Anthropic Launches Claude Science Beta: A Multi-Agent AI Workbench for Reproducible Genomics, Proteomics, and Cheminformatics Pipelines

This week, Anthropic was released Claude Science. An app for scientists, available in beta. It uses existing Anthropic Claude models, not a new model. The application guides researchers integrating databases, notebooks, and collection terminals. It conducts a multi-step study and records how each result is performed. The beta is available for Pro, Max, Team, and Enterprise plans.

Claude Science builds on Anthropic’s life science work from last fall. That previous job connected Claude to the environmental science program through MCPs and skills.

What is Claude Science?

Claude Science is an AI research bench. It includes the tools and packages that researchers use the most. Analyzes literature, conducts multi-step research, and produces detailed artifacts. You can refine calculations and manuscripts until they are ready for publication.

You talk to one generalist liaison agent in plain language. That agent has access to over 60 selected skills and connectors. This comes pre-prepared for genomics, single cell, proteomics, structural biology, and cheminformatics.

You can use it locally on macOS or Linux. You can also work on a remote machine via SSH or an HPC login. Everything that comes out has a readable history of how it was made.

How Multi-Agent Architecture Works

A generalist compounding agent receives your application in plain language. It can integrate other agents to manage the work. It may also involve special agents that users create for themselves. NVIDIA describes these as preconfigured, domain-specific agents. Each one knows the established workflow of their field.

A different updater agent is running as the pipeline starts. It checks the output step by step. It flags wrong quotes and numbers that it can’t track. It also flags statistics that do not match their underlying code. Then it corrects itself as it goes.

Reproduction and Provenance

Scientific research is evident in nature. So Claude Science produces calculations and manuscripts near the code of the elders. It natively provides 3D protein structures, genome browser tracks, chemical structures, and more.

When it generates a calculation, it records the exact code and location. It also records plain language description and full message history. This makes the work easier to verify and reproduce months later.

You can edit the math in simple language. For example, you can ask it to change the axis to scale the entry. The agent then edits its own code. You can also merge a session to compare two methods without losing the first one.

Calculate that Scale on Demand

Larger analyzes often require more than a laptop. Protein folding is one example. Claude Science writes a program before accessing new resources. It asks for approval and allows you to review or reverse any decision. It then logs and delivers the job to your infrastructure.

That means your HPC cluster over SSH or your Modal account. Analysis scales from one GPU to multiple as needed. Because agents hold context in memory, a large dataset loads only once.

The application works with your lab infrastructure. Very large or sensitive datasets do not need to leave their current systems. Only the necessary context for each step is sent to Claude.

Background Installation and NVIDIA BioNeMo

Scientific information is distributed in thousands of specialized sources. In biology, this includes UniProt, PDB, Ensemble, and Reactome. Also includes ClinVar, ChEMBL, GEO, journals, and preprint servers. Professional agents ask and connect you to all these sources.

Claude Science also uses capabilities from NVIDIA’s BioNeMo Agent Toolkit. The toolkit packs GPU-accelerated capabilities as hitable capabilities. This connects naturally to the Evo 2, Boltz-2, and OpenFold3. Evo 2 is a basic genomics model. Boltz-2 handles predicting biomolecular interactions. OpenFold3 handles protein structure prediction.

Use Cases with examples

Beta users initiated cell RNA sequencing analysis and CRISPR screen design. They also use protein structure prediction and cheminformatics.

• Target designation: Manifold Bio designs drugs that target tissues. It used Claude Science to select targets for its latest experiments. For each tissue and target, the application evaluated localization, trafficking, and safety. It then ranked the candidates against Manifold’s proprietary criteria. Nifold said the app did this to the end, unlike a regular coding assistant.
• Long-form literature review: Jérôme Lecoq at the Allen Institute developed a computerized review template. It contains about 20 custom skills for long form reviews. Sub-agents read thousands of documents in the federal evidence database. The pipeline then writes each section using agents that critique the character. Such an update has taken his team for a long time of two years. Now you have about 10 reviews, many more than 100 pages.
• Genomic Epidemiology: Stephen Francis at UCSF is studying the molecular epidemiology of glioma. Claude Science does germline workups about one-tenth of the time. His team independently verified the results.

Comparison table

Expanding Claude’s Science

Claude Science is an application, so it does not have a separate indexing API. You extend it with connectors and capabilities, which are continuous throughout the sessions.

You connect the tool to the lab through the Model Context Protocol (MCP) connector. This is the standard configuration format for an MCP client:

{
  "mcpServers": {
    "lab-eln": {
      "command": "npx",
      "args": ["-y", "@lab/eln-mcp-server"],
      "env": { "ELN_API_KEY": "REPLACE_ME" }
    }
  }
}

You save an existing pipeline as a reusable capability. The ability is a folder that contains a SKILL.md file:

---
name: rnaseq-qc
description: Run the lab's standard RNA-seq quality-control pipeline on a FASTQ directory.
---

# RNA-seq QC

1. Run `pipelines/qc.sh `.
2. Summarize the per-sample metrics.
3. Flag any sample below the QC threshold.

Future sessions inherit these connectors and capabilities automatically. So you keep your certified tools and data, while Claude organizes them.

Key Takeaways

• Claude Science is a beta app for macOS and Linux; uses Claude’s existing Anthropic models.
• A linking agent sends the messengers into action, while a separate review agent checks quotes, numbers, and statistics.
• Everyone comes with their own specific code, environment, description, and full message history.
• Compute runs locally, on HPC via SSH, or on Modal, scaling from one GPU to hundreds.
• It ships with 60+ insights and capabilities for NVIDIA BioNeMo (Evo 2, Boltz-2, OpenFold3) for life sciences.

Check it out Technical details here. Also, feel free to follow us Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to Our newspaper. Wait! are you on telegram? now you can join us on telegram too.

Need to work with us on developing your GitHub Repo OR Hug Face Page OR Product Release OR Webinar etc.? contact us