ChemMiniQ3-SAbRLo / evaluate_molecular_model.py
gbyuvd's picture
Update evaluate_molecular_model.py
c64a9dc verified
# evaluate_molecular_model.py
# now with Durrant's lab filtering in validity check
import os
import sys
import json
import argparse
import random
from typing import List, Optional
from tqdm import tqdm
import torch
from rdkit import Chem
from rdkit.Chem import AllChem
from rdkit import RDLogger
import selfies as sf
import pandas as pd
# Suppress RDKit warnings
RDLogger.DisableLog('rdApp.*')
# Add local path
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
from FastChemTokenizerHF import FastChemTokenizerSelfies
from ChemQ3MTP import ChemQ3MTPForCausalLM
# ----------------------------
# Robust Conversion & Validation (as per your spec)
# ----------------------------
def selfies_to_smiles(selfies_str: str) -> Optional[str]:
"""Convert SELFIES string to SMILES, handling tokenizer artifacts."""
try:
clean_selfies = selfies_str.replace(" ", "")
return sf.decoder(clean_selfies)
except Exception:
return None
def is_valid_smiles(smiles: str) -> bool:
"""
Check if a SMILES string represents a valid molecule.
FIXED: Now properly checks for heavy atoms (non-hydrogens) >= 3
and rejects disconnected/separated molecules
"""
if not isinstance(smiles, str) or len(smiles.strip()) == 0:
return False
smiles = smiles.strip()
# FAST CHECK: Reject separated molecules (contains dots)
if '.' in smiles:
return False # Disconnected components indicated by dots
try:
mol = Chem.MolFromSmiles(smiles)
if mol is None:
return False
# CRITICAL FIX: Check heavy atoms (non-hydrogens), not total atoms
heavy_atoms = mol.GetNumHeavyAtoms() # This excludes hydrogens
if heavy_atoms < 3:
return False
return True
except Exception:
return False
def passes_durrant_lab_filter(smiles: str) -> bool:
"""
Apply Durrant's lab filter to remove improbable substructures.
FIXED: More robust error handling, pattern checking, and disconnected molecule rejection.
Returns True if molecule passes the filter (is acceptable), False otherwise.
"""
if not smiles or not isinstance(smiles, str) or len(smiles.strip()) == 0:
return False
try:
mol = Chem.MolFromSmiles(smiles.strip())
if mol is None:
return False
# Check heavy atoms again (belt and suspenders approach)
if mol.GetNumHeavyAtoms() < 3:
return False
# REJECT SEPARATED/DISCONNECTED MOLECULES (double check here too)
fragments = Chem.rdmolops.GetMolFrags(mol, asMols=False)
if len(fragments) > 1:
return False # Reject molecules with multiple disconnected parts
# Define SMARTS patterns for problematic substructures
problematic_patterns = [
"C=[N-]", # Carbon double bonded to negative nitrogen
"[N-]C=[N+]", # Nitrogen anion bonded to nitrogen cation
"[nH+]c[n-]", # Aromatic nitrogen cation adjacent to nitrogen anion
"[#7+]~[#7+]", # Positive nitrogen connected to positive nitrogen
"[#7-]~[#7-]", # Negative nitrogen connected to negative nitrogen
"[!#7]~[#7+]~[#7-]~[!#7]", # Bridge: non-nitrogen - pos nitrogen - neg nitrogen - non-nitrogen
"[#5]", # Boron atoms
"O=[PH](=O)([#8])([#8])", # Phosphoryl with hydroxyls
"N=c1cc[#7]c[#7]1", # Nitrogen in aromatic ring with another nitrogen
"[$([NX2H1]),$([NX3H2])]=C[$([OH]),$([O-])]", # N=CH-OH or N=CH-O-
]
# Check for metals (excluding common biologically relevant ions)
metal_exclusions = {11, 12, 19, 20} # Na, Mg, K, Ca
for atom in mol.GetAtoms():
atomic_num = atom.GetAtomicNum()
# More precise metal detection
if atomic_num > 20 and atomic_num not in metal_exclusions:
return False
# Check for each problematic pattern
for pattern in problematic_patterns:
try:
patt_mol = Chem.MolFromSmarts(pattern)
if patt_mol is not None:
matches = mol.GetSubstructMatches(patt_mol)
if matches:
return False # Found problematic substructure
except Exception:
# If SMARTS parsing fails, continue to next pattern
continue
return True # Passed all checks
except Exception:
return False
def get_sa_label_and_confidence(selfies_str: str) -> tuple[str, float]:
"""Get SA label (Easy/Hard) and confidence from the model's SA classifier."""
