Anthropic/claude-protein-binder-design

TL;DR: Anthropic released the full experimental record of two Claude models acting as autonomous protein-design agents that generated 1,440 de novo miniprotein binders against 16 therapeutic and research targets — with 354 of 1,320 testable designs confirmed as binders by two independent CROs, plus raw kinetics, ten-predictor co-folds, and step-level agent provenance. This is the rare AI-for-science artifact that lets you audit where LLM-driven design actually works instead of trusting a press release.

What this is

Not a model, not a method paper — a data release. Two Claude systems (Mythos Preview, 900 designs; Opus 4.8, 540 designs) ran as agents that designed 50–120-residue binders. Every design was expressed and assayed at two contract research organizations using deliberately different methods:

• Adaptyv Bio — cell-free expression, SPR/BLI with the design immobilized.
• Twist Bioscience — Fc-fusion expression, capture SPR with a six-point antigen titration.

Two orthogonal assays with different expression systems and immobilization geometries is the credibility hook: a "binder" call requires agreement, not a single ambiguous sensorgram. The headline number is 354 binders out of 1,320 designs (~27%) across 15 targets (the 120 mature-GDF-8 designs were dropped — the antigen aggregated and bound surfaces non-specifically).

Targets span real drug-discovery territory: EGFR, PD-L1, TNF-α, VEGF-A, IL-7Rα, TrkA, TREM2, Nipah virus G, SpCas9, and others.

What's actually in the box

It is unusually complete. The core data/ tree is 9.9 GB / 129,003 files:

• 20 Parquet tables (each exposed as an HF viewer subset): design_summary (1,440 rows, one per design), wetlab/summary (two-vendor comparison and final call), Adaptyv replicates/reads/fits, Twist fits/titer/raw segments, epitope contacts, and a column_dictionary defining every column.
• Per-design folders: design model, seed-best co-fold from each of ten predictors, raw reads, fit-curve CSVs, and vendor sensorgram/report PNGs.
• A separate structure tier (74.5 GB): the design model and every seed of every co-fold — 10 predictors × 5 seeds — as mmCIF with the full PAE matrix (pae.npz) per model, 113,550 predictions.
• The prompts (1.16 GB): the multi-target campaign prompt, 17 single-target prompts, kickoff messages, and the external resource corpus the agents were given, with redistribution provenance per source.
• Step-level design provenance — the intermediate agent trajectory, which is what makes failure analysis possible.

Everything joins on uuid; full_name names the folder. Manifests with SHA-256 and a verify_manifest.py script are included. License is CC BY 4.0 for data/docs, MIT for scripts, with third-party terms preserved for structure-prediction outputs and vendor images.

Why it matters and what to watch

The interesting question isn't "can an LLM emit a plausible sequence" — it's whether an agent reasoning over literature, tools, and structure predictors can hit a ~27% two-vendor confirmed binder rate on hard targets. That's in the ballpark of dedicated diffusion/hallucination pipelines, achieved by a general model orchestrating tools. The release is built for skeptics: orthogonal assays, raw traces, negative controls, and the discarded GDF-8 set all shipped honestly.

Caveats the README is upfront about: this is a dataset, not a reusable designer — there's no released design agent or weights, so you can't rerun the pipeline. "Binder" is a two-vendor call with pooled K_D values; read DATA_NOTES.md before quoting affinities. Success is uneven across targets, and inspecting per-target rates (some targets clearly carry the 354) is the first analysis anyone should do. The nine confirmed binders arranged to spell ANTHROPIC is a fun flourish but also a reminder to weight the aggregate stats over cherry-picked structures.

Shortest path to trying it

from huggingface_hub import hf_hub_download
import pandas as pd

repo = "Anthropic/claude-protein-binder-design"
designs = pd.read_parquet(hf_hub_download(repo, "data/tables/design_summary.parquet", repo_type="dataset"))
wetlab  = pd.read_parquet(hf_hub_download(repo, "data/tables/wetlab/summary.parquet", repo_type="dataset"))

# per-target confirmed-binder rate
print(designs.groupby("target")["binder_final"].mean().sort_values())

For a quick look without the 10 GB of per-design files, grab the 33 MB protein_binder_design_data_release_docs_tables.zip (tables + docs + manifests only), or use hf download ... --include "data/tables/*".