Reproduce “PRiMeFlow: Capturing Complex Expression Heterogeneity in Perturbation Response Modelling” on OpenResearch

About this paper

Zichao Yan, Yan Wu, Mica Xu Ji, Chaitra Agrahar, Esther Wershof, Marcel Nassar, Mehrshad Sadria, Ridvan Eksi and 5 more

Altos Labs' PRiMeFlow directly predicts single-cell gene expression responses to perturbations using an end-to-end flow matching framework, operating in the high-dimensional gene expression space to preserve biological signals. The model demonstrates superior distribution-fitting capabilities and higher recall of ground-truth differentially expressed genes compared to methods relying on latent space compression, achieving the lowest average rank on the ARC Virtual Cell Challenge Generalist leaderboard.

Key insights

• Operating directly in the high-dimensional gene expression space, rather than a compressed latent space, significantly improves the model's ability to fit ground truth distributions and enhance differentially expressed gene recall.
• The U-Net architecture exhibits a strong inductive bias for modeling complex latent dependencies within raw gene expression data, outperforming simpler Multi-Layer Perceptron (MLP) variants for velocity field parameterization.
• Simpler independent coupling strategies in flow matching proved as effective as, or superior to, more complex optimal transport-based coupling methods for this task.

Summary by alphaXiv.