Why this matters
This release uniquely pairs AI-generated de novo binder designs with orthogonal experimental validation pipelines, letting researchers benchmark how often modern generative agents produce real, measurable binders. It includes 1,440 designs (50–120 residues) across 16 targets, with per-design raw traces, fitted kinetics, multiple co-fold predictions, and step-level provenance; 354 designs are called binders by a two-vendor consensus.
What Sets It Apart
- Linked design-to-experiment records: each design folder contains the model output, vendor raw traces, fitted kinetics and sensorgram images — so what? you can trace an individual design from generation through experimental readout without reconstruction.
- Two independent wet‑lab pipelines: Adaptyv Bio (cell-free expression; SPR/BLI) and Twist Bioscience (Fc‑fusion capture SPR) — so what? cross-vendor agreement enables more robust binder calls and assessment of assay-specific effects.
- Rich in-silico validation tier: seed-best co-folds from ten structure predictors (multiple seeds) and a large structure/PAE tier — so what? enables method comparisons and benchmarking of structure predictors on AI designs.
- Complete provenance and prompts: campaign prompts, kickoff messages and per-step provenance are included — so what? supports reproducibility and analysis of agent behavior that produced the designs.
Who it's for and tradeoffs
Great fit if you are building or evaluating generative protein-design models, benchmarking structure-prediction or affinity-prediction methods, or studying the gap between in-silico scores and experimental success. The dataset is large (many per-design files and a separate 74.5 GB structure tier), uses Parquet tables for tabular access, and contains third‑party predicted structures and vendor report images that retain their own terms. Look elsewhere if you need simple example code or a tiny curated benchmark: this release is a comprehensive experimental data package rather than a lightweight tutorial.
Additional practical notes
- License: data and documentation under CC BY 4.0; some third-party artifacts keep their own terms.
- Known caveat: the 120 designs for mature GDF‑8 had inconclusive wet‑lab results due to antigen aggregation and are included only with models and co-folds, not wet‑lab measurements.
