PertMind: Eliciting Emergent Biological Reasoning in LLM via Reinforcement Learning on Cellular Perturbation Data
Abstract
Large language models can describe mechanisms, yet scalable post-training still depends on costly, manually curated biological reasoning traces. Here we show that cellular perturbation atlases can instead become reinforcement-learning environments, where measured gene responses provide computable rewards for biological reasoning. We introduce PertMind, which combines trusted-trajectory supervised initialization with gene-, pathway-, and format-level reinforcement signals. Although trained only on forward perturbation-response prediction, PertMind improves response inference in unseen cellular contexts while retaining general language capabilities. It also transfers, without task-specific post-training, to reverse perturbation identification, double-perturbation reasoning, phenotypic-screen prioritization, and biological-process interpretation. PertMind further generates biological profiles that support competitive gene, cell, and donor representations across multiscale downstream tasks. These results support the hypothesis that reinforcement on experimental endpoints can concentrate reusable biological strategies already accessible to pretrained models. More broadly, perturbation-derived reinforcement learning offers a scalable route for transforming expanding experimental atlases into training environments for general-purpose biological reasoning.
Related
- CellChain: Benchmarking Cross-Scale Evidence Use for Perturbation Reasoning
- BioPhys-Bridge: A Benchmark for Interdisciplinary Scientific Reasoning in Physics-Grounded Biological Research
- A biological-response compound representation allows chemical perturbation prediction across cell lines
- Reinforcement Learning Reveals Spatial Strategies for Slowing Biological Invasions
- Reelin doses normalizing behavioural alterations induced by repeated corticosterone avoid eliciting neuroinflammation-associated morphology in hippocampal microglia
BibTeX
@misc{tang2026pertmind,
title = {{PertMind: Eliciting Emergent Biological Reasoning in LLM via Reinforcement Learning on Cellular Perturbation Data}},
author = {Zhenchao Tang and Xiaogang Xu and Jiafei Wu and Jiahui Guan and Bo Li and Tianxu Lv and Jiale Zhou and Haohuai He and Zhi Song and Hanbo Huang and Jiehui Huang and Xun Lin and Zhipeng Deng and Zhaoxing Li and Guanxing Chen and Yaokun Li and Mengran Li and Songming Zhang and Zhe Liu},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2608.16419},
eprint = {2608.16419},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2608.16419},
}
RIS
TY - UNPB
TI - PertMind: Eliciting Emergent Biological Reasoning in LLM via Reinforcement Learning on Cellular Perturbation Data
AU - Zhenchao Tang
AU - Xiaogang Xu
AU - Jiafei Wu
AU - Jiahui Guan
AU - Bo Li
AU - Tianxu Lv
AU - Jiale Zhou
AU - Haohuai He
AU - Zhi Song
AU - Hanbo Huang
AU - Jiehui Huang
AU - Xun Lin
AU - Zhipeng Deng
AU - Zhaoxing Li
AU - Guanxing Chen
AU - Yaokun Li
AU - Mengran Li
AU - Songming Zhang
AU - Zhe Liu
PY - 2026
DA - 2026/10/06
PB - arXiv
DO - 10.48550/arxiv.2608.16419
UR - https://arxiv.org/abs/2608.16419
AB - Large language models can describe mechanisms, yet scalable post-training still depends on costly, manually curated biological reasoning traces. Here we show that cellular perturbation atlases can instead become reinforcement-learning environments, where measured gene responses provide computable rewards for biological reasoning. We introduce PertMind, which combines trusted-trajectory supervised initialization with gene-, pathway-, and format-level reinforcement signals. Although trained only on forward perturbation-response prediction, PertMind improves response inference in unseen cellular contexts while retaining general language capabilities. It also transfers, without task-specific post-training, to reverse perturbation identification, double-perturbation reasoning, phenotypic-screen prioritization, and biological-process interpretation. PertMind further generates biological profiles that support competitive gene, cell, and donor representations across multiscale downstream tasks. These results support the hypothesis that reinforcement on experimental endpoints can concentrate reusable biological strategies already accessible to pretrained models. More broadly, perturbation-derived reinforcement learning offers a scalable route for transforming expanding experimental atlases into training environments for general-purpose biological reasoning.
ER -