 44. This approach is highly dependent on understanding the molecular mechanisms underlying the disease, as well as on the sensitivity of the read‐out in the assay that is used for detecting the abnormal phenotype. Testing a limited number of candidate drugs based on underlying disease mechanisms is already proving the fastest way to move forward to clinical application, since it is based on repurposing previously approved compounds for a new disease 45. In the cardiac field this has not yet led to rapid translation from the laboratory bench to patients, partly because cardiovascular diseases are often not as severe and untreatable as many neurodegenerative disorders. As an alternative to repurposing, hiPSC‐CMs can be used as a platform for high throughput drug testing 46, which is most valuable to pharmaceutical companies looking for new drug and disease targets since they often have technologies for automated measurements. In addition to drug screening and drug development, hiPSC‐CMs are now also beginning to demonstrate their value in revealing cardiotoxic effects. In particular, these cells are proving a valuable tool to identify electrophysiological and transcriptional changes related to HDAC inhibitor‐mediated cardiotoxicity 47. Furthermore, Burridge and colleagues have recently shown that patient‐specific hiPSC‐CMs can recapitulate the predisposition of some breast cancer patients to develop late heart failure after exposure to the chemotherapeutic drug doxorubicin 15. Although of significant interest, this study had some limitations: first, relatively few patients were included in each group (four in the in vitro doxorubicin cardiotoxicity assays and only three for the RNA‐seq analysis of the hiPSC‐CMs); second, the retrospective study design and coadministration of additional chemotherapeutic drugs in one patient group might have biased the outcome. Further validation in larger patient cohorts will be needed to determine whether different degrees of severity and early versus late cardiotoxic effects can be detected and whether the same approach proves valid for other patient groups such as those with tumours in other organs or paediatric patients also treated with doxorubicin 48, 49. Nevertheless, the work supports the idea that hiPSCs are able to capture complex genetic backgrounds of patients in a predictive way and therefore might contribute usefully to the realisation of the Precision Medicine Initiative 12. Pathological phenotypes and new mechanistic insights The successful generation of cardiac disease models with hiPSC‐CMs relies on their ability to recapitulate key aspects of CM biology, including their molecular, cellular and physiological properties, and on the scientist’ tools and ability to record and capture these specific features and changes upon pathological or cardiotoxic conditions. For this, appropriate and sensitive read‐out assays have been developed 13 and techniques are being continuously improved 22. Disease‐related phenotypes and read‐out assays During the first years that followed hiPSC discovery, their derivative CMs were used mostly to model known disease phenotypes to explore their potential value in recapitulating maladies and known pharmacological treatments 8. More recently hiPSC‐CMs proved helpful in providing novel mechanistic insights into inherited heart diseases with both known and unknown genetic cause using surprisingly small cohorts. The different assays used to characterise hiPSC‐CMs phenotypes examine parameters such as gene and protein expression, ultrastructural organisation, electrophysiological function, calcium handling, force of contraction and metabolic profile. Here, we discuss some of the latest examples. The various kinds of diseases that have been modelled using hiPSC are illustrated in Fig. 1. Analysis of not only gene expression in hiPSC‐CMs but also the changes that take place during hiPSC cardiac differentiation has offered hints on genes and potential pathways impaired in some inherited heart conditions. In an autosomal dominant form of DCM caused by mutations in the RNA‐binding motif protein 20 gene ( RBM20 ), for example, stage‐specific transcriptome profiling demonstrated early molecular perturbations during cardiogenesis in patient‐specific hiPSCs 50 ; these results suggested that this clinically aggressive form of DCM is a developmental disorder. In addition, using functional assays the authors demonstrated that RBM20‐dependent mis‐splicing of calcium‐handling genes contributed to alterations in the calcium homoeostasis and excitation–contraction coupling. Similarly, whole transcriptome sequencing led to the hypothesis that mitochondria were implicated in DMD cardiac pathogenesis 41 ; subsequent analysis of the metabolic profile demonstrated that indeed apoptosis in DMD hiPSC‐CMs is mainly induced by a mitochondrial network through the proteins DIABLO, XIAP and CASP3 rather than through cytochrome C and CASP9 cascade. The analysis of ultrastructural CM organisation has revealed phenotypes not only in several cardiomyopathies but also in glycogen storage diseases 51. Among these, Pompe disease was one of the first disorders characterised in depth using hiPSC‐CMs, in which both specific features of the cardiomyopathy and the efficacy of recombinant enzyme therapy in patients were faithfully recapitulated 37. However, only more recently did Raval and colleagues discover a specific deficit in the glycan synthesis in the Golgi that was initially revealed by the change in electrophoretic mobility of lysosomal‐associated membrane protein LAMP1 52. Likewise, ultrastructural analysis by electron microscopy revealed the presence of fragmented mitochondria within autophagosomes in hiPSC‐CMs carrying Danon disease, which is caused by lysosomal‐associated membrane protein LAMP2 deficiency 38 ; a consequent increase in oxidative stress and apoptosis was then demonstrated. This study was one of the first attempts to understand the molecular basis of some pathological features that are also characteristic of heart failure. Interestingly, genetic polymorphisms in the cardioprotective enzyme, aldehyde dehydrogenase 2 gene ( ALDH2 ), were studied by examining the metabolic profile of patient‐specific hiPSC‐CMs 53. A new function for this enzyme was demonstrated, namely modulation of cell survival decisions through changes in the oxidative stress in hiPSC‐CMs, although these finding were first observed in patient‐fibroblasts and then later examined in hiPSC‐CMs. Gene expression analysis highlighted overexpression of JUN and consequently the authors were able to restore ROS levels by the Jun N‐terminal kinase (JNK) inhibition. Importantly, no significant differences were identified under normoxic condition, while ischaemia simulation in vitro revealed the phenotype. Calcium influx into the cell triggers further calcium release from the sarcoplasmic reticulum to the cytosol and finally to the sarcomere resulting in cardiomyocyte contraction. The identification of calcium handling abnormalities in hPSC‐CMs harbouring alpha kinase 3 gene ( ALPK3 ) mutations allowed confirmation at the cellular and molecular level of strong genetic evidence that homozygous or bi‐allelic truncating mutations in ALPK3 can cause paediatric cardiomyopathy 54. However, in this casus, the specific role of ALPK3 remained unclear. Channelopathies are most often characterised by examining their electrophysiological and ion channel properties. Molecular profiling coupled with measurements of action potentials and the slow component of the delayed rectifier potassium current ( I Ks ) demonstrated a distinct molecular mechanism of action of two KCNQ1 mutations in JLNS hiPSC‐CMs 36. A recessive phenotype was associated with the amorphic mutation, while a gene dosage‐dependent ion channel protein reduction at the cell membrane explained the presence of a LQTS phenotype in the heterozygously mutated hiPSC‐CMs. Of note, however, the literature reports a wide range of values in the basic electrophysiological properties of hiPSC‐CMs, action potentials differing an order of magnitude and beating rates anywhere between 0. 5 and 1. 5 Hz, but also specific ion currents (e. g. the slow I Ks, and the rapid I Kr components of the delayed rectifier potassium currents, the sodium current I Na, and the L‐type calcium current I CaL ) varying widely even among wild‐type control hiPSC‐CMs 55, 56. Most notably, very different levels of I Ks have been described (ranging from ~ 0. 3 to ~ 2. 5 pA/pF 5, 57 ), variable observation leading to controversial conclusions: on the one hand, I Ks recapitulates physiological behaviour in playing a major role when repolarisation reserve is attenuated 58, 59 ; on the other, it seems to contribute to repolarisation in hiPSC‐CMs even in the absence of sympathetic stimulation 5, 36, 60, 61. The immature phenotype of all stem cell derivatives including hiPSC‐CMs is probably the reason for this variability but, independent of the cause, it is a limitation to extrapolating results obtained using hiPSC‐CMs to native – healthy and diseased – adult human CMs as discussed below. The variability in protocols used for cardiac differentiation and electrophysiology further contribute to making absolute conclusions on human cardiac physiology and disease. Nevertheless, hiPSC‐CMs with ion channel mutations have been able to contribute to understanding these diseases because in many cases they could recapitulate key disease features observed in patients and sometimes indicate underlying pathological molecular mechanism 56, 62. One example in which quantifiable hiPSC‐CM properties were used for drug screening purposes is diabetic cardiomyopathy 63, a complex metabolic condition affecting also the heart. Here the authors built two levels of disease models with hiPSC‐CMs: environmental, by modulating culture conditions to mimic the diabetes chemistry, and genetic, by deriving hiPSC‐CMs from two patients with different disease severities. Interestingly, in the patient‐specific cells, the diabetes phenotype appeared even in the absence of any diabetic trigger, suggesting that hiPSCs indeed capture and recapitulate genetic predisposition. A common limitation of these studies is the small number of patients analysed; to confirm this concept, it will be necessary to validate results independently across larger cohorts. Choosing the right controls The choice of controls is crucial to allow a proper definition and identification of normal versus abnormal phenotypes, including disease‐ and toxic‐specific molecular mechanisms. Each individual harbours many genetic variants in the genome (not only single nucleotide polymorphisms, copy number variations but also heterozygous and homozygous mutations 64, 65 ) that may be functionally interconnected with the genetic defect underlying a disease. Gene targeting enables isogenic hPSC lines to be created that differ only at specific loci, while the rest of the genome remains identical. The advantage of isogenic lines is that any difference in the phenotype is then most likely linked to genetic change since the only difference between the disease and control line is in principle the mutation of interest. With improvement in the methodologies that can be used for precise gene targeting 66, 67, genetically matched (isogenic) hiPSC lines are now becoming the first choice, although in the cardiac field only a few papers have adopted this approach 36, 40, 54, 68, 69, 70 (Fig. 2 ). Hinson and colleagues demonstrated that truncating mutations in the sarcomeric protein, titin, underlie DCM sarcomeric insufficiency 69 ; interestingly, when isogenic hiPSC‐CMs were used, the reduction in force of contraction was still detectable in the mutated CMs but to a lesser extent than when unrelated diseased and control cells were compared. These results confirmed earlier evidence that genetic background can modify disease phenotype. Similarly, we previously generated two pairs of LQTS and control hiPSCs and hESCs harbouring the same KCNH2 mutation 68 ; comparison of genetically matched CMs proved essential for neither under‐ nor overestimating the consequences of the mutation for the cardiac action potentials. Figure 2 summarises the controls used in all studies since first published in 2010. Of note, relatively few have used isogenic controls. Figure 2 Number of publications about hi PSC s and disease modelling using unrelated controls, family matched controls and isogenic controls from 2010 to mid 2016 in the cardiac field. PubMed Advanced Search Builder was used for the literature search using the following builder: [(human pluripotent stem cell) AND (cardiac disease model) NOT review]. Publications on heart regeneration were manually excluded. References from some of the most comprehensive reviews of the field 8, 13, 62, 87 were screened and manually added when not present in the above‐mentioned search. All the References were then screened and classified according to the control used. The complete list and analysis of references is provided in Table S1. Limitation of this representation relates to selection bias. Future challenges It is now clear that hiPSC‐CMs are useful for modelling inherited human cardiac diseases since there are many different examples in which these cells manifest pathogenic features of the disease. However, the predictive and instructive power of hiPSC‐CMs relies on comprehensive and accurate molecular and functional characterisation 13. Challenges that scientists are facing are the ability to model complex‐ and noncell‐autonomous disorders, predict clinical drug response, recapitulate maturation and ageing in vitro along with difficulties in actually recognising mature CMs in culture; for these issues, emerging solutions are discussed. Complex and noncell‐autonomous cardiovascular disorders Many cardiac diseases can be modelled using a single cell type, most often CMs. Ventricular CMs have been the cell type of choice for many diseases, although other cardiac subtypes might be necessary for studying different conditions, for example, nodal and Purkinje cells in conduction disease and atrial CMs in atrial fibrillation. Protocols are becoming available to derive some of these CM subtypes, 25, 26 and we expect that they will soon be used in studying both pathological and cardiotoxic changes. Furthermore, some maladies might benefit from advanced culturing techniques, since certain phenotypes might become evident only under optimised conditions. For example, contractile defects were only uncovered under specific metabolic culture conditions 71 or when engineered tridimensional (3D) microtissues were used 40, 69, 72. Importantly, the human heart is composed not only of CMs but also vascular, smooth muscle and epicardial cells; to better mimic its function, we predict that 3D cardiac tissue structures will be widely implemented, especially where interactions between different cell types might underlie the disease. As an example, ARVC has been modelled in hiPSC‐CMs and these are the major cellular players in the cardiac dysfunction in this disease 30, 34, 73 ; however, the suspected contribution of epicardial cells to fibro‐fatty substitution and the role of inflammation could not so far be studied in two‐dimensional monotypic cultures. The expectation is that complex multicellar structures will be necessary to reflect fully the pathology of the condition. We anticipate that in their second decade, iPSCs will find increasing utility when combined with cardiac tissue engineering. The necessity for better mimics of the multicellular and dynamic conditions of the cardiovascular system that can recapitulate diseases not only with both known or unknown genetic causes but also related to ageing and drug‐induced cardiac damage has already encouraged engineering of three‐dimensional cardiac microtissues. These are beginning to incorporate the different dynamics that reflect blood flow, mechanical stretch and strain and the electrical stimulation. Together with changes in energy substrates, it is expected that these will lead to structurally and functionally mature human myocardium into which biological and biophysical readouts can be built that allow high throughput, real‐time and quantitative measurement of cardiac (patho)physiological status. The hope is that a higher degree of complexity will advance the understanding of how the human heart responds to toxic compounds and disease, improve the integrity of the disease models, and refine the predictability of drug responses. Predicting clinical drug response One of the greatest promises of hiPSC technology, but at the same time its greatest challenge, is in predicting drug responses in a patient‐specific (personsalised) way that disease treatment and prevention can be tailored to the individual. Because hiPSCs capture the genetic background of the person from whom they are derived, they are excellent candidates for recapitulating ‘in a dish’ the variability found among single patients or subgroups of patients. It is one of the few ways forward in coupling genome‐wide association data, which associates disease risk with certain variants in the genome, to proof of causality in humans. This cannot be done in laboratory mice because of the genome differences. The ambition to implement hiPSC‐CMs in precision medicine partially relies on their ability to predict patients’ response to administered drugs. Recent studies provide optimism in this direction 15, 63, 74, although additional consent and focussed investigations will be needed to determine the extent to which individual variability can be distinguished in the hiPSC‐CMs, including mild or severe, acute, early or late responses. Of note, all studies so far have been based on a small number of patients per group and were conducted retrospectively. One goal in the coming years will be to demonstrate that hiPSC‐CMs can be used in prospective study designs, for example, by deriving them from a large cohort of patients (> 200) that are about to undergo a specific drug treatment, applying the same drug to their hiPSC‐CMs, and following‐up over time the patients to find out whether the in vitro responses matched the final clinical outcome. This approach could prove valuable especially in evaluating drug‐induced cardiotoxicity. In addition, a similar approach will ideally be applied in the evaluation of proarrhythmic risk. For example, if hiPSC‐CM‐based platforms for screening arrhythmic events can be combined with genetic‐ and FDA‐collected data for the generation of reliable patient‐specific arrhythmic scores, their real value will become clear in both the choice of individual patient treatment as well as in the drug development process. However, current challenges are not insignificant and suggest that expectations should be tempered in anticipation of more data. The ambition to reduce the incidence of sudden cardiac death as a result of drugs or inherent predisposition may however be a realisable goal in the coming decade. Maturation and ageing A relevant challenge in the field is to find ways to reproduce in vitro the physiological processes of maturation and ageing that the heart naturally experiences from its formation to birth and further during the lifespan of a human being. The heart contracts many millions of times over a lifetime so that defects that are minor in CMs at birth may only be revealed with ageing. The mechanism and process of postnatal CM maturation is incompletely understood and clearly requires environmental factors including hormones, exercise and CM growth by hypertrophy. Approaches used to address this issue include prolonged culture, metabolic manipulation, tissue engineering technologies, electromechanical pacing and other biophysical approaches 21, 22. Promoting adult patterns of metabolic activity already provided a more appropriate basal condition on which to model the response to pathological stimuli, such as in ARVC 34, HCM 71 and diabetic cardiomyopathy 63. Furthermore, anisotropic nanotopography was necessary to distinguish structural differences between control and DMD cardiomyopathy hiPSC‐CMs that were otherwise masked 70. Since both mechanical forces and molecular signalling from nonCM cell types are essential contributors to heart development, formation, ageing and disease progression 75, 76, we anticipate that a combinatorial application of different strategies will likely be most successful in promoting maturation in hiPSC‐CMs. Ideally 3D tissue structures will be developed, where hiPSC‐CMs (subtypes) and other cells are mixed together and organised in microtissues, with or without the addition of extracellular scaffolds, and will be subjected to electrical or mechanical stimulation 77. However, it is still unclear to what extent adult CM properties can be acquired in a culture dish. Nevertheless, some of these improvements in external parameters may contribute to the development of new and reliable methods for screening phenotypic changes also in response to drug treatments. Recognising a mature CM in culture An important question still remains: how do we recognise a mature CM in a culture dish? Our knowledge about human adult CMs relies on that of disease‐free primary tissue, which is scarce and technically challenging to isolate successfully 78, 79, 80. Nevertheless, there is consensus that adult CMs are elongated and rod shaped, the sarcomeres are highly organised, the resting membrane potential is quite negative (−80 to −90 mV), the upstroke velocity rapid (150–350 V·s −1 ), the sarcomeres organised in T‐tubules, the excitation–contraction coupling fast and efficient, the force of contraction relatively strong (10–50 mN·mm −2 ), the mitochondrial content high and the metabolism mainly based on fatty acids (reviewed in 21, 23 ). Ideally a combination of all of these parameters including the structural, molecular and electrophysiological characteristics associated with CM maturation should be assessed to determine whether hiPSC‐CMs resemble myocytes of the human adult heart. However, it is common practice to test only some of these parameters, usually only those that are most important for the disease phenotype to be assessed. We propose that the most informative assays are those based on assessing all aspects of functionality of the cells, including as a minimum the electrophysiology, calcium handling properties, patterns and force of contraction. If, for example, the expression of specific ion channel genes is examined, it is important to bear in mind that this is not always accompanied by a parallel change in the corresponding currents; there are several intermediate steps from transcription to function, including protein synthesis, post‐translational regulation, protein trafficking to the membrane, anchoring of the channel to the membrane, protein turnover and channel regulation by known and unknown accessory proteins and by intracellular signalling 81. Measuring the action potential would then seem more appropriate, since it can give some information on whether the CM population into question displays similar features to adult CMs. Drawbacks of electrophysiological measurements, especially of single‐cell patch clamp, are that they are time consuming and low throughput, and the skills and technology is not readily available in all laboratories. Furthermore, the resulting data refer to the subpopulation of CMs that survived dissociation into single cells; these are usually the most immature in the population. The calcium transients can be measured using calcium‐sensitive dyes and they usually reflect the action potentials, since they are closely related. In this case, complementary information is obtained from the kinetics of the calcium handling, variations of cytosolic calcium concentrations, and the extent of intracellular calcium stores. This type of analysis, much like patch clamp electrophysiology, is also low throughput although optical imaging of voltage and calcium might help increasing the measurement efficiencies. Finally, the force of contraction is another way to determine hPSC‐CM maturation, although it depends on the cell shape and on substrate stiffness 82. Techniques for measuring strain under controlled conditions have been developed 27, 83, 84, 85 and we expect this will increasingly become a parameter that will be evaluated, although specialist technical implementation is required. In summary, we believe that the intended application of hiPSC‐CMs should probably determine the evaluation method to be used for assessing their maturation, but we expect that development of automated methods to analyse voltage and calcium transients and force of contraction simultaneously in both 2D and 3D settings will become a useful tool for disease modelling and drug testing, as well as for testing conditions that may eventually enhance maturation. Additional variables will need to be determined that may play essential roles in modulating CM growth, such as substrate stiffness and specific molecular cues 86 and still it is unclear whether an adult phenotype will ever be completely achieved in vitro. Concluding remarks Patient‐specific models of cardiovascular diseases based on hiPSC‐CMs are proving valuable in advancing our understanding of the complex and sometimes unexpected molecular mechanisms underlying pathological changes. Recent findings provide optimism on the applicability of hiPSC technology to unravel complex disorders, identify cardiotoxic drug effects and ultimately to help defining patient subtypes towards tailored drug treatments. In the future, larger cohorts of patients will be needed from which derive hiPSC‐CMs and their phenotype analysis will tell until which point hiPSC in general, but in particular, their derived CMs can account for variables such as age, gender and medical treatments. Conflict of interest CLM is cofounder of Pluriomics b. v. Supporting information Table S1. Complete list of references used for Fig. 2. Click here for additional data file.

