Transcriptomics-Morphology Generation Via Treatment Conditioning With Rectified Flow

ICLR 2026 Conference Submission14294 Authors

Keywords: Multi-Modal, Generative AI, Transcriptomics, Cellular Imaging

TL;DR: Generating transcirptomics and morphology image with response to chemical perturbation.

Abstract: Predicting cellular responses to drug perturbations requires capturing complex dependencies between transcriptomic and morphological changes that single-modality approaches cannot adequately model. We introduce \textbf{PertFlow}, a unified framework that jointly predicts gene expression profiles and generates cellular morphology images in response to drug treatments, conditioned on control cellular states. Our method integrates control transcriptomic and imaging data through multi-head cross-modal attention mechanisms, learning a shared latent representation that incorporates drug compound features, background cellular profiles, and treatment specifications. From this unified representation, PertFlow employs a regression head for RNA-seq prediction and rectified flow dynamics for stable morphological image generation, with cross-modal consistency losses ensuring coherent molecular and phenotypic predictions. PertFlow enables accurate predictions from either complete multi-modal inputs or single-modality data alone, demonstrating robust cross-modal learning. Our evaluation on paired RNA-seq and Cell Painting fluorescent imaging datasets demonstrates that PertFlow achieves stronger cross-modal consistency and accurate prediction of drug-induced changes compared to diffusion baselines.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 14294