try:
from ChemQ3MTP.rl_utils import get_sa_classifier
classifier = get_sa_classifier()
if classifier is None:
return "Unknown", 0.0
# Get raw classifier output: [{'label': 'Easy', 'score': 0.9187200665473938}]
result = classifier(selfies_str, truncation=True, max_length=128)[0]
return result["label"], result["score"]
except Exception as e:
return "Unknown", 0.0
def get_morgan_fingerprint_from_smiles(smiles: str, radius=2, n_bits=2048):
mol = Chem.MolFromSmiles(smiles)
if mol is None:
return None
return AllChem.GetMorganFingerprintAsBitVect(mol, radius, nBits=n_bits)
def tanimoto_sim(fp1, fp2):
from rdkit.DataStructs import TanimotoSimilarity
return TanimotoSimilarity(fp1, fp2)
# ----------------------------
# Main Evaluation Function
# ----------------------------
def evaluate_model(
model_path: str,
train_data_path: str = "../data/chunk_5.csv",
n_samples: int = 1000,
seed: int = 42,
max_gen_len: int = 32
):
torch.manual_seed(seed)
random.seed(seed)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"🚀 Evaluating model at: {model_path}")
print(f" Device: {device} | Samples: {n_samples} | Seed: {seed}\n")
# Load tokenizer and model
tokenizer = FastChemTokenizerSelfies.from_pretrained("../selftok_core")
model = ChemQ3MTPForCausalLM.from_pretrained(model_path)
model.to(device)
model.eval()
# Load training set and normalize SELFIES (remove spaces)
print("📂 Loading and normalizing training set for novelty...")
train_df = pd.read_csv(train_data_path)
train_selfies_clean = set()
for s in train_df["SELFIES"].dropna().astype(str):
clean_s = s.replace(" ", "")
train_selfies_clean.add(clean_s)
print(f" Training set size: {len(train_selfies_clean)} unique (space-free) SELFIES\n")
# === MTP-AWARE GENERATION ===
print("GenerationStrategy: Using MTP-aware generation...")
all_selfies_raw = []
batch_size = 32
num_batches = (n_samples + batch_size - 1) // batch_size
with torch.no_grad():
for _ in tqdm(range(num_batches), desc="Generating"):
current_batch_size = min(batch_size, n_samples - len(all_selfies_raw))
if current_batch_size <= 0:
break
input_ids = torch.full(
(current_batch_size, 1),
tokenizer.bos_token_id,
dtype=torch.long,
device=device
)
if hasattr(model, 'generate_with_logprobs'):
try:
outputs = model.generate_with_logprobs(
input_ids=input_ids,
max_new_tokens=25,
temperature=1.0,
top_k=50,
top_p=0.95,
do_sample=True,
return_probs=True,
tokenizer=tokenizer
)
batch_selfies = outputs[0] # list of raw SELFIES (may have spaces)
except Exception as e:
print(f"⚠️ MTP generation failed: {e}. Falling back.")
