CellTarNet: Single-Cell Perturbation Prediction using Transformer based Normalizing Flow

Predicting the transcriptional response of cells to perturbations is a challenging task as perturbation datasets often include global gene expression shifts, which are difficult to seperate from individual expression. We present CellTarNet, a generative framework based on transformer based normalizing flows to learn transport from control cells to perturbed cells. The transformer encoder summarizes control-cell states into a context representation, while the normalizing flow learns a distribution over perturbed transcriptional profiles conditioned on the context. We employ contrastive matching to pull predicted samples toward the true perturbed distribution and separates them from mismatched perturbation-context pairs. We further show how to integrate this with gene interaction graphs to better model gene expressions.