Claude Science: Anthropic's AI Revolutionizes Research

Claude Science: A New Tool for Researchers

Anthropic, a company specializing in artificial intelligence, has recently unveiled an innovative tool for scientists: Claude Science. This platform aims to be an integrated solution, bringing together various tools that researchers use daily into a single interface. The goal is to simplify research processes by centralizing tasks.

For several years, artificial intelligence has gradually made its way into laboratories. However, its use often remains limited to specific tasks. For example, some AI models are designed to summarize articles, others to generate code, or to analyze data. This fragmentation forces scientists to juggle between various software and platforms, which can be time-consuming.

Anthropic intends to change this dynamic with Claude Science. The company presents this platform as a comprehensive research environment, specifically designed to meet the needs of scientists. Currently, Claude Science is available in beta for subscribers of the Claude Pro, Max, Team, and Enterprise plans.

Features of Claude Science

Claude Science is compatible with macOS and Linux operating systems and can be used directly on a personal computer. Researchers also have the option to connect it to a remote machine via SSH or to a high-performance computing (HPC) cluster.

With Claude Science, scientists can perform various tasks such as consulting scientific literature, conducting complex analyses, producing graphs, writing manuscripts, correcting figures, or preparing articles for publication. All of this is possible without leaving the platform's unique interface.

The platform goes well beyond a simple conversational assistant. It relies on a main agent that coordinates over 60 specialized skills. These skills cover a variety of fields such as genomics, proteomics, single-cell analysis, structural biology, and computational chemistry. If necessary, these agents can even create other specialized agents to accomplish specific tasks.

A proofreading agent is also integrated to automatically check citations, calculations, and potential inconsistencies. In case of an error, it can report and even correct it.

Claude Science integrates directly with the scientific models of NVIDIA's BioNeMo platform, including Evo 2, Boltz 2, and OpenFold 3. Researchers also have the option to connect their own analysis pipelines, internal tools, or private databases, allowing the AI to utilize these resources in future projects.

Ensuring Result Reproducibility

One of the major challenges of AI-assisted research is the reproducibility of results. Obtaining a graph is one thing, but being able to explain precisely how it was produced is another.

Anthropic has made this issue a priority with Claude Science. Every figure generated by the platform is accompanied by the code that created it, the computing environment used, and the history of exchanges that led to the final result. This allows a researcher to pick up the same work months later, understand each step, and reproduce the experiment without starting from scratch.

The platform also allows for modifying results using natural language. Requests such as "remove the grid lines," "switch the axis to logarithmic scale," or "change the color palette" can be made simply.

Claude Science is also capable of directly displaying complex scientific objects, such as three-dimensional protein structures, DNA sequences, or chemical molecules, without requiring additional software.

Managing Intensive Computations

Some scientific analyses require considerable computing power, such as protein folding, genomic pipelines, or analyzing millions of data points. Researchers must then prepare their tasks, reserve resources, verify proper functioning, and retrieve results, which can take hours or even days.

Claude Science automates this part of the process, which is often time-consuming. The application prepares the computation plan and requests permission to use new resources. It submits tasks to the infrastructures already used by the laboratory and automatically adjusts the necessary power, ranging from a simple GPU to several hundred.

The agents of Claude Science work within the same session, retaining all context in memory. Thus, large datasets are loaded only once, and sensitive data remains hosted on the laboratory's servers. Only the essential context is transmitted to Claude.

During execution, a control agent monitors the results to detect any erroneous citations, inconsistent values, or figures that no longer correspond to the code used.