CELLTARNET: SINGLE-CELL PERTURBATION PREDICTION USING TRANSFORMER BASED NORMALIZING FLOW

Predicting the transcriptional response to a genetic or chemical perturbation is a critical task for therapeutic design. However, building accurate models is challenged by the need to capture heterogeneous responses by a cell population to such perturbations. We propose a novel method that combines a transformer based generative model with a population-level contrastive objective to model the full distribution of perturbation effects. Furthermore, we integrate structured biological knowledge, to bias a transformer's self-attention mechanism. Initial experiments show that this enables robust, generalizable prediction of transcriptional responses.