10. 1002/1873-3468. 12559

Febs Letters

Identifying niche‐mediated regulatory factors of stem cell phenotypic state: a systems biology approach

Understanding how the cellular niche controls the stem cell phenotype is often hampered due to the complexity of variegated niche composition, its dynamics, and nonlinear stem cell–niche interactions. Here, we propose a systems biology view that considers stem cell–niche interactions as a many‐body problem amenable to simplification by the concept of mean field approximation. This enables approximation of the niche effect on stem cells as a constant field that induces sustained activation/inhibition of specific stem cell signaling pathways in all stem cells within heterogeneous populations exhibiting the same phenotype (niche determinants). This view offers a new basis for the development of single cell‐based computational approaches for identifying niche determinants, which has potential applications in regenerative medicine and tissue engineering.

Abbreviations MS, multiple sclerosis PCST, Prize Collecting Steiner Tree; NSC, neural stem cell Stem cells are indispensable for maintaining tissue homeostasis due to their unique ability to generate more specialized cell types in a well‐coordinated manner depending on the organismal needs. This function depends crucially on the ability of stem cells to make robust cell fate choices such as self‐renewal or differentiation. Multiple cell‐intrinsic and extrinsic factors control this decision‐making process. In this regard, interactions between stem cells and their microenvironment, also known as the niche, determine the stem cell phenotypic states such as quiescent and active stem cells 1. The cellular niche translates information from the neighborhood of the stem cell by transmitting external cues to intracellular signaling events that maintains its cellular state. Schofield in his description of hematopoiesis, proposed the concept of stem cell niche where, a stem cell must be associated ‘with other cells which determine its behavior’ in order to ‘prevent its maturation’; loss of this association was hypothesized to result in differentiation 2. This concept of stem cell niche has evolved over time, and now includes several different supportive stromal cell types, anatomical localization, soluble molecules, as well as physical factors, such as shear stress, oxygen tension, and temperature 3. Involvement of such disparate and stochastically fluctuating components, in addition to feedback regulation of the niche by stem cells, leads to the highly dynamic nature of the niche 1, 4, 5. Stem cells are known to remodel the niche by secreting ECM components and other diffusible factors in response to the signals received from the niche, thus giving rise to feedback regulation of niche–stem cell interactions 4. Such a bidirectional interplay between stem cells and niche is exemplified by the fact that daughter/progenitor cells can serve as niche cells for their parent stem cells in different tissue types 1. These feedback regulatory mechanisms, in addition to the complex bio‐physical characteristics of ECM, contribute to nonlinear stem cell–niche interactions 6. General physiological conditions of tissue and organismal requirements shape the niche effect on stem cell phenotype 7. For instance, healthy tissues under homeostatic conditions are characterized by the tight regulation of stem cell and progenitor cell turnover. However, this tissue‐level homeostasis is often disrupted in case of several diseases such as cancers, neurodegenerative diseases, and cardiac dysfunction. Furthermore, aging is known to contribute toward progressive decline in tissue homeostasis due to degenerative changes in niche‐mediated cues that regulate the stem cell activity 8. In general, complications often arise due to a lack of proper generation of progenitor cells, complete loss of stem cells, and uncontrolled growth of stem/progenitor cells. Deregulated niche components are known to be responsible for several of these defects 9. For such cases, regenerative medicine approaches that rely on transplanting or modulating endogenous stem cells hold immense potential 9. At present, a key challenge in this area includes the limited functional integration (or engraftment) of transplanted stem cells into the target tissue. This has been attributed to the negative regulatory effect of diseased niche on transplanted stem cells 10. In order to overcome this limitation, it is essential to understand those regulatory mechanisms that normally control stem cell functional state in response to the niche. However, the multifactorial complexity of the niche–stem cell interactions is a major roadblock in this direction. Therefore, the role of the niche in maintaining distinct stem cell phenotypic states, and how to influence the niche effect on stem cells to induce transitions among these states constitutes a fundamental problem in stem cell research. Recently, studies have begun to address this issue by explicit characterization of niche components and their interactions with stem cells 11, 12. Despite significant progress in identifying cells that comprise the niche, a comprehensive understanding of all niche components is not yet obtained. This lack of knowledge is predominantly due to the difficulty in obtaining and studying niche cells and factors in vivo. Furthermore, there is a lack of consensus on what actually constitutes the niche and the precise definition of niche components 13, 14, 15. In addition to experimental efforts, a few computational systems biology approaches that model population‐level dynamics of cell–cell interactions have been proposed to study niche regulation of stem cells 16, 17, 18, 19, 20. However, a complete description of stem cell–niche interactions that allows designing strategies for controlling the effect of niche on stem cells is still limited. This is mainly due to incomplete characterization of the niche, fluctuations of the niche components, and a large number of nonlinear interactions between the niche components and stem cells. In this article, we hypothesize that stem cell–niche interactions could be considered as a complex many‐body problem that can be simplified by the concept of mean field approximation. Such a view allows consideration of the net effect of all niche components on stem cells as a constant averaged effect or ‘mean field’. Most existing models consider the niche composition to model stem cell–niche interactions via rate equations. Our approach does not require this knowledge, precisely because it considers that stem cells interact with their niches via a mean field created by all niche components, which ultimately determines the sustained activation/inhibition of specific stem cell‐signaling pathways that maintain their phenotypic states. Application of our view allows the identification of niche‐mediated regulators of stem cell phenotypes by relying on single‐cell profiling data. To support this hypothesis, we use examples of different stem cell systems to illustrate how stem cells maintain their phenotypic state via constant activation or inhibition of certain pathways under homeostatic conditions. Such pathways that determine the stem cell states can be termed as niche determinants, and are expected to be constantly activated/inhibited in all cells within a population sharing the same phenotypic state despite the variability in their molecular profiles. Indeed, knowledge of these niche determinants should enable us to identify target genes whose perturbations can induce transitions between different phenotypic states. Mean field approximation: keeping it simple Mean field theory was initially developed by Pierre Curie and Pierre Weiss in physics for a simplified theory of ferromagnetism 21, 22. They considered a lattice composed of magnetic moments interacting with their nearest neighbors, and proposed to replace the actual interactions experienced by each magnetic moment with the mean interaction (provided by the mean magnetization) by setting the fluctuations around the mean equal to zero. Such an approximation that considers each magnetic moment to be influenced by a mean field created by all their neighboring moments enabled Curie and Weiss to effectively simplify the many‐body interaction problem to a two‐body problem without explicitly accounting for each pairwise interaction. Since its initial proposal, different interpretations of mean field theory have been applied to other disciplines, such as ecology, epidemiology, and protein structure prediction 23, 24, 25. Mean field approximation applied to the stem cell niche Despite the existence of different mean field concepts 24, here, we follow the definition proposed in ferromagnetism. In particular, we hypothesize that stem cells and niche components within a spatial compartment can be viewed as a many‐body interaction system that includes different types of interactions among them (Fig. 1 ). By stem cell niche we invoke the original concept of specialized microenvironment which supports stem cell survival and functions 1, 2. In this regard, even though individual components of the niche can fluctuate, their combinatorial effect on stem cells can be represented by a mean field, which is the average of all the molecular and cellular signals from the niche. A single component of the niche may be perturbed, but it does not form a defective field unless the perturbations spread and completely transforms the entire niche 1. The dynamic equilibrium between the niche and the stem cells is resilient and robust to small perturbations and noise in the individual niche components. Therefore, according to our hypothesis, it is not the interaction between stem cells and individual niche components that determines their state, but rather it is the constant interaction of each stem cell with the mean field that leads to a sustained activation or inhibition of specific stem cell intracellular signaling pathways. This ultimately dictates stem cell function and behavior, governing the choice between quiescence, proliferation, self‐renewal, or differentiation. In this way, not only are discrete fluctuations in niche signals buffered against, but so too are the epigenetic and gene expression heterogeneity that stem cell populations display. According to our view, a given stem cell population (sharing a common phenotype), although exposed to perturbations and noise due to fluctuations in individual niche components in addition to the presence of intrinsic molecular heterogeneity, nonetheless should share commonly activated/inhibited signaling pathways that determine their phenotypic state (Fig. 2 ). Such pathways that determine the stem cell state can be termed as niche determinants (Fig. 2 ). Figure 1 Mean field approximation of stem cell–niche interactions. The mean field approximation considers that each stem cell interacts with its niche via a ‘mean field’ created by all molecular and cellular signals from the niche. The figure depicts the complex nature of stem cell–niche interplay within a spatial compartment. Stem cells (red circles) are entangled in an intricate network of interactions (gray edges) with different niche components ( NC ) (yellow nodes of different shapes). Analyzing the effect of each individual component on stem cell would require consideration of a large number of interactions and fluctuations among them. In the right, the enlarged depiction of a stem cell shows a mean field (yellow cloud) created by the niche components around a stem cell. Figure 2 Niche determinants of stem cell phenotype. Representation of stem cell signaling and gene regulatory network states of a heterogeneous population of stem cells sharing a common phenotypic state. The figure depicts heterogeneity of gene expression at a single‐cell level (red and blue nodes) and the signaling pathways regulating the underlying gene regulatory network. According to the mean field hypothesis, in spite of molecular heterogeneity and fluctuations of niche signals, these cells should share commonly activated/inhibited signaling pathways (niche determinants) that determine their phenotypic state. Such pathways are depicted with red arrows, while the other transient signaling pathway activities not common to all cells in the population are depicted with dashed arrows. The underlying gene regulatory network that maintains the phenotype of these cells is depicted with red and blue nodes representing their expression status. As a consequence of approximating the niche components with an effective mean field, the focus is on identifying sustained signaling (shared within a cellular population) responsible for maintaining the specific stem cell phenotype instead of characterizing the niche explicitly. The proposed approach relies on single‐cell profiling data and works by first identifying the most conserved set of genes (based on the similarity of expression levels at single‐cell resolution) defining that particular phenotype. Subsequently, unique signaling pathways/networks that link the conserved receptors and transcription factors for specific stem cell phenotypes are inferred computationally by relying on network topology and expression levels. Case study: mean field approximation to identify niche determinants of NSCs Based on a mean field approximation hypothesis, we illustrate the applicability of this view of stem cell–niche interactions in order to identify niche determinants of quiescent and active neural stem cell (NSC) phenotypes based on a recently published single‐cell RNA sequencing data 26. The data were obtained from Gene Expression Omnibus ( GSE67833 ). Briefly, these data that we used in our approach are described as follows: mouse subventricular zone NSCs were isolated from their natural environment based on the expression of GLAST and Prom1. The transcriptome of 104 GLAST+/Prom1+ cells were analyzed by single‐cell RNA‐seq using Smartseq2 technology 27. These data were then subjected to principal component analysis followed by unsupervised hierarchical clustering of genes with the highest coordinates in the first four principal components (1844 genes) 26. This analysis partitioned the NSCs into two major clusters. One NSC cluster had Egfr expression (a known marker of active NSCs 28 ) in addition to the expression of cell cycle‐related genes. Based on these attributes, this cluster was defined as active NSCs. On the other hand the cluster that lacked the activation markers were classified as quiescent NSCs. Gene ontology and pathway enrichment analysis revealed that active NSCs were enriched in genes for cell cycle, protein synthesis, and mitosis, whereas glycolytic metabolism was found to be most enriched in quiescent NSCs. Gene ontology and pathway enrichment analysis further divided quiescent and active NSCs into two subpopulations each (quiescent NSC1/2 and active NSC1/2). In our current analysis for the sake of simplicity we considered only quiescent and active NSC populations as a whole without considering the further subpopulations. Our strategy relies on gene expression differences between stem cells displaying different niche‐dependent phenotypes, and aims to infer sustained signaling pathways that are required for stably maintaining their corresponding phenotypes. Moreover, despite the niche‐induced fluctuations in signaling, such pathways must be shared (or conserved) within the cells sharing a common phenotype. However, it must be mentioned that identification of conserved pathways can also result in housekeeping pathways that could be of general importance to a wide variety of cell populations (e. g. , pathways that are important for both quiescent and active NSCs) and therefore could lack cell type specificity. In order to overcome this issue, the approach focuses on uniquely conserved pathways within each population and is different across the populations. Single‐cell gene expression data offer the possibility to identify the set of genes whose expression pattern is conserved within a given phenotype. Such genes are more likely to play a dominant role in phenotype maintenance since their expression pattern is similar at single‐cell level. In the example of NSCs, we first identified the genes exhibiting similar expression pattern within quiescent or active phenotype. For this we employed Shannon entropy 29, which measures the disorder of a system, where lower values indicate similar expression pattern of a given gene. Entropy for each gene, X, is defined by: H X = − ∑ i = 1 n P ( x i ) log 2 p ( x i ) where P(x i ) represents probability of gene expression value x = x i. Entropy calculation was performed using data binning approach and the number of bins ( k ) was determined from the expression data using Sturges' rule 30, given by k = log 2 n + 1, where n is the sample size. After data binning, the computation of entropy was performed using maximum likelihood implementation (entropy. empirical) of the R entropy package. We used an entropy cutoff less than 1 and median expression (FPKM) value greater than 10 to classify the gene as having a conserved expression pattern. Entropy calculation for each gene allowed us to identify quiescent or active phenotype‐specific genes that showed similar expression pattern at a single‐cell level. Next, we sought to identify those signaling pathways that are more likely to be constantly active. For this, we first identified the set of receptors/ligands and transcription factors classified as conserved for quiescent and active NSCs. Entropy calculation based on single‐cell expression levels allowed us to identify the genes that shared a similar expression levels. From that list of genes, transcription factors and transcriptional regulators were identified based on annotation available at Animal TFDB ( http://www. bioguo. org/AnimalTFDB/ ). In the case of receptors, since a complete database of receptor molecules is currently unavailable, we used Gene Ontology classification of receptor activity and plasma membrane (GO:0004872, GO:0005886) to identify genes with possible receptor activity. For the case of secreted ligand molecules we utilized the classification of potential ligands reported in a recent study 31. About 90 and 128 receptors/ligands were identified for quiescent and active NSC phenotypes, respectively. From this, identifying the ones that are most likely to propagate the niche mediated signaling is a challenge. We made use of literature‐curated signaling database Reactome 32 as a background raw signaling network consisting of all reported signaling interactions and employed Prize Collecting Steiner Tree (PCST) formalism to infer the signaling