PerTurboAgent

Self-planning LLM agent from Genentech (Regev lab) that designs iterative Perturb-seq experiments by picking, at each round, which gene perturbations to screen next — choosing among prediction, reflection, and refinement actions over expression data and prior knowledge.

   
Affiliation Genentech, Tsinghua University, Stanford University (proceedings)
First introduced 2025-05 (bioRxiv 2025.05.25.656020; PMLR v311 MLCB 2025 hao25b)
Lifecycle stages Experiment design, Analysis — sequential gene-panel selection with in-loop analysis of control and perturbed gene-expression profiles
Autonomy level Semi-autonomous — human supplies a target phenotype defined by associated descriptive genes (ADGs); the agent then plans, analyzes, and selects each round’s perturbations
Domain focus Single-cell biology, specifically pooled CRISPR Perturb-seq screens
Availability Unknown — no code repository disclosed in the proceedings PDF

Approach

PerTurboAgent runs sequential rounds of Perturb-seq experiments. At each round it selects a set of m gene perturbations from the untested pool to maximize identification of “hit” perturbations — those that drive a strong expression change in the user-defined associated descriptive genes (ADGs) relative to control.

Within each round the agent operates as a multi-step planner with an explicit Action Pool:

  • Agent-based actions — Prediction (new perturbation candidates), Reflection (review results stored in memory), Refinement (refine existing predictions).
  • Data-driven actions — GSEA on control-cell expression; enrichment on positive- and negative-hit perturbation cells; enrichment on the negative-hit subset.
  • Prediction-model-driven actions — train a perturbation-prediction ML model, query perturbation embeddings, compute phenotype scores.

An Action Memory stores each (action, result) pair within the current round so subsequent steps can adapt the plan; at round end the agent emits its gene panel. The agent receives raw control-cell expression at round 0 and the raw expression of perturbed cells for previously selected perturbations in subsequent rounds. External knowledge (e.g., KEGG) is accessible as an additional tool.

Validation

Evaluated on eleven phenotypes drawn from genome-scale Perturb-seq data (Replogle et al. 2022 and related), against agent-based and active-learning baselines: BioDiscoveryAgent, GeneDisco, DiscoBAX, and IterPert. Tested with both closed-source and open-source LLMs as the underlying planner.

Notable results

  • Outperforms prior agent-based (BioDiscoveryAgent) and active-learning (GeneDisco, DiscoBAX, IterPert) baselines across the eleven phenotypes.
  • Compatible with both closed-source and open-source LLM backbones; more capable models give larger gains.
  • Action-frequency and action-memory analyses provide interpretable insight into which reasoning steps drive selection — a transparency property absent from fixed-plan agent baselines.

Primary paper

Hao, Lee, Wang, Scalia, Regev, “PerTurboAgent: An LLM-based Agent for Designing Iterative Perturb-Seq Experiments,” PMLR v311 (MLCB 2025); bioRxiv 2025.05.25.656020.

Other references

Code

Not released at proceedings time.