Directionality of the perturbation of the affected genes
Dear Dr Theodoris,
Thank you very much for your impressive work!
I am wondering if it would be possible to use your insilico perturbation analysis to extract the directionality of change of the affected genes.
If I understood your work correctly, you illustrate, among other things, how to quantify the gene embeddings change (cosine distance) in response to the deletion / over expression to illustrate the connectivity between the perturbed and affected gene. Would it be possible to use your model to understand the directionality of the perturbation (e.g. what genes become up or down regulated upon the deletion of MECP2) - and if so, would you have any pointers where one could start?
Best wishes,
Mateusz Kaczyński
Thank you for your interest in Geneformer. You are correct - the effect of an in silico perturbation is modeled by changes in the embeddings of potential downstream genes, which we quantify as cosine shifts. To infer the directionality of the gene expression changes one would expect, we would suggest potentially fine-tuning a model for this task with known perturbations as training data. One could also develop another model on top of Geneformer, taking the Geneformer embeddings as inputs to a model that explicitly infers directional connections between genes.
Thank you very much for your in-depth response, it was most informative!