From In-Silico to Wet-Lab: Evaluating AI Protein Design Performance

In this tutorial, we use Anthropic’s claude-protein-binder-design dataset, which contains 1,440 AI-designed miniprotein binders tested against 16 targets. Because the release includes both computational predictions and real wet-lab results from two independent labs, we can go beyond simply studying the designs. We evaluate how well structure predictors identify successful binders, whether combining predictions improves performance, how rankings translate into practical testing budgets, and how much disagreement comes from the assays themselves. Also, we train a target-aware classifier to test whether these signals can reliably predict experimental success.

We start by installing only what the runtime is actually missing, then enumerate the repository once and build a {subset to path} map instead of hard-coding file locations. This matters because the naming is not uniform; the subset wetlab_summary lives at data/tables/wetlab/summary.parquet, and a guessed path would fail silently. With the map in place we pull design_summary, one row per design, 1,440 rows wide enough to carry every join we need downstream.

We define the evaluable set by filtering on actual vendor calls rather than on binder_final, because that column is a bool and so records the 120 never-measured designs as False rather than missing. From there we compute hit rates by model, campaign, generator, and target, wrapping each in a Wilson interval since several subgroups sit in the small-n regime where the normal approximation misbehaves. The target plot is the one to read first: it shows antigen choice swamping every other factor we compare.

We score all ten predictors against the wet-lab label, on both ipSAE and self-consistency DockQ, with bootstrapped confidence intervals so we can see which differences are real. We then rank-normalize each column to percentiles and aggregate them, which keeps the comparison scale-free across metrics that live on different ranges and pile up differently at zero. The Spearman heatmap explains why the ensemble helps at all; if the predictors agreed perfectly there would be nothing left to harvest.

We convert ranking performance into precision@N, because no lab orders 1,300 constructs and AUC quietly hides how a score behaves at the top of the list. The enrichment table then tells us what triage actually buys at budgets of 25, 50, 100, and 200. We follow it with Cohen’s κ and a log-log KD comparison between vendors, which sets the ceiling: label noise bounds how high any AUC above can honestly climb.

We test whether expression titer alone discriminates binders, and if it does, we know part of every score from above is solubility riding along under another name. Re-running consensus on expressed-only designs isolates interface quality from biophysics. We then parse the epitope contact lists into residue sets and ask, per target and paired across targets, whether confirmed binders converge on a shared patch more than the failures do.