Anthropic has released experimental results demonstrating autonomous de novo protein binder design using its frontier Claude models, backed by physical wet-lab validation from two independent contract research organizations. In empirical testing against 15 target proteins, Claude-designed mini-binders successfully bound to 14 targets, delivering an overall hit rate of 26.8% and a 49% binding rate for its top-ranked candidates.
The campaign evaluated Claude Opus 4.8 and a preview build of Claude Mythos operating inside Claude Science, Anthropic's agentic research environment. Rather than training a specialized biological architecture from scratch, Anthropic configured general-purpose reasoning models to coordinate existing open-source structure generation, sequence design, and co-folding algorithms without human intervention during execution.
Campaign Setup and Tool Orchestration
Traditional computational protein design requires specialized bioinformaticians to manually configure docking interfaces, select scoring functions, and iterate through sequence-structure generation pipelines over weeks or months. In Anthropic's evaluation, Claude operated autonomously after receiving an initial 30,000-token prompt detailing design constraints, target specifications, and operational protocols.
The models were provided with:
- Compute allocations of up to 12,500 NVIDIA H100 GPU hours for multi-target campaigns and 2,500 H100 hours per target in single-target mode.
- Internet connectivity and scientific literature access via BioRxiv.
- Integration connectors for Google Drive, Gmail, and Slack.
- Direct execution access to publicly available open-source protein design and folding packages. Proprietary packages or models with restrictive licenses, including AlphaFold 3 weights, were excluded from the execution environment.
Claude orchestrated multi-step optimization loops: identifying binding epitopes on target proteins, generating backbone geometries, designing amino acid sequences, evaluating complex structures through co-folding models, and applying in silico solubility and expression filters. For each target, Claude generated and ranked candidate binders, delivering 30 ordered designs per target.
Wet-Lab Validation and Affinity Metrics
To evaluate functional efficacy, Anthropic partnered with Adaptyv Bio and Twist Bioscience. The laboratories synthesized all 1,320 candidate sequences and measured binding kinetics using high-throughput biophysical assays.
The empirical results showed:
- Broad Target Coverage: Claude generated confirmed binders for 14 out of 15 interpretable targets. The evaluated target suite included standard benchmark proteins from Adaptyv Bio's BenchBB collection as well as novel targets such as 15-PGDH and latent GDF-8.
- Hit Rate Superiority: In multi-target mode (48-hour continuous runtime), Mythos Preview and Opus 4.8 achieved overall hit rates of 26.7% and 22.6%, respectively. When executing in single-target mode (24 hours per target), Mythos Preview achieved a 35.1% hit rate. Industry baselines for unguided de novo computational campaigns typically range between 10% and 15%.
- Ranking Accuracy: The top-ranked candidate selected by Claude for each target achieved a 49% experimental hit rate, demonstrating effective self-screening and confidence calibration.
- Nanomolar Affinity: Out of 354 confirmed binders, 194 designs demonstrated dissociation constants (KD) below 100 nM, 90 designs measured below 10 nM, and 42 designs exhibited sub-nanomolar binding affinity (<1 nM). For multiple targets, including RBX1 and TNF-alpha, Claude's designs matched or exceeded the strongest published human-engineered binders.
Analytical Chemistry Benchmarks
Anthropic simultaneously evaluated Claude Opus 5 on automated analytical chemistry data interpretation. Given raw, proprietary instrument files from nuclear magnetic resonance (NMR) and liquid chromatography-mass spectrometry (LC-MS) assays alongside a two-sentence prompt, the model extracted spectra, picked peaks, and calculated purity metrics.
Opus 5 completed the analytical pipelines in 23 minutes for NMR and 19 minutes for LC-MS without specialized software plugins, matching contract laboratory ground truth on proton integration and sample purity (96.4% AI-calculated purity vs. 96.33% lab reference).
Anthropic has published the prompt templates, computational models, and experimental binding datasets on Hugging Face to enable external verification across the structural biology community.