pathways. Interactions reported in the Reactome database were used as the background network from where subsequent Steiner trees were inferred. Reactome consists of curated pathways with molecular interaction data from Reactome Functional Interaction Network and other databases such as IntAct, BioGRID, ChEMBL, iRefIndex, MINT, and STRING. We specifically used Reactome Functional Interaction Network ( http://www. reactome. org/pages/download-data/ ) as they contain information on direction and sign (positive of negative regulatory effect) of the interaction. We consider that the conserved receptors/ligands of a given stem cell phenotype are under the direct influence of the niche. Since the exact mechanisms of the niche effect on the signaling activity are not known, we represent the net effect of the niche by introducing a dummy niche node in the raw signaling network. The external dummy node is incorporated as a way to capture the topologically favorable receptors/ligands (from several expressed ones) that can link it to the TFs specific for quiescent and active NSCs. Furthermore, the dummy node is used as the root node which acts as the starting point for Steiner tree identification, consequently the receptors/ligands will be linked to the dummy node in the inferred Steiner trees. This dummy node is connected to all conserved receptors/ligands for each phenotype under consideration. Therefore, signal transduction from the niche to transcription factor must be propagated through at least one of the conserved receptors. The edges in the signaling interactome were weighted using the gene expression data, where the weights were calculated as, c e = 1 x i x j, where x i and x j are the expression levels of the interacting nodes. We specifically used such a weighting scheme since the objective of the PCST algorithm is to collect as many high prize nodes (genes with high expression) while minimizing the edge weights. Such an edge weighting scheme that inversely correlates with the expression levels will enable collecting those edges where both nodes are highly expressed. In such a weighted raw signaling network, that has a dummy niche node representing the net effect of the niche, we used PCST to infer subnetworks with the dummy niche node as the root or origin node and the conserved transcription factors as the terminal nodes. Steiner Tree formalism has been used earlier to reconstruct active signaling pathways 33, 34. Formally, the PCST problem is defined as, given a graph G = (V, E), representing the raw signaling interactome (where, V denotes the nodes and E denotes the edges), with defined edge costs (weights), c e and node prizes b v find a connected subgraph T = (V′, E′), V′ ⊆ V, E′ ⊆ E, that minimizes the following function: T = min ( E ′, V ′ ) connected ∑ e ∈ E ′ c e − λ ∑ v ∈ V ′ b v The node prizes are computed by b v = |log fold change ( V )| from the gene expression data and c e is the edge weights. The constant λ determines the tradeoff of adding new proteins to the inferred network by balancing the cost of new edges and the prize gained by adding a new protein. We chose λ = 0. 01 for our simulations and employed a heuristic method based on a message‐passing algorithm to infer the PCSTs 33. Basically, minimizing this function implies collecting the largest set of high prize nodes while minimizing the set of high cost edges in a tradeoff tuned by λ that results in a connected subgraph. Since the dummy node is connected only to the conserved receptors of a given cell type, the inferred subnetworks will encompass only those receptors that are both topologically favorable and maximize the expression values of the intermediate nodes. Therefore, from several conserved receptors, one could narrow down to the few linking the transcription factors based on their unique network topological features and expression levels. Employing the above strategy, we identified subnetworks that are likely to maintain the quiescent and active phenotypes of NSCs (Fig. S1). In the case of quiescent NSCs, we identified nine subnetworks with receptors as origins/sources responsible for controlling the expression status of the downstream terminal transcription factors (Fig. S2). Among such identified receptors, the role of Bmpr1b, Notch2, and S1pr1 are known in the case of quiescent NSCs. In fact, BMP signaling is known to maintain the NSC quiescence in an autocrine manner, and further this signaling must be downregulated for the subsequent activation of the quiescent NSCs 26. On the other hand, Notch signaling is known to be involved in a paracrine manner where Notch ligands are expressed by active NSCs and inhibition of Notch signaling increased the active stem cell population 26. Role of S1pr1 in maintaining NSC quiescence has been demonstrated in an independent study where addition of S1pr1 agonist sphingosine‐1‐phosphate significantly affected the activation of quiescent NSCs 28. In the case of active NSCs, we identified Egfr signaling in addition to five other receptor‐mediated signaling pathways(Fig. S3). Moreover, role of Egfr signaling for maintaining active NSCs is well established and in fact Egfr is used as a marker to isolate those cells 28. In principle, such an approach that focuses on sustained signaling pathways conserved within a cellular population could enable identification of niche‐mediated regulators of stem cell phenotypes without the knowledge of niche. Mean field approximation: caveats and comparisons to other models As a result of mean field approximation, transient fluctuations in signaling events that arise due to the dynamic nature of the niche are ignored, as they do not display any functional consequence for the maintenance of stem cell states. In this context, it must be noted that in addition to sustained signals, a cellular niche can also propagate transient, but functionally relevant signals induced by feedback mechanisms to robustly maintain tissue homeostasis 35. Other transient, yet functionally important signals could arise due to perturbations such as cellular injury or genomic mutations. The latter signals generally induce stem cell phenotypic transitions (i. e. , from quiescent to active/proliferative state 5 ), but are less likely to stably maintain the existing stem cell phenotype 36, 37. Therefore, it must be emphasized here that the mean field view of stem cell–niche interactions is valid for identifying the signaling pathways responsible for constant maintenance of cellular phenotypes and not for transient signals that can potentially trigger phenotypic transitions. Furthermore, identification of conserved signaling can provide accurate descriptions of individual cellular behavior only when heterogeneity within a defined population reflects functionally meaningless fluctuations around a single cellular state and not otherwise. Therefore, for the approach to yield accurate results, the characterization of the cellular populations needs to be accurate. Greater emphasis on the identification of sustained signaling pathways that are conserved within a cellular population exhibiting a common phenotype is a major outcome of the mean field approximation of the niche. Even though this outcome appears similar to pathway enrichment analysis that has been routinely utilized over the past decade 38 to identify deregulated (signaling or metabolic) pathways, in actual practice the idea has not been identification of sustained signaling pathways conserved within a cellular population. Moreover, several transient signaling pathways could be identified as deregulated due to indirect effects (of mutations, differences in the niche composition etc. ) and not as a cause for observed phenotypic difference. However, those signaling pathways that are constantly active are more likely to be the cause for stable maintenance of a specific cellular phenotype. Such a view offered by our hypothesis is fundamentally different from the prevailing view, and is often overlooked due to its apparent simplicity. Furthermore, it must be mentioned that computational analysis based on such a view enhances the utility of single‐cell omics data generation and adds value to current development of analytical methods 39 to decipher hidden patterns in such high‐resolution datasets. Given the complexity involved in stem cell–niche interactions, computational systems biology approaches have been useful in modeling their behavior. In fact, computational methods have been proposed to model interactions between stem cells and niche components 16, 17, 18, 19, 20, 40, 41, 42. These methods could be broadly classified into two major categories, (a) methods that aim to capture the population level behavior of stem cell–niche interactions by modeling cell–cell interaction dynamics and (b) construction of intercellular (cell–cell) interaction networks based on gene expression data. The first category of methods model the interaction dynamics of stem and progenitor cells using rate equations that describe the birth and death processes of each cell type and their interdependence on each other 16, 17, 18, 19, 20, 40. Such models are most commonly employed for studying stem cell–niche interaction dynamics and characterizing the system steady‐state properties in order to understand tissue homeostasis, and how perturbations (in the form of diseases) could affect the original steady states. A typical bottleneck in such dynamical models is the lack of knowledge of parameters or probabilities (such as, self‐renewal rate, synthesis rate of differentiated cells, death rates of stem and daughter cells) that govern the system dynamics. In addition to a lack of knowledge on parameters, even the precise composition of the cellular niche is far from being completely known, thereby rendering the development of such dynamical models difficult. Furthermore, these models tend to be powerful for a descriptive analysis of the system dynamics rather than being predictive in nature. In contrast, our proposed approach does not require the explicit knowledge of niche components or the parameters that govern the stem cell–niche interactions to identify niche‐mediated regulators of stem cell phenotype. The second category of models are based on construction of intercellular interaction networks based on gene expression data 41, 42. This approach attempts to build cell–cell interaction networks based on sorting of different cell populations followed by high‐throughput profiling, to define intercellular signaling between phenotypically defined populations of stem, progenitor, and mature cell types. This approach, although not affected by a lack of knowledge on parameters, nevertheless requires sorting and profiling of several cell types to construct the cell–cell interaction network. This is a major limitation since the cell types that truly serve as niche cells in several stem cell systems is not well characterized, and therefore cannot be sorted and profiled easily. However, our proposed strategy requires single‐cell gene expression profiling of only the stem cells with distinct phenotypes (like quiescent and active) and does not require expression profiling of the niche cells. This dramatically simplifies the isolation of the cells, data generation and further downstream analysis since only stem cells are required to be isolated and profiled without the necessity of profiling the niche cells. Although every stem cell system is unique in the way it is regulated by its niche 3, several recent studies in different stem cell systems have observed that stem cell states are determined by constant activation/inhibition of specific pathways by the constitutive influence of its niche 15, 28, 43. The presence of certain constantly activated/inhibited signaling pathways maintained by their niche appears to be the commonality in different stem cell systems. This offers possibilities to address the complexity of stem cell–niche interactions without the explicit niche characterization. Especially, the rapid advancements in single‐cell profiling technologies enable the dissection of cellular populations in greater detail. Moreover, the development of computational systems biology approaches based on the mean field approximation hypothesis finds a natural application of such increasingly available data for identifying signaling pathways that are constantly active in all cells within a population exhibiting the same phenotype. Importantly, identification of such niche determinants has several implications in regenerative medicine. Potential applications for regenerative medicine and tissue engineering The stem cell niche contains a rich and diverse set of cues that impinge constantly on stem cells that can be modulated for therapeutic gain 9, 10. Understanding and characterizing the niche determinants has potential applications in regenerative medicine and stem cell therapies for degenerative diseases of liver, heart, lung, and brain. Limited functional integration of transplanted stem cells into the target tissue possibly due to negative regulatory effect of diseased niche is currently a major challenge 10. In this regard, promoting regeneration by harnessing the latent regenerative potential of endogenous stem/progenitor cells has been used as an alternative regenerative medicine strategy in order to overcome the current translational bottlenecks associated with cell transplantation 44. For example, in the case of multiple sclerosis (MS), a demyelinating disease due to progressive failure of remyelination in the CNS due to aging, endogenous activation of oligodendrocyte precursors by mimicking a youthful microenvironment have been proven useful to promote remyelination in certain MS disease models 45, 46. In order to achieve this, identification of strategies for the activation of endogenous repair mechanisms to promote tissue regeneration in situations in which it does not occur normally is necessary 44. Within this context, the proposed approach for the identification of conserved signaling pathways under diseased and healthy niche conditions (determined by their physiological cues) can enable the development of potential strategies to modulate endogenous stem cell activity by either counteracting the effect of diseased niche or by mimicking the effect of healthy niche in the diseased counterpart. Such intervention strategies would be intended to make endogenous stem cells resistant to the perturbed signals in the diseased state and to sustain long‐term function. Another potential application where the knowledge of niche determinants can provide useful insights is in the area of tissue engineering. In particular, it is relevant in the context of ex vivo tissue engineering, where the main goal is to have the cells surviving and functioning in an optimal environment without necessarily having to replicate the in vivo conditions. In this regard, our proposed approach can enable identification of key factors that are responsible for maintaining a given cellular phenotype in vivo can aid defining better culture conditions for long‐term phenotype maintenance. For instance, long‐term maintenance of primary hepatocytes in a defined culture medium is still a challenge 47. Specifically, identification of a culture system that can facilitate long‐term maintenance of hepatocytes is advantageous for clinical applications such as drug screening and toxicity tests. Conclusions In general, cellular populations with the same functional phenotype exhibit a certain degree of heterogeneity in their molecular profiles due to intrinsic stochasticity in the transcriptional and translational program. Furthermore, the dynamical nature of the niche can perpetuate noisy fluctuations in stem cell signaling pathway activities. Therefore, stem cells face an acute challenge of robustly maintaining their state in the presence of intracellular and extracellular fluctuations, while responding precisely to developmental cues from the niche. The existence of a common stem cell phenotype within a spatial compartment of a tissue, despite the dynamic nature of the niche, seems contradictory. Our mean field view of stem cell–niche interactions provides an explanation for such a seemingly contradictory observation. By focusing on the net effect of the niche created by the mean field after disregarding internal and external fluctuations, it points to the existence of constantly activated/inhibited signaling pathways that maintains the stem cell state in response to the niche. In fact, identification of conserved signaling pathways that are constantly activated/inhibited in all cells in a stem cell population exhibiting the same phenotype will confirm our hypothesis. Furthermore, the development of single‐cell data‐based computational methods relying on a mean field view of the niche can aid in identification of niche determinants by simplifying the complexity of stem cell–niche interactions. Importantly, the knowledge of niche determinants will aid developing regenerative medicine strategies to enhance/modulate stem cell activity for the treatment of injury, disease, or age‐related dysfunctions. In addition, our approach is suitable for identifying factors that can facilitate long‐term maintenance of cells under culture conditions. Thus, combining recent developments in single‐cell technologies and stem cell research with the systems biology approaches discussed here should enable us to more accurately identify niche determinants, which in turn could lead to the implementation of more feasible strategies in regenerative medicine and tissue engineering. Author contributions AdS conceived the idea. SR performed the analysis. Both the authors wrote the manuscript. Supporting information Fig. S1. Inferred Steiner trees for quiescent and active NSCs. It can be seen that the dummy node in the center is the root node that connects with all receptors/ligands. The inverted triangles depict receptor molecules, circles depict signaling intermediates, and squares depict transcription factors. Click here for additional data file. Fig. S2. The figure shows the subnetworks of signaling pathways identified for quiescent NSCs. The inverted triangles depict receptor molecules, circles depict signaling intermediates, and squares depict transcription factors. The experimentally validated signaling pathways are highlighted. Click here for additional data file. Fig. S3. The figure shows the subnetworks of signaling pathways identified for active NSCs. The inverted triangles depict receptor molecules, circles depict signaling intermediates, and squares depict transcription factors. The experimentally validated signaling pathways are highlighted. Click here for additional data file.

