Anthropic has released Claude Science, which is essentially a workbench for AI researchers that tackles one of the least glamorous aspects of scientific research – the sheer burden of administration. In scientific research, there are many databases each having different schemas, incompatible file formats which require special pipelines and special viewers and jumping from tool to tool such as PubMed, Jupyter, R, and even the cluster terminal. All of this is combined into one in Claude Science.
Also read: Claude Sonnet 5 vs Opus 4.8: Is the flagship model still worth paying for
It can be used both locally on macOS and Linux or remotely using SSH or HPC login node similar to how Jupyter Notebook is used. The heart of it consists of a general purpose coordinating agent which has access to over 60 skills and connectors that have been configured for genomics, single cell analysis, proteomics, structural biology, cheminformatics and many other topics. These agents themselves can create sub-agents and a special reviewing agent ensures correct citations and computations.
There are two important aspects of Claude Science. First, everything is reproducible as all figures are delivered with the precise code and environment which were used to generate the figure, its description in simple language how it was done and the whole message history, so that months later one can verify the result. In addition, a scientist can ask the agent to adjust some aspects of the figure like removing grid lines or applying log-scale using natural language.
Also read: Fable 5, Mythos 5 coming back: Anthropic still hasn’t answered key question
Second, Claude Science computes by itself. It creates the computational plan, gets an approval before using any new resources and gives an opportunity for researchers to review or revoke any action before submitting a job to their lab’s HPC cluster or Modal account for on-demand computation, from 1 GPU to hundreds, as required. Confidential data do not need to be transferred outside of the current infrastructure, because only the context necessary for each step of the analysis is being transferred to Claude. As a unique feature, Claude Science uses NVIDIA BioNeMo Agent Toolkit to natively connect to life science models Evo 2, Boltz-2 and OpenFold3.
Examples of early adopters provide an idea of the variety. The biotech company Manifold Bio leveraged Claude Science for the nomination of drug targets for tissue-specific drugs using evaluation of surface expression, trafficking, and safety among other candidates. In the Allen Institute, Jérôme Lecoq (neuroscientist) has developed a 20-skills multi-agent pipeline that writes long-form literature reviews and reduces what would normally take up to two years into a matter of hours; it has so far written around 10 literature reviews, some more than 100 pages long. And at UCSF, the epidemiologist Stephen Francis employed it to accelerate glioma genetics to one-tenth of its earlier time, with validation from his lab of the results.
Claude Science is available in its beta version to Pro, Max, Team, and Enterprise customers on macOs and Linux systems, with reduced Team license prices for academic and non-profit laboratories and grants of up to $30,000 in credits for up to 50 AI for science projects selected, application for which closes July 15th.
Also read: Not thinking about storage in AI is a mistake: Dell’s Venkat Sitaram