The combination of pooled genetic perturbation screening with single-cell RNA-sequencing (scRNA-seq) has allowed for the large-scale profiling of single-cell transcriptional responses, providing valuable insights into how genes interact and influence cellular behaviors. There is a need for computational methods capable of modeling and predicting single-cell responses to novel genetic perturbations, thus allowing for the efficient exploration of the vast space of possible multi-gene perturbations. Optimal transport (OT) presents a compelling framework to model this type of data, given the lack of true control-perturbed cell pairings. However, existing neural OT models are generally constrained to perturbations previously observed during training, while the use of OT theory in methods capable of modeling novel genetic perturbations remains limited.
To bridge this gap, we present an OT-based method to predict single-cell transcriptional responses to novel genetic perturbations. Leveraging a flexible attention-based aggregation approach, our method incorporates gene representations from multiple sources of prior knowledge, spanning from functional descriptions of genes to gene-gene relationships, producing biologically informed representations of perturbations. Moreover, we utilize an OT-based loss to align predicted and observed perturbed cell populations, avoiding the need to assign random control-perturbed cell pairings. Experiments on three benchmark datasets demonstrate the highly competitive performance of our model compared to state-of-the-art methods. Further evaluations on differential expression analysis and genetic interaction modeling demonstrate the biological relevance and potential utility of our model in different applications.
