File size: 7,667 Bytes
a521a25 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 | from __future__ import annotations
import itertools
import json
from pathlib import Path
import matplotlib as mpl
mpl.use("Agg")
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
from sklearn.metrics import average_precision_score, brier_score_loss, roc_auc_score
ROOT = Path(__file__).resolve().parents[1]
SEED = 20260810
BOOTSTRAPS = 5000
def stratified_bootstrap(y: np.ndarray, p: np.ndarray, rng: np.random.Generator):
pos = np.flatnonzero(y == 1)
neg = np.flatnonzero(y == 0)
aucs, aps, briers = [], [], []
for _ in range(BOOTSTRAPS):
idx = np.r_[rng.choice(pos, len(pos), replace=True), rng.choice(neg, len(neg), replace=True)]
aucs.append(roc_auc_score(y[idx], p[idx]))
aps.append(average_precision_score(y[idx], p[idx]))
briers.append(brier_score_loss(y[idx], p[idx]))
return {
"auroc_ci_low": np.quantile(aucs, 0.025),
"auroc_ci_high": np.quantile(aucs, 0.975),
"auprc_ci_low": np.quantile(aps, 0.025),
"auprc_ci_high": np.quantile(aps, 0.975),
"brier_ci_low": np.quantile(briers, 0.025),
"brier_ci_high": np.quantile(briers, 0.975),
}
def exact_balanced_permutation(y: np.ndarray, p: np.ndarray):
n = len(y)
n_pos = int(y.sum())
observed_auc = roc_auc_score(y, p)
observed_ap = average_precision_score(y, p)
aucs, aps = [], []
for pos in itertools.combinations(range(n), n_pos):
yp = np.zeros(n, dtype=int)
yp[list(pos)] = 1
aucs.append(roc_auc_score(yp, p))
aps.append(average_precision_score(yp, p))
aucs = np.asarray(aucs)
aps = np.asarray(aps)
return {
"permutations": len(aucs),
"auroc_exact_p": float(np.mean(aucs >= observed_auc - 1e-12)),
"auprc_exact_p": float(np.mean(aps >= observed_ap - 1e-12)),
}
def summarize_external():
pred = pd.read_csv(ROOT / "results/external_GSE200546_predictions.csv")
rows = []
for i, (method, d) in enumerate(pred.groupby("method", sort=False)):
y = d.y.to_numpy(dtype=int)
p = d.prob.to_numpy(dtype=float)
point = {
"method": method,
"n": len(d),
"n_acid": int(y.sum()),
"n_control": int((1 - y).sum()),
"auroc": roc_auc_score(y, p),
"auprc": average_precision_score(y, p),
"brier": brier_score_loss(y, p),
}
point.update(stratified_bootstrap(y, p, np.random.default_rng(SEED + i)))
point.update(exact_balanced_permutation(y, p))
rows.append(point)
out = pd.DataFrame(rows)
out.to_csv(ROOT / "results/external_GSE200546_uncertainty.csv", index=False)
return out
def summarize_strict_cv():
d = pd.read_csv(ROOT / "results/strict_cv_metrics.csv")
out = (
d.groupby("method", sort=False)
.agg(
n_domains=("held_out", "nunique"),
auroc_mean=("auroc", "mean"),
auroc_min=("auroc", "min"),
auroc_max=("auroc", "max"),
auprc_mean=("auprc", "mean"),
auprc_min=("auprc", "min"),
auprc_max=("auprc", "max"),
brier_mean=("brier", "mean"),
)
.reset_index()
)
out.to_csv(ROOT / "results/strict_cv_summary.csv", index=False)
return d, out
def revised_figures(cv: pd.DataFrame, uncertainty: pd.DataFrame):
mpl.rcParams.update(
{
"font.family": "sans-serif",
"font.sans-serif": ["Arial", "DejaVu Sans"],
"svg.fonttype": "none",
"pdf.fonttype": 42,
"font.size": 7,
"axes.spines.top": False,
"axes.spines.right": False,
}
)
order = ["ExprLogReg", "MLP", "GCN", "AcidGNet", "SelfLoopGAT", "PermutedGraphGAT"]
