# src/features/ligand.py # # Ligand feature extraction — pure RDKit, zero ML models at inference. # All operations: O(N_atoms) or O(N_atoms²) at worst → microseconds/mol. # # Feature blocks: # # BINARY FINGERPRINTS (presence/absence of substructure) # ───────────────────────────────────────────────────── # ecfp2 1024d Morgan r=1 — ultra-local atom neighbourhoods # ecfp 1024d Morgan r=2 — standard local topology (ECFP4) # ecfp6 1024d Morgan r=3 — extended neighbourhoods # fcfp 1024d Functional class r=2 — pharmacophoric identity # maccs 167d 166 SMARTS pharmacophore keys # atom_pair 2048d All-pairs graph distance (global topology) # torsion 2048d 4-atom rotatable bond paths (conformational) # avalon 512d Avalon — completely different algorithm (Scitegic) # rdkit_pat 2048d RDKit layered — ring + aromaticity + bond order # # COUNT FINGERPRINTS (how many times each substructure appears) # ───────────────────────────────────────────────────────────── # ecfp_count 1024d Morgan r=2 counts — 3 benzenes != 1 benzene # ecfp6_count 1024d Morgan r=3 counts # # DENSE CONTINUOUS # ──────────────── # estate 79d EState sum indices — electrotopological signal # phys 217d RDKit full descriptor suite (RobustScaler normalised) # # Inference timing (HF Spaces free tier, 2 vCPU): # Per SMILES: ~3-5 ms total (all fingerprints + descriptors) # 1M compounds: ~50-80 min on single CPU core # No GPU, no transformer, no external calls. import numpy as np from rdkit import Chem, DataStructs from rdkit.Chem import AllChem, Descriptors, MACCSkeys, rdMolDescriptors from rdkit.Chem.EState import Fingerprinter as EStateFP from rdkit import RDLogger from sklearn.preprocessing import RobustScaler RDLogger.DisableLog('rdApp.*') _DESC_LIST = Descriptors._descList try: from rdkit.Avalon.pyAvalonTools import GetAvalonFP as _GetAvalonFP _AVALON_OK = True except ImportError: _AVALON_OK = False print(" WARNING: rdkit.Avalon not available — avalon features will be zeros. " "Reinstall RDKit with Avalon support if needed.") def smiles_to_features(smiles: str): """ Convert a SMILES string to the full ligand feature dict. Returns None if SMILES is invalid. """ mol = Chem.MolFromSmiles(smiles) if mol is None: return None # ── Binary Morgan fingerprints ───────────────────────────────────── def _bin(radius, nbits=1024): fp = AllChem.GetMorganFingerprintAsBitVect(mol, radius, nBits=nbits) arr = np.zeros(nbits, dtype=np.float32) DataStructs.ConvertToNumpyArray(fp, arr) return arr ecfp2 = _bin(1) ecfp = _bin(2) # ECFP4 ecfp6 = _bin(3) fp_fcfp = AllChem.GetMorganFingerprintAsBitVect(mol, 2, nBits=1024, useFeatures=True) fcfp = np.zeros(1024, dtype=np.float32) DataStructs.ConvertToNumpyArray(fp_fcfp, fcfp) # ── Morgan COUNT fingerprints ────────────────────────────────────── # Counts how many times each substructure hashes to each bit. # A drug with 3 chloro-phenyl groups looks different from one with 1. # Orthogonal to the binary versions above. def _cnt(radius, nbits=1024): fp = AllChem.GetHashedMorganFingerprint(mol, radius, nBits=nbits) arr = np.zeros(nbits, dtype=np.float32) DataStructs.ConvertToNumpyArray(fp, arr) return arr ecfp_count = _cnt(2) ecfp6_count = _cnt(3) # ── Avalon fingerprint (512d) ────────────────────────────────────── # Completely different algorithm from Morgan family. # Graph-invariant path enumeration — catches heteroaromatic scaffold # patterns Morgan misses. if _AVALON_OK: try: fp_av = _GetAvalonFP(mol, nBits=512) avalon = np.zeros(512, dtype=np.float32) DataStructs.ConvertToNumpyArray(fp_av, avalon) except Exception: avalon = np.zeros(512, dtype=np.float32) else: avalon = np.zeros(512, dtype=np.float32) # ── RDKit Pattern (Layered) fingerprint (2048d) ──────────────────── # Encodes atom connectivity WITH ring membership, aromaticity, bond # order layered in. Catches fused aromatic systems (indoles, purines, # quinolines) that ECFP treats as overlapping local neighbourhoods. try: fp_pat = Chem.RDKFingerprint(mol, fpSize=2048) rdkit_pat = np.zeros(2048, dtype=np.float32) DataStructs.ConvertToNumpyArray(fp_pat, rdkit_pat) except Exception: rdkit_pat = np.zeros(2048, dtype=np.float32) # ── MACCS keys (167d) ───────────────────────────────────────────── mk = MACCSkeys.GenMACCSKeys(mol) maccs = np.zeros(167, dtype=np.float32) DataStructs.ConvertToNumpyArray(mk, maccs) # ── AtomPair binary (2048d) ──────────────────────────────────────── fp_ap = rdMolDescriptors.GetHashedAtomPairFingerprintAsBitVect(mol, nBits=2048) atom_pair = np.zeros(2048, dtype=np.float32) DataStructs.ConvertToNumpyArray(fp_ap, atom_pair) # ── Topological Torsion binary (2048d) ──────────────────────────── fp_tt = rdMolDescriptors.GetHashedTopologicalTorsionFingerprintAsBitVect(mol, nBits=2048) torsion = np.zeros(2048, dtype=np.float32) DataStructs.ConvertToNumpyArray(fp_tt, torsion) # ── EState sum indices (79d dense continuous) ────────────────────── try: _, sum_e = EStateFP.FingerprintMol(mol) estate = np.array(sum_e, dtype=np.float64) estate = np.nan_to_num(estate, nan=0.0, posinf=0.0, neginf=0.0) estate = np.clip(estate, -1e6, 1e6).astype(np.float32) except Exception: estate = np.zeros(79, dtype=np.float32) # ── RDKit physicochemical descriptors (~217d) ────────────────────── phys = [] for _, func in _DESC_LIST: try: v = float(func(mol)) phys.append(v if (np.isfinite(v) and abs(v) < 1e15) else 0.0) except Exception: phys.append(0.0) return { 'ecfp2': ecfp2, 'ecfp': ecfp, 'ecfp6': ecfp6, 'fcfp': fcfp, 'maccs': maccs, 'atom_pair': atom_pair, 'torsion': torsion, 'avalon': avalon, 'rdkit_pat': rdkit_pat, 'ecfp_count': ecfp_count, 'ecfp6_count': ecfp6_count, 'estate': estate, 'phys': np.array(phys, dtype=np.float32), } def extract_ligand_features(smiles_list: list, scaler=None, fit_scaler: bool = False): """ Extract ligand features for a list of SMILES strings. Args: smiles_list: list of SMILES strings scaler: fitted RobustScaler (required if fit_scaler=False) fit_scaler: if True, fit a new scaler on the continuous features Returns: feats: dict of numpy arrays, one per feature type valid_idx: indices of successfully parsed SMILES scaler: fitted RobustScaler Note: Binary + count fingerprints are NOT scaled. GBMs are invariant to monotone transforms on binary features. Count fingerprints are log1p-transformed for numerical stability. """ ecfp2s, ecfps, ecfp6s, fcfps = [], [], [], [] maccss, aps, tors = [], [], [] avalons, rdkit_pats = [], [] ecfp_counts, ecfp6_counts = [], [] estates, physs = [], [] valid_idx = [] for i, smi in enumerate(smiles_list): r = smiles_to_features(smi) if r is None: continue ecfp2s.append(r['ecfp2']) ecfps.append(r['ecfp']) ecfp6s.append(r['ecfp6']) fcfps.append(r['fcfp']) maccss.append(r['maccs']) aps.append(r['atom_pair']) tors.append(r['torsion']) avalons.append(r['avalon']) rdkit_pats.append(r['rdkit_pat']) ecfp_counts.append(r['ecfp_count']) ecfp6_counts.append(r['ecfp6_count']) estates.append(r['estate']) physs.append(r['phys']) valid_idx.append(i) n_fail = len(smiles_list) - len(valid_idx) if n_fail: print(f" Ligand: {n_fail} SMILES failed to parse — dropped") # Continuous: clean then scale together phys_arr = np.nan_to_num( np.array(physs, dtype=np.float64), nan=0.0, posinf=0.0, neginf=0.0 ).astype(np.float32) estate_arr = np.array(estates, dtype=np.float32) continuous = np.concatenate([phys_arr, estate_arr], axis=1) if fit_scaler: scaler = RobustScaler() scaler.fit(continuous) continuous_scaled = scaler.transform(continuous) phys_scaled = continuous_scaled[:, :phys_arr.shape[1]] estate_scaled = continuous_scaled[:, phys_arr.shape[1]:] # Count FPs: log1p stabilises large int values without losing magnitude info ecfp_cnt_arr = np.log1p(np.array(ecfp_counts, dtype=np.float32)) ecfp6_cnt_arr = np.log1p(np.array(ecfp6_counts, dtype=np.float32)) feats = { 'ecfp2': np.array(ecfp2s, dtype=np.float32), 'ecfp': np.array(ecfps, dtype=np.float32), 'ecfp6': np.array(ecfp6s, dtype=np.float32), 'fcfp': np.array(fcfps, dtype=np.float32), 'maccs': np.array(maccss, dtype=np.float32), 'atom_pair': np.array(aps, dtype=np.float32), 'torsion': np.array(tors, dtype=np.float32), 'avalon': np.array(avalons, dtype=np.float32), 'rdkit_pat': np.array(rdkit_pats, dtype=np.float32), 'ecfp_count': ecfp_cnt_arr, 'ecfp6_count': ecfp6_cnt_arr, 'estate': estate_scaled, 'phys': phys_scaled, } total_dim = sum(v.shape[1] for v in feats.values()) print(f" Ligand: {len(valid_idx)} molecules | {total_dim}d total") print(f" Binary: ecfp2={feats['ecfp2'].shape[1]} ecfp={feats['ecfp'].shape[1]} " f"ecfp6={feats['ecfp6'].shape[1]} fcfp={feats['fcfp'].shape[1]} " f"maccs={feats['maccs'].shape[1]} ap={feats['atom_pair'].shape[1]} " f"tors={feats['torsion'].shape[1]} avalon={feats['avalon'].shape[1]} " f"rdkit_pat={feats['rdkit_pat'].shape[1]}") print(f" Counts: ecfp_cnt={feats['ecfp_count'].shape[1]} " f"ecfp6_cnt={feats['ecfp6_count'].shape[1]}") print(f" Dense: estate={feats['estate'].shape[1]} " f"phys={feats['phys'].shape[1]}") return feats, valid_idx, scaler