pubchem-faiss-library / code /spec_rag /cot_system2.py
YinkaiW's picture
Upload folder using huggingface_hub
db32e07 verified
Raw
History Blame Contribute Delete
14.1 kB
"""
System 2 CoT: Deductive Spectral Reasoning.
Generates structured, step-by-step reasoning traces deriving structure from spectrum.
Key components:
1. Precursor Analysis: ExactMass, Nitrogen Rule, DoU.
2. Neutral Loss Analysis: Delta mass attribution (Causal).
3. Fragment Analysis: Substructure validation (Anti-Hallucination).
4. Scaffold Inference: Aglycone/Core deduction.
5. Assembly: Coherent narrative.
"""
from __future__ import annotations
import json
import math
import re
from typing import Any, Optional, List, Dict, Tuple
# --- KNOWLEDGE BASES ---
# (Fragment Name, SMARTS, Approx m/z, Formula/Cation)
DIAGNOSTIC_FRAGMENTS = [
("Indole", "[nH]1cccc2ccccc12", 130, "C9H8N+"),
("3-alkyl-indole", "c1ccc2c(c1)[nH]cc2C", 130, "C9H8N+"),
("Tropylium", "[CH2]c1ccccc1", 91, "C7H7+"),
("Phenyl cation", "c1ccccc1", 77, "C6H5+"),
("Phenol", "c1ccc(O)cc1", 94, "C6H6O"),
("Aniline", "c1ccc(N)cc1", 93, "C6H7N"),
("Pyridine", "n1ccccc1", 79, "C5H5N+"),
("Imidazole", "c1cncn1", 68, "C3H4N2"),
("Quinoline/Isoquinoline", "c1ccc2ncccc2c1", 129, "C9H7N+"),
("Naphthalene", "c1ccc2ccccc2c1", 128, "C10H8+"),
("Acetyl", "CC(=O)", 43, "C2H3O+"),
("Methoxy", "CO", 31, "CH3O+"),
("Carboxyl", "C(=O)O", 45, "CHO2+"),
("Glucuronic Acid", "OC1C(O)C(O)C(O)C(O)O1", 176, "C6H9O6+"),
("Rhamnose", "CC1OC(O)C(O)C(O)C1O", 147, "C6H11O4+"),
("Hexose (Glucose/Galactose)", "OCC1OC(O)C(O)C(O)C1O", 163, "C6H11O5+"),
("Quercetin aglycone", "c1c(O)cc(O)c2c1(=O)c(O)c(c1ccc(O)c(O)c1)o2", 303, "C15H11O7+"),
("Kaempferol aglycone", "c1c(O)cc(O)c2c1(=O)c(O)c(c1ccc(O)cc1)o2", 287, "C15H11O6+"),
("Apigenin aglycone", "c1c(O)cc(O)c2c1(=O)cc(c1ccc(O)cc1)o2", 271, "C15H11O5+"),
("Luteolin aglycone", "c1c(O)cc(O)c2c1(=O)cc(c1ccc(O)c(O)c1)o2", 287, "C15H11O6+"),
("Coumarin", "O=C1C=Cc2ccccc2O1", 147, "C9H7O2+"),
("Ferulic acid moiety", "COc1cc(C=CC(=O)O)ccc1O", 195, "C10H11O4+"),
]
NEUTRAL_LOSSES_MAP = {
162.05: "hexose moiety (e.g., glucose/galactose, -162 Da)",
146.06: "deoxyhexose moiety (e.g., rhamnose, -146 Da)",
132.04: "pentose moiety (e.g., xylose/arabinose, -132 Da)",
176.03: "glucuronic acid moiety (-176 Da)",
308.11: "rutinose (rhamnose-glucose, -308 Da)",
18.01: "water (H2O, -18 Da)",
28.00: "CO (-28 Da)",
44.00: "CO2 (-44 Da)",
15.02: "methyl group (-15 Da)",
17.03: "ammonia (NH3, -17 Da)",
31.02: "methoxy group (-31 Da)",
32.03: "methanol (-32 Da)",
42.01: "acetyl group (ketene loss, -42 Da)",
46.01: "H2O + CO / formic acid (-46 Da)",
56.03: "C4H8 / retro-Diels-Alder fragment",
79.96: "sulfate / phosphate loss (approx 80 Da)",
# Halogen losses
36.00: "HCl (hydrogen chloride, -36 Da)",
38.00: "HCl (isotope, -38 Da)",
80.91: "HBr (hydrogen bromide, -81 Da)",
78.92: "HBr (isotope, -79 Da)",
127.90: "HI (hydrogen iodide, -128 Da)",
19.99: "HF (hydrogen fluoride, -20 Da)"
}
TOLERANCE = 0.5
# --- HELPER FUNCTIONS ---
def match_substructures(smiles: str) -> List[str]:
"""Return list of fragment names present in the molecule."""
