snap-stanford/Biomni

Biomni: a general-purpose biomedical AI agent

What it solves

Biomni is designed to automate complex research tasks across various biomedical subfields. It addresses the productivity gap in scientific research by helping scientists generate testable hypotheses and execute multi-step analysis workflows that would otherwise require manual coordination of multiple tools and data sources.

How it works

Biomni operates as an AI agent that combines large language model (LLM) reasoning with retrieval-augmented planning and code-based execution. It utilizes a "Know-How Library" of curated best practices and protocols, and can integrate with external tools via the Model Context Protocol (MCP). The system can be further enhanced by using Biomni-R0, a specialized reasoning model built on Qwen-32B and trained with reinforcement learning from agent interaction data.

Who it’s for

Biomni is built for biomedical researchers and scientists who need to automate data analysis, plan experimental designs (such as CRISPR screens), and perform complex biological problem-solving.

Highlights

  • General-Purpose Biomedical Agent: Capable of tasks ranging from scRNA-seq annotation to predicting ADMET properties of compounds.
  • Biomni-R0: A dedicated reasoning model optimized for tool use and iterative self-correction in biological contexts.
  • Biomni-Eval1: A comprehensive benchmark containing 433 instances across 10 biological reasoning tasks for performance assessment.
  • Know-How Library: An automatically retrieved collection of lab protocols and analysis best practices to provide domain expertise.
  • MCP Support: Ability to integrate external tools and servers for expanded functionality.

Related

  • Project
    exon-research/genomi

    Genomi is an open‑source AI‑agent runtime that lets local LLM agents answer personal genomics questions. It builds a private, queryable genome index from VCF/CSV/FASTQ files, ships with curated evidence libraries (ClinVar, GWAS, PharmGKB, etc.), and exposes `/genomi` commands via the Model Context Protocol (MCP). All raw DNA stays on the user’s machine, and the system logs citations and approvals for privacy‑first, evidence‑backed answers.

  • Project
    liruiw/GenSim

    GenSim is an LLM-powered pipeline that automatically generates robotic simulation tasks and environments to augment training data for multitask robotic policies.

  • Dispatch
    Qwen-AgentWorld release: language world model for seven domains and its impact on general agents

    Qwen releases Qwen‑AgentWorld, a language world model that simulates seven agent environments and improves general agents via controllable simulation and unified next‑state prediction.

  • Project
    mims-harvard/ToolUniverse

    An ecosystem for building AI scientist systems that provides a standardized protocol and access to over 1,000 scientific tools, ML models, and datasets.

  • Project
    yeahhe365/Prisma

    A visual deep multi-agent reasoning engine for Gemini and OpenAI-compatible models that decomposes complex queries into parallel expert tasks with a transparent reasoning workflow.