claude-protein-binder-design

# Claude protein binder design — data release v1.0 1,440 de novo miniprotein binders (50 to 120 residues) against 16 targets, designed by two Claude models operating as autonomous protein-design agents (**Mythos Preview**, 900 designs; **Opus 4.8**, 540 designs) and characterized at two contract research organizations, **Adaptyv Bio** (cell-free expression; SPR/BLI kinetics with the design immobilized) and **Twist Bioscience** (Fc-fusion expression; capture SPR with a six-point antigen titration). The release links, per design: binding calls and kinetics from both vendors with raw sensorgrams and report images, a per-design comparison and final assessment of the two, the design model, seed-best co-folded predictions from ten structure predictors with per-seed scores, and step-level design provenance. <video src="https://huggingface.co/datasets/Anthropic/claude-protein-binder-design/resolve/main/assets/claude_designed_binders_ANTHROPIC_1080p.mp4" poster="https://huggingface.co/datasets/Anthropic/claude-protein-binder-design/resolve/main/assets/claude_designed_binders_ANTHROPIC_poster.jpg" autoplay loop muted playsinline controls width="100%"></video> _{Nine experimentally confirmed de novo binders from this release, shown first as design models bound to their targets (grey) and then on their own, arranged to spell ANTHROPIC. Each letter is a real structure from `data/designs/…/insilico/` drawn as a cartoon, rotated and uniformly scaled only (no deformation); the BHRF1 “C” has no design-time complex in the release and is shown as its co-folded prediction. A = `mythos_preview_single_target_nipahg_rank15`, N = `opus_4_8_multi_target_trem2_rank11`, T = `mythos_preview_single_target_trka_rank06`, H = `mythos_preview_single_target_il7ra_rank20`, R = `mythos_preview_single_target_il7ra_rank19`, O = `mythos_preview_single_target_egfr_rank24`, P = `mythos_preview_multi_target_vegfa_rank26`, I = `mythos_preview_single_target_trem2_rank02`, C = `opus_4_8_multi_target_bhrf1_rank01`. K_D values in the captions are the pooled values in `design_summary`.} > **Dataset viewer:** the default subset is `design_summary` (one row per design, 1,440 rows). Every Parquet table in `data/tables/` is exposed as its own subset (e.g. `wetlab_summary`, `adaptyv_replicates`, `twist_fits`, `insilico_cofold_predictions`); pick it from the *Subset* menu or load it with `load_dataset("Anthropic/claude-protein-binder-design", "<subset>", split="full")`. Per-design folders under `data/designs/` (structures, sensorgram PNGs, per-curve CSVs) are files to browse/download, not viewer tables. Targets: 15-PGDH, BBF-14, BHRF1, SpCas9, EGFR, IL-7Rα, latent GDF-8 (pro-myostatin), mature GDF-8, MBP, Nipah virus G, PD-L1, RBX1, TNF-α, TREM2, TrkA, VEGF-A. Wet-lab measurements of the 120 mature GDF-8 designs were inconclusive (the antigen aggregated and bound assay surfaces non-specifically) and are not included; those designs ship with design models, co-folds and provenance only. Across the other 15 targets, 354 of 1,320 designs are binders by the release's two-vendor assessment. ## Repository layout ``` README.md this card data/ the data release, unpacked (129,003 files, 9.9 GB) README.md start here: layout, the two assays in brief, headline numbers docs/ DATA_NOTES.md, WETLAB.md, INSILICO.md, PROVENANCE.md, LOOKUP_TABLES.md, scripts/ tables/ 20 Parquet tables (design summary, wet-lab summary and measurements, Adaptyv Bio results/replicates/reads/fits, Twist Bioscience fits/titer/raw segments, antigens and controls, co-fold predictions, epitope contacts, provenance, target constructs, column dictionary) manifests/ file-level manifests for structures, sensorgram images and raw traces designs/<target>/<design>/ per design: design model, seed-best co-fold per predictor, Adaptyv Bio raw reads, fit curves and sensorgram images, Twist Bioscience raw traces and report images controls/ assay controls and reference binders measured alongside the designs MANIFEST.sha256, LICENSE.md, CITATION.cff protein_binder_design_data_release_docs_tables.zip README, docs/, tables/ and manifests/ only (33 MB), for a quick start without the per-design files structure_and_pae/ companion structure tier README.md layout and column definitions of the structure tier predictions.parquet one row per predicted model (113,550): predictor, seed, stoichiometry, scores, paths design_names.csv, design_models.csv, MANIFEST.sha256, LICENSE.md, CITATION.cff protein_binder_design_structure_and_pae_release.zip 74.5 GB: for every design, the design model and every seed of every co-fold (ten predictors, five seeds each) as mmCIF with the full PAE matrix (pae.npz) prompts/ the campaign prompts README.md how the prompt, kickoff message and resource corpus fit together prompts/ multi-target campaign prompt, 17 single-target prompts, kickoff messages, figures REDISTRIBUTION_NOTES.csv provenance and licence class of every corpus source protein_binder_design_prompts_release.zip 1.16 GB: the prompts plus the external resource corpus they reference (cited web pages, ProteinBase collections, method and model papers) ``` The complete data release is also what `data/` was unpacked from; its archive form (`protein_binder_design_data_release.zip`, 5.65 GB, 129,003 files with `MANIFEST.sha256`) can be rebuilt by zipping `data/`, and `python data/docs/scripts/verify_manifest.py` re-hashes every file against the manifest. SHA-256 of the archives in this repository: ``` bb7d45f9957bee70e715191566577d59e69363e42f18981aa96468b4b6547dec protein_binder_design_data_release_docs_tables.zip 2b39a3c07cda2823eedc66886aeeeb1d79c7a6788e2b8e979827262ed40ce359 structure_and_pae/protein_binder_design_structure_and_pae_release.zip ce86bae3d22922f3b1a06235688be79ccc209fcb1a61066d32471d256eb017c5 prompts/protein_binder_design_prompts_release.zip ``` ## Getting started ```python 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")) # one row per design (1,440) wetlab = pd.read_parquet(hf_hub_download(repo, "data/tables/wetlab/summary.parquet", repo_type="dataset")) # two-vendor comparison and final call print(designs[["full_name", "target", "design_model", "binder_final", "kd_nM_final"]].head()) ``` ```bash # tables and documentation only hf download Anthropic/claude-protein-binder-design --repo-type dataset --include "data/tables/*" "data/docs/*" "data/README.md" # one design's folder hf download Anthropic/claude-protein-binder-design --repo-type dataset --include "data/designs/EGFR/mythos_preview_single_target_egfr_rank02/*" ``` Every table joins on `uuid`; `full_name` (`<model>_<campaign>_<target>_rank<NN>`) names the per-design folder. `data/tables/column_dictionary.parquet` defines every column of every table; `data/docs/DATA_NOTES.md` lists the caveats that matter when interpreting the measurements. ## License and citation Data and documentation: [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). Scripts included in the archives: MIT. Third-party material (structure-prediction outputs, reference sequences, vendor report images, reagent names) keeps its own terms; see `LICENSE.md` in each archive. Please cite as given in `CITATION.cff`.