scripts/export_figure_data.py

"""Export all plotting CSVs for each figure into benchmark/figure_data/."""
from pathlib import Path
import pandas as pd
import numpy as np
import shutil

BENCH = Path(__file__).resolve().parents[1]
FIGDATA = BENCH / "figure_data"
FIGDATA.mkdir(exist_ok=True)

TARGETS = ["PDL1", "MDM2", "IL23R"]


def main():
    # ── fig_ablation_bo_tradeoff ──
    ablation = pd.read_csv(BENCH / "ablation" / "ablation_summary_pose_filtered.csv")
    ablation.to_csv(FIGDATA / "fig_ablation_bo_tradeoff_ablation_summary.csv", index=False)
    tradeoff = pd.read_csv(BENCH / "ablation" / "ablation_tradeoff_pose_filtered.csv")
    tradeoff.to_csv(FIGDATA / "fig_ablation_bo_tradeoff_tradeoff.csv", index=False)

    bad = pd.read_csv(BENCH / "buried_pose_inventory.csv")
    # Ablation Tier-1 reuses the dl D-peptide designs, so the ablation bad-pose
    # counts are the D_peptide buried poses per core target.
    bad_ab = bad[bad["peptide_mode"] == "D_peptide"]
    bad_counts = (
        bad_ab.groupby("target")
        .size()
        .reindex(TARGETS, fill_value=0)
        .rename("n_bad_poses")
        .reset_index()
    )
    bad_counts.to_csv(FIGDATA / "fig_ablation_bo_tradeoff_bad_pose_counts.csv", index=False)
    print(
        f"fig_ablation: {len(ablation)} summary rows + "
        f"{len(tradeoff)} trade-off rows + {len(bad_counts)} pose counts"
    )

    # ── fig_dl_score_distribution ──
    for t in TARGETS:
        for m in ["D_peptide", "L_peptide"]:
            src = BENCH / "dl_downsampled" / f"{t}_{m}_filtered.csv"
            dst = FIGDATA / f"fig_dl_score_dist_{t}_{m}.csv"
            shutil.copy2(src, dst)

    from scipy import stats as sp_stats

    rows = []
    for t in TARGETS:
        d = pd.read_csv(FIGDATA / f"fig_dl_score_dist_{t}_D_peptide.csv")
        l = pd.read_csv(FIGDATA / f"fig_dl_score_dist_{t}_L_peptide.csv")
        ks, p = sp_stats.ks_2samp(d["score"], l["score"])
        rows.append(
            {
                "target": t,
                "n": len(d),
                "D_median": d["score"].median(),
                "L_median": l["score"].median(),
                "D_mean": d["score"].mean(),
                "L_mean": l["score"].mean(),
                "KS_stat": ks,
                "KS_p": p,
            }
        )
    pd.DataFrame(rows).to_csv(FIGDATA / "fig_dl_score_dist_summary.csv", index=False)
    print("fig_dl_score_dist: 6 CSVs + summary")

    # ── fig_dl_sequence_comparison (t-SNE top-100) ──
    from sklearn.decomposition import PCA
    from sklearn.manifold import TSNE

    AAS = list("ACDEFGHIKLMNPQRSTVWY")
    TOP_N = 100
    aa2i = {a: i for i, a in enumerate(AAS)}

    def onehot(seqs, max_len):
        vecs = []
        for s in seqs:
            v = np.zeros(max_len * len(AAS))
            for pos, aa in enumerate(str(s)):
                if aa in aa2i:
                    v[pos * len(AAS) + aa2i[aa]] = 1.0
            vecs.append(v)
        return np.array(vecs)

    for t in TARGETS:
        d_df = pd.read_csv(FIGDATA / f"fig_dl_score_dist_{t}_D_peptide.csv")
        l_df = pd.read_csv(FIGDATA / f"fig_dl_score_dist_{t}_L_peptide.csv")
        d_top = d_df.nsmallest(TOP_N, "score").drop_duplicates("sequence").copy()
        l_top = l_df.nsmallest(TOP_N, "score").drop_duplicates("sequence").copy()

        ml = max(d_top["sequence"].str.len().max(), l_top["sequence"].str.len().max())
        dv = onehot(d_top["sequence"], ml)
        lv = onehot(l_top["sequence"], ml)
        X = np.vstack([dv, lv])
        npc = min(50, X.shape[1], X.shape[0] - 1)
        Xp = PCA(n_components=npc, random_state=42).fit_transform(X)
        Xt = TSNE(
            n_components=2, perplexity=30, random_state=42,
            learning_rate="auto", init="pca",
        ).fit_transform(Xp)

        nd = len(dv)
        d_out = d_top[["sequence", "score", "length", "pose"]].copy()
        d_out["tsne_1"] = Xt[:nd, 0]
        d_out["tsne_2"] = Xt[:nd, 1]
        d_out["type"] = "D"

        l_out = l_top[["sequence", "score", "length", "pose"]].copy()
        l_out["tsne_1"] = Xt[nd:, 0]
        l_out["tsne_2"] = Xt[nd:, 1]
        l_out["type"] = "L"

        combined = pd.concat([d_out, l_out], ignore_index=True)
        combined.to_csv(FIGDATA / f"fig_dl_tsne_{t}.csv", index=False)
        print(f"fig_dl_tsne_{t}: {nd} D + {len(l_out)} L = {len(combined)} rows")

    print(f"\nAll figure data saved to {FIGDATA}/")
    for f in sorted(FIGDATA.iterdir()):
        print(f"  {f.name}  ({f.stat().st_size/1024:.1f} KB)")


if __name__ == "__main__":
    main()