An AI agent that plans CRISPR screens and reads lab manuals

What it does

Biomni is a Python-based autonomous agent for biomedical research. It accepts natural language tasks—like planning CRISPR screens, annotating scRNA-seq data, or predicting ADMET properties for compounds—and executes them by combining LLM reasoning with retrieval-augmented planning and code-based execution. The system downloads an 11 GB biomedical datalake on first run and includes a curated Know-How Library of protocols and best practices that the agent retrieves automatically when relevant.

The interesting bit

The project ships with its own reasoning model, Biomni-R0, a Qwen-32B variant trained with reinforcement learning on agent interaction data specifically for biological tool use and multi-step reasoning. It also includes Biomni-Eval1, a benchmark of 433 instances across 10 biological tasks, which suggests the authors are serious about measuring whether the agent actually works rather than just demoing it.

Key highlights

• Supports multiple LLM backends including Anthropic, OpenAI, Azure, Gemini, Groq, and local custom models via an OpenAI-compatible API
• Integrates external tools through the Model Context Protocol (MCP), allowing the agent to query FDA databases and other resources
• Includes a Know-How Library that automatically retrieves lab protocols, troubleshooting guides, and metadata (licensing, commercial use) based on query relevance
• Can export conversation histories as PDF reports and exposes a Gradio web interface
• Provides a centralized configuration system via default_config to manage LLM sources and timeouts consistently across agents and database queries

Caveats

• The software environment is described as “massive” and some Python packages are excluded by default due to known dependency conflicts, requiring manual installation and possible code uncommenting
• The 11 GB datalake downloads automatically on first agent initialization unless explicitly skipped, which may surprise users with limited bandwidth or storage
• Using the project’s own Biomni-R0 reasoning model requires a local SGLang server with substantial GPU memory and tensor parallelism, making it a heavy self-hosted lift rather than a casual model swap

Verdict

Worth exploring for biologists tired of writing analysis pipelines, but the massive environment, 11 GB datalake, and known package conflicts make it a commitment, not a quick experiment.