File size: 11,058 Bytes
fc329a3 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 | """Run bulk deconvolution conformal prediction experiments.
Exp 2.1: Semi-synthetic bulk RNA-seq deconvolution.
- Load PBMC3K reference
- Generate pseudo-bulk samples with known cell type proportions (ONCE)
- NNLS deconvolution → residuals (ONCE)
- 200 reps = different random cal/test splits of the same residuals
"""
import argparse
import json
import logging
import time
from pathlib import Path
import sys
import numpy as np
import scanpy as sc
import yaml
sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
from src.dgp.deconv import nnls_deconv
from src.dgp.pseudobulk import generate_pseudobulk
from src.methods import (
jackknife_plus_conformal,
global_split_conformal,
oneshot_conformal,
partition_conformal,
trainres_conformal,
weighted_conformal,
twostage_conformal,
full_conformal,
)
from src.methods._knn_sigma import knn_sigma_hat
from src.metrics import (
coverage_variance,
marginal_coverage,
max_disparity,
stratified_coverage,
worst_stratum_coverage,
)
from src.metrics import mean_radius, radius_by_strata
from src.metrics.sscv import size_stratified_coverage_violation
from src.utils.simplex import aitchison_dist
from src.utils.strata import (
precompute_fixed_strata,
stratify_by_boundary,
stratify_by_entropy,
stratify_by_kmeans,
)
logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(message)s")
log = logging.getLogger(__name__)
STRATA_REGISTRY = {
"boundary": stratify_by_boundary,
"entropy": stratify_by_entropy,
"kmeans": stratify_by_kmeans,
}
def run_one_rep(
R: np.ndarray,
U: np.ndarray,
cfg: dict,
rep_idx: int,
base_seed: int,
fixed_labels: np.ndarray | None = None,
) -> dict:
"""Run one repetition: random cal/test split + conformal methods."""
alpha = cfg["conformal"]["alpha"]
n_strata = cfg["evaluation"]["n_strata"]
strata_method = cfg["evaluation"]["strata_method"]
cal_frac = cfg["conformal"]["cal_frac"]
n_samples = len(R)
seed = base_seed + rep_idx
rng = np.random.default_rng(seed)
# Random cal/test split
n_cal = int(n_samples * cal_frac)
idx = rng.permutation(n_samples)
idx_cal, idx_test = idx[:n_cal], idx[n_cal:]
R_cal, R_test = R[idx_cal], R[idx_test]
U_cal, U_test = U[idx_cal], U[idx_test]
# Stratification on test set
if fixed_labels is not None:
strata_test = fixed_labels[idx_test]
else:
strata_fn = STRATA_REGISTRY[strata_method]
strata_test = strata_fn(U_test, n_strata)
rep_results = {}
for method_name in cfg["conformal"]["methods"]:
start = time.perf_counter()
if method_name == "global":
result = global_split_conformal(R_cal, R_test, alpha)
elif method_name == "partition":
if fixed_labels is not None:
strata_cal = fixed_labels[idx_cal]
else:
strata_cal = strata_fn(U_cal, n_strata)
result = partition_conformal(R_cal, R_test, alpha, strata_cal, strata_test)
elif method_name == "twostage":
n_scale_est = len(R_cal) // 2
result = twostage_conformal(R_cal, R_test, alpha, U_cal, U_test, n_scale_est=n_scale_est)
elif method_name == "fullcp":
result = full_conformal(R_cal, R_test, alpha, U_cal, U_test)
elif method_name == "jackknife_plus":
result = jackknife_plus_conformal(R_cal, R_test, alpha, U_cal=U_cal, U_test=U_test)
elif method_name == "oneshot":
result = oneshot_conformal(R_cal, R_test, alpha, U_cal, U_test)
elif method_name == "weighted":
sigma_cal = knn_sigma_hat(U_cal, R_cal, U_cal)
sigma_test = knn_sigma_hat(U_cal, R_cal, U_test)
floor = float(np.mean(sigma_cal) * 0.1)
weights_cal = 1.0 / np.maximum(sigma_cal, floor)
weights_test = 1.0 / np.maximum(sigma_test, floor)
result = weighted_conformal(R_cal, R_test, alpha, weights_cal, weights_test)
elif method_name == "trainres":
train_perm = rng.permutation(n_samples)
idx_train = train_perm[:n_cal]
result = trainres_conformal(
R_cal, R_test, alpha, U_cal, U_test, R[idx_train], U[idx_train]
)
else:
continue
runtime_sec = time.perf_counter() - start
rep_results[method_name] = {
"marginal_coverage": float(marginal_coverage(result.covered)),
"max_disparity": float(max_disparity(result.covered, strata_test, alpha)),
"worst_stratum_coverage": float(worst_stratum_coverage(result.covered, strata_test)),
"stratified_coverage": {
str(k): float(v)
for k, v in stratified_coverage(result.covered, strata_test).items()
},
"mean_radius": float(mean_radius(result.radius)),
"sscv": float(size_stratified_coverage_violation(result.covered, result.radius, alpha)),
"coverage_variance": float(coverage_variance(result.covered, strata_test)),
"runtime_sec": float(runtime_sec),
"radius_by_strata": {
str(k): float(v)
for k, v in radius_by_strata(result.radius, strata_test).items()
},
}
return rep_results
def aggregate_reps(all_reps: list[dict]) -> dict:
"""Aggregate metrics across repetitions."""
