Delower's picture
Upload predict.py with huggingface_hub
2d29f1d verified
Raw History Blame Contribute Delete
3.48 kB
# predict.py — Kidney Arch5 feature-block ensemble deployment model.
# Predicts the probability of the kidney-toxic (nephrotoxicity-concern)
# class for one molecule.
#
# Usage:
# python3 predict.py "CCO" -> prints the probability
# from predict import predict; predict("CCO")
# from predict import predict_many; predict_many([smi1, smi2])
#
# Requires: scikit-learn, lightgbm, rdkit, joblib, numpy, torch,
# transformers, and deployment_common.py (shipped in this repo) on the
# import path — the shipped preprocessing objects (DescPrep) unpickle
# against that module. Component C uses ChemBERTa-77M embeddings
# (seyonec/ChemBERTa-zinc-base-v1, downloaded from the HF hub on first
# use; pretrained encoder only, never fine-tuned).
#
# Featurization is the deployment pipeline's own code: featurize /
# prep_matrix from deployment_common.py, embed() copied verbatim from
# the deployment prep_features.py (progress print removed; computation
# identical), block_X copied verbatim from train_arch5_deployment.py.
# Prediction replicates screen_leads.py's arch5_prob exactly.
import os, sys
HERE = os.path.dirname(os.path.abspath(__file__))
sys.path.insert(0, HERE)
import numpy as np
import joblib
from deployment_common import featurize, prep_matrix
_bundle = joblib.load(os.path.join(HERE, "model.joblib"))
_components, _weights = _bundle["components"], _bundle["weights"]
_prep_desc, _prep_phys = _bundle["prep_desc"], _bundle["prep_phys"]
def embed(smis, batch=64):
"""ChemBERTa-77M mean-pooled embeddings — the arm's own code, verbatim
(kidney_arch5_ensemble.py, pretrained encoder, never fine-tuned)."""
import torch
from transformers import AutoTokenizer, AutoModel
tok = AutoTokenizer.from_pretrained("seyonec/ChemBERTa-zinc-base-v1")
mdl = AutoModel.from_pretrained("seyonec/ChemBERTa-zinc-base-v1")
mdl.eval()
smis = [s if isinstance(s, str) else "" for s in smis]
E = np.zeros((len(smis), mdl.config.hidden_size), dtype=np.float32)
with torch.no_grad():
for i in range(0, len(smis), batch):
chunk = smis[i:i + batch]
enc = tok(chunk, padding=True, truncation=True, max_length=256, return_tensors="pt")
h = mdl(**enc).last_hidden_state
m = enc["attention_mask"].unsqueeze(-1).float()
E[i:i + len(chunk)] = ((h * m).sum(1) / m.sum(1)).numpy()
return E
def block_X(blocks, which, emb, prep_desc, prep_phys):
# verbatim from train_arch5_deployment.py
if which == "A":
return blocks["morgan_bits"]
if which == "B":
return prep_matrix(blocks, prep_desc, "desc_maccs_counts")
if which == "C":
return np.hstack([emb, prep_phys.transform(blocks["physchem"])]).astype(np.float32)
raise ValueError(which)
def predict_many(smiles_list):
"""Probabilities of the toxic class; None where unparseable."""
smiles_list = list(smiles_list)
blocks, valid = featurize(smiles_list)
emb = embed(smiles_list)
p = None
for which in ["A", "B", "C"]:
X = block_X(blocks, which, emb, _prep_desc, _prep_phys)
pw = _weights[which] * _components[which].predict_proba(X)[:, 1]
p = pw if p is None else p + pw
return [float(v) if ok else None for v, ok in zip(p, valid)]
def predict(smiles):
"""Probability of the toxic class; None if unparseable."""
return predict_many([smiles])[0]
if __name__ == "__main__":
print(predict(sys.argv[1]))