from __future__ import annotations import shutil from pathlib import Path import matplotlib as mpl import matplotlib.pyplot as plt import numpy as np import pandas as pd from matplotlib.patches import FancyArrowPatch, FancyBboxPatch ROOT = Path(__file__).resolve().parents[1] RESULTS = ROOT / "results" FIGURES = ROOT / "figures" SOURCE = FIGURES / "source_data_v5" SEED = 20260809 BLUE = "#2B6CB0" RED = "#C94C4C" TEAL = "#138A8A" GOLD = "#D69E2E" PURPLE = "#7656A8" GREY = "#68737D" LIGHT = "#EDF2F7" METHOD_COLORS = { "ExprLogReg": GREY, "MLP": GOLD, "GeneAwareDeepSet": PURPLE, "IDSelfLoopGAT": TEAL, "AcidGNetID": RED, } def setup_style() -> None: mpl.rcParams.update( { "font.family": "Arial", "font.size": 8, "axes.titlesize": 10, "axes.labelsize": 8.5, "axes.linewidth": 0.7, "xtick.labelsize": 7, "ytick.labelsize": 7, "legend.fontsize": 7, "svg.fonttype": "none", "pdf.fonttype": 42, "savefig.facecolor": "white", } ) def clean_axis(ax: plt.Axes) -> None: ax.spines["top"].set_visible(False) ax.spines["right"].set_visible(False) ax.tick_params(width=0.7, length=3) def panel(ax: plt.Axes, letter: str) -> None: ax.text(-0.13, 1.07, letter, transform=ax.transAxes, fontsize=13, fontweight="bold", va="top") def save(fig: plt.Figure, stem: str) -> None: FIGURES.mkdir(parents=True, exist_ok=True) fig.savefig(FIGURES / f"{stem}.png", dpi=600, bbox_inches="tight") fig.savefig(FIGURES / f"{stem}.pdf", bbox_inches="tight") fig.savefig(FIGURES / f"{stem}.svg", bbox_inches="tight") plt.close(fig) def copy_source(*filenames: str) -> None: SOURCE.mkdir(parents=True, exist_ok=True) for filename in filenames: shutil.copy2(RESULTS / filename, SOURCE / filename) def fig6_crc_external_biology() -> None: de = pd.read_csv(RESULTS / "v5_crc_aa_hct116_de.csv") con = pd.read_csv(RESULTS / "v5_crc_hct116_crossstudy_concordance.csv") con_summary = pd.read_csv(RESULTS / "v5_crc_hct116_crossstudy_summary.csv").iloc[0] fg = pd.read_csv(RESULTS / "v5_crc_aa_hct116_reactome_fgsea.csv") fig, axes = plt.subplots(2, 2, figsize=(11.5, 8.2), gridspec_kw={"hspace": 0.42, "wspace": 0.34}) ax = axes[0, 0] boxes = [ (0.02, 0.58, 0.28, 0.24, "Open author repository\ncommit + SHA256 locked\n63,152 genes", BLUE), (0.37, 0.58, 0.25, 0.24, "AA-HCT116 RNA-seq\n3 parental + 3 adapted\nindependent laboratory", TEAL), (0.69, 0.58, 0.28, 0.24, "Frozen external test\nno threshold/model tuning\n20 exact label allocations", RED), (0.20, 0.13, 0.26, 0.22, "limma-voom\n13,956 genes tested\n1,714 stringent DEGs", PURPLE), (0.55, 0.13, 0.29, 0.22, "Reactome GSEA +\ncross-study HCT116 audit\nheterogeneity retained", GOLD), ] for x, y, w, h, text, color in boxes: ax.add_patch(FancyBboxPatch((x, y), w, h, boxstyle="round,pad=0.02", fc=color, ec="none", alpha=0.14)) ax.text(x + w / 2, y + h / 2, text, ha="center", va="center", fontsize=7.3, color="#243447") for start, end in [((0.30, 0.70), (0.37, 0.70)), ((0.62, 0.70), (0.69, 0.70)), ((0.50, 0.58), (0.34, 0.35)), ((0.56, 0.58), (0.68, 0.35))]: ax.add_patch(FancyArrowPatch(start, end, arrowstyle="-|>", mutation_scale=10, lw=1, color=GREY)) ax.set_xlim(0, 