Imperial will test whether Claude can rediscover unpublished biology from pre-discovery evidence

A new partnership between Anthropic and Imperial College London contains a more informative scientific test than the usual promise that AI will make research faster. One Imperial group will ask Claude to reason from a deliberately time-bounded evidence set: only the experimental information that was publicly available before the laboratory discovered a previously unknown mechanism of gene transfer. The model will not be shown the unpublished answer. Its task is to see whether it can independently reach the mechanism—or generate a testable alternative—from the same pre-discovery record.[1][2]

That design matters because retrospective claims about AI discovery are vulnerable to contamination. A model can appear prescient if the answer, or a close description of it, entered training data, search results or a later paper. Freezing the evidence at a historical cutoff turns the exercise into a form of scientific forecasting. It still does not prove originality, but it makes the claim much more testable.[1][2]

Two very different scientific programmes

Imperial’s Department of Life Sciences will receive Claude access as part of Anthropic’s wider expansion of its AI for Science programme. The collaboration highlights two projects with different measures of success.[1][2]

The fungal project starts from a method that Imperial says predicts, with around 90% accuracy, how a class of fungal proteins that suppress host immunity interacts with host organisms. Researchers plan to combine that capability with Claude-assisted protein design. The long-term aim is ambitious: crops with greater disease resistance and new approaches to fungal infections in people. But the near-term scientific object is narrower and measurable—whether proposed binders can be designed, produced and tested against the relevant targets.[1][2]

Why the unpublished-mechanism test is unusually valuable

Dr Tiago da Costa’s group is investigating mechanisms involved in bacterial gene transfer and the spread of antimicrobial resistance. By withholding the laboratory discovery, the team can examine more than whether Claude summarises literature well. It can ask whether the model selects the decisive evidence, notices contradictions, proposes a mechanism that is not merely the most conventional one, and suggests experiments that could falsify its own account.[1][2]

A strong evaluation should record the model’s inputs, tools, prompts, intermediate hypotheses and revisions before the answer is revealed. It should also test multiple runs and models. A single successful output can be luck; a system that repeatedly identifies the same mechanism from the same evidence is more informative. Equally important, a useful model may fail to reproduce the exact discovery yet still propose a plausible mechanism that leads to a productive experiment.[1][2]

Access is not discovery

Anthropic’s programme supplies free Claude credits and partnerships to institutions including Imperial, the Wellcome Sanger Institute, the Francis Crick Institute, EMBL and the Space Telescope Science Institute. That broadens access, but compute and model tokens are inputs rather than scientific outcomes. The partnership announcement reports plans and early capabilities, not validated discoveries produced by the collaboration.[1][2]

There are also governance questions. Scientific models can produce confident but fragile explanations, and wet-lab work can be expensive enough that a persuasive false lead has a real cost. Researchers therefore need pre-registered evaluation criteria, independent replication where possible and a clear separation between model-generated hypotheses and experimentally established mechanisms.[1][2]

A better standard for AI-for-science claims

The most credible AI-for-science work will make the path from evidence to hypothesis auditable and then submit the hypothesis to the same experimental discipline as any human idea. The Imperial collaboration has the ingredients for that standard: a known but hidden answer, a historical evidence boundary and laboratories able to test what the model proposes.[1][2]

If Claude can recover the mechanism, the result will still need analysis of contamination, reproducibility and alternative explanations. If it cannot, that failure will also be useful, because it will show where literature-scale reasoning diverges from experimental insight. Either outcome can advance the field—provided the test remains more important than the demonstration.[1][2]

Report an error or suggest a correction