try:
from rdkit import Chem
mol = Chem.MolFromSmiles(smiles)
if mol is None: return []
found = []
for name, smarts, _, _ in DIAGNOSTIC_FRAGMENTS:
try:
pat = Chem.MolFromSmarts(smarts)
if pat and mol.HasSubstructMatch(pat):
found.append(name)
except:
continue
return found
except ImportError:
return []
def calc_dou(formula_str: str) -> float:
"""
Calculate Degree of Unsaturation.
DoU = C - H/2 - X/2 + N/2 + 1
(Where X = F, Cl, Br, I)
"""
if not formula_str: return 0.0
atoms = {}
# Parse elements and counts: "C16H17NO4Cl" -> [('C', '16'), ('H', '17')...]
for match in re.finditer(r"([A-Z][a-z]?)(\d*)", formula_str):
elem = match.group(1)
count = int(match.group(2)) if match.group(2) else 1
atoms[elem] = atoms.get(elem, 0) + count
c = atoms.get('C', 0)
h = atoms.get('H', 0)
n = atoms.get('N', 0)
p = atoms.get('P', 0) # P often treated similar to N in simple heuristics
# Halogens count as H for saturation
x = atoms.get('F', 0) + atoms.get('Cl', 0) + atoms.get('Br', 0) + atoms.get('I', 0)
# DoU = C + 1 - (H + X)/2 + (N + P)/2
return c + 1 - (h + x)/2 + (n + p)/2
def precursor_interpretation_deductive(
found_in_spectrum: bool,
mz: float,
exact_mw: float,
formula: str,
adduct: str = "[M+H]+"
) -> List[str]:
"""Generate deductive text about precursor, carefully distinguishing Neutral vs Ion."""
steps = []
nominal_mass = round(exact_mw)
# 1. Mass Match statement
if found_in_spectrum:
delta = mz - exact_mw
steps.append(f"The precursor ion is observed at m/z {mz:.4f}, consistent with an {adduct} adduct of a neutral molecule with mass {exact_mw:.4f} Da.")
else:
steps.append(f"The precursor ion is not observed in the spectrum. The theoretical mass is {exact_mw:.4f} Da.")
# 2. Nitrogen Rule & Complexity (DoU)
try:
dou = calc_dou(formula)
is_odd_mass = (nominal_mass % 2 != 0)
# Explicitly discuss the NEUTRAL molecule first
steps.append(f"The neutral nominal mass is {nominal_mass} Da.")
if "N" in formula:
steps.append(f"Based on the Nitrogen Rule, this {'odd' if is_odd_mass else 'even'} neutral mass indicates an {'odd' if is_odd_mass else 'even'} number of nitrogen atoms.")
else:
steps.append(f"The {'even' if not is_odd_mass else 'odd'} neutral mass is consistent with the absence of nitrogen.")
steps.append(f"The calculated Degree of Unsaturation (DoU) is {dou:.1f}, suggesting significant structural complexity (rings/double bonds).")
