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.Copy CodeCopiedUse a different Browserimport subprocess, sys, warnings, itertools, math warnings.filterwarnings("ignore") import importlib.

util needed = {"huggingfacehub": "huggingfacehub>=0.24", "pyarrow": "pyarrow", "pandas": "pandas", "sklearn": "scikit-learn", "matplotlib": "matplotlib", "scipy": "scipy"} missing = [pkg for mod, pkg in needed.items() if importlib.util.findspec(mod) is None] if missing: print("installing:", ", ".

join(missing)) subprocess.run([sys.executable, "-m", "pip", "install", "-q", missing], check=False) import numpy as np import pandas as pd import matplotlib.pyplot as plt from scipy import stats from huggingfacehub import HfApi, hfhubdownload from sklearn.

metrics import rocaucscore, cohenkappascore, averageprecisionscore from sklearn.modelselection import GroupKFold, StratifiedKFold from sklearn.ensemble import HistGradientBoostingClassifier from sklearn.inspection import permutationimportance SEED = 0 rngglobal = np.random.defaultrng(SEED) pd.

setoption("display.width", 200) pd.setoption("display.maxcolumns", 100) plt.rcParams.update({"figure.dpi": 110, "font.size": 9, "axes.grid": True, "grid.alpha": 0.25, "axes.spines.top": False, "axes.spines.

right": False}) REPO = "Anthropic/claude-protein-binder-design" BAR = "=" 78 def head(n, title): prefix = f"{n}." if str(n) else "" print(f"\n{BAR}\n {prefix}{title}\n{BAR}") head(1, "TABLE DISCOVERY") api = HfApi() repofiles = api.

listrepofiles(REPO, repotype="dataset") TABLES = {} for f in repofiles: if f.startswith("data/tables/") and f.endswith(".parquet"): key = f[len("data/tables/"): -len(".parquet")].

replace("/", "") TABLES[key] = f print(f"Found {len(TABLES)} Parquet tables:") for k in sorted(TABLES): print(f" - {k:38s} {TABLES[k]}") def loadtable(name: str) -> pd.DataFrame: """Load a subset by its viewer name, with a datasets-library fallback.""" if name in TABLES: return pd.

readparquet(hfhubdownload(REPO, TABLES[name], repotype="dataset")) from datasets import loaddataset return loaddataset(REPO, name, split="full").topandas() ds = loadtable("designsummary") print(f"\ndesignsummary: {ds.shape[0]:,} rows x {ds.

shape[1]} columns") 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 wetlabsummary lives at data/tables/wetlab/summary.parquet, and a guessed path would fail silently.With the map in place we pull designsummary, one row per design, 1,440 rows wide enough to carry every join we need downstream.

Copy CodeCopiedUse a different Browserhead(2, "SCHEMA + EVALUABLE SET") CALLS = {"binder", "nonbinder"} tested = ds["adaptyvbinding"].isin(CALLS) | ds["twistbinding"].isin(CALLS) ev = ds[tested].copy() ev["y"] = ev["binderfinal"].

astype(int) print(f"All designs : {len(ds):,}") print(f"Evaluable (>=1 vendor call): {len(ev):,}") print(f"Confirmed binders : {int(ev['y'].sum()):,} " f"({100 ev['y'].mean():.

1f}% base rate)") print(f"Never measured : {len(ds) - len(ev):,}") print("\nCategorical levels:") for c in ["designmodel", "campaign", "generator", "sequencedesignmethod", "vendoragreement"]: vals = ds[c].astype(str).valuecounts() print(f" {c:24s} ({len(vals)}): {', '.join(vals.index[:6])}" + (" ...

" if len(vals) > 6 else "")) print(f"\nTargets ({ds['target'].nunique()}): {', '.join(sorted(ds['target'].unique()))}") print(f"Binder length: {ds.binderlength.min()}-{ds.binderlength.max()} aa " f"(median {ds.binderlength.median():.0f})") head(3, "HIT-RATE LANDSCAPE") def wilson(k, n, z=1.

