"""Inference for Delower/herg-best-deployment (DEPLOYMENT REFIT). Predicts hERG blocker probability (10 uM, classification winner) and pIC50 (regression winner) from SMILES. Feature stack: Morgan ECFP4 bits+counts, RDKit 2D descriptors + QED + hERG-targeted physchem, ChemBERTa-77M pretrained embeddings (mean-pooled; downloads seyonec/ChemBERTa-zinc-base-v1 on first use). CDK/PaDEL descriptors are NOT used: a time-boxed PaDEL attempt stalled and the CDK slot was substituted by the RDKit 2D descriptor block (see FEATURE_NOTES.md). Requires: rdkit, numpy, scikit-learn, xgboost (if a tree is in a winner), tensorflow (if the LTN is in a winner), transformers + torch (embeddings). Usage: python predict_herg.py "CCOc1ccc..." ["SMILES2" ...] """ import sys, os, json, pickle import numpy as np HERE = os.path.dirname(os.path.abspath(__file__)) sys.path.insert(0, HERE) import feature_extraction as FE from feature_extraction import featurize_smiles CFG = json.load(open(os.path.join(HERE, "config.json"))) BLOCKS = CFG["blocks"] class ContPrep: def __init__(self, s): self.keep_ = np.array(s["keep"]); self.med_ = np.array(s["med"]) self.mu_ = np.array(s["mu"]); self.sd_ = np.array(s["sd"]); self.log1p = s["log1p"] def transform(self, X): X = np.where(np.isfinite(X), X, np.nan) if self.log1p: X = np.log1p(np.clip(X, 0, None)) X = X[:, self.keep_] X = np.where(np.isnan(X), self.med_, X) return ((X - self.mu_) / self.sd_).astype(np.float32) _CB = {} def chemberta_embed(smiles): if "model" not in _CB: import torch from transformers import AutoModel, AutoTokenizer _CB["tok"] = AutoTokenizer.from_pretrained("seyonec/ChemBERTa-zinc-base-v1") _CB["model"] = AutoModel.from_pretrained("seyonec/ChemBERTa-zinc-base-v1") _CB["model"].eval() _CB["torch"] = torch torch = _CB["torch"] enc = _CB["tok"](smiles, padding=True, truncation=True, max_length=160, return_tensors="pt") with torch.no_grad(): h = _CB["model"](**enc).last_hidden_state mask = enc["attention_mask"].unsqueeze(-1).float() return ((h * mask).sum(1) / mask.sum(1).clamp(min=1.0)).numpy().astype(np.float32) def featurize(smiles_list): raw = {b: [] for b in ["morgan", "morgan_counts", "desc", "physchem_extra", "herg"]} for smi in smiles_list: _, bl = featurize_smiles(smi) assert bl is not None, f"unparseable SMILES: {smi}" for b in raw: raw[b].append(np.asarray(bl[b], dtype=np.float32)) raw = {b: np.vstack(v) for b, v in raw.items()} raw["chemberta"] = chemberta_embed(smiles_list) return raw def assemble(raw, prep_states): preps = {b: ContPrep(s) for b, s in prep_states.items()} parts = [] for b in BLOCKS: X = np.asarray(raw[b], dtype=np.float32) if b in preps: X = preps[b].transform(X) parts.append(np.asarray(X, dtype=np.float32)) return np.hstack(parts) def predict(smiles_list): raw = featurize(smiles_list) out = {} for task in ("cls", "reg"): entry = CFG["tasks"][task] cand = entry["winner"] preds = {} if cand in ("tree", "ensemble"): with open(os.path.join(HERE, f"tree_{task}.pkl"), "rb") as fh: bundle = pickle.load(fh) X = assemble(raw, bundle["prep_states"]) mdl = bundle["model"] if task == "cls": preds["tree"] = mdl.predict_proba(X)[:, 1] elif mdl.__class__.__name__ == "Booster": import xgboost as xgb preds["tree"] = mdl.predict(xgb.DMatrix(X)) else: preds["tree"] = mdl.predict(X) if cand in ("ltn", "ensemble"): import tensorflow as tf meta = json.load(open(os.path.join(HERE, f"ltn_{task}_preps.json"))) X = assemble(raw, meta["prep_states"]) model = tf.keras.models.load_model(os.path.join(HERE, f"ltn_{task}.keras")) if task == "cls": preds["ltn"] = tf.nn.softmax(model(X, training=False), axis=1).numpy()[:, 1] else: preds["ltn"] = model(X, training=False).numpy()[:, 0] * meta["target_sd"] + meta["target_mu"] if cand == "ensemble": w = entry["ensemble_weight_on_tree"] out[task] = w * preds["tree"] + (1 - w) * preds["ltn"] else: out[task] = preds[cand] return out if __name__ == "__main__": smis = sys.argv[1:] res = predict(smis) for i, s in enumerate(smis): print(f"{s}\tprob_blocker={res['cls'][i]:.4f}\tpred_pIC50={res['reg'][i]:.3f}")