from __future__ import annotations from pathlib import Path import matplotlib as mpl mpl.use("Agg") import matplotlib.pyplot as plt import numpy as np import pandas as pd ROOT = Path(__file__).resolve().parents[1] OUT = ROOT / "figures" SOURCE = OUT / "source_data" SOURCE.mkdir(parents=True, exist_ok=True) ORDER = ["ExprLogReg", "MLP", "GCN", "AcidGNet", "SelfLoopGAT", "PermutedGraphGAT"] LABELS = ["Expression LR", "MLP", "GCN", "AcidGNet", "Self-loop GAT", "Permuted graph GAT"] COLORS = { "ExprLogReg": "#3F6FA0", "MLP": "#86A9C8", "GCN": "#78A996", "AcidGNet": "#C75D4B", "SelfLoopGAT": "#D5A056", "PermutedGraphGAT": "#9184A8", } def style() -> None: mpl.rcParams.update( { "font.family": "sans-serif", "font.sans-serif": ["Arial", "Helvetica", "DejaVu Sans"], "font.size": 7, "axes.spines.top": False, "axes.spines.right": False, "axes.linewidth": 0.8, "legend.frameon": False, "svg.fonttype": "none", "pdf.fonttype": 42, } ) def panel_label(ax, label: str) -> None: ax.text(-0.16, 1.08, label, transform=ax.transAxes, fontsize=9, fontweight="bold", va="top") def save(fig, stem: str) -> None: for ext in ["png", "pdf", "svg"]: fig.savefig(OUT / f"{stem}.{ext}", dpi=600 if ext == "png" else None, bbox_inches="tight") plt.close(fig) def heatmap(ax, table: pd.DataFrame, title: str, vmin=0, vmax=1) -> None: image = ax.imshow(table.to_numpy(), vmin=vmin, vmax=vmax, cmap="Blues", aspect="auto") for i in range(table.shape[0]): for j in range(table.shape[1]): value = table.iloc[i, j] ax.text(j, i, f"{value:.2f}", ha="center", va="center", fontsize=6, color="white" if value > 0.65 else "black") ax.set_xticks(range(table.shape[1]), table.columns) ax.set_yticks(range(table.shape[0]), [LABELS[ORDER.index(x)] for x in table.index]) ax.set_title(title) return image def strict_figure() -> None: cv = pd.read_csv(ROOT / "results/strict_cv_metrics.csv") cv.to_csv(SOURCE / "Fig3_strict_validation_source_data.csv", index=False) auc = cv.pivot(index="method", columns="held_out", values="auroc").loc[ORDER] apr = cv.pivot(index="method", columns="held_out", values="auprc").loc[ORDER] baseline = cv.loc[cv.method == "ExprLogReg", ["held_out", "auroc"]].set_index("held_out").auroc delta = cv.copy() delta["delta_auroc_vs_expression_lr"] = delta.apply(lambda r: r.auroc - baseline[r.held_out], axis=1) delta.to_csv(SOURCE / "Fig3_delta_source_data.csv", index=False) fig = plt.figure(figsize=(7.2, 5.1)) gs = fig.add_gridspec(2, 2, width_ratios=[1.05, 1.0], hspace=0.48, wspace=0.48) ax_a = fig.add_subplot(gs[0, 0]) heatmap(ax_a, auc, "AUROC by held-out cell line") panel_label(ax_a, "a") ax_b = fig.add_subplot(gs[0, 1]) heatmap(ax_b, apr, "AUPRC by held-out cell line") panel_label(ax_b, "b") ax_c = fig.add_subplot(gs[1, 0]) neural = [m for m in ORDER if m != "ExprLogReg"] offsets = np.linspace(-0.24, 0.24, 3) for j, held in enumerate(["FaDu", "HCT116", "SiHa"]): d = delta[(delta.held_out == held) & delta.method.isin(neural)].set_index("method").loc[neural] ax_c.scatter(np.arange(len(neural)) + offsets[j], d.delta_auroc_vs_expression_lr, s=28, label=held, alpha=0.9) ax_c.axhline(0, color="black", lw=0.8) ax_c.set_xticks(range(len(neural)), [LABELS[ORDER.index(m)] for