scripts/plot_ablation_dl_figures.py

"""Generate publication figures for ablation and D/L comparison results."""
from pathlib import Path
import re

import matplotlib.pyplot as plt
import numpy as np
import pandas as pd


ROOT = Path(__file__).resolve().parents[2]
BENCH = ROOT / "benchmark"
FIGDIR = BENCH / "figures"

PALETTE = {
    "blue_main": "#0F4D92",
    "blue_secondary": "#3775BA",
    "blue_light": "#7BAED4",
    "orange": "#E28E2C",
    "green": "#5BAA5B",
    "red_strong": "#B64342",
    "neutral_light": "#CFCECE",
    "neutral_mid": "#767676",
    "neutral_dark": "#4D4D4D",
}


MODE_LABELS = {
    "chroma+proteinmpnn": "Chroma+ProteinMPNN",
    "chroma+bo": "Chroma+BO",
    "chroma+proteinmpnn+bo": "Chroma+ProteinMPNN+BO",
}

MODE_COLORS = {
    "chroma+proteinmpnn": PALETTE["red_strong"],
    "chroma+bo": PALETTE["orange"],
    "chroma+proteinmpnn+bo": PALETTE["blue_main"],
}

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


def apply_nature_style(font_size=8, axes_linewidth=0.8):
    plt.rcParams["font.family"] = "sans-serif"
    plt.rcParams["font.sans-serif"] = ["Arial", "DejaVu Sans", "Liberation Sans"]
    plt.rcParams["svg.fonttype"] = "none"
    plt.rcParams["pdf.fonttype"] = 42
    plt.rcParams["axes.spines.right"] = False
    plt.rcParams["axes.spines.top"] = False
    plt.rcParams["axes.linewidth"] = axes_linewidth
    plt.rcParams["legend.frameon"] = False
    plt.rcParams["font.size"] = font_size
    plt.rcParams["axes.titlesize"] = font_size + 1
    plt.rcParams["axes.labelsize"] = font_size + 1
    plt.rcParams["xtick.labelsize"] = font_size
    plt.rcParams["ytick.labelsize"] = font_size
    plt.rcParams["legend.fontsize"] = font_size
    plt.rcParams["axes.grid"] = False
    plt.rcParams["xtick.direction"] = "out"
    plt.rcParams["ytick.direction"] = "out"
    plt.rcParams["xtick.major.width"] = axes_linewidth
    plt.rcParams["ytick.major.width"] = axes_linewidth
    plt.rcParams["xtick.major.size"] = 3
    plt.rcParams["ytick.major.size"] = 3


def add_panel_label(ax, label):
    ax.text(
        -0.10,
        1.05,
        label,
        transform=ax.transAxes,
        fontsize=9,
        fontweight="bold",
        color="black",
        ha="left",
        va="bottom",
    )


def save_pub_fig(fig, name):
    FIGDIR.mkdir(parents=True, exist_ok=True)
    fig.tight_layout(pad=1.2)
    fig.savefig(FIGDIR / f"{name}.svg", bbox_inches="tight")
    fig.savefig(FIGDIR / f"{name}.pdf", bbox_inches="tight")
    fig.savefig(FIGDIR / f"{name}.png", dpi=300, bbox_inches="tight")
    plt.close(fig)


def plot_ablation():
    apply_nature_style(font_size=8)
    df = pd.read_csv(BENCH / "ablation" / "ablation_summary_pose_filtered.csv")
    df["mode_label"] = df["ablation"].map(MODE_LABELS)

    fig, axes = plt.subplots(1, 2, figsize=(6.4, 2.7))
    x = np.arange(len(TARGETS))
    width = 0.24
    modes = ["chroma+proteinmpnn", "chroma+bo", "chroma+proteinmpnn+bo"]
    offsets = [-width, 0, width]

    ax = axes[0]
    for mode, offset in zip(modes, offsets):
        part = df[df["ablation"] == mode].set_index("target").loc[TARGETS]
        ax.bar(
            x + offset,
            part["best_score"],
            width=width,
            color=MODE_COLORS[mode],
            label=MODE_LABELS[mode],
        )
    ax.set_ylabel("Best ProteinMPNN score (lower is better)")
    ax.set_xticks(x)
    ax.set_xticklabels(TARGETS)
    ax.set_ylim(0, 2.9)
    add_panel_label(ax, "a")