10. 1002/1878-0261. 12323

Molecular Oncology

A combined tissue‐engineered/

Patient‐tailored therapy based on tumor drivers is promising for lung cancer treatment. For this, we combined in vitro tissue models with in silico analyses. Using individual cell lines with specific mutations, we demonstrate a generic and rapid stratification pipeline for targeted tumor therapy. We improve in vitro models of tissue conditions by a biological matrix‐based three‐dimensional (3D) tissue culture that allows in vitro drug testing: It correctly shows a strong drug response upon gefitinib (Gef) treatment in a cell line harboring an EGFR ‐activating mutation ( HCC 827), but no clear drug response upon treatment with the HSP 90 inhibitor 17 AAG in two cell lines with KRAS mutations (H441, A549). In contrast, 2D testing implies wrongly KRAS as a biomarker for HSP 90 inhibitor treatment, although this fails in clinical studies. Signaling analysis by phospho‐arrays showed similar effects of EGFR inhibition by Gef in HCC 827 cells, under both 2D and 3D conditions. Western blot analysis confirmed that for 3D conditions, HSP 90 inhibitor treatment implies different p53 regulation and decreased MET inhibition in HCC 827 and H441 cells. Using in vitro data (western, phospho‐kinase array, proliferation, and apoptosis), we generated cell line‐specific in silico topologies and condition‐specific (2D, 3D) simulations of signaling correctly mirroring in vitro treatment responses. Networks predict drug targets considering key interactions and individual cell line mutations using the Human Protein Reference Database and the COSMIC database. A signature of potential biomarkers and matching drugs improve stratification and treatment in KRAS ‐mutated tumors. In silico screening and dynamic simulation of drug actions resulted in individual therapeutic suggestions, that is, targeting HIF 1A in H441 and LKB 1 in A549 cells. In conclusion, our in vitro tumor tissue model combined with an in silico tool improves drug effect prediction and patient stratification. Our tool is used in our comprehensive cancer center and is made now publicly available for targeted therapy decisions.

