How Much Can AI Virtual Cells Gain from Unified Perturbational Modeling?

Unified perturbational modeling (UPM) integrated into a VAE-based AIVC predictor achieves the highest PES and gene-wise response correlation on genetic perturbation benchmarks.

• Why it matters: Heterogeneous Perturb-seq datasets differ in gene sets, perturbation types, and cellular contexts, limiting how well existing AIVC predictors transfer across settings.
• What they did: UPM adds gene-channel alignment, shared perturbation encoding, and joint embedding modules, training on 89.7 million cells from 71 contexts within a 110 million-cell, 22,000-gene dataset.
• The result: UPM improves all four metrics on held-out Tahoe-100M cell lines and Stack cytokine benchmarks, boosts combinatorial prediction versus specialized models, and yields a 40.0% accuracy gain in an HSC knockdown recapitulation task.