EpiADR-Net / utils.py
ADjayantan
Fix GitHub Actions CI: Resolved ruff linter check and updated CI workflow flags
5d4afe2
Raw
History Blame Contribute Delete
14.4 kB
import random
import re
from typing import Any
import numpy as np
import torch
# Suppress RDKit C++ stderr noise
try:
from rdkit import RDLogger
RDLogger.DisableLog('rdApp.*')
except Exception:
pass
# ─────────────────────────────────────────────────────────────────
# Atom & Bond Lookup Tables
# ─────────────────────────────────────────────────────────────────
ATOM_NUMBERS = {
'H': 1, 'C': 6, 'N': 7, 'O': 8, 'F': 9,
'P': 15, 'S': 16, 'Cl': 17, 'Br': 35, 'I': 53,
'Si': 14, 'B': 5, 'Se': 34, 'Na': 11, 'K': 19, 'Pt': 78,
}
ATOM_MASSES = {
'H': 1.008, 'C': 12.011, 'N': 14.007, 'O': 15.999, 'F': 18.998,
'P': 30.974, 'S': 32.06, 'Cl': 35.45, 'Br': 79.904, 'I': 126.90,
'Si': 28.085, 'B': 10.811, 'Se': 78.971, 'Na': 22.990, 'K': 39.098, 'Pt': 195.08,
}
ELECTRONEGATIVITY = {
'H': 2.20, 'C': 2.55, 'N': 3.04, 'O': 3.44, 'F': 3.98,
'P': 2.19, 'S': 2.58, 'Cl': 3.16, 'Br': 2.96, 'I': 2.66,
'Si': 1.90, 'B': 2.04, 'Se': 2.55, 'Na': 0.93, 'K': 0.82, 'Pt': 2.28,
}
VDW_RADIUS = {
'H': 1.20, 'C': 1.70, 'N': 1.55, 'O': 1.52, 'F': 1.47,
'P': 1.80, 'S': 1.80, 'Cl': 1.75, 'Br': 1.85, 'I': 1.98,
'Si': 2.10, 'B': 1.92, 'Se': 1.90, 'Na': 2.27, 'K': 2.75, 'Pt': 1.75,
}
HALOGENS = {'F', 'Cl', 'Br', 'I'}
HETEROATOMS = {'P', 'S', 'Se', 'B', 'Si', 'Pt'}
def smiles_to_graph(smiles: str) -> tuple[torch.Tensor, torch.Tensor, list[str]]:
"""
Converts a SMILES string into a rich molecular graph with 24-dimensional atom features x_i in R^24
and edge_index E in R^(2 x M).
Node Features x_i in R^24:
[0] Atomic Number / 100.0
[1] Atomic Mass / 200.0
[2] Degree / 6.0
[3] Formal Charge clamped [-2, +2] / 2.0
[4] Hybridization code (1=sp, 2=sp2, 3=sp3, 4=sp3d, 5=sp3d2, 0=other) / 5.0
[5] Is Aromatic (binary)
[6] Implicit Valence / 6.0
[7] Is In Ring (binary)
[8] Is Stereocenter / Chiral Flag (binary)
[9] Total Hydrogen Count / 4.0
[10-13] Element Group One-Hot: [C/N/O, Halogens, Heteroatoms, Other]
[14] Pauling Electronegativity / 4.0
[15] vdW Radius / 3.0
[16] Is Ring Size 3 (binary)
[17] Is Ring Size 4 (binary)
[18] Is Ring Size 5 (binary)
[19] Is Ring Size 6 (binary)
[20] Is Ring Size 7 (binary)
[21] Is Ring Size 8 (binary)
[22] Is Conjugated Atom (binary)
[23] Gasteiger Charge Proxy (normalized [-1, +1])
"""
def _atom_to_features(symbol, num, deg, chg, hyb, aromatic,
imp_val, in_ring, mass, chiral, h_count,
ring_sizes, conjugated, charge_proxy) -> list[float]:
group = [0.0, 0.0, 0.0, 0.0]
if symbol in {'C', 'N', 'O'}:
group[0] = 1.0
elif symbol in HALOGENS:
group[1] = 1.0
elif symbol in HETEROATOMS:
group[2] = 1.0
else:
group[3] = 1.0
en = ELECTRONEGATIVITY.get(symbol, 2.0) / 4.0
vdw = VDW_RADIUS.get(symbol, 1.7) / 3.0
r3 = 1.0 if 3 in ring_sizes else 0.0
r4 = 1.0 if 4 in ring_sizes else 0.0
r5 = 1.0 if 5 in ring_sizes else 0.0
r6 = 1.0 if 6 in ring_sizes else 0.0
r7 = 1.0 if 7 in ring_sizes else 0.0
r8 = 1.0 if 8 in ring_sizes else 0.0
return [
float(num) / 100.0, # [0]
float(mass) / 200.0, # [1]
min(float(deg), 6.0) / 6.0, # [2]
max(-2.0, min(2.0, float(chg))) / 2.0, # [3]
float(hyb) / 5.0, # [4]
float(aromatic), # [5]
min(float(imp_val), 6.0) / 6.0, # [6]
float(in_ring), # [7]
float(chiral), # [8]
min(float(h_count), 4.0) / 4.0, # [9]
*group, # [10-13]
en, # [14]
vdw, # [15]
r3, r4, r5, r6, r7, r8, # [16-21]
float(conjugated), # [22]
max(-1.0, min(1.0, float(charge_proxy))), # [23]
]
try:
from rdkit import Chem
mol = Chem.MolFromSmiles(smiles)
if mol is not None:
Chem.SanitizeMol(mol)