m in neural], rotation=32, ha="right") ax_c.set_ylabel("ΔAUROC versus expression LR") ax_c.set_title("Graph models provide no consistent gain") ax_c.legend(ncol=3, loc="lower left", fontsize=6) panel_label(ax_c, "c") ax_d = fig.add_subplot(gs[1, 1]) for i, method in enumerate(ORDER): vals = cv.loc[cv.method == method, "brier"].to_numpy() ax_d.scatter(np.full_like(vals, i, dtype=float), vals, color=COLORS[method], s=28, zorder=3) ax_d.plot([i - 0.18, i + 0.18], [vals.mean(), vals.mean()], color="black", lw=1) ax_d.axhline(0.25, color="0.6", ls="--", lw=0.8, label="uninformative 0.5 probability") ax_d.set_xticks(range(len(ORDER)), LABELS, rotation=32, ha="right") ax_d.set_ylabel("Brier score (lower is better)") ax_d.set_title("Ranking does not imply calibration") ax_d.legend(fontsize=6, loc="upper left") panel_label(ax_d, "d") fig.suptitle("Strict leave-one-cell-line-out evaluation", fontsize=10, y=0.995) save(fig, "Fig3_strict_validation_optimized") def external_figure() -> None: unc = pd.read_csv(ROOT / "results/external_GSE200546_uncertainty.csv").set_index("method").loc[ORDER].reset_index() pred = pd.read_csv(ROOT / "results/external_GSE200546_predictions.csv") unc.to_csv(SOURCE / "Fig4_external_uncertainty_source_data.csv", index=False) pred.to_csv(SOURCE / "Fig4_external_predictions_source_data.csv", index=False) fig = plt.figure(figsize=(7.2, 5.0)) gs = fig.add_gridspec(2, 2, hspace=0.48, wspace=0.42) for ax, metric, label, panel in [ (fig.add_subplot(gs[0, 0]), "auroc", "AUROC", "a"), (fig.add_subplot(gs[0, 1]), "auprc", "AUPRC", "b"), ]: vals = unc[metric].to_numpy() lo = vals - unc[f"{metric}_ci_low"].to_numpy() hi = unc[f"{metric}_ci_high"].to_numpy() - vals ax.bar(range(len(ORDER)), vals, color=[COLORS[m] for m in ORDER], width=0.72) ax.errorbar(range(len(ORDER)), vals, yerr=np.vstack([lo, hi]), fmt="none", ecolor="black", capsize=2, lw=0.8) for i, row in unc.iterrows(): ax.text(i, 0.02, f"p={row[f'{metric}_exact_p']:.2f}", ha="center", va="bottom", rotation=90, fontsize=5) ax.axhline(0.5, color="0.6", ls="--", lw=0.8) ax.set_ylim(0, 1.08) ax.set_xticks(range(len(ORDER)), LABELS, rotation=32, ha="right") ax.set_ylabel(label) ax.set_title(f"{label} with 95% bootstrap CI") panel_label(ax, panel) ax_c = fig.add_subplot(gs[1, 0]) vals = unc.brier.to_numpy() ax_c.bar(range(len(ORDER)), vals, color=[COLORS[m] for m in ORDER], width=0.72) ax_c.axhline(0.25, color="0.4", ls="--", lw=0.8) ax_c.set_xticks(range(len(ORDER)), LABELS, rotation=32, ha="right") ax_c.set_ylabel("Brier score") ax_c.set_title("External probability calibration") panel_label(ax_c, "c") ax_d = fig.add_subplot(gs[1, 1]) wide = pred.pivot(index=["sample", "y"], columns="method", values="prob").reset_index() for y, color, label in [(0, "#4C78A8", "pH 7.4"), (1, "#C75D4B", "pH 6.6")]: d = wide[wide.y == y] ax_d.scatter(d.ExprLogReg, d.AcidGNet, s=38, color=color, label=label, edgecolor="white", linewidth=0.5) ax_d.axhline(0.5, color="0.75", lw=0.7, ls="--") ax_d.axvline(0.5, color="0.75", lw=0.7, ls="--") ax_d.set_xlabel("Expression LR score") ax_d.set_ylabel("AcidGNet score") ax_d.set_title("Model outputs in eight external