Research map / Multi-agent systems and games / LLM agents

BioDiscoveryAgent: An AI Agent for Designing Genetic Perturbation Experiments

ICLR 2025 · Multi-agent systems and games › LLM agents · approach: Agentic benchmarks and orchestration · 16 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces BioDiscoveryAgent, an LLM agent that designs genetic perturbation (CRISPR) experiments in closed loop by reasoning over prior knowledge and experimental results, optionally with literature, gene-search, and AI-critic tools, outperforming Bayesian optimization baselines on six single-gene datasets and extending to gene-pair perturbation.

Problem

Selecting small batches of genes to perturb in CRISPR screens so as to maximize the number of phenotype-relevant hits over few experimental rounds, where experiments are costly and training ML models on small data is limited.

Approach

Replaces trained ML models and hand-designed acquisition functions with a single LLM prompted with the task and prior round results, using biological knowledge for cold start, plus optional tools (PubMed literature search, Reactome-based gene search, LLM critic) and a structured Reflection/Research Plan/Solution response format.

Key results

• Claude 3.5 Sonnet agent (no tools) achieves on average about 21% higher hit ratio at round 5 than the best ML/Bayesian baseline across six datasets, and about 46% higher on non-essential genes.
• On 2-gene perturbation prediction in K562, the Claude 3.5 Sonnet agent finds about 170% more hits than random sampling (40.0 vs 16.4 cumulative hits at round 5).
• Tool use helps smaller models (Claude 3 Haiku with all tools gains about 21-31% on several datasets) but did not help Claude 3.5 Sonnet, which was hurt on some datasets (e.g. -28% on Scharenberg with all tools).

Measured effects reported in the paper

Surprising findings

• Tool use hurt larger models such as Claude 3.5 Sonnet and GPT-4o, while it improved smaller models like Claude 3 Haiku and Claude v1.
• The best baseline's first-round predictions combined with the agent did not outperform the agent alone.

Limitations

• Performance benefits concentrated in early rounds; gains shrink in longer runs.
• Model performance varies across cell types, likely tied to literature coverage.
• Unpublished CAR-T dataset is not publicly available, limiting reproducibility.

Experimental setup

Datasets: Schmidt1 (IFNG CRISPR screen) Schmidt2 (IL-2 CRISPR screen) CAR-T (unpublished) Scharenberg (lysosomal choline recycling) Carnevale (T cell proliferation) Sanchez (tau expression) K562 gene pair screen (100,576 pairs)

Benchmarks: Schmidt1 Schmidt2 CAR-T Scharenberg Carnevale Sanchez K562 gene pair screen

Metrics: hit ratio cumulative hits Jaccard similarity

Baselines: Random Human Soft Uncertain Top Uncertain Margin Sample Coreset Badge K-Means DiscoBax

Builds on: GeneDisco DiscoBAX ReAct RLAIF Reflexion

Closest papers by meaning

• Contextualizing biological perturbation experiments through language (ICLR 2025 · similarity 0.88)
• AutoScientists: Self-Organizing Agent Teams for Long-Running Scientific Experimentation (NeurIPS 2026 · similarity 0.86)
• PerturbReason: A Knowledge-Grounded Benchmark and Framework for Cell-State–Conditioned Mechanistic Reasoning of Perturbation Effects (NeurIPS 2026 · similarity 0.85)
• BioAgent Bench: An AI Agent Evaluation Suite for Bioinformatics (ICML 2026 · similarity 0.85)
• Protein Design with Agent Rosetta: A Case Study for Specialized Scientific Agents (ICML 2026 · similarity 0.84)
• Efficient Biological Data Acquisition through Inference Set Design (ICLR 2025 · similarity 0.84)