    ax = axes[1]
    for mode, offset in zip(modes, offsets):
        part = df[df["ablation"] == mode].set_index("target").loc[TARGETS]
        ax.bar(
            x + offset,
            part["best_synthesis_penalty"],
            width=width,
            color=MODE_COLORS[mode],
        )
    ax.axhline(0.8, color=PALETTE["neutral_mid"], lw=0.8, ls="--")
    ax.axhline(1.6, color=PALETTE["neutral_mid"], lw=0.8, ls=":")
    ax.set_ylabel("Best synthesis penalty (lower is better)")
    ax.set_xticks(x)
    ax.set_xticklabels(TARGETS)
    ax.set_ylim(0, 2.7)
    ax.text(2.55, 0.82, "low", fontsize=7, color=PALETTE["neutral_mid"])
    ax.text(2.45, 1.62, "moderate", fontsize=7, color=PALETTE["neutral_mid"])
    add_panel_label(ax, "b")

    handles, labels = axes[0].get_legend_handles_labels()
    fig.legend(handles, labels, loc="upper center", ncol=3, bbox_to_anchor=(0.5, 1.08))
    save_pub_fig(fig, "fig_ablation_bo_tradeoff")


def _parse_top_diff(diff_text):
    rows = []
    for item in str(diff_text).split(";"):
        item = item.strip()
        if not item:
            continue
        match = re.match(r"([A-Z]):D([0-9.]+)/L([0-9.]+)", item)
        if match:
            aa, d_value, l_value = match.groups()
            rows.append((aa, float(d_value), float(l_value)))
    return rows


def plot_dl_comparison():
    apply_nature_style(font_size=8)
    from sklearn.manifold import TSNE

    fig, axes = plt.subplots(1, 3, figsize=(9.2, 3.0))

    AAS_ORDER = list("ACDEFGHIKLMNPQRSTVWY")
    TOP_N = 100

    for idx, target in enumerate(TARGETS):
        ax = axes[idx]
        d_df = pd.read_csv(BENCH / "dl_downsampled" / f"{target}_D_peptide_filtered.csv")
        l_df = pd.read_csv(BENCH / "dl_downsampled" / f"{target}_L_peptide_filtered.csv")

        d_top = d_df.nsmallest(TOP_N, "score").drop_duplicates("sequence")
        l_top = l_df.nsmallest(TOP_N, "score").drop_duplicates("sequence")

        def _onehot_padded(sequences, max_len):
            aa_to_idx = {aa: i for i, aa in enumerate(AAS_ORDER)}
            vecs = []
            for seq in sequences:
                vec = np.zeros(max_len * len(AAS_ORDER))
                for pos, aa in enumerate(str(seq)):
                    if aa in aa_to_idx:
                        vec[pos * len(AAS_ORDER) + aa_to_idx[aa]] = 1.0
                vecs.append(vec)
            return np.array(vecs)

        all_seqs = pd.concat([d_top, l_top])["sequence"]
        max_len = all_seqs.str.len().max()

        d_vecs = _onehot_padded(d_top["sequence"], max_len)
        l_vecs = _onehot_padded(l_top["sequence"], max_len)

        X = np.vstack([d_vecs, l_vecs])
        n_d = len(d_vecs)

        # PCA init to reduce dimensionality before t-SNE
        from sklearn.decomposition import PCA
        n_pca = min(50, X.shape[1], X.shape[0] - 1)
        X_pca = PCA(n_components=n_pca, random_state=42).fit_transform(X)

        tsne = TSNE(n_components=2, perplexity=30, random_state=42,
                     learning_rate="auto", init="pca")
        X_embed = tsne.fit_transform(X_pca)

        ax.scatter(
            X_embed[n_d:, 0], X_embed[n_d:, 1],
            s=28, alpha=0.6, color=PALETTE["orange"], label="L-peptide", rasterized=True,
        )
        ax.scatter(
            X_embed[:n_d, 0], X_embed[:n_d, 1],
            s=28, alpha=0.6, color=PALETTE["blue_main"], label="D-peptide", rasterized=True,
        )