samples") ax_d.legend(fontsize=6) panel_label(ax_d, "d") fig.suptitle("Independent cross-cancer technical validation (GSE200546; n=8; not CRC)", fontsize=9.5, y=0.995) save(fig, "Fig4_external_validation_optimized") def graph_diagnostic_figure() -> None: graphs = pd.read_csv(ROOT / "results/strict_cv_graphs.csv") cv = pd.read_csv(ROOT / "results/strict_cv_metrics.csv") ext = pd.read_csv(ROOT / "results/external_GSE200546_metrics.csv").set_index("method") abl = pd.read_csv(ROOT / "results/external_node_ablation.csv") graph_auc = graphs.merge(cv[cv.method == "AcidGNet"][["held_out", "auroc"]], left_on="fold", right_on="held_out") graph_auc.to_csv(SOURCE / "Fig5_graph_domain_source_data.csv", index=False) abl_summary = abl.groupby("symbol", as_index=False).agg(mean_abs_effect=("abs_effect_all", "mean"), sd=("abs_effect_all", "std")) abl_summary.to_csv(SOURCE / "Fig5_node_ablation_source_data.csv", index=False) fig = plt.figure(figsize=(7.2, 4.7)) gs = fig.add_gridspec(2, 2, hspace=0.5, wspace=0.42) ax_a = fig.add_subplot(gs[0, 0]) ax_a.bar(graphs.fold, graphs.directed_edges, color="#8A9AAA") for i, row in graphs.iterrows(): ax_a.text(i, row.directed_edges + 5, f"{int(row.nodes)} nodes", ha="center", fontsize=6) ax_a.set_ylabel("Directed STRING edges") ax_a.set_title("Fold-specific graphs are sparse") panel_label(ax_a, "a") ax_b = fig.add_subplot(gs[0, 1]) for _, row in graph_auc.iterrows(): ax_b.scatter(row.directed_edges, row.auroc, s=42, color="#C75D4B") ax_b.text(row.directed_edges + 4, row.auroc, row.fold, va="center", fontsize=6) ax_b.axhline(0.5, color="0.65", ls="--", lw=0.8) ax_b.set_xlim(70, 240) ax_b.set_ylim(0, 1.03) ax_b.set_xlabel("Directed STRING edges") ax_b.set_ylabel("Held-out AcidGNet AUROC") ax_b.set_title("Edge count does not rescue transfer") panel_label(ax_b, "b") ax_c = fig.add_subplot(gs[1, 0]) controls = ["ExprLogReg", "GCN", "AcidGNet", "SelfLoopGAT", "PermutedGraphGAT"] vals = [ext.loc[m, "auroc"] for m in controls] ax_c.bar(range(len(controls)), vals, color=[COLORS[m] for m in controls]) ax_c.axhline(0.5, color="0.65", ls="--", lw=0.8) ax_c.set_xticks(range(len(controls)), [LABELS[ORDER.index(m)] for m in controls], rotation=30, ha="right") ax_c.set_ylim(0, 1.02) ax_c.set_ylabel("External AUROC") ax_c.set_title("Specific STRING topology adds no external gain") panel_label(ax_c, "c") ax_d = fig.add_subplot(gs[1, 1]) top = abl_summary.nlargest(12, "mean_abs_effect").sort_values("mean_abs_effect") lower = np.minimum(top["sd"].fillna(0).to_numpy(), top.mean_abs_effect.to_numpy()) upper = top["sd"].fillna(0).to_numpy() ax_d.barh(top.symbol, top.mean_abs_effect, xerr=np.vstack([lower, upper]), color="0.55") ax_d.ticklabel_format(axis="x", style="sci", scilimits=(0, 0)) ax_d.set_xlabel("Mean absolute score change") ax_d.set_title("Chance-level model yields negligible ablations") panel_label(ax_d, "d") fig.suptitle("Diagnostic evidence for graph-prior failure", fontsize=10, y=0.995) save(fig, "Fig5_graph_prior_diagnostics") def main() -> None: style() strict_figure() external_figure() graph_diagnostic_figure() print("Optimized Figs. 3-5 exported as PNG/PDF/SVG with source-data CSVs.") if __name__ == "__main__": main()