Machine Learning ADMET Prediction: Methods and Tools

What Is ADMET Prediction?

ADMET prediction refers to computational methods that estimate the pharmacokinetic and toxicological properties of drug candidates—specifically Absorption, Distribution, Metabolism, Excretion, and Toxicity. These properties determine whether a compound can reach its target in sufficient concentration and remain in the body long enough to be effective, without causing harmful side effects.

Machine learning-based ADMET prediction has become an indispensable tool in early drug discovery because approximately 60% of drug candidates fail in clinical trials due to inadequate pharmacokinetic properties. By identifying problematic compounds before they enter expensive preclinical and clinical stages, ADMET prediction saves both time and resources.

Data: ADMET in Drug Development

Source: Combined data from ADMETlab 2.0 benchmark and Nature Reviews Drug Discovery (2023).

The economic impact is significant: each late-stage clinical failure costs an estimated $100-800 million, and ADMET-related failures account for the majority of attrition after Phase I.

How: ML-Based ADMET Prediction Pipeline

Step 1: Data Preparation

1. Collect experimental ADMET data from databases:
   • ChEMBL (bioactivity data)
   • PubChem (compound properties)
   • FDA Orange Book (approved drug data)
   • TOX21/ToxCast (toxicity screening)
2. Standardize molecular structures (SMILES normalization, salt removal)
3. Handle missing data and class imbalance (SMOTE, weighted sampling)

Step 2: Feature Engineering

1. 2D descriptors: Molecular weight, logP, TPSA, hydrogen bond donors/acceptors, rotatable bonds
2. Fingerprints: ECFP4 (2048-bit), MACCS keys (166-bit)
3. 3D descriptors: Molecular volume, dipole moment, polar surface area
4. Graph-based: GNN embeddings capturing molecular topology

Step 3: Model Training

1. Classical ML: Random Forest, XGBoost, SVM (strong for smaller datasets)
2. Deep learning: Multi-layer perceptrons, GNNs, transformers (better for large datasets)
3. Multi-task learning: Predict multiple ADMET endpoints simultaneously
4. Cross-validation: Scaffold-based splitting to avoid data leakage

Step 4: Model Evaluation and Deployment

1. Evaluate on external test sets (not just cross-validation)
2. Deploy as web services or API endpoints
3. Provide uncertainty estimates (conformal prediction, ensemble methods)
4. Integrate into drug design workflows

Comparison: ADMET Prediction Tools

Summary: Key Takeaways

1. ADMET prediction is critical because 60% of clinical failures stem from poor pharmacokinetic properties.
2. ML models achieve AUC/R² values of 0.75-0.85 for most endpoints, with solubility and permeability being most predictable.
3. Graph-based representations and multi-task learning represent the current state of the art.
4. No single tool covers all ADMET endpoints—best practice is to use multiple tools and integrate predictions.
5. Uncertainty quantification is emerging as essential for actionable ADMET predictions in drug discovery.

References

1. Xiong, G. et al. "ADMETlab 2.0: an integrated online platform for accurate and comprehensive predictions of ADMET properties." Nucleic Acids Research (2024).
2. Daina, A. et al. "SwissADME: a free web tool to evaluate pharmacokinetics, drug-likeness and medicinal chemistry friendliness." Scientific Reports 7, 42717 (2017).
3. Pires, D.E.V. et al. "pkCSM: Predicting Small-Molecule Pharmacokinetic and Toxicity Properties Using Graph-Based Signatures." J. Med. Chem. 58, 4066-4072 (2015).
4. Mayr, A. et al. "DeepTox: toxicity prediction using deep learning." Frontiers in Environmental Science 3, 80 (2016).
5. Sun, M. et al. "ADMET-AI: a machine learning ADMET prediction platform." Nature Computational Science (2024).