OpenScientist
An autonomous AI scientist that generates and tests hypotheses from scientific data.
OpenScientist is an open-source, domain-agnostic autonomous discovery agent led by Justin Reese (Lawrence Berkeley National Laboratory) and released under the Apache-2.0 license. Given a set of data files and a research question, the system runs an iterative loop — generating hypotheses, testing them through sandboxed Python code execution, and searching PubMed for supporting literature — before producing a final report of findings and mechanistic interpretations. The project documents applications across genomics, transcriptomics, proteomics, and metabolomics data, and supports common scientific file formats (CSV, TSV, Excel, Parquet, JSON, PDB, mmCIF, FASTA, images, among others).
OpenScientist Agent
The underlying agent is built on an agentic coding-assistant framework, exposed via Model Context Protocol (MCP) tools for code execution, literature search, and knowledge-state tracking. An optional module adds structural biology capabilities through integration with Phenix, supporting structure comparison and superposition, geometry and validation metrics, and AlphaFold confidence analysis. The project also includes a “skills” system separating general workflow logic (hypothesis generation, result interpretation, prioritization, stopping criteria) from domain-specific skills (metabolomics, genomics/transcriptomics, structural biology, statistics).
OpenScientist can connect to several model-hosting providers — CBORG (Lawrence Berkeley National Laboratory’s API service), Google Vertex AI, AWS Bedrock, or Azure AI Foundry — with per-project cost tracking against each provider’s billing APIs. It is deployed via Docker with a browser-based (NiceGUI) interface for submitting and monitoring jobs, and supports multiple concurrent jobs through an internal job manager. A live instance is hosted at openscientist.io. As of this writing, the project is at an early release stage (v0.1.0, March 2026).