        ax.set_xlabel(f"{target} (top {TOP_N}) — t-SNE 1")
        if idx == 0:
            ax.set_ylabel("t-SNE 2")
        ax.set_xticks([])
        ax.set_yticks([])

        panel_labels = ["a", "b", "c"]
        add_panel_label(ax, panel_labels[idx])

    handles = [
        plt.Line2D([0], [0], marker="o", color="w", markerfacecolor=PALETTE["blue_main"],
                   markersize=8, label="D-peptide"),
        plt.Line2D([0], [0], marker="o", color="w", markerfacecolor=PALETTE["orange"],
                   markersize=8, label="L-peptide"),
    ]
    fig.legend(handles=handles, loc="upper center", ncol=2, bbox_to_anchor=(0.5, 1.04),
               fontsize=8, handletextpad=0.5, columnspacing=1.5)

    save_pub_fig(fig, "fig_dl_sequence_comparison")


def _load_bad_pose_keys(peptide_mode):
    """Buried (inside-target) pose keys from the geometric surface filter."""
    path = BENCH / "buried_pose_inventory.csv"
    if not path.exists():
        return set()
    df = pd.read_csv(path)
    sub = df[df["peptide_mode"] == peptide_mode]
    return set(zip(sub["target"], sub["length"].astype(int), sub["pose"].astype(int)))


def _load_filtered_dl_df(target, peptide_mode):
    csv_path = BENCH / target / peptide_mode / "all_results.csv"
    df = pd.read_csv(csv_path)
    bad_keys = _load_bad_pose_keys(peptide_mode)
    if bad_keys:
        keys = list(zip(
            [target] * len(df),
            df["length"].astype(int),
            df["pose"].astype(int),
        ))
        keep = [k not in bad_keys for k in keys]
        df = df.loc[keep].copy()
    return df.dropna(subset=["score"])


def plot_dl_score_distribution():
    apply_nature_style(font_size=8)
    from scipy import stats as sp_stats

    fig, axes = plt.subplots(1, 3, figsize=(9.2, 2.8))
    rng = np.random.default_rng(42)

    for idx, target in enumerate(TARGETS):
        ax = axes[idx]
        d_df = _load_filtered_dl_df(target, "D_peptide")
        l_df = _load_filtered_dl_df(target, "L_peptide")

        # Downsample L to match D count
        n_d = len(d_df)
        n_l = len(l_df)
        if n_l > n_d:
            l_df = l_df.sample(n=n_d, random_state=rng).copy()

        d_scores = d_df["score"].values
        l_scores = l_df["score"].values

        bins = np.linspace(
            min(d_scores.min(), l_scores.min()),
            max(np.percentile(d_scores, 99), np.percentile(l_scores, 99)),
            50,
        )

        ax.hist(
            d_scores, bins=bins, density=True, alpha=0.55,
            color=PALETTE["blue_main"], label="D-peptide",
        )
        ax.hist(
            l_scores, bins=bins, density=True, alpha=0.55,
            color=PALETTE["orange"], label="L-peptide",
        )

        ks_stat, ks_p = sp_stats.ks_2samp(d_scores, l_scores)

        ax.axvline(np.median(d_scores), color=PALETTE["blue_main"], ls="--", lw=0.8)
        ax.axvline(np.median(l_scores), color=PALETTE["orange"], ls="--", lw=0.8)

        ax.set_xlabel(f"{target} — ProteinMPNN score (NLL)")
        if idx == 0:
            ax.set_ylabel("Density")
        ax.text(
            0.97, 0.95,
            f"$n$ = {n_d}\nKS = {ks_stat:.3f}\np < {ks_p:.1e}" if ks_p < 0.001 else f"$n$ = {n_d}\nKS = {ks_stat:.3f}\np = {ks_p:.3f}",
            transform=ax.transAxes, ha="right", va="top",
            fontsize=7, color=PALETTE["neutral_dark"],
            bbox=dict(boxstyle="round,pad=0.3", fc="white", ec=PALETTE["neutral_light"], lw=0.5),
        )
        if idx == 0:
            ax.legend(loc="upper left", fontsize=7)

    save_pub_fig(fig, "fig_dl_score_distribution")


def main():
    plot_ablation()
    plot_dl_comparison()
    plot_dl_score_distribution()


if __name__ == "__main__":
    main()