AI Agent Designs 354 Protein Binders in 48 Hours, Doubling Industry Success Rates

Anthropic has published results from an August 2024 protein binder design campaign in which its AI models autonomously generated 1,320 protein designs across 15 targets, yielding 354 confirmed binders—a hit rate of 26.7% to 35.1% that the company says doubles or triples the current 10-15% industry standard. The campaign, validated through surface plasmon resonance assays by Adaptyv Bio and Twist Bioscience, represents a strategic pivot for the AI company: rather than developing proprietary protein modeling algorithms, Anthropic demonstrated that frontier language models can orchestrate existing computational biology tools to achieve expert-level results.

Autonomous Orchestration at Scale

Two Anthropic models—Opus 4.8 and the unreleased Mythos Preview—were given a 30,000-token design prompt, internet access, a library of protein design papers, GPU compute resources, and the ability to spawn sub-agents. Human involvement was limited to approving network requests, maintaining infrastructure, and ordering DNA sequences. The models autonomously selected and orchestrated publicly available structure design, sequence design, and co-folding tools across two configurations: a multi-target 48-hour run consuming up to 12,500 H100 GPU hours, and single-target 24-hour parallel sessions using up to 2,500 H100 hours each.

The campaign targeted 16 proteins initially, with 15 reported after mature GDF-8 was excluded due to aggregation issues. Seven targets have oncology relevance: PD-L1, EGFR, VEGF-A, TrkA, IL-7Rα, RBX1, and 15-PGDH. Mythos Preview achieved a 26.7% hit rate in multi-target mode and 35.1% in single-target mode, while Opus 4.8 reached 22.6% in multi-target configuration. Per-target success varied dramatically, from approximately 90% down to zero for maltose-binding protein, which yielded no confirmed binders across 90 designs.

Validation and Limitations

Adaptyv Bio tested 1,296 designs at five concentrations in duplicate, classifying 336 as binders, while Twist Bioscience tested 1,260 designs. The two laboratories agreed on 89% of the 1,235 overlapping designs, with a Cohen's kappa of 0.71 indicating substantial concordance. After reconciliation, the final binder count reached 354, with 95% of all designs successfully expressed. High-affinity binders with dissociation constants below 10 nanomolar were obtained against at least six targets. For 15-PGDH, the best affinity improved from 1.7 micromolar to 33.4 nanomolar; for RBX1, from 25.7 nanomolar to 3.9 nanomolar.

However, Anthropic emphasizes these are "first-pass" results. The campaign measured only binding affinity to purified antigens via surface plasmon resonance—no data exists on cell-surface engagement, receptor blockade, epitope competition with approved biologics, immunogenicity, or in vivo behavior. Two assay plates carried caveats: the SpCas9 plate at Twist lacked a positive control in capture format, and the RBX1 plate used a CUL1–RBX1 complex in which CUL1 occludes part of the RBX1 surface. The campaign was also open-loop, with no iterative refinement based on experimental feedback.

Industry Reactions and Access Constraints

Adaptyv Bio, a paid participant, stated the models demonstrated "expert-level skill" and would have won five of six Adaptyv-run protein design competitions. XenoSpectrum noted the significance lies in architectural orchestration—a single agent coordinating existing tools—rather than molecular innovation. Martin Shkreli, a former pharmaceutical executive, dismissed the work as "not impressive," arguing the affinities are "quite low for peptidics" and that extracellular targets are better addressed with monoclonal antibodies.

Anthropic has released prompts, sequences, design files, and raw data on Hugging Face, with Proteinbase deposition announced. However, protein design capabilities remain blocked for general access in the company's Claude 3.5 model due to dual-use concerns. A scientist access program is planned but not yet launched, limiting independent replication to re-analysis of released data rather than experimental validation. The thousands of H100 hours per target also raise questions about cost-effectiveness for resource-constrained organizations, even as the approach could compress early discovery timelines if validated beyond binding affinity measurements.

Sources

1. Anthropic's AI Protein Design Run: 354 Binders, 15 Targets — OncoDaily