Cell embeddings and cosine shift
Hi Christina, could you clarify something for me?
Below is a plot of the cell embeddings extracted from my model fine-tuned using OA data. Can the cosine shift be explained as how far on the coordinate plane the diseased cell embedding moves towards the centroid of the healthy cells, as indicated in the first plot below? Would a better approach be imitating the natural progression of diseased cells, but in reverse, as in the second figure?
Thank you for your question! That is an interesting idea and something that one could certainly implement. Keep in mind these visualizations are compressing 256 dimensions into 2, but yes it could be of interest to determine whether some perturbations shift towards an intermediate state in addition to those that directly shift the cells towards the goal end state, especially in cases where one would think multiple stages of shifting the cells’ identities would be needed (similarly to multiple stages of factors directing a differentiation, for example).