# -*- coding: utf-8 -*- import os script_dir = os.path.dirname(os.path.abspath(__file__)) import pandas as pd import numpy as np from datetime import datetime import pickle import os import multiprocessing from tqdm import tqdm from sklearn.svm import SVC from sklearn.linear_model import SGDClassifier from sklearn.model_selection import cross_val_predict, KFold from skmultilearn.problem_transform import BinaryRelevance from sklearn.metrics import accuracy_score, f1_score, precision_score, recall_score, hamming_loss aspect_type = "" dataset_type = "" representation_dataframe = "" representation_name = "" detailed_output = False def warn(*args, **kwargs): pass import warnings warnings.warn = warn def check_for_at_least_two_class_sample_exits(y): for column in y: column_sum = np.sum(y[column].array) if column_sum < 2: print('At least 2 positive samples are required for each class {0} class has {1} positive samples'.format(column,column_sum)) return False return True def create_valid_kfold_object_for_multilabel_splits(X,y,kf): check_for_at_least_two_class_sample_exits(y) sample_class_occurance = dict(zip(y.columns,np.zeros(len(y.columns)))) for column in y: for fold_train_index,fold_test_index in kf.split(X,y): fold_col_sum = np.sum(y.iloc[fold_test_index,:][column].array) if fold_col_sum > 0: sample_class_occurance[column] += 1 for key in sample_class_occurance: value = sample_class_occurance[key] if value < 2: random_state = np.random.randint(1000) print("Random state is changed since at least two positive samples are required in different train/test folds.\ \nHowever, only one fold exits with positive samples for class {0}".format(key)) print("Selected random state is {0}".format(random_state)) kf = KFold(n_splits=5, shuffle=True, random_state=random_state) create_valid_kfold_object_for_multilabel_splits(X,y,kf) else: return kf def MultiLabelSVC_cross_val_predict(representation_name, dataset, X, y, classifier): #dataset split, estimator, cv clf = classifier Xn = np.array(np.asarray(X.values.tolist()), dtype=float) kf_init = KFold(n_splits=5, shuffle=True, random_state=42) kf = create_valid_kfold_object_for_multilabel_splits(X,y,kf_init) y_pred = cross_val_predict(clf, Xn, y, cv=kf) if detailed_output: ont_path = r"../results/Ontology_based_function_prediction_{1}_{0}_model.pkl".format(representation_name,dataset.split(".")[0]) with open(os.path.join(script_dir, ont_path),"wb") as file: pickle.dump(clf,file) acc_cv = [] f1_mi_cv = [] f1_ma_cv = [] f1_we_cv = [] pr_mi_cv = [] pr_ma_cv = [] pr_we_cv = [] rc_mi_cv = [] rc_ma_cv = [] rc_we_cv = [] hamm_cv = [] for fold_train_index,fold_test_index in kf.split(X,y): acc = accuracy_score(y.iloc[fold_test_index,:],y_pred[fold_test_index]) acc_cv.append(np.round(acc,decimals=5)) f1_mi = f1_score(y.iloc[fold_test_index,:],y_pred[fold_test_index],average="micro") f1_mi_cv.append(np.round(f1_mi,decimals=5)) f1_ma = f1_score(y.iloc[fold_test_index,:],y_pred[fold_test_index],average="macro") f1_ma_cv.append(np.round(f1_ma,decimals=5)) f1_we = f1_score(y.iloc[fold_test_index,:],y_pred[fold_test_index],average="weighted") f1_we_cv.append(np.round(f1_we,decimals=5)) pr_mi = precision_score(y.iloc[fold_test_index,:],y_pred[fold_test_index],average="micro") pr_mi_cv.append(np.round(pr_mi,decimals=5)) pr_ma = precision_score(y.iloc[fold_test_index,:],y_pred[fold_test_index],average="macro") pr_ma_cv.append(np.round(pr_ma,decimals=5)) pr_we = precision_score(y.iloc[fold_test_index,:],y_pred[fold_test_index],average="weighted") pr_we_cv.append(np.round(pr_we,decimals=5)) rc_mi = recall_score(y.iloc[fold_test_index,:],y_pred[fold_test_index],average="micro") rc_mi_cv.append(np.round(rc_mi,decimals=5)) rc_ma = recall_score(y.iloc[fold_test_index,:],y_pred[fold_test_index],average="macro") rc_ma_cv.append(np.round(rc_ma,decimals=5)) rc_we = recall_score(y.iloc[fold_test_index,:],y_pred[fold_test_index],average="weighted") rc_we_cv.append(np.round(rc_we,decimals=5)) hamm = hamming_loss(y.iloc[fold_test_index,:],y_pred[fold_test_index]) hamm_cv.append(np.round(hamm,decimals=5)) means = list(np.mean([acc_cv,f1_mi_cv,f1_ma_cv,f1_we_cv,pr_mi_cv,pr_ma_cv,pr_we_cv,rc_mi_cv,rc_ma_cv,rc_we_cv,hamm_cv], axis=1)) means = [np.round(i,decimals=5) for i in means] stds = list(np.std([acc_cv,f1_mi_cv,f1_ma_cv,f1_we_cv,pr_mi_cv,pr_ma_cv,pr_we_cv,rc_mi_cv,rc_ma_cv,rc_we_cv,hamm_cv], axis=1)) stds = [np.round(i,decimals=5) for i in stds] return ([representation_name+"_"+dataset,acc_cv,f1_mi_cv,f1_ma_cv,f1_we_cv,pr_mi_cv,pr_ma_cv,pr_we_cv,rc_mi_cv,rc_ma_cv,rc_we_cv,hamm_cv],\ [representation_name+"_"+dataset]+means,\ [representation_name+"_"+dataset]+stds,\ y_pred) def ProtDescModel(): #desc_file = pd.read_csv(r"protein_representations\final\{0}_dim{1}.tsv".format(representation_name,desc_dim),sep="\t") datasets = os.listdir(os.path.join(script_dir, r"../data/auxilary_input/GO_datasets")) if dataset_type == "All_Data_Sets" and aspect_type == "All_Aspects": filtered_datasets = datasets elif dataset_type == "All_Data_Sets": filtered_datasets = [dataset for dataset in datasets if aspect_type in dataset] elif aspect_type == "All_Aspects": filtered_datasets = [dataset for dataset in datasets if dataset_type in dataset] else: filtered_datasets = [dataset for dataset in datasets if aspect_type in dataset and dataset_type in dataset] cv_results = [] cv_mean_results = [] cv_std_results = [] for dt in tqdm(filtered_datasets,total=len(filtered_datasets)): print(r"Protein function prediction is started for the dataset: {0}".format(dt.split(".")[0])) dt_file = pd.read_csv(os.path.join(script_dir, r"../data/auxilary_input/GO_datasets/{0}".format(dt)),sep="\t") dt_merge = dt_file.merge(representation_dataframe,left_on="Protein_Id",right_on="Entry") dt_X = dt_merge['Vector'] dt_y = dt_merge.iloc[:,1:-2] if check_for_at_least_two_class_sample_exits(dt_y) == False: print(r"No funtion will be predicted for the dataset: {0}".format(dt.split(".")[0])) continue #print("raw dt vs. dt_merge: {} - {}".format(len(dt_file),len(dt_merge))) #print("Calculating predictions for " + dt.split(".")[0]) #model = MultiLabelSVC_cross_val_predict(representation_name, dt.split(".")[0], dt_X, dt_y, classifier=BinaryRelevance(SVC(kernel="linear", random_state=42))) cpu_number = multiprocessing.cpu_count() model = MultiLabelSVC_cross_val_predict(representation_name, dt.split(".")[0], dt_X, dt_y, classifier=BinaryRelevance(SGDClassifier(n_jobs=cpu_number, random_state=42))) cv_results.append(model[0]) cv_mean_results.append(model[1]) cv_std_results.append(model[2]) predictions = dt_merge.iloc[:,:6] predictions["predicted_values"] = list(model[3].toarray()) if detailed_output: predictions.to_csv(os.path.join(script_dir, r"../results/Ontology_based_function_prediction_{1}_{0}_predictions.tsv".format(representation_name,dt.split(".")