File size: 10,908 Bytes
ebaa0d1 | 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 | #!/usr/bin/env python3
"""Build measured Poucher substantivity regression targets and gate baselines."""
from __future__ import annotations
import json
import math
from collections import Counter
from pathlib import Path
import numpy as np
DATA = Path("data")
ARTIFACTS = Path("artifacts")
FEATURE_NAMES = [
"log10_vapor_pressure_pa",
"boiling_point_k",
"molecular_weight",
"logp",
]
def load_dataset_cas() -> set[str]:
cas_set: set[str] = set()
with open(DATA / "empirical_dataset_v8.jsonl") as f:
for line in f:
record = json.loads(line)
if record.get("is_control"):
continue
for comp in record.get("formula", []):
cas = comp.get("cas")
if cas:
cas_set.add(cas)
return cas_set
def load_poucher_targets() -> dict[str, dict]:
rows = {}
with open(DATA / "poucher_substantivity.jsonl") as f:
for line in f:
row = json.loads(line)
rows[row["cas"]] = row
return rows
def load_aroma_features() -> dict[str, dict]:
features = {}
with open(DATA / "aroma_chemicals.jsonl") as f:
for line in f:
row = json.loads(line)
cas = row.get("cas")
if cas:
features[cas] = row
return features
def feature_row(raw: dict | None) -> dict:
if raw is None:
return {name: None for name in FEATURE_NAMES}
vp = raw.get("vapor_pressure_pa")
return {
"log10_vapor_pressure_pa": math.log10(vp) if isinstance(vp, (int, float)) and vp > 0 else None,
"boiling_point_k": raw.get("boiling_point_k"),
"molecular_weight": raw.get("molecular_weight"),
"logp": raw.get("logp"),
}
def write_targets() -> dict:
dataset_cas = load_dataset_cas()
poucher = load_poucher_targets()
aroma = load_aroma_features()
all_rows = []
dataset_rows = []
for cas, target in sorted(poucher.items()):
raw_features = aroma.get(cas)
features = feature_row(raw_features)
coeff = float(target["poucher_coefficient"])
row = {
"cas": cas,
"name": target["name"],
"target": {
"raw_poucher_coefficient": coeff,
"log10_poucher_coefficient": math.log10(coeff),
"source": target["source"],
"source_page": target["source_page"],
"target_type": target["target_type"],
},
"features": features,
"feature_role": "molecular_features_only_not_target_derivation",
"in_empirical_dataset_v8": cas in dataset_cas,
"feature_source": "data/aroma_chemicals.jsonl",
"cas_match_source": target.get("cas_match_source"),
"coefficient_conflict": target.get("coefficient_conflict", False),
}
all_rows.append(row)
if row["in_empirical_dataset_v8"]:
dataset_rows.append(row)
(DATA / "substantivity_targets_poucher.jsonl").write_text(
"\n".join(json.dumps(r, sort_keys=True) for r in all_rows) + "\n"
)
(DATA / "substantivity_targets_dataset.jsonl").write_text(
"\n".join(json.dumps(r, sort_keys=True) for r in dataset_rows) + "\n"
)
values = np.array([r["target"]["raw_poucher_coefficient"] for r in all_rows], dtype=float)
dataset_values = np.array([r["target"]["raw_poucher_coefficient"] for r in dataset_rows], dtype=float)
transform = {
"target_source": "data/poucher_substantivity.jsonl",
"raw_target": "Poucher measured duration-of-evaporation coefficient, 1-100",
"model_target": "log10_poucher_coefficient",
"transform": "log10(raw_poucher_coefficient)",
"inverse_transform": "10 ** model_prediction",
"reason": "positive bounded coefficient with long upper tail; features remain separate and are never used to derive the target",
"tier_band_boundaries_for_presentation_only": {"top": [1, 14], "mid": [15, 60], "base": [61, 100]},
}
ARTIFACTS.mkdir(exist_ok=True)
(ARTIFACTS / "substantivity_transform.json").write_text(json.dumps(transform, indent=2))
hist_counts, hist_edges = np.histogram(values, bins=[1, 15, 31, 46, 61, 81, 101])
summary = {
"poucher_rows": len(all_rows),
"empirical_dataset_unique_cas": len(dataset_cas),
"empirical_dataset_measured_coverage": len(dataset_rows),
"empirical_dataset_measured_coverage_pct": len(dataset_rows) / len(dataset_cas),
"raw_summary": summarize(values),
"dataset_raw_summary": summarize(dataset_values) if len(dataset_values) else None,
"raw_histogram": {
"bins": ["1-14", "15-30", "31-45", "46-60", "61-80", "81-100"],
"counts": hist_counts.tolist(),
},
"tier_band_distribution_all": dict(Counter(r["target"]["raw_poucher_coefficient"] <= 14 and "top" or r["target"]["raw_poucher_coefficient"] <= 60 and "mid" or "base" for r in all_rows)),
"feature_completeness_all": feature_completeness(all_rows),
"feature_completeness_dataset": feature_completeness(dataset_rows),
}
(ARTIFACTS / "substantivity_target_summary.json").write_text(json.dumps(summary, indent=2))
return summary
def summarize(values: np.ndarray) -> dict:
return {
"n": int(len(values)),