96): if n == 0: return (np.nan, np.nan, np.nan) p = k / n d = 1 + z2 / n c = (p + z2 / (2 n)) / d h = z math.sqrt(p (1 - p) / n + z2 / (4 * n2)) / d return p, max(0.0, c - h), min(1.0, c + h) def ratetable(df, by): rows = [] for key, g in df.groupby(by, dropna=False): p, lo, hi = wilson(int(g.y.

sum()), len(g)) rows.append({by: key, "n": len(g), "hits": int(g.y.sum()), "rate": p, "lo": lo, "hi": hi}) return pd.DataFrame(rows).sortvalues("rate", ascending=False).

resetindex(drop=True) for dim in ["designmodel", "campaign", "generator", "sequencedesignmethod"]: t = ratetable(ev, dim) print(f"\n--- hit rate by {dim} ---") print(t.tostring(index=False, formatters={"rate": "{:.3f}".format, "lo": "{:.3f}".format, "hi": "{:.3f}".

format})) tt = ratetable(ev, "target") fig, ax = plt.subplots(figsize=(9, 4.2)) ax.bar(tt.target, tt.rate, color="#4C72B0") ax.errorbar(tt.target, tt.rate, yerr=[(tt.rate - tt.lo).clip(lower=0), (tt.hi - tt.rate).clip(lower=0)], fmt="none", ecolor="0.25", capsize=3, lw=1) ax.axhline(ev.y.

mean(), ls="--", c="crimson", lw=1, label=f"pooled {ev.y.mean():.2f}") ax.setylabel("experimental hit rate"); ax.settitle("Hit rate by target (Wilson 95% CI)") ax.tickparams(axis="x", rotation=55); ax.legend(); plt.tightlayout(); plt.

show() print("\nRead this plot as the dominant effect size in the dataset: target choice " "swamps generator choice.Any model comparison that does not stratify by " "target is mostly measuring which targets that model was pointed at.

") We define the evaluable set by filtering on actual vendor calls rather than on binderfinal, 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.Copy CodeCopiedUse a different Browserhead(4, "PER-PREDICTOR DISCRIMINATIVE POWER") PREDICTORS = sorted({c[len("ipsaemin"):] for c in ds.columns if c.

startswith("ipsaemin")}) print(f"Predictors ({len(PREDICTORS)}): {', '.join(PREDICTORS)}") def aucci(y, s, nboot=300, seed=SEED): s = np.asarray(s, dtype=float); y = np.asarray(y, dtype=int) m = ~np.isnan(s) y, s = y[m], s[m] if len(y) < 30 or len(np.unique(y)) < 2: return dict(auc=np.nan, lo=np.

nan, hi=np.nan, n=len(y), ap=np.nan) base = rocaucscore(y, s) ap = averageprecisionscore(y, s) rng = np.random.defaultrng(seed) idx, boots = np.arange(len(y)), [] for in range(nboot): b = rng.choice(idx, len(idx), replace=True) if len(np.unique(y[b])) > 1: boots.

append(rocaucscore(y[b], s[b])) lo, hi = (np.percentile(boots, [2.5, 97.5]) if boots else (np.nan, np.nan)) return dict(auc=base, lo=lo, hi=hi, n=len(y), ap=ap) rows = [] for p in PREDICTORS: for metric in ["ipsaemin", "scdockq"]: col = f"{metric}{p}" if col in ev.columns: r = aucci(ev.

y, ev[col]) rows.append({"predictor": p, "metric": metric, **r}) perf = pd.DataFrame(rows) piv = perf.pivot(index="predictor", columns="metric", values="auc").sortvalues("ipsaemin", ascending=False) print("\nAUC vs experimental binderfinal:") print(perf.sortvalues("auc", ascending=False).

tostring( index=False, formatters={c: "{:.3f}".format for c in ["auc", "lo", "hi", "ap"]})) fig, ax = plt.subplots(figsize=(9, 4.2)) x = np.arange(len(piv)); w = 0.38 for i, (metric, colr) in enumerate([("ipsaemin", "#4C72B0"), ("scdockq", "#DD8452")]): sub = perf[perf.metric == metric].

setindex("predictor").reindex(piv.index) loe