Biomni Integrates 105 Tools and 59 Databases to Enable AI‑Driven End‑to‑End Life‑Science Discovery

Biomni Overview

Biomni is a general‑purpose biomedical AI agent introduced in Science (2026‑07‑09). It is designed to act as an autonomous research assistant that can understand problems, locate tools, design workflows, execute analyses, and suggest experiments.

Environment (Biomni‑E1)

The Biomni‑E1 environment was built by mining a large corpus of papers from 25 biomedical domains to extract experimental tasks, tools, databases, and software, creating a unified research‑tool space.

It integrates 105 widely used bioinformatics software packages and 59 curated databases, including protein‑structure, disease, and variant repositories.

Agent Architecture (Biomni‑A1)

Dynamic resource selection – the AI automatically determines which databases, software, and methods are needed for a given goal.

Code as a universal execution language – database queries, data processing, model predictions, and analysis steps are expressed as code and executed as a coherent workflow.

Dynamic planning – the plan is iteratively revised based on intermediate results, allowing adaptive optimization of the research pathway.

Benchmark (Biomni‑Eval1)

Biomni‑Eval1 contains 443 questions covering ten typical biomedical tasks such as CRISPR delivery, causal‑gene identification, variant prioritisation, database querying, DNA‑sequence analysis, rare‑disease diagnosis, and experimental design.

Biomni achieved an average accuracy of 57%.

Comparative results:

Claude Sonnet 4.5: 30% accuracy.

TxAgent: 25% accuracy.

Claude Code: 43% accuracy.

ReAct system combined with the Biomni environment: 44% accuracy.

Expert‑Level Task Evaluations

Single‑cell annotation: 45.8% accuracy (within senior‑expert range 40.5%–50.9%); runtime reduced from ~230 minutes to 75 minutes.

Rare‑disease diagnosis: 60% accuracy; runtime reduced from 110 minutes to 3 minutes.

GWAS causal‑gene detection: 80% accuracy (expert level); runtime reduced from 90 minutes to 4 minutes.

Integration with Wet‑Lab Automation

Biomni was connected to the PyLabRobot framework. Natural‑language experiment requests are translated into executable liquid‑handling robot code, enabling protocols such as gradient dilutions and cell‑viability assays.

Case Studies

In a multi‑omics research case, Biomni autonomously analysed heterogeneous data, identified key regulatory networks from gene‑expression changes, and proposed follow‑up validation experiments.

In a protein‑engineering task, Biomni invoked structure‑prediction tools and relevant databases, evaluated protein stability changes, and suggested design strategies to improve performance.

Implications

The combination of a unified tool space, code‑based action interface, and adaptive planning demonstrates a viable path toward AI‑augmented laboratories where scientists pose questions, the AI plans and runs experiments, and iteratively refines hypotheses based on feedback.

Paper link: https://www.science.org/doi/10.1126/science.adz4351

Code example

来源：ScienceAI
本文
约2000字
，建议阅读
5
分钟
如果AI不只是帮助科学家分析某一步，而是能够像一名研究助理一样，自主理解问题、寻找工具、设计流程、执行分析，甚至提出实验方案，会发生什么？