colors = dict(zip(order, ["#3F6FA0", "#82A6C8", "#76A996", "#C65D4B", "#D5A059", "#9182AB"]))
fig, axes = plt.subplots(1, 2, figsize=(7.2, 2.75))
for ax, metric in zip(axes, ["auroc", "auprc"]):
for j, method in enumerate(order):
vals = cv.loc[cv.method == method, metric].to_numpy()
ax.scatter(np.full_like(vals, j, dtype=float), vals, s=24, color=colors[method], zorder=3)
ax.plot([j - 0.18, j + 0.18], [vals.mean(), vals.mean()], color="black", lw=1.2)
ax.axhline(0.5, ls="--", lw=0.8, color="0.65")
ax.set_xticks(range(len(order)), order, rotation=35, ha="right")
ax.set_ylim(-0.03, 1.05)
ax.set_ylabel(metric.upper())
ax.set_title("Strict leave-one-cell-line-out\n(each point is one held-out domain)")
fig.tight_layout()
for ext in ["png", "pdf", "svg"]:
fig.savefig(ROOT / f"figures/Fig4_strict_ablation.{ext}", dpi=600 if ext == "png" else None, bbox_inches="tight")
plt.close(fig)
u = uncertainty.set_index("method").loc[order].reset_index()
fig, axes = plt.subplots(1, 2, figsize=(7.2, 2.8))
for ax, metric, label in zip(axes, ["auroc", "auprc"], ["AUROC", "AUPRC"]):
vals = u[metric].to_numpy()
lo = vals - u[f"{metric}_ci_low"].to_numpy()
hi = u[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 j, row in u.iterrows():
ax.text(j, 0.03, f"p={row[f'{metric}_exact_p']:.2f}", ha="center", va="bottom", rotation=90, fontsize=5)
ax.axhline(0.5, ls="--", lw=0.8, color="0.5")
ax.set_xticks(range(len(order)), order, rotation=35, ha="right")
ax.set_ylim(0, 1.08)
ax.set_ylabel(label)
ax.set_title(f"GSE200546 CLS1 external ranking\n95% stratified bootstrap CI; exact p")
fig.suptitle("Cross-cancer technical validation only (n=8; not CRC)", y=1.04, fontsize=8)
fig.tight_layout()
for ext in ["png", "pdf", "svg"]:
fig.savefig(ROOT / f"figures/Fig5_external_validation.{ext}", dpi=600 if ext == "png" else None, bbox_inches="tight")
plt.close(fig)
abl = pd.read_csv(ROOT / "results/external_node_ablation.csv")
a = (
abl.groupby("symbol")
.agg(abs_effect=("abs_effect_all", "mean"), abs_effect_sd=("abs_effect_all", "std"))
.sort_values("abs_effect")
.tail(20)
)
fig, ax = plt.subplots(figsize=(5.2, 3.7))
ax.barh(a.index, a.abs_effect, xerr=a.abs_effect_sd.fillna(0), color="0.55")
ax.set_xlabel("Mean absolute probability change")
ax.set_title("Exploratory node ablation is non-interpretable\nExternal AcidGNet AUROC = 0.50; effects are <0.002")
ax.ticklabel_format(axis="x", style="sci", scilimits=(0, 0))
fig.tight_layout()
for ext in ["png", "pdf", "svg"]:
fig.savefig(ROOT / f"figures/Fig6_node_ablation.{ext}", dpi=600 if ext == "png" else None, bbox_inches="tight")
plt.close(fig)
def main():
uncertainty = summarize_external()
cv, cv_summary = summarize_strict_cv()
revised_figures(cv, uncertainty)
report = {
"external_n": 8,
"external_is_crc": False,
"bootstrap_replicates": BOOTSTRAPS,
"balanced_exact_permutations": 70,
"strict_cv": cv_summary.set_index("method").round(4).to_dict("index"),
"external": uncertainty.set_index("method").round(4).to_dict("index"),
"interpretation": "No AcidGNet superiority; node-ablation ranks are not driver evidence.",
}
(ROOT / "logs/postprocess_validation_summary.json").write_text(json.dumps(report, indent=2), encoding="utf-8")
print(json.dumps(report, indent=2))
if __name__ == "__main__":
main()
|