tzcfly/PertMind
PertMind is a 4 billion parameter biological language model developed by tzcfly, initialized from Qwen/Qwen3-4B-Base. It uniquely reorganizes public cellular perturbation atlases into reinforcement-learning environments, using measured gene responses as reward signals for biological reasoning. This model specializes in perturbation-centered inference and biological profile generation, combining supervised initialization with perturbation-derived reinforcement learning.
Loading preview...
PertMind: A Biological Language Model for Perturbation Reasoning
PertMind is a 4 billion parameter biological language model, initialized from the Qwen3-4B-Base family, developed by tzcfly. Its core innovation lies in treating cellular perturbation atlases as reinforcement learning environments, where gene responses provide computable reward signals for biological reasoning. This approach allows PertMind to combine trusted-trajectory supervised initialization with perturbation-derived reinforcement learning, focusing on perturbation-centered inference and biological profile generation.
Key Capabilities
- Perturbation-Response Prediction: Predicts how specific genes respond to perturbations in various cell lines, based on experimentally observed outcomes.
- Reinforcement Learning from Biological Data: Utilizes a three-stage process involving gene-centered queries from the Tahoe-100M perturbation atlas, supervised initialization with trusted trajectories, and Group Relative Policy Optimization with a composite reward.
- Operational Emergence: Demonstrates transfer capabilities to tasks beyond its post-training objective, such as reverse perturbation-condition inference, phenotypic-screen prioritization, and biological-process interpretation, suggesting it learns reusable biological strategies.
- Biological Profile Generation: Creates PertMind-derived biological profiles for downstream molecular, cellular, and donor-level representations.
Good for
- Research Use: Ideal for hypothesis generation and method development in biological research.
- Forward Perturbation-Response Reasoning: Predicting outcomes in held-out biological contexts.
- Reverse Perturbation-Condition Inference: Inferring conditions from observed perturbations.
- Biological Process Interpretation: Aiding in understanding and naming biological processes.
- Screen-Oriented Biological Briefing: Prioritizing hits in phenotypic screens.
Note: PertMind is intended for research and development; it is not for clinical diagnosis, treatment selection, or replacing wet-lab validation.