Abbreviations 17AAG 17‐ N ‐allylamino‐17‐demethoxygeldanamycin 2D two‐dimensional 3D three‐dimensional ADME absorption, distribution, metabolism, excretion AICAR 5‐aminoimidazole‐4‐carboxamide ribonucleotide BioVaSc ® Biological Vascularized Scaffold COSMIC Catalogue Of Somatic Mutations In Cancer DRPs differentially regulated proteins DrumPID Drug‐minded Protein Interaction Database Gef gefitinib HE hematoxylin and eosin HPRD Human Protein Reference Database HSP heat shock protein NGS next‐generation sequencing NSCLC non‐small cell lung cancer PK phospho‐kinase RTK receptor tyrosine kinase SMILES Simplified Molecular Input Line Entry Specification SQUAD Stardardized Qualitative Dynamical Modelling Suite TKI tyrosine kinase inhibitor 1 Introduction In the highly mortal lung cancer, next‐generation sequencing (NGS) approaches successfully reveal driver mutations to stratify lung cancer patients for targeted therapies (Buettner et al. , 2013 ). Tyrosine kinase inhibitor (TKI) treatment shows remarkable response rates, exemplified by EGFR inhibitors in patients with activating EGFR mutations (Ciardiello et al. , 2004 ; Paez et al. , 2004 ; Russo et al. , 2015 ). However, often the therapy is only initially successful and then followed by secondary resistance. Unfortunately, tumors with KRAS mutations are primarily resistant to targeted therapies and comprise about 30–40% of all patients (Sequist et al. , 2011 ). Due to poor correlations of preclinical in vitro data to drug efficacy in patients, particularly in the field of cancer (Bhattacharjee, 2012 ), new 3D tumor models arise, such as spheroids, microfluidic devices, organoids, and matrix‐based approaches (Alemany‐Ribes and Semino, 2014 ; Edmondson et al. , 2014 ; Xu et al. , 2014 ). The generally high proliferation rate in 2D cell cultures is one reason for false‐positive predictions of cytostatic compounds (Cree et al. , 2010 ). Decreased proliferation of tumor cells corresponding to clinical specimens was demonstrated on our scaffold (Göttlich et al. , 2016 ; Nietzer et al. , 2016 ; Stratmann et al. , 2014 ) originating from the Biological Vascularized Scaffold (BioVaSc ® ) (Linke et al. , 2007 ; Schanz et al. , 2010 ). It maintains the extracellular matrix, including structures of the basement membrane, enabling physiological anchorage of epithelial cells. Earlier, we combined the tissue‐engineered lung tumor model with its in silico representation to investigate tumor and, thereby, drug‐relevant dependencies – also in the context of resistance (Göttlich et al. , 2016 ; Stratmann et al. , 2014 ). In this study, we introduce a patient stratification tool according to tumor drivers as a promising decision tool for precision medicine in lung cancer. This is exemplified here by studying individual in vitro cell lines and their differing drug responses in 2D and 3D, and by integrating these data in corresponding in silico analyses for target predictions. The tool is generic and provides a rapid stratification pipeline that can support tumor boards to utilize more and more clinically available NGS data from individual patients. We studied how a biological matrix‐based 3D tissue culture allows in vitro drug testing of relevant lung cancer subgroups. To unravel signal cascade outputs in more detail, we investigated apoptosis and proliferation as drug responses. Regarding the EGFR inhibition with the TKI gefitinib (Gef) in a cell line carrying the corresponding biomarker, we observed an enhancement in apoptosis induction compared to 2D. Moreover, we exemplified our stratification tool by looking at responses of two further cell lines (A549, H441) harboring KRAS mutations to the HSP90 inhibitor 17AAG. In contrast to the EGFR inhibition, in this setting only the 3D system could predict no drug efficiency in line with clinical findings. Therefore, we analyzed differences in signaling changes upon treatment between cell lines and between 2D and 3D conditions. Using the experimental data of the 3D tissue model, we created (a) in silico cell line‐specific topologies of the centrally involved proteins including their logical connectivity. Based on these data, (b) dynamic in silico simulations mirrored the differences in cellular responses apparent in the experiments. Considering protein neighbors of central important signaling cascades and cell‐specific mutations from databases resulted (c) in larger in silico networks which were next screened in silico for individual therapeutic options for each cell line. Resulting drug suggestions reflect clinical experiences and include comprehensive FDA‐approved treatment options. In its unique combination, the tool raises hopes of efficiently exploiting upcoming sequence information of patient tumors in the near future for targeted therapy. 2 Results 2. 1 Analysis path To exemplify the process of how a single patient's sequence could be integrated into our new preclinical prediction tool, we chose three cell lines representing different patient subgroups regarding KRAS mutation (HCC827: KRAS wild‐type; A549: KRAS mutant, independent; H441: KRAS mutant, dependent) (Singh et al. , 2009 ). To unravel the complex interdependent signaling network in lung cancer in different mutational backgrounds, we experimentally measured, in a global approach using phospho‐arrays signaling, changes to Gef and 17‐allylamino‐17‐demethoxygeldanamycin (17AAG) treatment in our three different cell lines (Figs 1, 2, 3 ; Table 1 ; Figs S1–S3 ; first simulation is in S2 ; further in silico analyses in S4–S7 ). Firstly, we recognized that with the EGFR inhibitor Gef affected proteins are roughly the same in HCC827 (EGFR mutated) in 2D and 3D models (Table 1 A, Fig. S1 ). By analyzing signaling changes upon the HSP90 inhibitor treatment by phospho‐arrays and western blot in all three cell lines in 2D and in 3D, it became obvious that besides MET, changes between the 2D and the 3D models concern mostly p53 and HSP60 (Table 1 B; Figs 4 A and S3 ). Figure 1 Improved reflection of tumor characteristics by the 3D tissue‐engineered lung tumor model. Tested tumor cell line populations display more homogeneous marker expression in 3D as well as reduced proliferation correlating to tumors. (A) Cells cultured in 2D and 3D conditions shown with immunofluorescence by double stain for E‐cadherin and β‐catenin. Green arrows indicate positive cells and white arrows negative. Scale bars are 50 μm. (B) Paraffin‐embedded adenocarcinoma from patient biopsy was immunofluorescence‐double‐stained against E‐cadherin and β‐catenin. Scale bar is 100 μm. (C) The expression of the proliferation marker Ki67 was detected by immunofluorescence staining of 2D and 3D cultured HCC 827 cells, as well as in vivo tissue from a patient biopsy. Scale bar is 100 μm. Figure 2 Biomarker‐dependent response upon EGFR inhibition is improved in 3D and can also be simulated in silico. (A) Cells cultured in 3D that were either treated with 1 μ m Gef or used as untreated controls were paraffin‐embedded and HE ‐stained. Scale bar is 100 μm. (B) The proliferation rate (proliferative cells per total cell number) was determined by counting Ki67‐positive cells from immunofluorescence staining in 10 images per sample. Total cell number was quantified by DAPI counterstaining. *** P < 0. 001, n ≥ 4. (C) Apoptosis was investigated by M30 CytoDeath™ ELISA. Therefore, supernatants of treated and untreated samples were collected prior to and at 24, 48, and 72 h after treatment. Concentrations of M30 in samples after treatment were normalized to T0 values from samples taken before treatment and related to untreated samples (red line). *** P < 0. 001, n ≥ 4. (D) In silico simulation of the Gef treatment (right, pink curve full on at 1. 0) shows reduced proliferation (right, black curve) only in HCC 827 cells and higher apoptosis (right, gray curve), as compared to untreated cells (left, pink curve switched off at 0. 0). Figure S2 A shows the in silico topology and Fig. S2 B the simulations for A549 and H441. * P < 0. 05, ** P < 0. 01 Figure 3 Effects of the HSP 90 inhibitor 17 AAG diminish in 3D and cannot be aligned to the biomarker KRAS (A549/H441). Strong treatment responses regarding viability, proliferation, and apoptosis can be observed only in 2D conditions. (A) Cells cultured in 2D conditions were treated with different concentrations of the HSP 90 inhibitor 17 AAG. Viability was determined after 3 days of treatment by a CellTiter‐Glo ® Luminescent Cell Viability Assay. n ≥ 4. (B) 3D cultured cells were treated with 0. 25 μ m 17 AAG, paraffin‐embedded, and HE ‐stained. Scale bar is 100 μm. (C) The proliferation rate in 2D and 3D was determined by counting Ki67‐positive cells from immunofluorescence staining in 10 images per sample. Total cell number was quantified by DAPI counterstaining. * P < 0. 05, n ≥ 4. (D) Apoptosis was investigated by M30 CytoDeath™ ELISA. Therefore, supernatants of treated and untreated samples were collected prior to and at 24, 48, and 72 h after treatment. Concentrations of M30 in samples after treatment were normalized to T0 values from samples taken before treatment and related to untreated samples (red line). * P < 0. 05, *** P < 0. 001, n ≥ 4. ** P < 0. 01. Table 1 Comparison of the phosphorylation data showing different regulation between the cell lines in the 2D and 3D system for Gef and 17AAG HCC827 A549 H441 2D 3D 2D 3D 2D 3D (A) Gef treatment a pEGFR ↓ ↓ 0 0 const. const. pErbB2 ↓ 0 0 0 0 0 pMET ↓ ↓ 0 0 const. const. (B) 17AAG treatment b pEGFR ↓ ↓ 0 0 ↓↓↓ ↓↓ pErbB2 ↓ ↓ 0 0 ↓ ↓ pErbB3 ↓ ↓ 0 0 ↓ ↓ pMET ↓↓ ↓ 0 0 ↓↓↓ ↓ pc‐Ret ↓ ↓ 0 0 ↓ ↓ pVEGFR2 0 0 0 0 ↓ ↓ pFGFR3 0 0 0 0 0 ↓ p‐p53 (S46) ↑↑↑ const. 0 0 const. ↑↑↑ HSP60 const. const. ↑ ↑↑ ↑ const. a Based on the RTK array data, this is a qualitative summary of all proteins measured, showing a phosphorylation difference in at least one cell line upon Gef treatment (0 reflects no activation, and const. means no activation change after treatment). Experimental data are shown in Fig. S1. b Based on the western blots (semiquantitative, more than one arrow possible) and RTK array data (qualitative, only one arrow possible), this is a summary of all proteins measured, showing a phosphorylation difference in at least one cell line upon 17AAG treatment (detailed experimental data shown in Figs 4 A and S3A; 0 reflects no activation, const. means no activation change after treatment). John Wiley & Sons, Ltd Figure 4 Signaling changes after HSP 90 inhibition differ between 2D and 3D and between the different cell lines and are integrated into in silico topologies. (A) Cells cultured in 2D and 3D were treated with 0. 25 μ m 17 AAG for 24 h (2D) or 72 h (3D). The signaling changes of different phospho‐proteins were analyzed by western blot. The same lysates were used for the pEGFR and ph‐p53(S46) blots of all three cell lines in 2D and 3D HCC 827 and for ph‐p53(S46) and pMET blots in 3D H441; thus, the same β‐actin loading control is shown below these phospho‐proteins. (B) DRPs from the in vitro 3D system are connected in silico to the central tumor signaling cascade. Here, we show the topology shared between all three cell lines. Colors reflect important input (treatment), signaling proteins, and cellular output (proliferation and apoptosis). Proteins (‘nodes’) from the topology of Stratmann et al. ( 2014 ) are bold rimmed and have an olive background; proteins added specifically to the in silico topology are presented as simple boxes; protein node colors are as in the simulation curves; cell line‐specific proteins (‘nodes’) appear as plus (+). Specific topologies and simulation results for each cell line are given in the Supporting information. Regarding in silico analyses, we first set up cell line‐specific in silico topologies by integrating important signaling nodes that distinguished the cell lines upon Gef and 17AAG treatment into our basic in silico topology (Table 2 ; Stratmann et al. , 2014 ). The nodes of this basic topology are marked in all newly generated topologies with bold printed borders. After the generation of these cell line‐specific in silico topologies, we mirrored the in vitro treatment response of Gef and 17AAG, by applying semiquantitative Boolean simulations using the software squad (Stardardized Qualitative Dynamical Modelling Suite). Based on the logical connectivity of each cellular topology, this software models the dynamic evolution of the included signaling cascades using exponential functions (Di Cara et al. , 2007 ). Furthermore, different activation strengths for each node of the signaling cascade are considered in the simulations that were necessary to adapt the in silico simulation results to the in vitro results for differences of 3D and 2D cultures. Input into the topology of Fig. 4 B is listed in Table 3 A and B for 3D conditions (further network analyses in Tables S1–S3 ) and in Box S2 for 2D conditions. Simulations’ output of 3D conditions is presented in Figs 2 D and S2 for Gef and in Fig. 5 for 17AAG treatment. Simulation results in 2D conditions of KRAS‐mutated cell lines are represented in Fig. S6 for Gef and in Fig. S7 for 17AAG treatment. Table 2 Cell line‐specific proteins introduced in addition to the original topology. a Cell line Gef 17AAG HCC827 (3D) MET cascade + MET cascade + ; Erb2 cascade + ; Erb3 cascade + ; c‐RET cascade + ; HSP90; HSP60; HIF1A; p53 A549 (3D) – HSP90; HSP60 HIF1A; p53 H441 (3D) MET cascade + MET cascade + ; HSP90; HSP60; HIF1A; p53; Erb2 cascade + ; VEGF2 cascade + ; Erb3 cascade + ; c‐RET cascade + ; FGFR3 cascade + a Listed are the proteins extending the network of Stratmann et al. ( 2014 ) that responds upon Gef or 17AAG treatment and which were measured in arrays and western blots. According to interaction analysis p53, HSP60, HIF1A and HSP90 are added as cascades around 17AAG. A plus (+) indicates cell line‐specific protein nodes added according to the experimental data. Further cell line‐specific protein nodes according to COSMIC and relevant to our in silico network as being close to or in our signaling cascades are listed in Table S3 (9 in A549, 18 in H441). Cell‐specific mutations analyzed in detail are shown in Figs 6 and 7. A complete list of all cell line‐specific mutations known is given in Table S2. John Wiley & Sons, Ltd Table 3 Different activation strengths for each node for in silico simulations. Cell line‐specific differences in pathway activities on (A) Gef a and (B) 17AAG a and (C) AMPK activator and HIF1A inhibitor b Cell line Parameter (−) gef (+) gef (A) HCC827 (3D) MET 0. 22 0. 15 EGFR 0. 22 0. 12 EGF‐EGFR 0. 22 0. 12 FLIP 0. 6 0. 5 A549 (3D) KRAS c 0. 413 0. 413 FLIP 0. 6 0. 6 H441 (3D) KRAS c 0. 43 0. 43 EGFR 0. 205 0. 205 MET 0. 205 0. 205 FLIP 0. 4 0. 4 Cell line Parameter (−) 17AAG (+) 17AAG (B) HCC827 (3D) EGFR 0. 25 0. 2 MET 0. 25 0. 2 Stress 0. 6 0. 6 HSP60 0. 4 0. 4 FLIP 0. 7 0. 6 p53 0. 4 0. 4 (Erb2/Erb3/c‐RET) 0. 07 0. 07 Erb2/Erb3/c‐RET 0. 1 0. 07 (EGF‐EGFR) 0. 1935 0. 1935 (MET) 0. 1935 0. 1935 A549 (3D) KRAS c 0. 352 0. 352 Stress 0. 5 0. 5 FLIP 0. 7 0. 33 p53 0. 0 0. 0 HSP60‐act 0. 4 0. 4 H441 (3D) KRAS c 0. 43 0. 43 EGFR 0. 05 0. 01 Erb2/Erb3/c‐RET/FGFR3 0. 04 0. 03 MET 0. 05 0. 02 Stress 0. 7 0. 7 p53‐act 0. 65 0. 65 HIF1‐act 0. 65 0. 65 HSP60‐act 0. 05 0. 05 VEGFR2 0. 35 0. 33 FLIP 0. 75 0. 75 PTEN 0. 41 0. 41 Cell line Parameter (−) AICAR (+) AICAR (−) PX‐478 (+) PX‐478 (C) A549 (3D) KRAS c 0. 345 0. 345 Stress 0. 5 0. 5 FLIP 0. 6 0. 2 p53 0. 0 – p53‐act – 0. 1 low glucose 0. 3 0. 3 HIF1‐act 0. 8 0. 8 mTOR‐act 0. 65 0. 65 H441 (3D) KRAS c 0. 43 0. 43 EGFR 0. 05 0. 05 Erb2/Erb3/c‐RET/FGFR3 0. 04 0. 04 MET 0. 05 0. 05 Stress 0. 7 0. 7 p53‐act 0. 65 0. 65 HIF1‐act 0. 65 0. 65 HSP60‐act 0. 05 0. 05 VEGFR2 0. 35 0. 05 FLIP 0. 75 – casp3‐act – 0. 75 PTEN 0. 41 0. 41 a Cell line‐specific receptor or pathway activity of proteins according to the experimentally determined differences in response behavior (apoptosis, proliferation, RTK, and western blot data); all other proteins were simulated with no specific activation. (−) Treatment activation at stage 0; (+) treatment activation at stage 1. b For the simulation of the AMPK activator AICAR in A549 and the HIF1A inhibitor PX‐478 in H441, we used the cell line‐specific activity from the untreated cells of the 17AAG treatment (Table 3 C); all other proteins were simulated with no specific activation. (−) Treatment activation at stage 0; (+) treatment activation at stage 1. c Constant activation, as there is a KRAS mutation in these cell lines. John Wiley & Sons, Ltd Figure 5 Cell line‐specific in silico simulations for 17 AAG treatment according to data from the 3D system. Simulations of the 17 AAG treatment reflect the in vitro data. Coloring of the curves is according to the network node colors shared for all three cell lines shown in Fig. 4 B. Cell line‐specific pathway differences included are given in Table 2. Top: Simulation of the 17 AAG treatment in HCC 827 cells (right, red curve at full activation) results in slightly induced apoptosis (gray curve at 0. 