Anthropic Claude AI Designs Protein Binders for 14 of 15 Targets in Lab-Validated Breakthrough

Claude AI Achieves Breakthrough in Protein Design

Anthropic Claude designed functional protein binders that succeeded against 14 of 15 targets in wet-lab testing, marking a significant advance in AI-driven scientific discovery1

1

. The campaign produced 354 confirmed binders from 1,320 designs, with Anthropic Claude achieving hit rates between 22.6% and 35.1% depending on the approach used2

2

. These results exceed the typical 10-15% success rate seen in protein design campaigns, demonstrating how AI in scientific discovery could accelerate drug development timelines that traditionally require months of work1

1

.

Two independent firms, Twist Bioscience and Adaptyv Bio, validated the AI-generated proteins through physical testing3

3

. This external verification addresses concerns about self-reported results and establishes credibility for Claude's capabilities in synthetic biology applications. The validation process still takes weeks regardless of software advances, highlighting the continued importance of lab testing infrastructure1

1

.

How Claude Opus 4.8 and Mythos Preview Orchestrated Protein Design

Anthropic deployed two models for the protein design work: Claude Opus 4.8 and a preview of its Mythos model, running them inside the Claude Science research workbench1

1

. The models received extensive GPU compute resources, with up to 12,500 Nvidia H100 hours allocated over 48-hour sessions in one configuration, and up to 2,500 H100 hours per target in another mode1

1

. This substantial computational investment enabled the AI designed functional protein binders to work autonomously, selecting binding sites, orchestrating existing design and folding models, and screening candidates without human scientific guidance during execution2

2

.

The Mythos Preview achieved a 35.1% hit rate when designing against individual targets in separate 24-hour sessions, compared with 26.7% when tackling all targets simultaneously2

2

. This finding suggests that focused, single-target approaches yield better results than parallel processing across multiple targets. The models began with an initial prompt of approximately 30,000 tokens written by a human expert, then leveraged publicly available specialist protein design models to complete the work2

2

.

Outstanding Performance Against RBX1 and Variable Results Across Targets

Against the protein target RBX1, Mythos Preview achieved a 40% success rate, dramatically outperforming the 3.7% rate achieved by human participants in an Adaptyv Bio competition1

1

3

3

. The top design from Claude beat the winning entry from that competition, demonstrating superior affinity levels in certain contexts1

1

. For four targets overall, the strongest designed proteins matched or exceeded previously published affinity levels, with some binding several times more tightly than the best prior results4

4

.

However, protein binders performance varied significantly across different targets. Claude Opus 4.8 successfully designed minibinders against TNFα—the protein targeted by anti-inflammatory drugs such as Humira—that worked across human, monkey, and mouse versions1

1

. Curiously, the more advanced Mythos Preview failed on this same target, and Anthropic acknowledged uncertainty about why the less capable model succeeded where the stronger one did not1

1

. Against maltose-binding protein, a notoriously smooth target, none of 90 designs were confirmed to bind1

1

3

3

.

Autonomous Chemical Analysis Completed in Minutes

Beyond protein design, Anthropic tested Claude Opus 5 on autonomous chemical analysis, a routine but time-consuming task in drug discovery1

1

. Given only raw files from NMR and LC-MS instruments and a two-sentence prompt, the model processed results in 23 and 19 minutes respectively, running both analyses in parallel1

1

3

3

. Its measurements closely matched the laboratory's results, calculating purity at 96.4% versus the lab's 96.33%, with hydrogen counts within rounding margins1

1

.

This represents a dramatic acceleration compared to traditional workflows, where chemists typically spend 30 minutes to an hour per sample, and complete lab reports can take four days1

1

. Claude demonstrated error-correction capabilities, catching and fixing its own mistake after initially overstating peak shifts in a follow-up reading1

1

. The model also independently proposed the same confirmatory test the lab had run three days earlier, and worked out the encoding of an undocumented vendor format for the LC-MS file before analyzing the data1

1

.

Market Response and Dual-Use Risks Shape Future Access

Twist Bioscience shares jumped approximately 17% following the announcement, reaching their highest level since 2021, with the stock more than quadrupling since the start of the year4

4

. The market reaction reflects expectations that pharmaceutical labs may increase demand for synthetic biology infrastructure as they adopt AI to generate new drug candidates4

4

. This complementarity between AI models and biotechnology companies positions firms like Twist Bioscience and Adaptyv Bio as critical validation partners in the AI-driven drug discovery ecosystem4

4

.

Yet Anthropic acknowledged that protein design and other life-science research tasks remain restricted in its most capable models while the company develops safety measures for dual-use risks2

2

. The company emphasized that protein binders are not drugs, and designing high-affinity binders represents just the first step in developing drug-like molecules2

2

. Critics, including Martin Shkreli, noted limitations such as relatively low affinities for peptidics and the absence of intracellular protein targets among Claude's successful designs2

2

. Anthropic plans to launch an access program for scientists to use its most capable models and aims to have Claude eventually run the full drug-development process across all drug modalities2

2

. CEO Dario Amodei has previously stated that AI could help cure most human diseases within five to 10 years, arguing the industry must deliver real breakthroughs in medicine and biology rather than rely on optimistic messaging3

3

.