gen_tokens = model.generate(
input_ids,
max_length=max_gen_len,
do_sample=True,
top_k=50,
top_p=0.95,
temperature=1.0,
pad_token_id=tokenizer.pad_token_id,
eos_token_id=tokenizer.eos_token_id
)
batch_selfies = [
tokenizer.decode(seq, skip_special_tokens=True)
for seq in gen_tokens
]
else:
gen_tokens = model.generate(
input_ids,
max_length=max_gen_len,
do_sample=True,
top_k=50,
top_p=0.95,
temperature=1.0,
pad_token_id=tokenizer.pad_token_id,
eos_token_id=tokenizer.eos_token_id
)
batch_selfies = [
tokenizer.decode(seq, skip_special_tokens=True)
for seq in gen_tokens
]
all_selfies_raw.extend(batch_selfies)
if len(all_selfies_raw) >= n_samples:
break
all_selfies_raw = all_selfies_raw[:n_samples]
print(f"\n✅ Generated {len(all_selfies_raw)} raw SELFIES strings.\n")
# Process: SELFIES → clean SELFIES → SMILES → valid molecules
valid_records = []
print("🧪 Processing SELFIES and converting to SMILES...")
for i, raw_selfies in enumerate(tqdm(all_selfies_raw, desc="Converting")):
# Clean the SELFIES (remove spaces as tokenizer uses whitespace)
clean_selfies = raw_selfies.replace(" ", "")
# Convert to SMILES
smiles = selfies_to_smiles(clean_selfies)
if smiles is not None and is_valid_smiles(smiles) and passes_durrant_lab_filter(smiles):
valid_records.append({
"raw_selfies": raw_selfies,
"selfies_clean": clean_selfies,
"selfies": clean_selfies, # canonical version
"smiles": smiles.strip()
})
# >>> DEBUG: Print multiple examples and SA label analysis <<<
if valid_records:
print("\n🔍 DEBUG: Sample generated molecules")
print("-" * 70)
for i in range(min(5, len(valid_records))):
example = valid_records[i]
print(f"Example {i+1}:")
print(f" Raw SELFIES : {example['raw_selfies'][:80]}{'...' if len(example['raw_selfies']) > 80 else ''}")
print(f" SMILES : {example['smiles']}")
# Get SA label and confidence
label, confidence = get_sa_label_and_confidence(example['raw_selfies'])
print(f" SA Label : {label} (confidence: {confidence:.3f})")
if i == 0:
# Test SA classifier with simple molecules
simple_label, simple_conf = get_sa_label_and_confidence('[C]')
benzene_label, benzene_conf = get_sa_label_and_confidence('[c] [c] [c] [c] [c] [c] [Ring1] [=Branch1]')
print(f" 🧪 SA Test - Simple molecule: {simple_label} ({simple_conf:.3f})")
print(f" 🧪 SA Test - Benzene: {benzene_label} ({benzene_conf:.3f})")
# Check molecule properties
mol = Chem.MolFromSmiles(example['smiles'])
if mol:
print(f" Atoms : {mol.GetNumAtoms()}")
print(f" Bonds : {mol.GetNumBonds()}")
print()
print("-" * 70)
# SA Label distribution analysis
sa_labels = []
for r in valid_records[:100]:
label, _ = get_sa_label_and_confidence(r["raw_selfies"])
sa_labels.append(label)
easy_count = sa_labels.count("Easy")
hard_count = sa_labels.count("Hard")
unknown_count = sa_labels.count("Unknown")
print(f"🔍 SA Label Analysis (first 100 molecules):")
print(f" Easy to synthesize: {easy_count}/100 ({easy_count}%)")
print(f" Hard to synthesize: {hard_count}/100 ({hard_count}%)")
if unknown_count > 0:
print(f" Unknown/Failed: {unknown_count}/100 ({unknown_count}%)")
else:
print("\n⚠️ WARNING: No valid molecules generated in sample!")