2) and unchanged proliferation (black curve), as compared to untreated cells (left, red curve at 0. 0, no treatment). Middle: The in silico simulation of the 17 AAG treatment for A549 shows only low apoptosis induction (0. 2); we see no therapeutic effect on proliferation (black curve, dots) compared to untreated cells. However, HSP 60 (black curve, squares) is induced after 17 AAG treatment, similar to the in vitro data. Bottom: In H441 cells, apoptosis is not elevated over time and no effect on proliferation can be obtained. p53 (pink curve) is induced after 17 AAG treatment and correlates with the in vitro data. To reveal – in a systemic approach – further relevant cell line‐specific drug targets in KRAS ‐mutated conditions in the 3D system, we reconstructed two larger in silico networks for A549 and H441 cells. Therefore, we searched in Human Protein Reference Database (HPRD; Table S1) the interacting neighbors of the nine upon 17AAG treatment between A549 and H441 differentially regulated proteins (DRPs) (Table 1 B, Fig. 6 ) and identified individual promising drug targets by mapping these to cell‐specific mutations in COSMIC (Catalogue Of Somatic Mutations In Cancer) generating thereby two cell‐specific networks for A549 and H441 cells. From networks analyses, we expanded the in this study created topology from Fig. 4 B (marked with olive gray background in Fig. 7 A, C) further with gain or loss of function mutations and other important factors by adding activated or inhibited nodes. In subsequent simulations we could predict optimal drug targets in a specific mutational background of KRAS ‐mutated tumors. In Table 3 C, input into topologies and subsequent simulations in Fig. 7 are given. Matching drugs were suggested by screening of our DrumPID (Drug‐minded Protein Interaction Database) (Kunz et al. , 2016 ) screening tool for available target‐specific test substances. Figure 6 KRAS signature development and individual target predictions by generation of HPRD networks. We generated a network around KRAS, according to the experimentally validated DRPs between both KRAS‐mutated cell lines (H441, A549) in the 3D system (Table 1 B; 17 AAG treatment), and included their direct protein interaction partners using the genomewide HPRD. The resulting larger KRAS interaction network includes 556 proteins (= nodes) and 680 protein–protein interactions (= edges), around nine strongly DRPs ( EGFR, ErbB2, ErbB3, MET, FGFR 3, c‐Ret, VEGFR 2, p53, and HSP 60). (A) A Venn diagram compares cell line‐specific mutations. Mapping of cell line‐specific protein mutations (573 for H441 (blue) and 361 for A549 (yellow) from the COSMIC database) against the 556 proteins from the network around KRAS results in 18 H441‐specific mutations and in nine A549‐specific mutations which were included in each cell line‐specific in silico topology to yield the network. Details are given in the Supporting information, and key network differences are shown in B and C. (B) A549‐specific network: represents neighbor proteins that we could target if we consider the experimental data and directly interacting protein neighbors (from HPRD ; functional clusters in Fig. S5A ). As drug targets do not appear for these small modules from key signaling proteins, we considered experimental derived proteins (red) with all first‐degree neighbors, HSP 90 (orange rectangle), and additionally direct neighbors to cell line‐specific mutations (in yellow, suspected ‘driver mutations’). Direct neighbor proteins are labeled in lavender, in cyan are neighbors from neighbors, which are also mutated. The black square ( AMPK, interactor of p53 and LKB 1) indicates a promising drug target (screening procedure given in Box S1 ). (C) H441‐specific network: shows neighbor proteins that we could target, if we consider the experimental data, HSP 90 (orange rectangle) and directly interacting protein neighbors from HPRD (functional clusters in Fig. S5B ). Directly interacting neighbors are shown (lavender, labeling binary interactions). As drug targets do not appear for these small modules from key signaling proteins, we considered all experimental determined nodes (red) with all first‐degree neighbors integrating cell line‐specific H441 mutations (in blue, suspected ‘driver mutations’; EGFR and p53 labeled in red with blue circles as they are array nodes and mutated). Protein interactors according to HPRD are labeled in lavender; in cyan are neighbors from driver mutations, also showing a mutation in H441. The square ( HIF 1A) indicates a promising drug target (screening procedure given in Box S1 ). Figure 7 In silico topologies and simulations of AMPK and HIF 1A treatment. Cell‐specific network extensions according to the experimental data (Table 2 ) are mapped into the shared topology (bold nodes from basic topology from Stratmann et al. ( 2014 ), olive shade for topology nodes from Fig. 4 B). Furthermore, AMPK as a relevant target for A549 (network in (A)) and HIF 1A as a target for H441 (topology in (C)) are included (nodes equivalent to the 17 AAG treatment are deposited in olive, protein node colors are the same as in the simulation curves). Both protein targets were integrated with their direct interacting protein neighbors in the cell‐specific networks to mirror in silico individual therapy. In (B) and (D), the cell‐specific topologies are next simulated dynamically, and selected trajectories of protein node activities were plotted, showing the effects of the potential drug candidate AICAR as an AMPK activator for A549 (B), and the HIF 1A inhibitor PX ‐478 for H441 in (D) to illustrate the in silico screen of different drugs in the two cell line‐specific topologies. (B) Simulation of AMPK activation in A549 cells (right, red curve at stage 1) results in higher apoptosis (pink curve) and reduced proliferation (salmon curve), as compared to untreated cells (left, red curve at 0. 0, no activation). (D) The in silico simulation of the HIF 1A inhibition for H441 (right, olive curve at full activation) shows higher apoptosis (black curve) and reduced proliferation (salmon curve), as compared to untreated cells (left, olive curve at 0. 0, no activation). 2. 2 Tissue‐engineered lung tumor models resemble tumor specimens Firstly, we looked at molecular markers for tumors and tissue differentiation and their variances. We observed homogenous E‐cadherin/β‐catenin localization in fluorescence staining of 3D models as well as tumor specimens, whereas 2D models showed high variation ranging from strongly stained to completely negative tumor cells (Fig. 1 A, B). By comparison to 2D models, we demonstrated with Ki67‐staining the reduction in proliferation rate in 3D models to levels that correlate to lung adenocarcinoma samples (Fig. 1 C). 2. 3 Enhanced biomarker‐dependent drug response to EGFR inhibition in the 3D model As a test for clinically applied biomarker‐guided anti‐EGFR therapy, we compared A549 and H441 (EGFR wild‐type) with HCC827 cells (activating EGFR mutation). Responses upon 3 days of Gef treatment were biomarker‐dependent, as represented by hematoxylin and eosin staining (HE staining) in 3D models (Fig. 2 A), proliferation reduction (Fig. 2 B), and apoptosis induction (Fig. 2 C). Although the proliferation in 3D conditions was reduced to in vivo like rates (Stratmann et al. , 2014 ), treatment with Gef decreased the proliferation further by about 80%, as also observed in 2D. However, biomarker‐related apoptosis induction upon Gef treatment in HCC827 was significantly enhanced in 3D conditions (about 3. 5 to 6‐fold increase), compared to 2D conditions (about 2. 5‐fold increase), which suggests better specificity for the 3D system. Signaling analyses by receptor tyrosine kinase (RTK) and phospho‐kinase (PK) array experiments are provided for the 2D and 3D systems (Fig. S1, Table 1 A). Complementing these biomarker‐dependent drug responses to Gef, we set up an in silico network of key pathways for the proliferative and apoptotic response, to model the observed in vitro responses of each of the three cell lines. An in silico topology was previously developed for HCC827 and A549 (Stratmann et al. , 2014 ). This was extended by those cell line‐specific proteins and pathways (Fig. S2A ; previous network proteins are in bold and with olive background) which showed signaling changes upon the Gef treatment in the experiments (Table 2 ). For the newly investigated H441 cell line, we used the A549 in silico topology as basis. Specifically, we integrated the MET signal transduction cascade for HCC827 and H441. We then applied the squad software to simulate the Gef treatment responses for all cell lines, using initial node stimulations based on the mutational background and the experimental results on protein phosphorylation (prestimulation in Table 3 A; method in the Supporting information). Our simulation of HCC827 with Gef treatment (Fig. 2 D) compared to untreated cells demonstrates, as for the in vitro results, reduced proliferation and higher apoptosis over time. Results of the simulations for A549 and H441 are represented in Fig. S2B. 2. 4 Chemoresistance against HSP90 inhibition in 3D models align to clinical observations 2D models and animal experiments predict HSP90 inhibitor efficiency in KRAS ‐mutated tumors (Acquaviva et al. , 2012 ; Sos et al. , 2009 ). As known from 2D in vitro screens, HCC827, A549, and H441 exhibit different sensitivities to the HSP90 inhibitor 17AAG (Ciocca and Calderwood, 2005 ; Sos et al. , 2009 ). We observed that about 50% of the H441 cells died from 0. 25 μ m 17AAG, which decreased the viability of A549 to 5% and of HCC827 to 35%, as shown by the cell viability assay CellTiter‐Glo ® in 2D (Fig. 3 A). However, the failure of HSP90 inhibitor treatment in a clinical setting of KRAS ‐mutated tumors was reflected in 3D tissue cultures. From HE staining of 3D tumor models, after three days of 0. 25 μ m 17AAG treatment only slight effects were visible in A549 and HCC827, whereas H441 cells were completely unresponsive (Fig. 3 B). Proliferation analysis of the 2D systems predicts KRAS ‐mutated cells to be more responsive to 17AAG than KRAS wild‐type HCC827 cells. This is contrasted by only weak changes between both cell types in 3D tissue culture (Fig. 3 C). A strong apoptotic response upon 17AAG is only observed in KRAS ‐mutated A549 cells in 2D (4 to 6‐fold) but not in 3D models (1 to 2‐fold) (Fig. 3 D). All phosphorylation data from arrays and western blot experiments are summarized and compared in 2D and 3D models in Table 1 for Gef (A) and 17AAG (B) treatment. Protein nodes for in silico topology which were applied later are given in Table 2. 2. 5 Differences in signaling between 2D and 3D conditions upon HSP90 inhibitor treatment as a basis for in silico analyses Signaling responses upon application of the HSP90 inhibitor 17AAG were analyzed by comparing 2D and 3D conditions. Protein activation as observed by RTK arrays was confirmed by western blot (Fig. 4 A) and quantified (Fig. S3B ). Data indicated an inhibition of the EGFR and of MET in HCC827 and in H441 in 2D as well as 3D conditions. In western blot analysis, inhibition of MET was weaker in 3D than in 2D cultures in both cell lines. Interestingly, p53 (S46) was activated in HCC827 with 17AAG treatment in 2D, but stayed constant in 3D conditions. Vice versa, in H441 p53 was activated only in 3D conditions and remained unchanged in the 2D culture. Furthermore, HSP60 was clearly upregulated only in A549 cells under 3D conditions upon 17AAG application. Regulated proteins identified in 3D conditions upon 17AAG treatment include EGFR, ErbB2, ErbB3, MET, c‐Ret, VEGFR2, FGFR3, p53, and HSP60 (Table 1 B, Figs 4 A, and S3A ). Similar to the Gef treatment, we extended our in silico network and topology adding these experimentally measured cell‐specific proteins (Fig. 4 B). Particularly, we included for mirroring 17AAG treatment effects – next to the MET protein – ErbB2, ErbB3, and c‐RET cascade in HCC827 and H441, and also in all three cell line‐specific in silico topologies p53, HSP60, HIF1A, and HSP90, as part of the 17AAG treatment cascade (Table 2 ). For H441, we included further VEGFR2 and FGFR3, as they were downregulated in the arrays in the 3D model upon treatment with 17AAG, in contrast to the other two cell lines (Fig. S3A, Table 1 B). We show only responses for key proteins of all three cells, but we simulated the complete network responses looking at all proteins of the topology. Important aspects of the 3D tissue model upon 17AAG application (red curve at 1) are reflected by in silico simulations (Fig. 5 ): (a) In HCC827 (top), cell proliferation is unchanged and apoptosis is slightly induced compared to the untreated control, (b) in A549 (middle), proliferation is regarded as unchanged and apoptosis is only slightly induced in 3D conditions, and (c) in H441 (bottom), proliferation is unchanged and apoptosis is not induced. Notably, HSP60 is exclusively induced in A549, whereas p53 is upregulated only in H441. Moreover, based on the in silico topology connectivity, in our in silico simulation we found that beside p53 HIF1A is also upregulated in H441. For comparison, the in silico simulations can also be modified to appropriately reflect results of 2D culture. To illustrate this, we focused on the A549 and H441 cell lines and applied the same topology as for 3D, but adjusted activation levels for Gef and 17AAG treatment according to the 2D in vitro conditions (Fig. 4 B and S2A ; Table 1 ; Box S2 ): Essentially, we elevated the value of Raf to simulate higher basic proliferation in 2D, and furthermore, we changed FLIP for the higher apoptotic response in 2D upon 17AAG as this is reported to be important for higher apoptotic resistance when cells grow on collagen (Philippi et al. , 2009 ). Whereas Gef treatment simulation resulted in both KRAS ‐mutated cells in no change of proliferation and apoptosis over time (Fig. S6 ), HSP90 inhibition simulation of 2D conditions revealed in contrast to 3D a lower proliferation and an induced apoptosis over time in A549 (Fig. S7 ). However, the established tool allows us now to test and screen in silico in a systems perspective for tailored therapies according to the cell line‐specific mutational profile and tumor drivers, as detailed in the following section. 2. 6 Generation of in silico protein–protein interaction networks for cell‐specific drug target predictions in KRAS ‐mutated cells Next, as we observed signaling differences between the KRAS ‐mutated A549 and H441 cell lines, we sought to identify a KRAS complementing signature of further potential biomarkers and resultant drug targets for each cell line. For this purpose, we combined experimental data and the cell line‐specific mutational backgrounds with integrated systems biology analysis (Kunz et al. , 2017 ; Naseem et al. , 2014 ), considering direct interacting proteins and available drugs to modulate this extended network. We generated a network around KRAS by considering the DRPs of both KRAS ‐mutated cell lines (A549 and H441) upon 17AAG treatment (Table 1 B) in the 3D system and included their direct interacting proteins according to the genomewide HPRD. The resulting KRAS interaction network includes 556 proteins (= nodes) and 680 protein–protein interactions (= edges) around the nine experimentally DRPs (EGFR, ErbB2, ErbB3, MET, FGFR3, c‐Ret, VEGFR2, p53, and HSP60; Fig. S4A ). Comparing all cell line‐specific mutations known from the genomewide COSMIC database (573 for H441, blue circle, and 361 for A549, yellow circle; Table S2; Fig. 6 A) with this KRAS interaction network, we could match 18 H441‐specific mutations and nine A549‐specific mutations, as parts of our KRAS interaction network (Table S3 ). The two reconstructed cell‐specific KRAS interaction networks for A549 and H441 included these specific mutations, HSP90 as a target of 17AAG, and their direct interaction partners from HPRD (Fig. S4B–D ), which were then analyzed for functional clusters. In the A549‐specific network (322 proteins and 371 protein‐protein interactions, Fig. S4B ; extended network with 795 nodes and 1034 interactions in Fig. S4C ), we found two functional protein clusters with a strong network effect (so‐called hubs) around proteins VEGFR2, MET (experimental measurements) and CBL (mutated), and p53 (experimental) and ARID3A (mutated; Fig. S5A ). Similarly, for the H441‐specific network (903 proteins and 1119 protein–protein interactions; Fig. S4D ), we found two clusters around the proteins PRKACA (mutated) and p53 (experimental and mutated) as well as HSP90AA1, ACTA, and HIF1A (mutated; Fig. S5B ). We compared potential targets in the two cell line‐specific KRAS networks, in terms of their distance and usefulness to modulate cell‐specific signaling cascades. This yielded a highly connected network between interesting tumor drivers (Fig. 6 B for A549, driver mutations in yellow; Fig. 6 C for H441, driver mutations in blue), and cell‐specific biomarker signatures (Table 4, Box S1 ). Other cell‐specific mutations close to the central cascade are indicated by cyan (neighbors of neighbors), and unmutated interactors in lavender circles for A549 and H441, respectively. Regarding ranking of drug targets for potential clinical application, we considered proteins and connections and assigned the priority to direct neighbors, if they could be targeted easily by existing medical drugs, for example, AMPK for A549 and HIF1A for H441. All targets are ranked in Box S1. This drug‐search strategy was made possible by applying our DrumPID (Kunz et al. , 2016 ). Table 4 Overview of potential predictive markers and new therapeutic targets for the KRAS ‐mutated cell lines a John Wiley & Sons, Ltd 2. 