1); ax.set_ylim(0, 1); ax.axis("off") ax.set_title("Independent CRC chronic-acidosis evidence layer") panel(ax, "a") ax = axes[0, 1] sig = (de.fdr <= 0.05) & (de.log2fc.abs() >= 1) ax.scatter(de.loc[~sig, "log2fc"], -np.log10(de.loc[~sig, "fdr"].clip(1e-300)), s=5, c="#BBC4CC", alpha=0.45, linewidth=0) colors = np.where(de.loc[sig, "log2fc"] > 0, RED, BLUE) ax.scatter(de.loc[sig, "log2fc"], -np.log10(de.loc[sig, "fdr"].clip(1e-300)), s=7, c=colors, alpha=0.7, linewidth=0) ax.axvline(-1, color=GREY, lw=0.7, ls="--"); ax.axvline(1, color=GREY, lw=0.7, ls="--") ax.axhline(-np.log10(0.05), color=GREY, lw=0.7, ls="--") label = de.loc[sig & de.symbol.notna()].nsmallest(6, "fdr") for row in label.itertuples(): ax.text(row.log2fc, -np.log10(max(row.fdr, 1e-300)) + 0.25, row.symbol, fontsize=6, ha="center") ax.set_xlabel("log2 fold change (acid-adapted / parental)") ax.set_ylabel("−log10 FDR") ax.set_title("AA-HCT116 transcriptional remodeling") ax.text(0.02, 0.98, "1,714 DEGs", transform=ax.transAxes, va="top", color=RED, fontweight="bold") clean_axis(ax); panel(ax, "b") ax = axes[1, 0] both = (con.fdr_Pellegrini2018 <= 0.05) & (con.log2fc_Pellegrini2018_adapted_vs_parental.abs() >= 1) & (con.fdr_GSE253054 <= 0.05) & (con.log2fc_GSE253054_HCT116_acid_vs_control.abs() >= 1) ax.scatter(con.loc[~both, "log2fc_GSE253054_HCT116_acid_vs_control"], con.loc[~both, "log2fc_Pellegrini2018_adapted_vs_parental"], s=5, c="#BBC4CC", alpha=0.35, linewidth=0) ax.scatter(con.loc[both, "log2fc_GSE253054_HCT116_acid_vs_control"], con.loc[both, "log2fc_Pellegrini2018_adapted_vs_parental"], s=10, c=RED, alpha=0.65, linewidth=0) ax.axhline(0, color=GREY, lw=0.6); ax.axvline(0, color=GREY, lw=0.6) ax.set_xlabel("GSE253054 HCT116 log2FC") ax.set_ylabel("Pellegrini 2018 AA-HCT116 log2FC") ax.set_title("Cross-study overlap exceeds chance, but effects diverge") ax.text(0.03, 0.97, f"ρ={con_summary.spearman_log2fc:.2f}\nDEG overlap OR={con_summary.significant_overlap_fisher_or:.2f}\nP={con_summary.significant_overlap_fisher_p:.1e}", transform=ax.transAxes, va="top", fontsize=7) clean_axis(ax); panel(ax, "c") ax = axes[1, 1] selected_names = [ "Fatty acid metabolism", "Aerobic respiration and respiratory electron transport", "Respiratory electron transport", "Mitochondrial translation", "Mitochondrial biogenesis", "Selective autophagy", "Glutamate and glutamine metabolism", "Glutathione conjugation", ] focus = fg.set_index("pathway").reindex(selected_names).dropna().reset_index() focus = focus.sort_values("NES") colors = np.where(focus.padj <= 0.05, np.where(focus.NES > 0, RED, BLUE), "#AAB3BB") ax.barh(np.arange(len(focus)), focus.NES, color=colors, alpha=0.85) ax.axvline(0, color=GREY, lw=0.7) ax.set_yticks(np.arange(len(focus))) ax.set_yticklabels(focus.pathway.str.replace("Aerobic respiration and ", "Aerobic respiration +\n", regex=False), fontsize=6.5) ax.set_xlabel("Normalized enrichment score") ax.set_title("Metabolic programs are remodeled, not uniformly activated") for i, row in focus.reset_index(drop=True).iterrows(): x = row.NES - 0.05 if row.NES >= 0 else row.NES + 0.08 ax.text(x, i, f"FDR={row.padj:.3g}", va="center", ha="right" if row.NES >= 0 else "left", fontsize=6) clean_axis(ax); panel(ax, "d") fig.suptitle("Fig. 6 | Independent CRC chronic-acidosis data reveal a reproducible but heterogeneous response", y=0.995, fontsize=12, fontweight="bold") save(fig, "Fig6_v5_independent_CRC_biology") copy_source( "v5_crc_aa_hct116_de.csv", "v5_crc_aa_hct116_reactome_fgsea.csv", "v5_crc_hct116_crossstudy_concordance.csv", "v5_crc_hct116_crossstudy_summary.csv", ) def fig7_crc_external_model() -> None: metrics = pd.read_csv(RESULTS / "v5_crc_external_metrics.csv") panels = pd.read_csv(RESULTS / "v5_crc_external_panel_sensitivity.csv") perms = pd.read_csv(RESULTS / "v5_crc_external_label_permutations.csv") perm_summary = pd.read_csv(RESULTS / "v5_crc_external_label_permutation_summary.csv") curve = pd.read_csv(RESULTS / "v5_crc_external_learning_curve.csv") fig, axes = plt.subplots(2, 2, figsize=(11.5, 8.2), gridspec_kw={"hspace": 0.48, "wspace": 0.32}) ax = axes[0, 0] methods = ["ExprLogReg", "MLP", "GeneAwareDeepSet", "IDSelfLoopGAT", "AcidGNetID"] y = np.arange(len(methods)) for offset, scenario, marker in [(-0.12, "all_source_domains", "o"), (0.12, "exclude_all_HCT116_training", "s")]: use = metrics.set_index(["scenario", "method"]).loc[scenario] vals = [use.loc[m, "auroc"] for m in methods] ax.scatter(vals, y + offset, s=45, marker=marker, c=[METHOD_COLORS[m] for m in methods], edgecolor="white", linewidth=0.5, label=scenario.replace("_", " ")) ax.axvline(0.5, color=GREY, lw=0.7, ls="--") ax.set_yticks(y); ax.set_yticklabels(methods); ax.invert_yaxis() ax.set_xlim(-0.03, 1.05); ax.set_xlabel("External AA-HCT116 AUROC (n=3+3)") ax.set_title("Frozen external discrimination is baseline-led") ax.legend(loc="lower left", fontsize=6.5) clean_axis(ax); panel(ax, "a") ax = axes[0, 1] for scenario, ls in [("all_source_domains", "-"), ("exclude_all_HCT116_training", "--")]: for method in ["ExprLogReg", "AcidGNetID"]: use = panels[(panels.scenario == scenario) & (panels.method == method)].sort_values("panel_size") ax.plot(use.panel_size, use.auroc, marker="o", lw=1.4, ls=ls, color=METHOD_COLORS[method], alpha=0.9, label=f"{method}; {'all' if scenario.startswith('all') else 'HCT116 excluded'}") ax.axhline(0.5, color=GREY, lw=0.7, ls=":") ax.set_xticks([100, 200, 300, 500]); ax.set_ylim(-0.03, 1.05) ax.set_xlabel("Training-fold variance panel size") ax.set_ylabel("External AUROC") ax.set_title("AcidGNet is sensitive to panel size and lineage exclusion") ax.legend(fontsize=6, ncol=2, loc="lower left") clean_axis(ax); panel(ax, "b") ax = axes[1, 0] groups = [] labels = [] observed = [] colors = [] for scenario in ["all_source_domains", "exclude_all_HCT116_training"]: for method in ["ExprLogReg", "AcidGNetID"]: use = perms[(perms.scenario == scenario) & (perms.method == method)] groups.append(use.loc[use.kind.ne("observed"), "auroc"].to_numpy()) labels.append(f"{method}\n{'all' if scenario.startswith('all') else 'no HCT116'}") observed.append(float(use.loc[use.kind.eq("observed"), "auroc"].iloc[0])) colors.append(METHOD_COLORS[method]) parts = ax.violinplot(groups, positions=np.arange(1, 5), showextrema=False, widths=0.75) for body, color in zip(parts["bodies"], colors): body.set_facecolor(color); body.set_edgecolor("none"); body.set_alpha(0.24) rng = np.random.default_rng(20260809) for i, (values, obs, color) in enumerate(zip(groups, observed, colors), 1): ax.scatter(i + rng.uniform(-0.16, 0.16, len(values)), values, s=8, c=color, alpha=0.35, linewidth=0) ax.scatter(i, obs, marker="*", s=95, c=color, edgecolor="black", linewidth=0.4, zorder=5) row = perm_summary[(perm_summary.scenario == ("all_source_domains" if i <= 2 else "exclude_all_HCT116_training")) & (perm_summary.method == ("ExprLogReg" if i % 2 == 1 else "AcidGNetID"))].iloc[0] ax.text(i, 1.04, f"P={row.permutation_p_ge_observed:.2f}", ha="center", va="bottom", fontsize=6.5) ax.axhline(0.5, color=GREY, lw=0.7, ls="--") ax.set_xticks(np.arange(1, 5)); ax.set_xticklabels(labels, fontsize=6.5) ax.set_ylim(-0.03, 1.12); ax.set_ylabel("External AUROC") ax.set_title("Training-label nulls do not support a graph-model advantage") clean_axis(ax); panel(ax, "c") ax = axes[1, 1] for method in ["ExprLogReg", "AcidGNetID"]: use = curve[curve.method == method] for k, group in use.groupby("n_training_domains"): jitter = np.random.default_rng(SEED + int(k)).uniform(-0.08, 0.08, len(group)) ax.scatter(np.full(len(group), int(k)) + jitter, group.auroc, s=12, color=METHOD_COLORS[method], alpha=0.25, linewidth=0) summary = use.groupby("n_training_domains").auroc.agg(["mean", "std", "count"]).reset_index() err = summary["std"].fillna(0) ax.errorbar(summary.n_training_domains, summary["mean"], yerr=err, marker="o", lw=1.5, capsize=2.5, color=METHOD_COLORS[method], label=method) ax.axhline(0.5, color=GREY, lw=0.7, ls="--") ax.set_xticks([2, 3, 4, 5, 6]); ax.set_ylim(-0.03, 1.08) ax.set_xlabel("Number of source domains") ax.set_ylabel("AA-HCT116 AUROC across domain combinations") ax.set_title("More domains do not monotonically improve AcidGNet") ax.legend() clean_axis(ax); panel(ax, "d") fig.suptitle("Fig. 7 | Independent CRC validation exposes model and topology instability", y=0.995, fontsize=12, fontweight="bold") save(fig, "Fig7_v5_CRC_external_model_robustness") copy_source( "v5_crc_external_metrics.csv", "v5_crc_external_panel_sensitivity.csv", "v5_crc_external_label_permutations.csv", "v5_crc_external_label_permutation_summary.csv", "v5_crc_external_learning_curve.csv", ) def fig_s1_topology_null() -> None: topology = pd.read_csv(RESULTS / "v5_crc_external_topology_sensitivity.csv") nulls = pd.read_csv(RESULTS / "v5_crc_external_coherent_graph_nulls.csv") summary = pd.read_csv(RESULTS / "v5_crc_external_coherent_graph_null_summary.csv") graphs = pd.read_csv(RESULTS / "v5_crc_external_topology_graphs.csv") fig, axes = plt.subplots(1, 3, figsize=(12.2, 3.6), gridspec_kw={"wspace": 0.34}) ax = axes[0] topologies = ["full_STRING", "training_acid_coherent_STRING", "no_PPI_edges"] labels = ["Full STRING", "Acid-coherent\nSTRING", "No PPI edges"] x = np.arange(3) width = 0.34 for