[0])),sep="\t",index=None) return (cv_results, cv_mean_results,cv_std_results) #def pred_output(representation_name, desc_dim): def pred_output(): model = ProtDescModel() cv_result = model[0] df_cv_result = pd.DataFrame({"Model": pd.Series([], dtype='str') ,"Accuracy": pd.Series([], dtype='float'),"F1_Micro": pd.Series([], dtype='float'),\ "F1_Macro": pd.Series([], dtype='float'),"F1_Weighted": pd.Series([], dtype='float'),"Precision_Micro": pd.Series([], dtype='float'),\ "Precision_Macro": pd.Series([], dtype='float'),"Precision_Weighted": pd.Series([], dtype='float'),"Recall_Micro": pd.Series([], dtype='float'),\ "Recall_Macro": pd.Series([], dtype='float'),"Recall_Weighted": pd.Series([], dtype='float'),"Hamming_Distance": pd.Series([], dtype='float')}) for i in cv_result: df_cv_result.loc[len(df_cv_result)] = i if detailed_output: df_cv_result.to_csv(os.path.join(script_dir, r"../results/Ontology_based_function_prediction_5cv_{0}.tsv".format(representation_name)),sep="\t",index=None) cv_mean_result = model[1] df_cv_mean_result = pd.DataFrame({"Model": pd.Series([], dtype='str') ,"Accuracy": pd.Series([], dtype='float'),"F1_Micro": pd.Series([], dtype='float'),\ "F1_Macro": pd.Series([], dtype='float'),"F1_Weighted": pd.Series([], dtype='float'),"Precision_Micro": pd.Series([], dtype='float'),\ "Precision_Macro": pd.Series([], dtype='float'),"Precision_Weighted": pd.Series([], dtype='float'),"Recall_Micro": pd.Series([], dtype='float'),\ "Recall_Macro": pd.Series([], dtype='float'),"Recall_Weighted": pd.Series([], dtype='float'),"Hamming_Distance": pd.Series([], dtype='float')}) #pd.DataFrame(columns=["Model","Accuracy","F1_Micro","F1_Macro","F1_Weighted","Precision_Micro","Precision_Macro","Precision_Weighted",\ # "Recall_Micro","Recall_Macro","Recall_Weighted","Hamming_Distance"]) for j in cv_mean_result: df_cv_mean_result.loc[len(df_cv_mean_result)] = j df_cv_mean_result.to_csv(os.path.join(script_dir, r"../results/Ontology_based_function_prediction_5cv_mean_{0}.tsv".format(representation_name)),sep="\t",index=None) #save std deviation of scores to file cv_std_result = model[2] df_cv_std_result = pd.DataFrame({"Model": pd.Series([], dtype='str') ,"Accuracy": pd.Series([], dtype='float'),"F1_Micro": pd.Series([], dtype='float'),\ "F1_Macro": pd.Series([], dtype='float'),"F1_Weighted": pd.Series([], dtype='float'),"Precision_Micro": pd.Series([], dtype='float'),\ "Precision_Macro": pd.Series([], dtype='float'),"Precision_Weighted": pd.Series([], dtype='float'),"Recall_Micro": pd.Series([], dtype='float'),\ "Recall_Macro": pd.Series([], dtype='float'),"Recall_Weighted": pd.Series([], dtype='float'),"Hamming_Distance": pd.Series([], dtype='float')}) #pd.DataFrame(columns=["Model","Accuracy","F1_Micro","F1_Macro","F1_Weighted","Precision_Micro","Precision_Macro","Precision_Weighted",\ # "Recall_Micro","Recall_Macro","Recall_Weighted","Hamming_Distance"]) for k in cv_std_result: df_cv_std_result.loc[len(df_cv_std_result)] = k df_cv_std_result.to_csv(os.path.join(script_dir, r"../results/Ontology_based_function_prediction_5cv_std_{0}.tsv".format(representation_name)),sep="\t",index=None) print(datetime.now()) # tcga = pred_output("tcga","50") # protvec = pred_output("protvec","100") # unirep = pred_output("unirep","5700") # gene2vec = pred_output("gene2vec","200") # learned_embed = pred_output("learned_embed","64") # mut2vec = pred_output("mut2vec","300") # seqvec = pred_output("seqvec","1024") #bepler = pred_output("bepler","100") # resnet_rescaled = pred_output("resnet-rescaled","256") # transformer_avg = pred_output("transformer","768") # transformer_pool = pred_output("transformer-pool","768") # apaac = pred_output("apaac","80") #ksep = pred_output("ksep","400")