AI Molecule Property Prediction System (ADMET)

Developing AI System for Molecule Property Prediction (ADMET)

ADMET (Absorption, Distribution, Metabolism, Excretion, Toxicity) — set of pharmacokinetic properties determining drug fate in body. ~50% clinical trial failures — due to ADMET problems that could have been predicted earlier.

Critical ADMET Properties

Absorption

• Aqueous solubility: poor solubility → inconsistent bioavailability
• Lipophilicity (logP/logD): determines membrane penetration, solubility
• Caco-2 / MDCK permeability: intestinal absorption
• P-glycoprotein (P-gp) efflux: active cell export, reduces bioavailability
• Oral bioavailability (F%): what fraction of dose reaches systemic circulation

Distribution

• Volume of distribution (Vd): how distributes across tissues
• Blood-brain barrier permeability (BBB): needed for CNS drugs, undesirable for peripheral
• Plasma protein binding (PPB): albumin binding, only free drug active

Metabolism

• CYP450 inhibition (CYP3A4, CYP2D6, CYP2C9, CYP2C19, CYP1A2): slows metabolism of other drugs → interactions
• CYP450 substrate: which isoforms metabolize compound
• Half-life (T½): how fast cleared from body
• Hepatotoxicity (DILI): liver damage

Excretion

• Renal clearance: rate of kidney elimination

Toxicity

• hERG inhibition: blocking cardiac K⁺ channel → QT prolongation → potentially fatal arrhythmia. Major reason for drug withdrawal
• Ames test: mutagenicity / genotoxicity
• DILI (Drug-Induced Liver Injury): hepatotoxicity
• Skin sensitization: contact dermatitis
• Reproductive toxicity: teratogenicity

Prediction Models

Molecular Fingerprints + ML

ECFP4/6 (circular fingerprints 1024–2048 bits) + XGBoost/Random Forest. Fast, interpretable, good on small datasets.

Graph Neural Networks

Molecule as graph → GNN learns structural patterns. MPNN, AttentiveFP, D-MPNN (chemprop). On most TDC benchmarks GNN exceeds fingerprint+ML.

Multitask Learning

One model predicts 20+ ADMET properties simultaneously. Advantage: shared representations improve prediction for properties with small dataset through information from related tasks.

from chemprop import args, data, featurizers, models, train

# Chemprop — state-of-the-art for molecular ADMET
arguments = [
    '--data_path', 'admet_train.csv',
    '--dataset_type', 'regression',
    '--target_columns', 'solubility logP hERG_inhibition caco2_permeability',
    '--smiles_columns', 'smiles',
    '--epochs', '50',
    '--batch_size', '64',
    '--ffn_num_layers', '3',
    '--dropout', '0.1',
    '--save_dir', 'admet_model',
]
args.parse_train_args(arguments)
train.cross_validate(...)

Uncertainty Quantification

ADMET prediction: know not just value but model confidence. For molecules outside applicability domain — warning about unreliable prediction.

Methods: Monte Carlo Dropout, Deep Ensembles, Conformal Prediction. Conformal Prediction gives statistically rigorous prediction intervals.

Datasets

Data Problem: many biological datasets small and noisy. Transfer learning (pretraining on large chemical corpus → fine-tuning on specific task) helps with small datasets.

Applicability Domain

Model reliable only for molecules similar to training data. AD evaluation:

• Tanimoto similarity to nearest neighbors in training set
• Leverage hat matrix (Williams plot)
• k-NN distance in embedding space

When exiting AD → explicit warning "low confidence prediction".

Integration: REST API, Jupyter-friendly Python API, KNIME nodes for chemist workflows. Visualization: 2D property map with color-coding drug-likeness violations (Lipinski Rule of 5, Veber rules).