Datasets:
Formats:
parquet
Languages:
English
Size:
10M - 100M
Tags:
biology
chemistry
drug-discovery
clinical-trials
protein-protein-interaction
gene-essentiality
License:
File size: 14,286 Bytes
6d1bbc7 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 | """LLM response evaluation for GE benchmark tasks GE-L1 through GE-L4.
Mirrors src/negbiodb_ppi/llm_eval.py structure:
- parse_*_answer/response: Extract structured data from raw LLM output
- evaluate_*: Compute task-specific metrics
- compute_all_ge_llm_metrics: Dispatch to task-specific evaluator
"""
from __future__ import annotations
import json
import logging
import re
from collections import Counter
import numpy as np
logger = logging.getLogger(__name__)
# Valid labels for each task
_L1_VALID = {"A", "B", "C", "D"}
_L4_VALID = {"tested", "untested"}
# ── GE-L1: 4-way classification ──────────────────────────────────────────
def parse_ge_l1_answer(raw: str) -> str | None:
"""Extract single-letter answer (A-D) from LLM response.
Tries:
1. Single letter on first line
2. Pattern like "Answer: B" or "(B)"
3. First A-D word token in full text
"""
if not raw:
return None
if raw.startswith("ERROR:"):
return None
raw = raw.strip()
first_line = raw.split("\n")[0].strip()
# Single letter
if first_line.upper() in _L1_VALID:
return first_line.upper()
# "Answer: B" or "(B)" or "B)"
m = re.search(r"(?:answer[:\s]*|[\(\[])\s*([A-D])\s*[)\]]?", first_line, re.IGNORECASE)
if m:
return m.group(1).upper()
# First A-D as a whole word/token (not embedded in words)
m = re.search(r"\b([A-D])\b", raw, re.IGNORECASE)
if m:
return m.group(1).upper()
return None
def evaluate_ge_l1(
predictions: list[str],
gold_labels: list[str],
) -> dict:
"""Compute GE-L1 metrics: accuracy, weighted_f1, macro_f1, MCC."""
from sklearn.metrics import (
accuracy_score,
f1_score,
matthews_corrcoef,
)
parsed = [parse_ge_l1_answer(p) for p in predictions]
valid_mask = [p is not None for p in parsed]
n_valid = sum(valid_mask)
n_total = len(predictions)
if n_valid == 0:
return {
"accuracy": 0.0, "weighted_f1": 0.0, "macro_f1": 0.0,
"mcc": 0.0, "valid_rate": 0.0, "n_valid": 0, "n_total": n_total,
}
y_pred = [p for p, v in zip(parsed, valid_mask) if v]
y_true = [g for g, v in zip(gold_labels, valid_mask) if v]
labels = sorted(_L1_VALID)
return {
"accuracy": accuracy_score(y_true, y_pred),
"weighted_f1": f1_score(y_true, y_pred, labels=labels, average="weighted", zero_division=0),
"macro_f1": f1_score(y_true, y_pred, labels=labels, average="macro", zero_division=0),
"mcc": matthews_corrcoef(y_true, y_pred),
"valid_rate": n_valid / n_total,
"n_valid": n_valid,
"n_total": n_total,
}
# ── GE-L2: Information extraction ────────────────────────────────────────
def parse_ge_l2_response(raw: str) -> dict | None:
"""Parse JSON response from GE-L2 extraction task.
Tries:
1. Direct JSON parse
2. Extract JSON from markdown code block
3. Regex for JSON object
"""
if not raw:
return None
raw = raw.strip()
# Try direct parse
try:
return json.loads(raw)
except json.JSONDecodeError:
pass
# Try markdown code block
m = re.search(r"```(?:json)?\s*([\s\S]*?)```", raw)
if m:
try:
return json.loads(m.group(1).strip())
except json.JSONDecodeError:
pass
# Try extracting JSON object
m = re.search(r"\{[\s\S]*\}", raw)
if m:
try:
return json.loads(m.group(0))
except json.JSONDecodeError:
pass
return None
def _get_gene_list(extraction: dict) -> list[dict]:
"""Extract gene list from either 'genes' or legacy 'essentiality_findings' key."""
return extraction.get("genes") or extraction.get("essentiality_findings") or []
def evaluate_ge_l2(
predictions: list[str],
gold_extractions: list[dict],
) -> dict:
"""Compute GE-L2 metrics: parse_rate, schema_compliance, field_f1,
essentiality_accuracy.
Normalised gold schema uses 'genes' key (legacy 'essentiality_findings'
is also accepted).