atoms, atom_symbols = [], []
for atom in mol.GetAtoms():
symbol = atom.GetSymbol()
num = atom.GetAtomicNum()
deg = atom.GetDegree()
chg = atom.GetFormalCharge()
hyb_val = int(atom.GetHybridization())
hyb = {2: 1, 3: 2, 4: 3, 5: 4, 6: 5}.get(hyb_val, 0)
aromatic = 1.0 if atom.GetIsAromatic() else 0.0
try:
imp_val = float(atom.GetValence(Chem.ValenceType.IMPLICIT))
except Exception:
imp_val = float(atom.GetImplicitValence())
in_ring = 1.0 if atom.IsInRing() else 0.0
mass = float(atom.GetMass())
chiral = 1.0 if (atom.HasProp('_ChiralityPossible') or atom.GetChiralTag() != Chem.ChiralType.CHI_UNSPECIFIED) else 0.0
h_count = float(atom.GetTotalNumHs())
ring_sizes = [size for size in range(3, 9) if atom.IsInRingSize(size)]
conjugated = 1.0 if atom.GetIsAromatic() or any(b.GetIsConjugated() for b in atom.GetBonds()) else 0.0
charge_proxy = float(chg) + (0.1 if symbol in {'N', 'O'} else (-0.1 if symbol in {'C'} else 0.0))
feats = _atom_to_features(
symbol, num, deg, chg, hyb, aromatic,
imp_val, in_ring, mass, chiral, h_count,
ring_sizes, conjugated, charge_proxy
)
atoms.append(feats)
atom_symbols.append(symbol)
edges = []
for bond in mol.GetBonds():
i = bond.GetBeginAtomIdx()
j = bond.GetEndAtomIdx()
edges.extend([[i, j], [j, i]])
if not edges:
edges = [[0, 0]]
node_feats = torch.tensor(atoms, dtype=torch.float32)
edge_index = torch.tensor(edges, dtype=torch.long).t().contiguous()
return node_feats, edge_index, atom_symbols
except Exception:
pass
# ── Regex Fallback Parser ───────────────────────────────────
tokens = re.findall(r'Cl|Br|Si|Se|Pt|[A-Z][a-z]?|[a-z]|[\=\#\-\+\(\)]', smiles)
atoms, atom_symbols, edges = [], [], []
stack, prev_idx = [], None
for tok in tokens:
sym = tok.upper() if tok.isalpha() else tok
if sym in ATOM_NUMBERS or tok in ATOM_NUMBERS:
key = sym if sym in ATOM_NUMBERS else tok
num = ATOM_NUMBERS.get(key, 6)
mass = ATOM_MASSES.get(key, 12.0)
aromatic = 1.0 if tok.islower() else 0.0
idx = len(atoms)
feats = _atom_to_features(
key, num, deg=2 if aromatic else 1, chg=0, hyb=2 if aromatic else 3,
aromatic=aromatic, imp_val=0.0, in_ring=aromatic, mass=mass,
chiral=0.0, h_count=1.0, ring_sizes=[6] if aromatic else [],
conjugated=aromatic, charge_proxy=0.0
)
atoms.append(feats)
atom_symbols.append(key)
if prev_idx is not None:
edges.extend([[prev_idx, idx], [idx, prev_idx]])
prev_idx = idx
elif tok == '(':
if prev_idx is not None:
stack.append(prev_idx)
elif tok == ')':
if stack:
prev_idx = stack.pop()
if not atoms:
atoms = [_atom_to_features('C', 6, 1, 0, 3, 0, 0, 0, 12.011, 0, 1, [], 0, 0.0)]
atom_symbols = ['C']
edges = [[0, 0]]
if not edges:
edges = [[0, 0]]
node_feats = torch.tensor(atoms, dtype=torch.float32)
edge_index = torch.tensor(edges, dtype=torch.long).t().contiguous()
return node_feats, edge_index, atom_symbols
# ─────────────────────────────────────────────────────────────────
# Bemis-Murcko Scaffold Splitter
# ─────────────────────────────────────────────────────────────────
def get_bemis_murcko_scaffold(smiles: str) -> str:
try:
from rdkit import Chem
from rdkit.Chem.Scaffolds import MurckoScaffold
mol = Chem.MolFromSmiles(smiles)
if mol is not None:
return MurckoScaffold.MurckoScaffoldSmiles(mol=mol, includeChirality=False)
except Exception:
pass
rings = re.findall(r'c1[a-z0-9\=\#\-]+1|C1[A-Za-z0-9\=\#\-]+1', smiles)
if rings:
return "-".join(sorted(rings))
c_count = smiles.upper().count('C')
return f"Framework_C{c_count}"
def bemis_murcko_scaffold_split(
dataset,
smiles_list: list[str],
frac_train: float = 0.8,
frac_val: float = 0.1,
frac_test: float = 0.1,
seed: int = 42,
