Accepted_test

Single cell transcriptomics is a method for profiling complex and heterogenous systems, focusing on cell dynamics and responses to perturbations. Generative modeling of perturbation response allows for in silico data generation. The ability to generate previously unseen data makes it a promising approach. The study evaluates the performance of stVAE, a deep learning model using Conditional Variational Autoencoder and Y-Autoencoder training approaches. The model was also tested on a single cell dataset of peripheral blood mononuclear cells from pericarditis patients. The study aimed to train a model to predict treatment effects for specific cell types. Different training data approaches were tested: training on data from all cell types or from a subset of cell types. Results showed high correlations between predicted gene expressions and ground truth. The accuracy of perturbation predictions was evaluated using metrics like differentially expressed genes expression correlation, fold changes correlation, and enriched pathways intersection. The results demonstrate that stVAE is able to accurately predict single-cell perturbation response out-of-sample.