pubchem-faiss-library / code /scripts /generate_synthetic_thoughts.py
YinkaiW's picture
Upload folder using huggingface_hub
db32e07 verified
Raw
History Blame Contribute Delete
5.94 kB
#!/usr/bin/env python
"""
Generate synthetic "Ground Truth Thought" for CoT training.
Two modes:
- **System 2 (default when formula/precursor_mz available):** Deductive trace:
precursor analysis (formula, DoU, Nitrogen rule) → fragment logic (peak → substructure)
→ neutral loss analysis → core reconstruction → assembly. Uses RDKit for
substructure matching and formula.
- **Fallback:** Short peak-based hints (m/z 91 → tropylium, etc.) when data or RDKit missing.
"""
from __future__ import annotations
import argparse
import json
import sys
from pathlib import Path
ROOT = Path(__file__).resolve().parents[1]
if str(ROOT) not in sys.path:
sys.path.insert(0, str(ROOT))
# Fallback fragment hints (no RDKit required)
FRAGMENT_HINTS = [
(91, "m/z 91 (tropylium) suggests benzyl or aromatic ring"),
(77, "m/z 77 suggests benzene ring"),
(65, "m/z 65 suggests cyclopentadienyl or aromatic fragment"),
(43, "m/z 43 often indicates acetyl or C3H7+"),
(57, "m/z 57 suggests butyl or C4H9+"),
(41, "m/z 41 suggests allyl or C3H5+"),
(130, "m/z 130 is diagnostic for indole / 3-alkyl-indole cation"),
(18, "loss of 18 Da suggests water (e.g. -OH)"),
(17, "loss of 17 Da suggests ammonia or -OH"),
(28, "loss of 28 Da suggests CO or C2H4"),
(44, "loss of 44 Da suggests CO2"),
(15, "loss of 15 Da suggests methyl"),
]
def get_top_peaks(peaks: list, top_k: int = 10) -> list[tuple[float, float]]:
"""Return top-k peaks by intensity."""
sorted_peaks = sorted(peaks, key=lambda x: float(x[1]), reverse=True)
return [(float(p[0]), float(p[1])) for p in sorted_peaks[:top_k]]
def generate_thought_fallback(peaks: list, precursor_mz: float | None = None) -> str:
"""Simple peak-based hints when System 2 is not used."""
top = get_top_peaks(peaks, top_k=8)
parts = []
seen = set()
for mz, _ in top:
mz_round = round(mz)
for frag_mz, hint in FRAGMENT_HINTS:
if abs(mz_round - frag_mz) <= 2 and frag_mz not in seen:
parts.append(hint)
seen.add(frag_mz)
if precursor_mz is not None and precursor_mz > 0:
parts.insert(0, f"Precursor m/z {precursor_mz:.1f}.")
if not parts:
parts = [f"Key peaks at m/z {', '.join(f'{m:.0f}' for m, _ in top[:5])}."]
return " ".join(parts)
def _load_cot_system2():
"""Load cot_system2 module without importing full spec_rag (avoids numpy etc)."""
import importlib.util
p = ROOT / "spec_rag" / "cot_system2.py"
spec = importlib.util.spec_from_file_location("cot_system2", p)
mod = importlib.util.module_from_spec(spec)
spec.loader.exec_module(mod)
return mod
def generate_thought_for_record(
obj: dict, use_system2: bool = True, _system2_mod=None, output_format: str = "text"
) -> str:
"""
Generate thought for one JSONL record. Prefer System 2 when formula or
precursor_mz present and use_system2 is True.
"""
peaks = obj.get("peaks", [])
smiles = obj.get("smiles", "")
formula = obj.get("formula") or None
precursor_mz = obj.get("precursor_mz")
if precursor_mz is not None:
try:
precursor_mz = float(precursor_mz)
except (TypeError, ValueError):
precursor_mz = None
if use_system2 and (formula or precursor_mz or smiles):
try:
if _system2_mod is None:
_system2_mod = _load_cot_system2()
thought = _system2_mod.build_system2_thought(
smiles=smiles,
peaks=peaks if peaks else [[0, 0]],
precursor_mz=precursor_mz,
formula=formula,
max_peaks=10,
output_format=output_format,
)
if thought and len(thought.strip()) > 50:
return thought.strip()
print(f"Short thought ({len(thought)}): {thought}")
except Exception as e:
print(f"System 2 failed: {e}")
pass
if peaks:
return generate_thought_fallback(peaks, precursor_mz)
return "Analyzing spectrum for structural features."
def parse_args() -> argparse.Namespace:
p = argparse.ArgumentParser(
description="Generate synthetic thoughts for CoT (System 2 or fallback)."
)
p.add_argument("--input-jsonl", required=True, help="JSONL with peaks, smiles, optional formula, precursor_mz")
p.add_argument("--output-jsonl", required=True, help="Same + 'thought' field")
p.add_argument("--no-system2", action="store_true", help="Use only simple peak hints, no RDKit/System 2")
p.add_argument("--format", choices=("text", "json"), default="text", help="System 2 thought format: sectioned text or JSON")
p.add_argument("--precursor-col", default="precursor_mz", help="Column name for precursor m/z if any")
return p.parse_args()
def main() -> None:
args = parse_args()
use_system2 = not args.no_system2
try:
system2_mod = _load_cot_system2() if use_system2 else None
except Exception as e:
print(f"Failed to load System 2 module: {e}")
system2_mod = None
use_system2 = False
out_lines = []
with open(args.input_jsonl) as f:
for line in f:
line = line.strip()
if not line:
continue
obj = json.loads(line)
thought = generate_thought_for_record(
obj,
use_system2=use_system2,
_system2_mod=system2_mod,
output_format=getattr(args, "format", "text"),
)
obj["thought"] = thought
out_lines.append(json.dumps(obj) + "\n")
Path(args.output_jsonl).parent.mkdir(parents=True, exist_ok=True)
with open(args.output_jsonl, "w") as f:
f.writelines(out_lines)
print(f"Wrote {len(out_lines)} records to {args.output_jsonl}")
if __name__ == "__main__":
main()