科學 AI 與模型後訓練
PertMind Trains a 4B Model on Cellular Perturbation Endpoints, Turning 100 Million Single-Cell Profiles into Reinforcement Learning Rewards
Starting from Qwen3-4B, PertMind combines gene responses, pathway direction, and output formatting into computable rewards, eliminating the need to manually author large numbers of biological reasoning traces. The model transfers zero-shot to inverse perturbation identification and dual-perturbation tasks, but its natural-language mechanistic explanations still cannot be treated as causal proof.
PertMind attempts to address a core bottleneck in post-training scientific language models: experimental datasets typically record only treatment conditions and final measurements, without complete natural-language reasoning processes. The team organized precomputed differential-expression statistics from roughly 100 million single-cell profiles in Tahoe-100M into queries consisting of a cell line, a drug perturbation, and a target gene, then asked the model to predict whether gene expression increased, decreased, or showed no reliable change.
The system starts from Qwen3-4B Base. In the first stage, it retrieves context from public biological knowledge graphs and the training split, samples candidate explanations, and retains only traces whose conclusions are correct, whose formats are parseable, and whose entity relationships are supported by sources. It then performs one round of weighted supervised fine-tuning (SFT). The second stage uses Group Relative Policy Optimization (GRPO), combining three signals: experimentally observed gene endpoints, pathway direction calculated from the transcriptional responses of other pathway members, and format compliance. The label for the current query and the pathway proxy label exist only on the reward side and are not included in the prompt.
To reduce leakage, the authors split the data by cell line and held out five complete cell lines for testing. Differential-expression values and labels for the current query, along with results from the test split, are excluded from retrieval. The paper reports that, on differential-expression detection and direction prediction in unseen cell lines, PertMind on average matched or slightly outperformed the VCWorld reference pipeline using Gemini 2.5 Flash. Although trained only on forward-response tasks, the model could, without additional fine-tuning, infer perturbations from cellular states, rank the correct choice among 69 candidates, and handle 131 dual-gene perturbations. It showed only small declines on MMLU and CMMLU.
What is truly notable is not any particular leaderboard position, but the conversion of high-throughput experimental endpoints into a verifiable reinforcement learning environment. This approach could extend to scientific fields that have abundant measurements but lack human-authored reasoning annotations. Engineers should next examine whether the reward proxies induce shortcuts and evaluate out-of-distribution performance across experimental platforms and novel drugs. The authors also explicitly warn that a correct endpoint constrains only the conclusion; it does not prove that the model’s generated intermediate mechanism is causally faithful.