"min": float(np.min(values)),
"p25": float(np.percentile(values, 25)),
"median": float(np.median(values)),
"mean": float(np.mean(values)),
"p75": float(np.percentile(values, 75)),
"max": float(np.max(values)),
}
def feature_completeness(rows: list[dict]) -> dict:
return {
name: sum(r["features"].get(name) is not None for r in rows)
for name in FEATURE_NAMES
}
def matrix_from_rows(rows: list[dict]) -> tuple[np.ndarray, np.ndarray]:
y = np.array([r["target"]["log10_poucher_coefficient"] for r in rows], dtype=float)
raw = []
missing = []
for r in rows:
vals = [r["features"].get(name) for name in FEATURE_NAMES]
raw.append([np.nan if v is None else float(v) for v in vals])
missing.append([1.0 if v is None else 0.0 for v in vals])
x = np.array(raw, dtype=float)
miss = np.array(missing, dtype=float)
return np.concatenate([x, miss], axis=1), y
def impute_standardize(train_x: np.ndarray, test_x: np.ndarray) -> tuple[np.ndarray, np.ndarray]:
med = np.nanmedian(train_x, axis=0)
med = np.where(np.isnan(med), 0.0, med)
train = np.where(np.isnan(train_x), med, train_x)
test = np.where(np.isnan(test_x), med, test_x)
mean = train.mean(axis=0)
std = train.std(axis=0)
std = np.where(std == 0, 1.0, std)
return (train - mean) / std, (test - mean) / std
def r2_score(y: np.ndarray, pred: np.ndarray) -> float:
ss_res = float(np.sum((y - pred) ** 2))
ss_tot = float(np.sum((y - y.mean()) ** 2))
return 1.0 - ss_res / ss_tot if ss_tot else float("nan")
def folds(n: int, k: int = 5, seed: int = 20260710) -> list[np.ndarray]:
rng = np.random.default_rng(seed)
idx = np.arange(n)
rng.shuffle(idx)
return np.array_split(idx, k)
def ridge_cv(x: np.ndarray, y: np.ndarray) -> tuple[float, float]:
parts = folds(len(y))
pred = np.zeros_like(y)
alpha = 1.0
for test_idx in parts:
train_idx = np.setdiff1d(np.arange(len(y)), test_idx)
xt, xv = impute_standardize(x[train_idx], x[test_idx])
yt = y[train_idx]
xt1 = np.c_[np.ones(len(xt)), xt]
xv1 = np.c_[np.ones(len(xv)), xv]
reg = np.eye(xt1.shape[1]) * alpha
reg[0, 0] = 0.0
beta = np.linalg.pinv(xt1.T @ xt1 + reg) @ xt1.T @ yt
pred[test_idx] = xv1 @ beta
return r2_score(y, pred), float(np.corrcoef(y, pred)[0, 1])
def stump_gbm_cv(x: np.ndarray, y: np.ndarray) -> tuple[float, float]:
parts = folds(len(y))
pred = np.zeros_like(y)
for test_idx in parts:
train_idx = np.setdiff1d(np.arange(len(y)), test_idx)
xt, xv = impute_standardize(x[train_idx], x[test_idx])
yt = y[train_idx]
train_pred = np.full(len(yt), yt.mean())
test_pred = np.full(len(test_idx), yt.mean())
lr = 0.05
for _ in range(160):
residual = yt - train_pred
best = None
for j in range(xt.shape[1]):
thresholds = np.unique(np.quantile(xt[:, j], np.linspace(0.1, 0.9, 9)))
for threshold in thresholds:
left = xt[:, j] <= threshold
if left.sum() == 0 or (~left).sum() == 0:
continue
lv = residual[left].mean()
rv = residual[~left].mean()
update = np.where(left, lv, rv)
sse = float(np.sum((residual - update) ** 2))
if best is None or sse < best[0]:
best = (sse, j, threshold, lv, rv)
if best is None:
break
_, j, threshold, lv, rv = best
train_pred += lr * np.where(xt[:, j] <= threshold, lv, rv)
test_pred += lr * np.where(xv[:, j] <= threshold, lv, rv)
pred[test_idx] = test_pred
return r2_score(y, pred), float(np.corrcoef(y, pred)[0, 1])
def run_gate() -> dict:
rows = [json.loads(line) for line in open(DATA / "substantivity_targets_poucher.jsonl")]
feature_ready = [
r for r in rows
if any(r["features"].get(name) is not None for name in FEATURE_NAMES)
]
x, y = matrix_from_rows(feature_ready)
ridge_r2, ridge_corr = ridge_cv(x, y)
gbm_r2, gbm_corr = stump_gbm_cv(x, y)
result = {
"target": "log10_poucher_coefficient",
"n_total_poucher_targets": len(rows),
"n_with_any_molecular_feature": len(feature_ready),
"features": FEATURE_NAMES,
"missing_value_handling": "train-fold median imputation plus missing indicators",
"ridge_linear_5fold_r2": ridge_r2,
"ridge_linear_5fold_pearson": ridge_corr,
"stump_gbm_5fold_r2": gbm_r2,
"stump_gbm_5fold_pearson": gbm_corr,
"gate_interpretation": interpret(max(ridge_r2, gbm_r2)),
}
(ARTIFACTS / "substantivity_non_circularity_gate.json").write_text(json.dumps(result, indent=2))
return result
def interpret(best_r2: float) -> str:
if best_r2 >= 0.95:
return "STOP: target is essentially reconstructed by simple physicochemical features"
if best_r2 < 0.2:
return "STOP_AND_ESCALATE: simple features have very low signal for measured target"
return "PROCEED: moderate signal, not a closed-form feature formula"
def main() -> None:
summary = write_targets()
gate = run_gate()
print(json.dumps({"summary": summary, "gate": gate}, indent=2))
if __name__ == "__main__":
main()
|