PerturbCellRL: Verifier-Guided Reinforcement Learning for Single-Cell Perturbation Prediction

2026-06-26

AI Summary

Current single-cell perturbation prediction models lack explicit constraints to ensure biological consistency when generating individual cells, leading to unreliable outputs. This work proposes a reinforcement learning–based post-training framework that, for the first time, incorporates multidimensional cell-level validators—including Pearson top-k similarity, RMSE top-k proximity, Spearman correlation of differentially expressed genes, and pathway activity—as reward signals to guide a pretrained generator toward biologically plausible perturbation responses. The method significantly improves alignment with these validators and enhances held-out evaluation metrics across multiple genetic and chemical perturbation benchmarks, while maintaining population-level performance comparable to state-of-the-art approaches, thereby advancing a more trustworthy paradigm for single-cell perturbation prediction.

Abstract

Single-cell perturbation models can reduce costly wet-lab screening by predicting how cells respond transcriptionally to interventions. While recent generative models improve population-level prediction, individual generated cells are not explicitly checked for biological consistency. We introduce PerturbCellRL, a reinforcement learning (RL) framework that post-trains a pretrained single-cell transcriptomic generator using a suite of cell-level verifiers as rewards. These verifiers define four rewards: Pearson top-k similarity, RMSE top-k proximity, DE Spearman, and Pathway activity. The Pathway activity verifier rewards cells whose pathway responses match known perturbation biology. We evaluate PerturbCellRL on multiple genetic and chemical perturbation benchmarks. Across these benchmarks, PerturbCellRL improves over the pretrained flow-matching generator on reward-aligned evaluation metrics and a held-out evaluation metric. Moreover, PerturbCellRL remains competitive with state-of-the-art methods on population-level metrics. Together, these results frame trustworthy single-cell prediction as verifier-guided generative alignment, moving beyond matching expression distributions toward predictions whose single-cell perturbation effects are explicitly checked for biological consistency.

Problem

• single-cell perturbation
• biological consistency
• generative models
• transcriptional response
• verifier-guided

Innovation

• reinforcement learning
• single-cell perturbation prediction
• verifier-guided generation
• biological consistency
• pathway activity