methods = all_reps[0].keys()
agg = {}
for method in methods:
scalar_keys = [
"marginal_coverage",
"max_disparity",
"worst_stratum_coverage",
"mean_radius",
"sscv",
"coverage_variance",
"runtime_sec",
]
agg[method] = {}
for key in scalar_keys:
values = [rep[method][key] for rep in all_reps]
agg[method][key] = {"mean": float(np.mean(values)), "std": float(np.std(values))}
# Aggregate per-strata coverage (some reps may lack certain strata)
all_strata_keys: set[str] = set()
for rep in all_reps:
all_strata_keys.update(rep[method]["stratified_coverage"].keys())
agg[method]["stratified_coverage"] = {}
for s in sorted(all_strata_keys):
vals = [
rep[method]["stratified_coverage"][s]
for rep in all_reps
if s in rep[method]["stratified_coverage"]
]
if vals:
agg[method]["stratified_coverage"][s] = {
"mean": float(np.mean(vals)),
"std": float(np.std(vals)),
"n_reps": len(vals),
}
# Aggregate per-strata radius
all_radius_keys: set[str] = set()
for rep in all_reps:
all_radius_keys.update(rep[method]["radius_by_strata"].keys())
agg[method]["radius_by_strata"] = {}
for s in sorted(all_radius_keys):
vals = [
rep[method]["radius_by_strata"][s]
for rep in all_reps
if s in rep[method]["radius_by_strata"]
]
if vals:
agg[method]["radius_by_strata"][s] = {
"mean": float(np.mean(vals)),
"std": float(np.std(vals)),
"n_reps": len(vals),
}
return agg
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--config", required=True)
args = parser.parse_args()
with open(args.config) as f:
cfg = yaml.safe_load(f)
exp_name = cfg["experiment"]
log.info(f"Running experiment: {exp_name}")
# ── Step 1: Load PBMC3K reference ──
log.info("Loading PBMC3K reference data...")
adata = sc.datasets.pbmc3k()
log.info(f" Loaded {adata.n_obs} cells, {adata.n_vars} genes")
celltype_key = cfg["data"]["celltype_key"]
expr = adata.X
if hasattr(expr, "toarray"):
expr = expr.toarray()
expr = np.asarray(expr, dtype=np.float64)
if celltype_key not in adata.obs.columns:
log.info(f" '{celltype_key}' not found, adding via KMeans clustering...")
from sklearn.decomposition import PCA
from sklearn.cluster import KMeans
pca = PCA(n_components=30, random_state=42)
X_pca = pca.fit_transform(expr)
kmeans = KMeans(n_clusters=8, random_state=42, n_init=10)
adata.obs[celltype_key] = "ct_" + kmeans.fit_predict(X_pca).astype(str)
log.info(f" Created {adata.obs[celltype_key].nunique()} cell type clusters")
cell_type_names = sorted(np.unique(adata.obs[celltype_key].values))
gene_names = adata.var_names.tolist()
labels = adata.obs[celltype_key].values
# ── Step 2: Generate pseudo-bulk + deconvolve (ONCE) ──
base_seed = cfg["seed"]
log.info("Generating pseudo-bulk dataset (once)...")
pb = generate_pseudobulk(
expr=expr,
labels=labels,
cell_type_names=cell_type_names,
gene_names=gene_names,
n_samples=cfg["data"]["n_samples"],
cells_per_sample=cfg["data"]["cells_per_sample"],
concentration=cfg["data"]["concentration"],
noise_sd=cfg["data"]["noise_sd"],
seed=base_seed,
)
log.info(f" Generated {pb.bulk.shape[0]} samples, {len(cell_type_names)} types")
log.info("Running NNLS deconvolution (once)...")
U = nnls_deconv(pb.bulk, pb.signature)
log.info(f" Deconvolved {U.shape[0]} samples")
R = aitchison_dist(pb.proportions, U)
log.info(f" Computed residuals: mean={R.mean():.3f}, std={R.std():.3f}")
# ── Step 3: 200 reps = different random cal/test splits ──
n_reps = cfg["conformal"]["n_reps"]
log.info(f"Running {n_reps} conformal reps (split-only)...")
fixed_labels = None
if cfg["evaluation"].get("fixed_strata", True):
fixed_labels = precompute_fixed_strata(
U,
cfg["evaluation"]["strata_method"],
cfg["evaluation"]["n_strata"],
seed=base_seed,
)
all_reps = []
for i in range(n_reps):
rep_results = run_one_rep(R, U, cfg, i, base_seed, fixed_labels=fixed_labels)
all_reps.append(rep_results)
if (i + 1) % 50 == 0:
log.info(f" Completed {i + 1}/{n_reps} reps")
agg = aggregate_reps(all_reps)
# ── Save results ──
out_dir = Path("results/tables")
out_dir.mkdir(parents=True, exist_ok=True)
out_path = out_dir / f"{exp_name}.json"
with open(out_path, "w") as f:
json.dump({"config": cfg, "aggregated": agg, "raw": all_reps}, f, indent=2)
log.info(f"Results saved to {out_path}")
for method, metrics in agg.items():
cov = metrics["marginal_coverage"]
disp = metrics["max_disparity"]
log.info(f" {method}: coverage={cov['mean']:.3f}±{cov['std']:.3f}, "
f"disparity={disp['mean']:.3f}±{disp['std']:.3f}")
if __name__ == "__main__":
main()
|