7 AMPK and HIF1A targeting in cell‐specific in silico simulations for A549 and H441 cells Subsequently, we investigated in silico the potential therapeutic effect of AMPK as a relevant target for A549 and HIF1A as a target for H441. For this, we integrated the LKB1 cascade for A549 into the cell‐specific in silico topology that simulated the 17AAG therapy (Fig. 7 A), and further considered the connectivity of HIF1A in H441 (Fig. 7 C). Proteins in the network that correspond to the basic topology of Stratmann et al. ( 2014 ) are bold rimmed and proteins that match to the topology from Fig. 4 B have an olive background (Fig. 7 A, C). From our drug screening (Box S1 ), we identified 5‐aminoimidazole‐4‐carboxamide ribonucleotide (AICAR) in A549 as the potential activator of AMPK that is directly modulated by its interactor LKB1, which is specifically mutated in A549 (synonym: STK11; see DrumPID pathway ko04152). Similarly, we found PX‐478 as a selective HIF1A inhibitor for the H441 cell line. Based on this, we simulated their potential therapeutic effect by applying the SQUAD algorithm (Fig. 7 B, D; prestimulations in Table 3 C). For in silico simulations of AMPK activation and inhibition of HIF1A, we found induced apoptosis and reduced proliferation in each cell line over time as a desirable drug effect. 3 Discussion 3. 1 Motivation Personalized treatment strategies have to cope with highly redundant tumor pathways resulting in resistance, whereas combination therapies often show severe toxic side effects (Tannock and Hickman, 2016 ). Therefore, it is necessary to reconsider the design of clinical studies with targeted anticancer approaches. It is critical to understand the underlying dependencies in signaling networks, and to provide tools for exploiting now frequently available sequencing data from patients. Moreover, in the field of oncology, the success rate of preclinical testing is at under 5%, generating enormous financial costs (Bhattacharjee, 2012 ). Even though animal models predict toxicity quite convincingly, they tend to fail in efficacy testing (Greaves et al. , 2004 ; Kubinyi, 2003 ). In particular, for signaling analyses, mice are not adequate models due to inappropriate ligands to some centrally connected human receptors, such as MET (Francone et al. , 2007 ). Next to ethical concerns, these aspects underline the urgent need to develop novel human tumor test systems. Here, we introduce a new concept of in vitro tissue tumor models and in silico analyses to design and test individual biomarker profiles and intervention strategies. This prepares the floor for patient‐tailored clinical studies, required for personalized cancer medicine (Tannock and Hickman, 2016 ). As our main aim is to develop a powerful tool that can be implemented into the clinic by analyzing the patient's sequence data, we investigate as proof of concept in this work three individual lung cancer cell lines with known genome sequence information. However, here we show only exemplary data and not a large validation series. We are aware that exact quantitative estimates require a higher number of experiments and more cell lines with similar driver mutations. Exploring its clinical implementation, currently our in silico tool advises on a case‐by‐case basis the molecular tumor board in our comprehensive cancer center. In particular it supports to identify alternative protein targets when resistance to treatment occurs. Our human 3D tumor model generated by tissue engineering technologies should reduce preclinical failure, as it reflects tumor characteristics better and shows higher predictive accuracy than conventional 2D cultures: it retains the tissue architecture, extracellular matrix components and structures of the basement membrane as unique features for cellular interactions. These are important modifiers of cellular responses (Linke et al. , 2007 ; Philippi et al. , 2009 ; Schanz et al. , 2010 ). In detail, we observed (a) more homogenous staining of E‐cadherin/β‐catenin and lower proliferation rates according to tumor specimens, (b) a biomarker‐dependent apoptosis induction and proliferation reduction by the EGFR inhibitor Gef, and (c) in contrast to other preclinical findings, a reduced response upon HSP90 inhibitor treatment in KRAS ‐mutated tumor cells, which matches observations from clinical studies. Thus, we believe that our in vitro model resolves interpathway dependencies more reliably than 2D or animal models. Individual KRAS in silico networks were established by integrating relevant proteins from in vitro experiments and their interaction partners from HPRD. Our simulations start from a general in silico network for lung cancer, which is refined here to reveal the most relevant protein clusters. By matching cell line‐specific mutations from the COSMIC database, we derived individual drug targets and by screening our custom‐made, protein–drug interaction database DrumPID appropriate drugs (Kunz et al. , 2016 ). 3. 2 HSP90 inhibition in KRAS ‐mutated tumors and correlation of our 3D tissue models and other preclinical models to clinical findings In a previous study of a lung cancer model, we were able to demonstrate a stronger apoptosis induction in the 3D model by Gef, compared to conventional 2D culture (Stratmann et al. , 2014 ). After setting up standard operating procedures (Göttlich et al. , 2016 ), we could predict the clinical failure of HSP90 inhibitor treatment in the context of KRAS mutation, in contrast to other in vitro and in vivo models (Acquaviva et al. , 2012 ; Sos et al. , 2009 ). Heat shock proteins have gained attention in recent years as therapeutic tools, as they are involved in tumor cell proliferation, invasion, and cell death. Their high expression was observed in several cancer entities in clinical settings (Ciocca and Calderwood, 2005 ). Specifically, HSP90 belongs to a family of chaperons important for the function of relevant oncogenic drivers in lung adenocarcinomas. From a genomewide screening of 84 cell lines, KRAS mutation was identified to confer sensitivity to HSP90 inhibition that could also be verified in murine models (Sos et al. , 2009 ). In this screening, the geldanamycin derivatives 17AAG, and 17‐dimethylaminoethylamino‐17‐demethoxygeldanamycin in mice experiments were applied for HSP90 inhibition. However, geldanamycin and its derivates turned out to have safety and pharmacological limitations (Jhaveri and Modi, 2015 ). Another in vitro study showed the effectiveness of HSP90 inhibition in several KRAS ‐mutated non‐small cell lung cancer (NSCLC) lines by ganetespib – a non‐geldanamycin analog with less toxic side effects (Acquaviva et al. , 2012 ). However, single agent HSP90 inhibition by ganetespib failed in NSCLC patients with KRAS ‐mutated tumors. Combination therapy trials with docetaxel (GALAXY 1 and 2) led to better outcomes in patients with adenocarcinomas, than docetaxel single agent therapy, but not in the subgroup of KRAS ‐mutated tumors (Bhattacharya et al. , 2015 ). Recently, ALK, ROS1, and RET kinase gene rearrangements have been suggested to predict efficacy by targeting HSP90 (Rothschild, 2015 ; Sang et al. , 2013 ; Socinski et al. , 2013 ). 3. 3 In silico simulations of cell responses and development of a predictive KRAS signature In this study, in silico analyses for KRAS signature development are executed in three steps: Set up of cell‐specific in silico topology with logical Boolean connectivity (software tool celldesigner ; http://www. celldesigner. org ; Funahashi et al. , 2008 ) Cell‐specific dynamic in silico simulations of tumor cell responses (software tool squad ) For systematic drug‐target identification we generate larger cell‐specific protein–protein interaction networks considering neighbors of the central cascades (using data from HPRD) and cell line‐specific mutations (using data from COSMIC). For drug suggestions we apply the database tool DrumPID. In detail, we explain here the three above mentioned distinct types of in silico analyses: With the term ‘ in silico topology’, we considered a previously published knowledge‐based network which focuses mainly on kinase cascades (Stratmann et al. , 2014 ) and integrated here cell‐specific differences as additional nodes (proteins) derived from experimental data to specifically mirror the effects of Gef treatment and HSP90 inhibition. For this, we measured by phospho‐arrays and western blot signaling changes as drug responses as well as differences in proliferation and apoptosis in the different cell lines in 2D and 3D conditions. Missing parts of the cascade or modulatory crosstalk are filled in according to expert knowledge and public databases. These represent only the key parts of the signaling cascades. We used the tool ‘CellDesigner’ to set up the in silico topologies and to bring them into a machine‐readable format as done before for other cell types (Schlatter et al. , 2012 ). ‘ in silico simulations’ with the SQUAD tool predict the systemic response of a tumor cell upon a specific treatment which depends on the tumor cell topology and the activation/inactivation of its integrated nodes. As input the activation level of certain nodes can be set between zero as an inhibitory effect (inactivation) and one as an activating effect (activation). Furthermore, mutations can be integrated that stay independent from upstream signaling events at a certain value in case of gain or loss of function mutations. Also differences in 2D and 3D conditions can be simulated by adjusting the nodes’ values to experimentally measured levels, that is, phosphorylation determined by western blot. Some of the nodes summarize also global cellular responses, for example, ‘stress’. Importantly, also the drug responses proliferation and apoptosis are integrated in the topology as nodes. Values of the other nodes must be adjusted until the level of proliferation and apoptosis comply with the in vitro observations. Traditionally, differential equations for detailed kinetic modeling look at biological responses (Di Cara et al. , 2007 ; Dwivedi et al. , 2015 ; Robubi et al. , 2005 ). However, this requires then detailed kinetic information on individual kinases. This is not necessary in our approach, as the squad modeling software interpolates automatically exponential functions between our protein network nodes fitting signal transmission and logical connectivity (Di Cara et al. , 2007 ). We previously applied this combination to study cancer (Göttlich et al. , 2016 ; Stratmann et al. , 2014 ), infection biology (Audretsch et al. , 2013 ; Naseem et al. , 2012 ), and different tissues (Brietz et al. , 2016 ; Czakai et al. , 2017 ; Philippi et al. , 2009 ). For drug targeting, we looked systematically at larger protein–protein interaction networks; in particular, we collected all neighbors of upon 17AAG treatment between DRPs of both KRAS ‐mutated cell lines. To this network, we matched cell‐specific mutations from the COSMIC database. These larger networks we term here ‘ in silico networks’. The cell‐specific networks were then scrutinized to identify most promising treatment targets considering their relation to highly connected proteins in the network that are called ‘hubs’. A robust drug prediction algorithm collates information from several large‐scale databanks including chemical information according to Simplified Molecular Input Line Entry Specification (SMILES) notation and basic drug pharmacokinetics ADME (absorption, distribution, metabolism, excretion) rules (DrumPID, Kunz et al. , 2016 ). We used our reconstructed cell‐specific networks (Fig. 6 B, C) and screened which drugs according to DrumPID (Kunz et al. , 2016 ) influence apoptosis and proliferation in a cell‐specific manner. Targets are ranked by the effect strength, closeness to central cascades and druggability (Box S1 ). Subsequently, we simulated the potential therapeutic effect on apoptosis and proliferation focusing on AICAR (AMPK activator) and PX‐478 (HIF1A inhibitor) as top candidates and integrated their specific connectivity to the central cascades in the in silico topology. However, other drug target candidates can also be simulated, but for each simulation the individual targets and side targets of the drug has to be considered. Further testing of predictions is required to confirm suggested targets regarding clinical relevance. So far, neither HCC827, nor A549, nor H441 lung cancer cell lines have been analyzed by such a comprehensive in silico approach. 3. 4 Experimentally measured differences between the 2D and 3D system and in silico analyses Besides a higher chemoresistance in the case of HSP90 inhibitor treatment, we observed in 3D lower reduction of MET upon 17AAG treatment and an inverse regulation of p53, when compared to 2D conditions. Next to semiquantitatively evaluated western blot experiments, we present data from two phospho‐array screens (RTK, PK) as a starting point for further analyses (Figs S1 and S3A ). Importantly, in our 3D experiments we observed in contrast to 2D conditions upon 17AAG treatment an upregulation of HSP60 exclusively in A549, and an activation of p53 only in H441, which we were able to achieve also in our in silico simulations. However, the literature reports HSP60 inhibition by HSP90 and p53 inactivation by HSP60 (Ghosh et al. , 2008 ), which would explain our experimental observations in A549 and H441. The reason why HSP60 upregulation and a lack of p53 expression in A549 has a small effect on apoptosis in this setting, could be due to reduced HIF1A activation upon 17AAG treatment, as predicted by our in silico simulation. This reduced activation could stem from the inhibition of HSP90. HIF1A is not completely silenced in the simulation, due to its connection to LKB1 via mTOR in the in silico topology, as according to the COSMIC database, in A549 LKB1 carries a loss of function mutation. Furthermore, our in silico topology illustrates that this LKB1 mutation should lead to reduced AMPK activation and, thereby, also reflects the nonproliferative effect of 17AAG treatment via the mTOR signaling pathway. On the other hand, apoptosis induction is also blocked in H441. Induction of p53 upon inhibition of HSP90 should have no apoptotic effect due to the loss of function mutation of p53 identified in the COSMIC database. Furthermore, in our simulation we can see that HIF1A is still activated in H441 following HSP90 inhibition. This could be due to a mutation inside this gene that leads to an inhibitory effect on apoptosis and favors proliferation. Box S2 shows that we can use the same topology to simulate 2D results in silico, but differences in protein activation have to be taken into account to appropriately simulate the stronger apoptotic as well as proliferative responses upon 17AAG treatment observed in 2D cultures. As we could correctly simulate the observed responses upon Gef and 17AAG treatment for 2D and 3D conditions we could support that nodes in our topology are so far connected correctly. The in silico screen can reveal new dependencies, as high quality databases consider cancer‐subtype‐specific mutations and their interacting proteins along with all available drugs to directly attack the mutated protein or one of its neighbor. AICAR and PX‐478 are given as attractive examples (top ranked; see Supporting information) and their therapeutic effect on apoptosis and proliferation is simulated. However, other drugs can also be used by integrating other drug target candidates by considering its individual targets and side targets. Subsequently, the targets and drugs can be integrated in the in silico topology by considering its specific connectivity to the central cascades and further in silico simulated with SQUAD. 3. 