except Exception as e:
pass
return steps
def neutral_loss_analysis(
precursor_mz: float,
peaks: List[Tuple[float, float]],
formula: Optional[str]
) -> List[str]:
"""Identify direct neutral losses from precursor with Formula Validation."""
steps = []
# Parse formula to check for allowed losses
# Simple count of atoms in the parent formula
atom_counts = {}
if formula:
import re
for match in re.finditer(r"([A-Z][a-z]?)(\d*)", formula):
elem = match.group(1)
count = int(match.group(2)) if match.group(2) else 1
atom_counts[elem] = atom_counts.get(elem, 0) + count
# Helper: Can this formula lose this group?
def is_loss_allowed(loss_name):
if not formula: return True # Fallback if no formula
loss_lower = loss_name.lower()
# Halogen Checks
if "chloride" in loss_lower or "hcl" in loss_lower:
if atom_counts.get("Cl", 0) < 1: return False
if "bromide" in loss_lower or "hbr" in loss_lower:
if atom_counts.get("Br", 0) < 1: return False
if "iodide" in loss_lower or "hi" in loss_lower:
if atom_counts.get("I", 0) < 1: return False
if "fluoride" in loss_lower or "hf" in loss_lower:
if atom_counts.get("F", 0) < 1: return False
# Sugar Checks (Sugars are O-rich)
# Pentose/Hexose/Rhamnose usually require ~3-4+ Oxygens
if any(x in loss_lower for x in ["hexose", "pentose", "rhamnose", "glucuronic", "rutinose"]):
if atom_counts.get("O", 0) < 3: return False
# Ammonia check
if "ammonia" in loss_lower or "nh3" in loss_lower:
if atom_counts.get("N", 0) < 1: return False
return True
# Analysis Loop
sorted_peaks = sorted(peaks, key=lambda x: x[1], reverse=True)[:10]
found_losses = []
for mz, _int in sorted_peaks:
if mz >= precursor_mz - 1.0: continue
loss = precursor_mz - mz
best_match = None
min_diff = 999.0
for ref_loss, name in NEUTRAL_LOSSES_MAP.items():
diff = abs(loss - ref_loss)
# Check tolerance AND chemical feasibility
if diff < TOLERANCE and diff < min_diff:
if is_loss_allowed(name):
min_diff = diff
best_match = name
if best_match:
steps.append(f"A neutral loss of {loss:.2f} Da (m/z {precursor_mz:.2f} -> {mz:.2f}) corresponds to the loss of a {best_match}.")
found_losses.append(mz)
return steps, found_losses
# --- MAIN BUILDER ---
def build_system2_structured(
smiles: str,
peaks: List[List[float]],
precursor_mz: Optional[float] = None,
formula: Optional[str] = None,
max_peaks: int = 20,
) -> Dict[str, Any]:
"""
Build a structured dictionary containing the reasoning steps.
