Claude designed protein binders that worked in lab tests, but the results do not prove AI can create medicine #
Claude AI designed 1,320 protein binders and sent them into an automated lab pipeline, where 354 were confirmed to bind their intended targets, raising fresh questions about how far AI can now go on its own in biological design.
The work, led by Adaptyv Bio using Anthropic's Claude models, effectively showed an AI system assembling and running much of a protein-design workflow, with humans and lab robots handling only the physical experiments.
The system achieved a 26.8% average hit rate, with successful binders across 14 of 15 targets. One target produced no binders, and another was excluded due to low-quality measurements.
The setup used Claude Science, Anthropic's research platform, which connects a general-purpose model to specialist tools for protein structure prediction and design. Anthropic supplied an initial expert-written protocol. Claude then used available tools to generate, evaluate and refine designs before sequences were sent for lab testing.
Adaptyv Bio converted sequences into DNA, produced the proteins and measured binding. It reported a 95% expression success rate.
What the Lab Tests Showed #
Performance varied sharply by target. Claude reached an 80% hit rate against TREM2, up from 38.3% in a prior Adaptyv Bio competition. The strongest designs also outperformed winners from five of six previous competitions on binding affinity.
For 15-PGDH, the best Claude binder measured 33.4 nanomolar versus 1.7 micromolar previously. For RBX1, it was 3.9 nanomolar versus 25.7 nanomolar. Adaptyv Bio said the results would have won five of its six past protein-design competitions and described the system as operating at an expert level when using existing design tools. The comparison is internal to Adaptyv Bio's competitions and does not reflect pharmaceutical industry benchmarks.
Shkreli Questions the Value #
Martin Shkreli, the former pharmaceutical executive, criticised the significance of the results. 'This is not impressive work,' he said, arguing the reported affinities were weak for peptide binders. He also questioned the lack of intracellular targets.
'These aren't useful probe molecules because none of them are intracellular,' he said. 'If I needed an extracellular probe, that's what a monoclonal antibody is for.' He added that existing antibody fragments could often serve the same function without designing new binders.
His comments underline a key gap: binding a target is not the same as producing a useful drug candidate. Anthropic makes the same distinction.
'Protein binders are not drugs,' the company said, describing them as an early step toward therapeutics that still require extensive optimisation and testing.
How Autonomous Was Claude's Design Loop? #
The main technical result is not just the lab validation, but how much of the design process was automated.
Adaptyv Bio said Claude combined general reasoning with specialist protein-design tools in an automated workflow.
The model interpreted objectives, generated candidates and iterated using computational feedback. The system reflects what Adaptyv Bio calls an 'agentic science loop': AI generates designs, selects candidates and feeds results back into further design cycles.
Claude did not perform physical lab work. DNA synthesis, protein production and measurements were handled by automated systems and human operators. But a large portion of the computational design loop was executed by the model itself.
A Binder Is Not a Medicine #
The results do not show drug discovery. A protein binder is only an early-stage component. Any therapeutic candidate must still pass through optimisation for stability, selectivity, safety, delivery and pharmacokinetics.
This experiment measured binding only. It did not assess whether any molecule could function as a drug. Anthropic has said its longer-term goal is to extend Claude across more stages of drug development, including antibodies and small molecules.
Where the Safety Concerns Begin #
The experiment does not demonstrate a biological threat, nor does it show autonomous drug creation. It does show that an AI system can assemble existing biological design tools into a working pipeline that produces experimentally validated molecules.
That shifts the question from whether AI can design proteins to how much of the design process it can automate. In conventional workflows, researchers make repeated decisions across target selection, design strategy and iteration.
In this case, Claude handled much of that computational layer after receiving initial instructions. The physical lab remains a constraint. So do synthesis, testing and downstream development.
The study suggests that parts of the expertise required for protein design can now be embedded in an AI-driven system. It showed that Claude could autonomously orchestrate existing protein-design tools and produce candidates that were later validated in the laboratory.
It did not show that the system can independently discover a drug or complete the wider development process. Questions over the use of increasingly autonomous AI in biological research remain open.
© Copyright IBTimes 2025. All rights reserved.