"""
n_total = len(predictions)
parsed = [parse_ge_l2_response(p) for p in predictions]
n_parsed = sum(1 for p in parsed if p is not None)
required_fields = {"genes", "total_genes_mentioned", "screen_type"}
n_compliant = 0
field_scores: list[float] = []
essentiality_correct = 0
essentiality_total = 0
for pred, gold in zip(parsed, gold_extractions):
if pred is None:
field_scores.append(0.0)
continue
# Schema compliance — normalise pred key (accept either)
pred_norm_keys = set(pred.keys())
if "essentiality_findings" in pred_norm_keys and "genes" not in pred_norm_keys:
pred_norm_keys = (pred_norm_keys - {"essentiality_findings"}) | {"genes"}
if required_fields.issubset(pred_norm_keys):
n_compliant += 1
# Field-level F1 against normalised gold keys
gold_norm = {}
if gold:
gold_norm = dict(gold)
if "essentiality_findings" in gold_norm and "genes" not in gold_norm:
gold_norm["genes"] = gold_norm.pop("essentiality_findings")
if "total_gene_count" in gold_norm and "total_genes_mentioned" not in gold_norm:
gold_norm["total_genes_mentioned"] = gold_norm.pop("total_gene_count")
gold_fields = set(gold_norm.keys())
pred_fields = pred_norm_keys
if not gold_fields and not pred_fields:
field_scores.append(1.0)
elif not gold_fields or not pred_fields:
field_scores.append(0.0)
else:
tp = len(gold_fields & pred_fields)
precision = tp / max(len(pred_fields), 1)
recall = tp / max(len(gold_fields), 1)
f1 = 2 * precision * recall / max(precision + recall, 1e-10)
field_scores.append(f1)
# Essentiality accuracy — compare per-gene essentiality_status
pred_genes = _get_gene_list(pred)
gold_genes = _get_gene_list(gold_norm)
if gold_genes:
# Build gold lookup: gene_name → essentiality_status
gold_lookup = {
g.get("gene_name", "").upper(): g.get("essentiality_status", "")
for g in gold_genes
if isinstance(g, dict)
}
for pg in pred_genes:
if not isinstance(pg, dict):
continue
gname = pg.get("gene_name", "").upper()
pred_status = pg.get("essentiality_status", "").lower().strip()
if gname in gold_lookup:
essentiality_total += 1
gold_status = gold_lookup[gname].lower().strip()
if pred_status == gold_status:
essentiality_correct += 1
return {
"parse_rate": n_parsed / max(n_total, 1),
"schema_compliance": n_compliant / max(n_total, 1),
"field_f1": float(np.mean(field_scores)) if field_scores else 0.0,
"essentiality_accuracy": (
essentiality_correct / essentiality_total
if essentiality_total > 0
else 0.0
),
"essentiality_n": essentiality_total,
"n_parsed": n_parsed,
"n_total": n_total,
}
# ── GE-L3: Reasoning evaluation ──────────────────────────────────────────
GE_L3_JUDGE_PROMPT = (
"You are evaluating a scientific explanation for why a gene is "
"NON-ESSENTIAL in a specific cancer cell line.\n\n"
"GENE-CELL LINE CONTEXT:\n{context_text}\n\n"
"RESPONSE TO EVALUATE:\n{response_text}\n\n"
"Score the response on 4 dimensions (1-5 each):\n"
"1. biological_plausibility: Are biological reasons (gene function, "
"pathway role, tissue context) scientifically sound?\n"
"2. pathway_reasoning: Does the explanation address pathway redundancy, "
"compensatory mechanisms, or lineage-specific dispensability?\n"
"3. context_specificity: Are claims specific to this gene in this cell "
"line/lineage or generic?\n"
"4. mechanistic_depth: Are multiple relevant factors considered "
"(expression, copy number, mutation status, tissue of origin)?\n\n"
'Return ONLY a JSON object: {{"biological_plausibility": N, '
'"pathway_reasoning": N, "context_specificity": N, "mechanistic_depth": N}}'
)
def parse_ge_l3_judge_scores(raw: str) -> dict[str, float] | None:
"""Parse judge model scores for GE-L3 reasoning evaluation.
