PertMind: Eliciting Emergent Biological Reasoning in LLM via Reinforcement Learning on Cellular Perturbation Data

August 18, 2026 | Source: arXiv AI

Tags: reinforcement learning, biology, LLM, cellular perturbation, drug discovery, RLHF

PertMind trains LLMs on cellular perturbation prediction using gene-response data as RL rewards — no manual reasoning traces required — and the resulting models transfer without task-specific fine-tuning to reverse perturbation identification, double-perturbation reasoning, and phenotypic screen prioritisation.

Details

Researchers introduce PertMind, a framework that treats cellular perturbation atlases as reinforcement learning environments for training LLMs on biological reasoning. The key motivation is scalability: existing biological reasoning LLMs depend on manually curated reasoning traces, which are expensive to produce. PertMind replaces this with measurable gene-response data from perturbation experiments as computable RL rewards. The training combines supervised initialisation on trusted trajectories with gene-level, pathway-level, and format-level RL signals. The model is trained only on forward perturbation-response prediction, yet transfers without task-specific post-training to four related tasks: reverse perturbation identification, double-perturbation reasoning, phenotypic screen prioritisation, and biological process interpretation. PertMind also generates biological profiles competitive with specialised representations for gene, cell, and donor classification tasks. The authors interpret this as evidence that RL on experimental endpoints concentrates reusable biological strategies already latent in pretrained models — similar in spirit to how RLHF surfaces human preference capabilities. The broader claim is that expanding perturbation atlases (CRISPR screens, drug panels) could serve as a scalable, annotation-free training environment for general biological AI. Code, model, and project page are all publicly available.