5 Exemplified target and drug candidate prediction for A549 cells From the newly established KRAS network, based on proteins that exhibit changes in signaling between A549 and H441 in 3D conditions and cell‐specific mutations, the for A549 unique mutation LKB1 stands out. Our drug–protein interaction database DrumPID (Kunz et al. , 2016 ) identifies drugs that modulate the query protein directly, or one of its directly interacting neighbor proteins. This database tool identifies AMPK in our analyses as a potential drug target in A549 cells which can be modulated by the drug AICAR. AMPK protein is a direct interaction partner of LKB1 (Fig. 6 B) (Fay et al. , 2009 ; Rattan et al. , 2005 ; Tang et al. , 2011 ). AMPK activation using the drug AICAR, an analog of AMP, leads to tumor growth arrest in our in silico simulation for A549 cells (Fig. 7 A, B; nodes from first topology are olive‐shaded in 7 A). Moreover, AICAR shows promising results in the clinical phase 1/2 for chronic lymphatic leukemia (Van Den Neste et al. , 2013 ). In addition, the approved anticancer agent pemetrexed is known to indirectly activate AMPK by the accumulation of ZMP in LKB1‐null lung cancer (Rothbart et al. , 2010 ). 3. 6 Exemplified target and drug candidate prediction for H441 cells Regarding H441 cells, we also screened the H441 protein interaction network around the KRAS signature for potential drugs targeting either the protein or its direct neighbor. HIF1A was the highest‐ranked target (Box S1 ), as it is altered in H441 cells according to COSMIC data and is involved in a signaling loop (Greijer and van der Wall, 2004 ) (Fig. 6 C). Inhibition of HIF1A using PX‐478 shows an antitumor effect in our in silico simulations (Fig. 7 C, D; nodes from first topology are olive‐shaded in 7C). Studies demonstrated that inhibition of HIF1A shows promising therapeutic effects in human xenograft models (Welsh et al. , 2004 ). 3. 7 Application of the combined in vitro / in silico tool For clinical application, patient tumors have to be sequenced first, or at least tested by PCR or microarrays, to confirm that the driver mutation profile matches those in our cell lines. Notably, primary tumor cell culture is still challenging and has to be optimized for its utilization in routine personalized approaches. 4 Conclusion Predictive gene signatures were identified in a combined, tissue‐engineered, 3D lung tumor model with improved clinical correlation and a Boolean in silico approach that integrated measured cell‐specific differences in drug responses. We established cell line‐specific networks that depend on individual mutation patterns. This enabled better understanding of the interdependencies between single signaling cascades to prevent treatment resistance. Exemplified by the KRAS ‐mutated cell lines A549 and H441, we demonstrated how our analysis tool could lead to individual signature development, based on in vitro / in silico investigations on signaling, interaction partners from the HPRD, and sequence data from COSMIC. The limited number of direct interference points with the proliferative and/or apoptosis signaling cascade suggests and ranks best cell line‐specific targets, implying future therapies according to NGS data, tailored to the individual cancer mutation profile. Translated into clinical application, our lung cancer cell line‐specific examples suggest for patient stratification to determine not only the KRAS mutation status, but also to test for LKB1, p53, and HIF1A. Such cancer‐specific prescreening could distinguish among individual mutational subgroups to improve patient stratification and the design of clinical studies. 5 Materials and methods 5. 1 Cell culture HCC827 and A549 cell lines were purchased from DSMZ (Braunschweig, Germany), H441 from ATCC (LGC Standards GmbH – Germany Office, Wesel, Germany). A549 and H441 cells were cultured in RPMI + 10% FBS, HCC827 cells in RPMI + 20% FBS. Cells were monitored for pathogen infections at regular intervals. For a 2D culture, cells were either grown on glass coverslips in well plates until they had reached a confluency of about 70% or were cultured for 5 days in 12‐well plates or 6‐cm petri dishes. For a 3D culture, 1 × 10 5 tumor cells were grown for 14 days on the SISmuc (see Section 5. 3 ) that was fixed between two metal rings, as described in the literature (Göttlich et al. , 2016 ; Moll et al. , 2013 ; Stratmann et al. , 2014 ). Both 2D and 3D cultures were performed under standard conditions (37 °C, 5% CO 2 ). 5. 2 Treatment with Gef and 17AAG After 1 day in a 2D and 11 days in a 3D culture, cells were treated with either 1 μ m Gef (Iressa™, AstraZeneca, Wedel, Germany; Selleckchem) or 0. 01, 0. 05, 0. 1, 0. 25, 0. 5 or 1 μ m 17AAG (17‐ N ‐allylamino‐17‐demethoxygeldanamycin, Tanespimycin; Selleckchem) for 72 h, with a medium change after the first 48 h of treatment. 5. 3 Porcine material The SISmuc consisting of porcine small intestine submucosa (SIS) and mucosa (muc) was used as a scaffold for all 3D culture experiments. It was prepared from the BioVaSc ® as described in the literature (Linke et al. , 2007 ; Schanz et al. , 2010 ). All explantations were in compliance with the German Animal Protection Laws (§4(3), supervised by the institute's animal protection officer, all animals received proper care according to the National Institute of Health standards (NIH publication no. 85e23, revised 1996)), and as approved by the institutional animal protection board. 5. 4 Human material Human lung tumor tissue was provided by the Department of Thoracic Surgery of the University Hospital of Wuerzburg (local ethics approval: 182/10, 25. 11. 2015). 5. 5 Histology and immunofluorescence Cells cultured on glass slides in 2D were fixed in 4% paraformaldehyde for 10 min, cells in a 3D culture for 2 h, and the human lung tumor tissue overnight at 4 °C. The SISmuc samples, as well as the tumor tissue, were embedded in paraffin and sectioned at 3 μm thickness for hematoxylin–eosin (HE) and immunofluorescence staining. The primary antibodies E‐cadherin (#610181; BD Transduction Laboratories, Heidelberg, Germany), β‐catenin (#ab32572; Abcam, Cambridge, UK), and Ki67 (#ab16667; Abcam) were diluted 1 : 100 and incubated overnight at 4 °C. Secondary antibodies conjugated with fluorescent dyes Alexa 555 or 647 were diluted 1 : 400 and incubated for 1 h at room temperature. Nuclei were counterstained by DAPI dissolved in a Mowiol embedding solution. Pictures were taken with a digital microscope (BZ‐9000; Keyence Deutschland GmbH, Neu‐Isenburg, Germany). 5. 6 Cell proliferation To determine the proliferation rate, cells cultured in 2D and 3D were stained against Ki67. Ten nonoverlapping images of 3D sections and five nonoverlapping images of 2D cultures were taken. Quantification of the proliferation rate was performed as described in the literature (Göttlich et al. , 2016 ). 5. 7 M30‐elisa Apoptosis was determined from supernatants taken from untreated and treated tumor models during the last 4 days of the culture. M30 CytoDeath™ ELISA (Peviva) was performed according to the manufacturer's instructions. All samples were measured in duplicates. 5. 8 Western blot and phospho‐RTK and PK arrays Cells were lysed in modified RIPA buffer (137 m m NaCl, 50 m m NaF, 20 m m Tris/HCl pH 8. 0, 2 m m EDTA, 10% (v/v) glycerol, 1% (v/v) NP‐40, 0. 5% (w/v) DCA, 0. 1% (w/v) SDS, 1 m m Na 3 VO 4, and 1× protease inhibitor cocktail (Sigma‐Aldrich, Darmstadt, Germany)), or in the provided lysis buffers of the respective array kit. For western blot analysis, protein samples (27 μg per lane) were separated electrophoretically in a 10% SDS/gel and blotted on a 0. 2‐μm nitrocellulose membrane (Whatman, Fisher Scientific GmbH, Schwerte, Germany). The primary antibodies pEGFR (#ab32430; Abcam), pMet (#3077; Cell Signaling Technology, Frankfurt a. Main, Germany), phospho‐p53 (S46) (#2521; Cell Signaling Technology), HSP60 (#ab46798; Abcam), and β‐actin (#3700; Cell Signaling Technology) were incubated in NFDM or a 1% BSA overnight at 4 °C. Secondary anti‐mouse or anti‐rabbit IgG antibodies conjugated to horseradish peroxidase (#JAC‐111035144 or #JAC‐115035146; Jackson ImmunoResearch, Cambridgeshire, UK) were incubated for 1 h at room temperature. Bands were visualized using the Pierce ECL Western Blotting kit (Thermo Scientific, Breda, Netherlands). Phospho‐RTK and PK arrays were performed according to the manufacturer's instructions. Western blot and array membranes were imaged at the imaging station FluorChem Q (Biozym Scientific, Hessisch Oldendorf, Germany). Gray values were determined with the related image acquisition and analysis software alphaview (version 3. 2. 2. 0; Proteinsimple, San Jose, CA, USA). 5. 9 Statistical analysis of the experimental data The nonparametric Kruskal–Wallis test and post hoc Wilcoxon rank‐sum test were used for statistical analysis of proliferation and apoptosis results. P < 0. 05 was considered as significant. Statistical analysis was carried out with the open‐source software r (The Comprehensive R Archive Network). 5. 10 Bioinformatics analysis 5. 10. 1 Network analysis Bioinformatics analyses combined cell culture array and western blot data for the Gef and 17AAG treatment in the 3D system with information from databases to build up an individual network for each cell line. We extended the original in silico topology (Stratmann et al. , 2014 ) by integrating proteins listed in Table 2. 5. 10. 2 Dynamic simulation For the 3D in vitro system, we simulated the Gef and 17AAG treatment using the squad software (Di Cara et al. , 2007 ), by taking the pathway activity differences into account (Table 2 ) while running the simulation (prestimulations in Table 3 ). For the 2D system we focused on the A549 and H441 cell lines (prestimulations in Box S2 ). SQUAD represents the network topology (activation, inhibition) using logical Boolean operator (AND, OR, NOT) and interpolates them by applying mathematical e‐functions. The resulting network effects are visualized in a graph as changes of state over an arbitrary time, allowing in silico simulations of different network scenarios. Simulation protocols were written using the SQUAD function ‘perturbator’ (prestimulation option in the simulation menu of the software), in which the value for the drug Gef and 17AAG were set to an initial state of 0 and 1 (reflecting either no treatment or standard treatment, respectively), and experimental nodes and mutations were adjusted (Table 3 and Box S2 ). All parameters for the proteins (‘nodes’) in the network without experimental or mutational regulation were set as an active node pulse (state = 0 and time = 0) that changes, depending on interconnectivity in the cell‐specific network. 5. 10. 3 Software for visualization To set up the silico topology we used the celldesigner software tool. For visualizing the network, we used cytoscape version 2. 8. 3 (Shannon et al. , 2003 ). The cytoscape software is an open‐source platform for visualization and analysis of biological networks using several plug‐ins (Saito et al. , 2012 ; Shannon et al. , 2003 ). We analyzed the reconstructed cell line‐specific networks for functional modules (‘clusters’) using the Cytoscape plug‐in MCODE (Bader and Hogue, 2008 ). Potentially available drugs were selected using our previously developed DrumPID (Kunz et al. , 2016 ). The following methods were applied, as detailed below: An in silico signaling network is invariably a simplified view of the biological complexity. We focus here on the major cascades relevant for the output. The bioinformatics analysis combined the cell culture array and western blot data with the systems biology network analysis approaches and information from databases. The robustness of the simplified networks of central tumor cascades in cell‐specific in silico networks was also verified by considering removing and adding protein nodes at the rim of the network. This did not affect signaling responses, whereas changing central hub protein nodes strongly affected the results. Cell compartmentalization (e. g. , divergent cytosolic and mitochondrial processes), multiphosphorylation processes and complex formations were not included. This limits such approaches to semiquantitative descriptions of the sequential order, strength, and respective duration of events (Brietz et al. , 2016 ; Göttlich et al. , 2016 ). However, the simulations allow an in silico overview of the lung tumor topology and important drug responses, such as changes in individual cell line‐specific responses. 5. 10. 4 Cell line‐specific network reconstruction For establishing cell line‐specific signaling networks, we always used the same central cascades based on our previously published in silico topology (Göttlich et al. , 2016 ; Stratmann et al. , 2014 ) for HCC827 and A549 cell lines. In contrast, we extended the in silico network for the additionally introduced KRAS ‐mutated H441 cell line. 5. 10. 5 Simulation protocol For dynamic simulation, we first looked at the effects of Gef and next 17AAG treatment in the tissue in vitro system, revealed by phospho‐RTK arrays and western blots. As the model is semiquantitative, weaker or stronger biological activation has to be taken into account with values between 0 (reflects no activation) and 1 (reflects full activation) to model key input. We fitted parameters to the results obtained from experiments (data‐driven modeling) and optimized the fit in iterative cycles of new simulations and new experiments. This included prestimulations according to mutations known from their pharmacological behavior. We hence simulated the proteins as network nodes with the parameters given in Table 3 and Box S2. Additional information on the details of the bioinformatics analysis is given in the Doc. S1 and the Supporting information figures and boxes. Author contributions CG performed all the experiments in the study, supported in some by SLN. MK performed all the bioinformatics work including simulations, network analysis, and large‐scale data analysis. In several places and in particular for the initial model set up, CZ worked on the bioinformatics analysis. GD and SLN supervised the experimental work. HW gave expert analysis and advice on all experimental work. HW, GD, and TD supervised CG, and TD and MK supervised CZ. TD and GD supervised the bioinformatics analysis, analyzed all data, and made suggestions for new experiments, simulations, data comparisons, and other analyses. TD and GD drafted the manuscript together, and all authors contributed to the iterations and agreed to the final version of the manuscript. TD and GD led and guided the study. Data availability All data and simulation protocols for the study are made available with the publication (paper plus all Supporting information). Supporting information Fig. S1. Signaling is unchanged in gefitinib responsive HCC827 cells in 2D and 3D. Fig. S2. In silico model and simulation for the gefitinib treatment in A549 and H441. Fig. S3. Signaling changes in 2D and 3D after treatment of different cell lines with the HSP90 inhibitor 17AAG. Fig. S4. Biological network analyses on the KRAS‐mutated cell lines for 17AAG in the 3D system. Fig. S5. Functional cluster analyses of the cell line‐specific networks. Fig. S6. Cell line‐specific in silico simulations for gefitinib treatment in A549 and H441 according to data from the 2D system. Fig. S7. In silico simulations for 17AAG treatment in A549 and H441 according to data from the 2D system. Box S1. Ranking and comparison of all cell‐specific mutations for KRAS signature development and individual target predictions. Box S2. Cell line‐specific differences modeled in 2D. Click here for additional data file. Doc. S1. Additional information for the bioinformatics analyses Click here for additional data file. Table S1. For the generation of networks we downloaded the HPRD which contains 9620 protein nodes and 39185 protein–protein interaction edges (release 9 from April 13, 2010). Table S2. For the identification of a KRAS signature of potential markers we downloaded cell line‐specific mutations from the COSMIC database (A549: Sample Name: A549, Sample ID: 905949; H441: Sample Name: NCI‐H441, Sample ID: 908460). Table S3. Mapping of the COSMIC mutations to the KRAS‐mutated network results in 18 H441‐ and 9 A549‐specific overlapping proteins (nodes). Click here for additional data file.

10. 1002/1878-0261. 13037

Molecular Oncology

"Lineage‐specific mechanisms and drivers of breast cancer chemoresistance revealed by 3D biomimeti(...TRUNCATED)

"To improve the success rate of current preclinical drug trials, there is a growing need for more co(...TRUNCATED)

"Abbreviations ABC ATP‐binding cassette Ct threshold cycle DEGs differentially expressed genes ECM(...TRUNCATED)

10. 1002/2211-5463. 12250

FEBS Open Bio

Role of the TGF‐β pathway in dedifferentiation of human mature adipocytes

"Dedifferentiation of adipocytes contributes to the generation of a proliferative cell population th(...TRUNCATED)

"Abbreviations COL1A1 collagen type I alpha 1 COL1A2 collagen type I alpha 2 COL6A3 collagen type 6 (...TRUNCATED)