) -> tuple[list[int], list[int], list[int]]:
scaffolds: dict[str, list[int]] = {}
for idx, smi in enumerate(smiles_list):
sc = get_bemis_murcko_scaffold(smi)
scaffolds.setdefault(sc, []).append(idx)
scaffold_sets = sorted(scaffolds.values(), key=len, reverse=True)
rng = random.Random(seed)
rng.shuffle(scaffold_sets)
total = len(dataset)
train_cut = int(frac_train * total)
val_cut = int((frac_train + frac_val) * total)
train_idx, val_idx, test_idx = [], [], []
for cluster in scaffold_sets:
if len(train_idx) + len(cluster) <= train_cut:
train_idx.extend(cluster)
elif len(train_idx) + len(val_idx) + len(cluster) <= val_cut:
val_idx.extend(cluster)
else:
test_idx.extend(cluster)
return train_idx, val_idx, test_idx
def random_split(
dataset,
frac_train: float = 0.8,
frac_val: float = 0.1,
frac_test: float = 0.1,
seed: int = 42,
) -> tuple[list[int], list[int], list[int]]:
indices = list(range(len(dataset)))
random.Random(seed).shuffle(indices)
n = len(indices)
train_cut = int(frac_train * n)
val_cut = int((frac_train + frac_val) * n)
return indices[:train_cut], indices[train_cut:val_cut], indices[val_cut:]
# ─────────────────────────────────────────────────────────────────
# XAI — Toxic Hotspot Highlighting
# ─────────────────────────────────────────────────────────────────
def highlight_toxic_subgraph(
smiles: str,
attention_scores: torch.Tensor,
top_k: int = 3,
) -> dict[str, Any]:
_node_feats, _edge_index, atom_symbols = smiles_to_graph(smiles)
num_nodes = len(atom_symbols)
if attention_scores is None or len(attention_scores) == 0:
scores = np.ones(num_nodes) / max(num_nodes, 1)
else:
scores = attention_scores.detach().cpu().numpy()
if len(scores) < num_nodes:
scores = np.pad(scores, (0, num_nodes - len(scores)), 'constant')
elif len(scores) > num_nodes:
scores = scores[:num_nodes]
max_s = np.max(scores) if np.max(scores) > 0 else 1.0
norm = scores / max_s
top_idxs = np.argsort(norm)[::-1][:min(top_k, num_nodes)].tolist()
hotspots = [
{
"atom_index": int(i),
"atom_symbol": atom_symbols[i],
"attention_score": round(float(norm[i]), 4),
"is_toxic_hotspot": True,
}
for i in top_idxs
]
return {
"smiles": smiles,
"total_atoms": num_nodes,
"atom_symbols": atom_symbols,
"attention_weights": [round(float(s), 4) for s in norm],
"top_toxic_hotspots": hotspots,
"plot_title": "GAT Layer Attention Distribution Map (Highlight = High Attention Weight)",
}
def calculate_tanimoto_applicability_domain(
query_smiles: str,
training_smiles_list: list[str]
) -> dict[str, Any]:
"""
Computes maximum Tanimoto similarity between query molecule and training dataset.
Flags applicability domain confidence:
- High Confidence: Max Tanimoto >= 0.70
- Moderate Confidence: 0.40 <= Max Tanimoto < 0.70
- Low Confidence (Out of Domain): Max Tanimoto < 0.40
"""
try:
from rdkit import Chem, DataStructs
from rdkit.Chem import RDKFingerprint
q_mol = Chem.MolFromSmiles(query_smiles)
if q_mol is None:
return {"max_tanimoto": 0.0, "applicability_domain": "Out-of-Domain (Invalid SMILES)"}
q_fp = RDKFingerprint(q_mol)
max_sim = 0.0
for tr_smi in training_smiles_list:
tr_mol = Chem.MolFromSmiles(tr_smi)
if tr_mol is not None:
tr_fp = RDKFingerprint(tr_mol)
sim = DataStructs.TanimotoSimilarity(q_fp, tr_fp)
max_sim = max(max_sim, sim)
max_sim = round(float(max_sim), 4)
if max_sim >= 0.70:
domain = "High Confidence (In-Domain)"
elif max_sim >= 0.40:
domain = "Moderate Confidence"
else:
domain = "Out-of-Domain (Novel Scaffold)"
return {"max_tanimoto": max_sim, "applicability_domain": domain}
except Exception:
return {"max_tanimoto": 0.50, "applicability_domain": "Unknown Domain"}