GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
← Back to feed

Claude Runs a Full Protein Engineering Campaign: 1,440 Binders, 10 Targets, All Models Open-Source

Anthropic has published a paper documenting what may be the most complete autonomous protein engineering campaign run by an AI to date. Claude managed the full pipeline across 10 protein targets: reviewing literature, specifying design objectives, ranking computational outputs from open-source structure prediction models, and submitting candidates to two contract research organisations (CROs) for binding measurements.

The result: 1,440 de novo binder designs, with reliable binding data collected for 1,320 of them.

What Claude actually did

The campaign was end-to-end autonomous. Claude did not hand off to human scientists between design cycles. It used open-source models throughout, meaning any laboratory with API access can run the same protocol. The prompts, design decisions, per-design provenance, and CRO binding data are all released.

Of the designs subsequently tested against the mouse ortholog of their respective targets, 233 also bound — demonstrating cross-species generalization without that being an explicit design objective.

Why this is different from prior protein AI work

Most protein AI coverage focuses on structure prediction (AlphaFold, ESMFold) or single-target wet-lab validations. This campaign covers 10 targets simultaneously, runs the ranking and triage logic in Claude rather than in a purpose-built ML system, and uses only open-source computational models. Every design decision is traceable in the prompts.

The benchmark dataset — 1,320 designs with ground-truth binding measurements and computational predictions — gives the field a new resource for calibrating filtering and ranking methods.

Key numbers

  • 10 protein targets
  • 1,440 de novo binder designs
  • 1,320 designs with reliable binding measurements (released)
  • 233 binders also bound the mouse ortholog when tested
  • 2 CROs used for independent binding validation
  • All computational models: open-source

What it implies for agentic AI

The paper is as much about agentic workflow design as protein biology. Claude was given the full task specification and operated as a director, not a search engine. The prompts — which Anthropic releases — show how a general-purpose model can manage a multi-stage scientific workflow with real physical outputs.

The release as an open benchmark is deliberate: Anthropic is positioning this as infrastructure for the field, not a proprietary capability. Any lab can fork the protocol.