PerturbCellRL Enhances Single-Cell Perturbation Prediction with RL and Verifiers.

Predicting how individual cells respond to genetic or chemical interventions is crucial for drug discovery and biological research, potentially reducing the need for expensive lab screenings. While current generative models can predict population-level responses, they often lack explicit checks for biological consistency at the individual cell level. A new framework, PerturbCellRL, addresses this by employing reinforcement learning (RL) to post-train pre-existing single-cell transcriptomic generators. This framework uses a suite of "cell-level verifiers" as reward signals during RL training. These verifiers assess various biological aspects, including gene similarity, expression proximity, differential expression, and pathway activity, ensuring that generated cells align with known biological responses to perturbations. Evaluations across multiple benchmarks for genetic and chemical perturbations demonstrate that PerturbCellRL significantly improves reward-aligned metrics and a held-out evaluation metric compared to its pre-trained base generator. It also maintains competitive performance on population-level metrics against state-of-the-art methods. This approach marks a shift towards more trustworthy single-cell predictions by explicitly verifying biological consistency at the individual cell level.