Conditional Normalizing Flows for the Design of T Cell Therapies

ICML 2025 Workshop FM4LS Submission43 Authors

Keywords: single-cell foundation model, gene expression anaylsis, CAR T, normalizing flows

TL;DR: Conditional normalizing flows can effectively decouple patient-specific variation from biologically relevant features from clinical correlative datasets to accurately predict therapeutic outcomes and optimize drug design by nominating perturbations.

Abstract: Designing effective cell therapies requires understanding how transcriptional regulation within infusion products influence patient outcomes. Here, we propose a generative model that leverages single-cell RNA sequencing (scRNA-seq) data paired with clinical outcomes to learn the gene regulatory networks within engineered Chimeric Antigen Receptor (CAR) T cells. Using conditional normalizing flows, our model captures the high-dimensional distribution of gene activity while conditioning on patient and response-specific features. This approach enables patient response prediction with 73\% accuracy and accurate simulations of gene knockdown, knockouts, and over-expression experiments. Our model identified function-recovering genetic modifications for CAR T infusion products, which were validated experimentally in the context of a genetic screen.

Submission Number: 43