Perturbation-derived reinforcement learning

Perturbation data can teach language models to reason about biology.

PertMind turns measured gene responses into computable reward signals. Instead of depending on manually curated reasoning traces, it uses perturbation observations to shape structured biological reasoning at scale.

Explore the model View source
  • Gene, pathway, and format rewards grounded in perturbation observations.
  • Capability transfer beyond the explicit post-training task.
  • Molecular, cellular, and donor representations from generated profiles.

Core discovery

From perturbation atlas to a verifiable biological RL environment.

Measured terminal outcomes can supervise learning without pretending to reveal every hidden mechanism.

PertMind operationalizes cellular perturbation observations as cell line, perturbation, target gene queries with experimentally observed responses. These endpoints provide scalable reward signals for learning structured biological outputs.

The claim boundary is explicit: correct perturbation endpoints do not by themselves prove that every generated intermediate explanation is causally faithful. The model's reasoning remains a hypothesis-bearing scaffold for research, not a proof of complete biological mechanism.

01

Perturbation-derived RL

Gene, pathway, and format rewards stay tied to perturbation observations rather than free-form preference signals.

02

Emergent biological reasoning

“Emergent” here means usable capability on a task that was absent from post-training, not a claim of omniscience.

03

One model, multiple scales

PertMind-generated profiles can support downstream molecular, cellular, and donor representations alongside reasoning tasks.

Capability overview

Figure 1 maps the intended capability space.

Figure 1 schematic showing PertMind reasoning-based and embedding-based capability scope across perturbation-response tasks and biological representations.
Figure 1. Schematic scope of the capabilities enabled by PertMind. The figure describes the intended application space, while the quantitatively evaluated subset currently covers forward perturbation-response prediction, reverse perturbation-condition inference in single- and double-perturbation settings, phenotypic-screen hit prioritization, biological-process naming, and molecular, cellular, and donor representation tasks.

How it works

Three training stages, one restrained objective.

01

Structure the perturbation queries

Organize cellular perturbation observations into cell-line, perturbation, and target-gene prompts with measurable response labels.

02

Initialize with trusted trajectories

Use supervised fine-tuning on trusted structured outputs so the model starts from a legible reasoning format.

03

Optimize experimentally grounded rewards

Run GRPO over gene outcomes, pathway directions, and output format consistency, all tied back to perturbation evidence.

Evidence and boundaries

What is supported today, and what remains intentionally bounded.

Evidence-supported transfer areas

  • Forward perturbation-response prediction
  • Reverse perturbation-condition inference in single- and double-perturbation settings
  • Phenotypic-screen prioritization
  • Biological-process naming
  • Molecular, cellular, and donor representation tasks

“Emergent” is used operationally: a capability counts as emergent when it appears on a task absent from post-training.

Generated reasoning should be read as a hypothesis-bearing explanation that helps users inspect possible biological logic. It is not proof of a complete causal mechanism, and Figure 1 should not be read as a benchmark summary for every depicted use case.

Model access

Public model, source, and quick-start path.

Quick start

git clone https://github.com/shapsider/PertMind
cd PertMind
conda create -n pertmind python=3.11 -y
conda activate pertmind
git lfs install
git clone https://huggingface.co/tzcfly/PertMind model-release
cd model-release
pip install -r requirements.txt
python inference_vllm.py \
  --model . \
  --prompt "In HCT116, predict how LMBR1 changes expression of SLC26A6."