Expected format (from judge prompt):
biological_plausibility: 4
pathway_reasoning: 3
context_specificity: 5
mechanistic_depth: 4
"""
if not raw:
return None
dimensions = [
"biological_plausibility", "pathway_reasoning",
"context_specificity", "mechanistic_depth",
]
scores = {}
for dim in dimensions:
pattern = rf"{dim}\s*[:=]\s*(\d(?:\.\d)?)"
m = re.search(pattern, raw, re.IGNORECASE)
if m:
scores[dim] = float(m.group(1))
# Also try JSON format (with or without markdown code fence)
if not scores:
json_str = raw
m = re.search(r"```(?:json)?\s*([\s\S]*?)```", raw)
if m:
json_str = m.group(1).strip()
try:
data = json.loads(json_str)
if isinstance(data, dict):
for dim in dimensions:
if dim in data:
scores[dim] = float(data[dim])
except (json.JSONDecodeError, TypeError, ValueError):
pass
return scores if scores else None
def evaluate_ge_l3(
judge_outputs: list[str],
) -> dict:
"""Compute GE-L3 metrics from judge model outputs."""
dimensions = [
"biological_plausibility", "pathway_reasoning",
"context_specificity", "mechanistic_depth",
]
all_scores: dict[str, list[float]] = {d: [] for d in dimensions}
n_parsed = 0
for output in judge_outputs:
scores = parse_ge_l3_judge_scores(output)
if scores is None:
continue
n_parsed += 1
for dim in dimensions:
if dim in scores:
all_scores[dim].append(scores[dim])
result: dict = {"n_parsed": n_parsed, "n_total": len(judge_outputs)}
for dim in dimensions:
vals = all_scores[dim]
result[f"{dim}_mean"] = float(np.mean(vals)) if vals else 0.0
result[f"{dim}_std"] = float(np.std(vals)) if vals else 0.0
# Overall mean
all_vals = [v for vals in all_scores.values() for v in vals]
result["overall_mean"] = float(np.mean(all_vals)) if all_vals else 0.0
result["overall_std"] = float(np.std(all_vals)) if all_vals else 0.0
return result
# ── GE-L4: Tested/Untested discrimination ────────────────────────────────
def parse_ge_l4_answer(raw: str) -> str | None:
"""Extract 'tested' or 'untested' from GE-L4 response."""
if not raw:
return None
if raw.startswith("ERROR:"):
return None
raw = raw.strip().lower()
first_line = raw.split("\n")[0].strip()
if first_line in _L4_VALID:
return first_line
# Search for keywords
if "untested" in first_line:
return "untested"
if "tested" in first_line:
return "tested"
# Try full text
if "untested" in raw:
return "untested"
if "tested" in raw:
return "tested"
return None
def evaluate_ge_l4(
predictions: list[str],
gold_labels: list[str],
) -> dict:
"""Compute GE-L4 metrics: accuracy, MCC, temporal contamination gap."""
from sklearn.metrics import accuracy_score, matthews_corrcoef
parsed = [parse_ge_l4_answer(p) for p in predictions]
valid_mask = [p is not None for p in parsed]
n_valid = sum(valid_mask)
n_total = len(predictions)
if n_valid == 0:
return {
"accuracy": 0.0, "mcc": 0.0, "valid_rate": 0.0,
"n_valid": 0, "n_total": n_total,
}
y_pred = [p for p, v in zip(parsed, valid_mask) if v]
y_true = [g for g, v in zip(gold_labels, valid_mask) if v]
return {
"accuracy": accuracy_score(y_true, y_pred),
"mcc": matthews_corrcoef(y_true, y_pred),
"valid_rate": n_valid / n_total,
"n_valid": n_valid,
"n_total": n_total,
"prediction_distribution": dict(Counter(y_pred)),
"gold_distribution": dict(Counter(y_true)),
}
# ── Dispatch ──────────────────────────────────────────────────────────────
def compute_all_ge_llm_metrics(
task: str,
predictions: list[str],
gold_data: list,
) -> dict:
"""Dispatch to task-specific evaluator.
Args:
task: 'ge-l1', 'ge-l2', 'ge-l3', or 'ge-l4'
predictions: Raw LLM responses
gold_data: Gold labels/extractions/judge outputs — may be either
plain values (str/dict) or full record dicts with a 'gold_answer'
or 'gold_extraction' key.
Returns:
Metrics dict
"""
# Normalise gold_data: if records are dicts with dataset keys, extract
# the relevant field for each task.
def _extract(records, field):
if records and isinstance(records[0], dict) and field in records[0]:
return [r.get(field) for r in records]
return records
if task == "ge-l1":
return evaluate_ge_l1(predictions, _extract(gold_data, "gold_answer"))
elif task == "ge-l2":
return evaluate_ge_l2(predictions, _extract(gold_data, "gold_extraction"))
elif task == "ge-l3":
return evaluate_ge_l3(predictions)
elif task == "ge-l4":
return evaluate_ge_l4(predictions, _extract(gold_data, "gold_answer"))
else:
raise ValueError(f"Unknown task: {task}")
|