# <<< END DEBUG >>>
# Now compute metrics...
validity = len(valid_records) / n_samples
unique_valid = list({r["selfies_clean"]: r for r in valid_records}.values())
uniqueness = len(unique_valid) / len(valid_records) if valid_records else 0.0
novel_count = sum(1 for r in unique_valid if r["selfies_clean"] not in train_selfies_clean)
novelty = novel_count / len(unique_valid) if unique_valid else 0.0
# SA Label Counts (using model's SA classifier)
sa_labels_all = []
for r in unique_valid:
label, _ = get_sa_label_and_confidence(r["raw_selfies"])
sa_labels_all.append(label)
easy_total = sa_labels_all.count("Easy")
hard_total = sa_labels_all.count("Hard")
unknown_total = sa_labels_all.count("Unknown")
total_labeled = len(sa_labels_all)
# Internal Diversity (on SMILES)
if len(unique_valid) >= 2:
fps = []
for r in unique_valid:
fp = get_morgan_fingerprint_from_smiles(r["smiles"])
if fp is not None:
fps.append(fp)
if len(fps) >= 2:
total_sim, count = 0.0, 0
for i in range(len(fps)):
for j in range(i + 1, len(fps)):
total_sim += tanimoto_sim(fps[i], fps[j])
count += 1
internal_diversity = 1.0 - (total_sim / count)
else:
internal_diversity = 0.0
else:
internal_diversity = 0.0
# ----------------------------
# Final Summary
# ----------------------------
print("\n" + "="*55)
print("📊 MOLECULAR GENERATION EVALUATION SUMMARY")
print("="*55)
print(f"Model Path : {model_path}")
print(f"Generation Mode : {'MTP-aware' if hasattr(model, 'generate_with_logprobs') else 'Standard'}")
print(f"Samples Generated: {n_samples}")
print("-"*55)
print(f"Validity : {validity:.4f} ({len(valid_records)}/{n_samples})")
print(f"Uniqueness : {uniqueness:.4f} (unique valid)")
print(f"Novelty (vs train): {novelty:.4f} (space-free SELFIES)")
print(f"Synthesis Labels : Easy: {easy_total}/{total_labeled} ({easy_total/max(1,total_labeled)*100:.1f}%) | Hard: {hard_total}/{total_labeled} ({hard_total/max(1,total_labeled)*100:.1f}%)")
if unknown_total > 0:
print(f" Unknown: {unknown_total}/{total_labeled} ({unknown_total/max(1,total_labeled)*100:.1f}%)")
print(f"Internal Diversity: {internal_diversity:.4f} (1 - avg Tanimoto)")
print("="*55)
results = {
"model_path": model_path,
"generation_mode": "MTP-aware" if hasattr(model, 'generate_with_logprobs') else "standard",
"n_samples": n_samples,
"validity": validity,
"uniqueness": uniqueness,
"novelty": novelty,
"sa_easy_count": easy_total,
"sa_hard_count": hard_total,
"sa_easy_percentage": easy_total/max(1,total_labeled)*100,
"sa_hard_percentage": hard_total/max(1,total_labeled)*100,
"internal_diversity": internal_diversity,
"valid_molecules_count": len(valid_records)
}
if unknown_total > 0:
results["sa_unknown_count"] = unknown_total
results["sa_unknown_percentage"] = unknown_total/max(1,total_labeled)*100
output_json = os.path.join(model_path, "evaluation_summary.json")
with open(output_json, "w") as f:
json.dump(results, f, indent=2)
print(f"\n💾 Results saved to: {output_json}")
return results
# ----------------------------
# CLI
# ----------------------------
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="Evaluate molecular generative model with MTP-aware generation")
parser.add_argument("--model_path", type=str, required=True, help="Path to model checkpoint")
parser.add_argument("--n_samples", type=int, default=1000, help="Number of molecules to generate")
parser.add_argument("--seed", type=int, default=42, help="Random seed")
parser.add_argument("--train_data", type=str, default="../data/chunk_5.csv", help="Training data CSV")
args = parser.parse_args()
evaluate_model(
model_path=args.model_path,
train_data_path=args.train_data,
n_samples=args.n_samples,
seed=args.seed
)