How AI from Stanford Helps Biologists Plan Experiments

Imagine you need to plan a CRISPR screen to find genes responsible for T-cell exhaustion. Previously, this meant weeks of reading papers, manually selecting targets, and endless spreadsheets. Now you can just write a single line of Python code.

A team from Stanford (the SNAP group) has released Biomni as open source — a universal agent based on large language models, tailored for biomedical tasks. It can answer questions, write code for data analysis, plan experiments, and even generate testable hypotheses.

What Biomni Can Do

The project is interesting because it's not just a wrapper around GPT-4 or Claude. It's a system that combines LLM logic with search tools (RAG) and a code execution environment. The agent "understands" when it needs to look up a database and when to run a script for scRNA-seq processing.

Here are a few real examples of what it does:

• Experiment planning. You describe the task in plain language, and the agent outputs a list of genes to edit or a research protocol.
• On-the-fly data analysis. You can feed it a path to single-cell sequencing results, and it will perform annotation and suggest hypotheses.
• Compound property prediction. The agent works with chemical formulas (SMILES) and can evaluate ADMET parameters (absorption, distribution, metabolism, excretion, and toxicity).

Technical Implementation

The developers went beyond simply using off-the-shelf models. The repository includes Biomni-R0 — a specialized model based on Qwen-32B, fine-tuned using reinforcement learning (RL). It's optimized for tool use and multi-step reasoning in a biological context.

An interesting detail: on first launch, the agent downloads a "data lake" of about 11 GB. This is a local knowledge base that allows the model to avoid hallucinations and rely on verified protocols and publications.

To evaluate quality, the authors created the Biomni-Eval1 benchmark. It contains 433 scenarios: from identifying causal genes in GWAS to selecting CRISPR delivery methods. If you work in bioinformatics, this dataset alone could be useful for testing your models.

How to Run the Project

Installation requires patience, as the environment is quite heavy. The authors prepared a setup script:

conda activate biomni_e1
pip install biomni --upgrade

You'll need API keys from Anthropic or OpenAI. Initializing the agent looks concise:

from biomni.agent import A1

# Инициализируем агента, он сам подтянет нужные данные
agent = A1(path='./data', llm='claude-sonnet-4-20250514')

# Пример задачи: планирование CRISPR скрининга
agent.go("Plan a CRISPR screen to identify genes that regulate T cell exhaustion, generate 32 genes that maximize the perturbation effect.")

If you don't want to write code, the project has a Gradio interface. You can spin it up locally with a single command agent.launch_gradio_demo(), and you'll get a full-featured chat in your browser.

Security and Considerations

There are a couple of things to keep in mind before deploying this to production:

1. Code execution. The agent generates and executes Python code with full system privileges. The README honestly warns: run this only in isolated sandboxes or containers.
2. Licenses. Biomni itself is under Apache 2.0, but it uses third-party databases and tools that may have their own restrictions on commercial use.

Verdict: Is It Worth Trying?

Biomni is a great example of how LLMs are evolving from "smart parrots" into genuinely useful working tools. It's useful for bioinformaticians who want to automate routine tasks and researchers who need quick access to synthesis of knowledge from hundreds of publications.

Currently, the team is building a community for the next version — Biomni-E2. If you have experience in biomedical software development or know great open protocols, this is a good opportunity to contribute to a serious open source project from Stanford.

You can try the web version without installation at the project's official website.