offset, scenario, color in [(-width / 2, "all_source_domains", BLUE), (width / 2, "exclude_all_HCT116_training", RED)]: use = topology.set_index(["scenario", "topology"]).loc[scenario] ax.bar(x + offset, [use.loc[t, "auroc"] for t in topologies], width=width, color=color, alpha=0.8, label="All domains" if scenario.startswith("all") else "HCT116 excluded") ax.axhline(0.5, color=GREY, lw=0.7, ls="--") ax.set_xticks(x); ax.set_xticklabels(labels, fontsize=6.5); ax.set_ylim(0, 1.08) ax.set_ylabel("External AUROC"); ax.set_title("Topology sensitivity") ax.legend(fontsize=6.5); clean_axis(ax); panel(ax, "a") ax = axes[1] for i, scenario in enumerate(["all_source_domains", "exclude_all_HCT116_training"], 1): values = nulls[(nulls.scenario == scenario) & (nulls.kind == "degree_preserving")].auroc.to_numpy() parts = ax.violinplot([values], positions=[i], showextrema=False, widths=0.75) color = BLUE if i == 1 else RED parts["bodies"][0].set_facecolor(color); parts["bodies"][0].set_alpha(0.25); parts["bodies"][0].set_edgecolor("none") rng = np.random.default_rng(20260809 + i) ax.scatter(i + rng.uniform(-0.16, 0.16, len(values)), values, s=9, color=color, alpha=0.4, linewidth=0) obs = float(nulls[(nulls.scenario == scenario) & (nulls.kind == "observed_acid_coherent")].auroc.iloc[0]) p = float(summary.loc[summary.scenario == scenario, "exact_p_ge_observed"].iloc[0]) ax.scatter(i, obs, marker="*", s=100, color=color, edgecolor="black", linewidth=0.4) ax.text(i, 1.03, f"P={p:.2f}", ha="center", fontsize=7) ax.axhline(0.5, color=GREY, lw=0.7, ls="--") ax.set_xticks([1, 2]); ax.set_xticklabels(["All domains", "HCT116 excluded"]) ax.set_ylim(-0.03, 1.12); ax.set_ylabel("External AUROC") ax.set_title("Acid-coherent graph vs degree-matched nulls") clean_axis(ax); panel(ax, "b") ax = axes[2] use = graphs[graphs.topology.isin(["full_STRING", "training_acid_coherent_STRING"])].copy() labels = [f"{'All' if r.scenario.startswith('all') else 'No HCT116'}\n{'Full' if r.topology == 'full_STRING' else 'Coherent'}" for r in use.itertuples()] ax.bar(np.arange(len(use)), use.undirected_edges, color=[BLUE, TEAL, RED, GOLD], alpha=0.8) ax.set_xticks(np.arange(len(use))); ax.set_xticklabels(labels, fontsize=6.5) ax.set_ylabel("Undirected edges") ax.set_title("The coherent graph is very sparse") for i, row in enumerate(use.itertuples()): ax.text(i, row.undirected_edges + 2, f"{row.undirected_edges}\niso={row.isolated_fraction:.0%}", ha="center", fontsize=6) clean_axis(ax); panel(ax, "c") fig.suptitle("Supplementary Fig. S1 | Acid-coherent topology remains exploratory", y=1.01, fontsize=11, fontweight="bold") save(fig, "FigS1_v5_acid_coherent_graph_null") copy_source( "v5_crc_external_topology_sensitivity.csv", "v5_crc_external_topology_graphs.csv", "v5_crc_external_coherent_graph_nulls.csv", "v5_crc_external_coherent_graph_null_summary.csv", ) def main() -> None: setup_style() fig6_crc_external_biology() fig7_crc_external_model() fig_s1_topology_null() print("Generated Fig6, Fig7 and FigS1 in PNG/PDF/SVG with source-data copies.") if __name__ == "__main__": main()