"""
from rdkit import Chem
from rdkit.Chem import Descriptors
from rdkit.Chem.rdMolDescriptors import CalcMolFormula
mol = Chem.MolFromSmiles(smiles)
if not mol:
return {"error": "Invalid SMILES"}
exact_mw = Descriptors.ExactMolWt(mol)
calc_formula = formula or CalcMolFormula(mol)
# Filter and Sort Peaks
max_int = max([p[1] for p in peaks]) if peaks else 1.0
valid_peaks = [(p[0], p[1]) for p in peaks if p[1]/max_int > 0.01]
valid_peaks.sort(key=lambda x: x[1], reverse=True)
top_peaks = valid_peaks[:max_peaks]
base_peak = top_peaks[0] if top_peaks else (0,0)
# 1. Precursor Identification
precursor_info = {
"found": False,
"mz": 0.0,
"adduct": "",
"deduction": []
}
# Check spectrum for [M+H]+ or [M+Na]+
found_in_spectrum = False
observed_mz = 0.0
if any(abs(mz - (exact_mw + 1.0078)) < TOLERANCE for mz, _ in top_peaks):
observed_mz = next(mz for mz, _ in top_peaks if abs(mz - (exact_mw + 1.0078)) < TOLERANCE)
precursor_info["mz"] = observed_mz
precursor_info["adduct"] = "[M+H]+"
found_in_spectrum = True
elif any(abs(mz - (exact_mw + 22.989)) < TOLERANCE for mz, _ in top_peaks):
observed_mz = next(mz for mz, _ in top_peaks if abs(mz - (exact_mw + 22.989)) < TOLERANCE)
precursor_info["mz"] = observed_mz
precursor_info["adduct"] = "[M+Na]+"
found_in_spectrum = True
# If not found in spectrum, use metadata or calculation, but flag as not found
if not found_in_spectrum:
if precursor_mz and precursor_mz > 0:
precursor_info["mz"] = precursor_mz
precursor_info["adduct"] = "[M+H]+" # Assume protonated if provided
else:
precursor_info["mz"] = exact_mw + 1.0078
precursor_info["adduct"] = "[M+H]+"
precursor_info["found"] = found_in_spectrum
# Generate text for precursor
precursor_info["deduction"] = precursor_interpretation_deductive(
found_in_spectrum,
precursor_info["mz"],
exact_mw,
calc_formula,
precursor_info["adduct"]
)
# 2. Neutral Loss Analysis
loss_steps, loss_mzs = neutral_loss_analysis(precursor_info["mz"], top_peaks, calc_formula)
# 3. Fragment Analysis (Strict Causal)
present_substructures = match_substructures(smiles)
fragment_steps = []
for mz, _int in top_peaks:
# Check if peak matches a diagnostic fragment
for name, _smarts, hint_mz, hint_formula in DIAGNOSTIC_FRAGMENTS:
if abs(mz - hint_mz) < 1.0:
# CRITICAL: Only claim it if substructure is in molecule
if name in present_substructures:
fragment_steps.append(f"The peak at m/z {mz:.2f} is diagnostic for {name}" + (f" ({hint_formula})." if hint_formula else "."))
break
fragment_steps = list(dict.fromkeys(fragment_steps))
# 4. Scaffold/Base Peak
scaffold_text = ""
if base_peak[0] > 0:
if found_in_spectrum and abs(base_peak[0] - precursor_info["mz"]) < 1.0:
scaffold_text = "The base peak corresponds to the intact precursor, suggesting a stable molecular ion."
elif loss_mzs and abs(base_peak[0] - loss_mzs[-1]) < 1.0:
scaffold_text = "The base peak represents the core scaffold after the loss of labile groups."
else:
scaffold_text = f"The base peak at m/z {base_peak[0]:.2f} represents the most stable fragment ion."
return {
"precursor": precursor_info,
"losses": loss_steps,
"fragments": fragment_steps,
"scaffold": scaffold_text
}
def format_structured_cot_as_text(structure: Dict[str, Any]) -> str:
"""Convert structured dict to narrative string."""
parts = []
# Precursor
if structure.get("error"):
return structure["error"]
prec = structure["precursor"]
parts.extend(prec["deduction"])
# Losses
if structure["losses"]:
parts.append("Fragmentation analysis reveals characteristic neutral losses:")
parts.extend(structure["losses"])
else:
parts.append("No common neutral losses were clearly identified.")
# Fragments
if structure["fragments"]:
parts.append("Diagnostic fragment ions confirm the presence of specific substructures:")
parts.extend(structure["fragments"])
# Scaffold
if structure["scaffold"]:
parts.append(structure["scaffold"])
return " ".join(parts)
def build_system2_thought(
smiles: str,
peaks: list[list[float]],
precursor_mz: Optional[float] = None,
formula: Optional[str] = None,
max_peaks: int = 20,
output_format: str = "text",
) -> str:
"""
Entry point for generating CoT.
"""
struct = build_system2_structured(smiles, peaks, precursor_mz, formula, max_peaks)
if output_format == "json":
return json.dumps(struct, indent=2)
return format_structured_cot_as_text(struct)