Anthropic’s claude-protein-binder-design dataset dropped with 1,440 AI-designed miniproteins tested against 16 targets, and the setup is unusually honest for this space: you get both the computational predictions and actual wet-lab results from two independent labs. That dual-source design lets researchers ask the uncomfortable question most papers skip — do the structure predictors actually identify which designs will bind?
The tutorial, published as a notebook walkthrough, starts by enumerating the HuggingFace repository and building a path map rather than hard-coding file locations. That’s a practical detail that matters because the naming isn’t uniform; the wetlab_summary subset lives at data/tables/wetlab/summary.parquet, and a guessed path would fail silently. The design_summary table carries 1,440 rows, one per design, with enough columns to join everything downstream.
What the analysis reveals is sobering. Only a fraction of designs get a definitive binder or non-binder call from at least one vendor, and the base rate for confirmed binders is modest — the tutorial computes Wilson confidence intervals to quantify hit rates across design models, campaigns, generators, and sequence design methods rather than cherry-picking the best-looking number. The binder lengths span a wide range, and the targets include 16 distinct proteins, giving enough diversity to test whether any single predictor generalizes.
The deeper question the tutorial chases is whether combining multiple structure predictors improves discrimination, and whether a target-aware classifier trained on these signals can reliably predict experimental success. That last point is the one to watch: if a gradient-boosted model can’t beat the base rate using every computational feature available, it’s a strong signal that current in-silico binding predictions are nowhere near replacing wet-lab screening — they’re just a triage filter with a high false-positive rate. The two-lab assay disagreement also gets quantified, which tells you how much noise comes from the experiments themselves rather than the models. For protein design teams allocating limited testing budgets, that’s the number that actually determines how many designs you can afford to test before a hit.
