[Paper Review] Accurate ADMET Prediction with XGBoost

Hao Tian, Rajas Ketkar+1 more|arXiv (Cornell University)|Apr 15, 2022

Computational Drug Discovery Methods4 citations

TL;DR

This paper presents XGBoost-based ADMET prediction using an ensemble of molecular fingerprints and descriptors, achieving state-of-the-art performance on the Therapeutics Data Commons (TDC) ADMET benchmark. The model ranks first in 11 out of 22 tasks and top 3 in 21, with results accessible via the publicly available ADMETboost web server at https://ai-druglab.smu.edu/admet.

ABSTRACT

The absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties are important in drug discovery as they define efficacy and safety. In this work, we applied an ensemble of features, including fingerprints and descriptors, and a tree-based machine learning model, extreme gradient boosting, for accurate ADMET prediction. Our model performs well in the Therapeutics Data Commons ADMET benchmark group. For 22 tasks, our model is ranked first in 18 tasks and top 3 in 21 tasks. The trained machine learning models are integrated in ADMETboost, a web server that is publicly available at https://ai-druglab.smu.edu/admet.

Motivation & Objective

• To develop a high-accuracy machine learning model for predicting ADMET properties in drug discovery.
• To address the challenge of inconsistent model evaluation due to disparate datasets by using the unified TDC benchmark.
• To integrate diverse molecular features—fingerprints and descriptors—into a single robust predictive model.
• To improve prediction performance over existing methods by leveraging XGBoost's ensemble learning and feature selection capabilities.
• To provide a publicly accessible web server for rapid ADMET prediction using trained models.

Proposed method

• The model uses an ensemble of six featurizers: MACCS, Extended Connectivity, Mol2Vec, PubChem, Mordred, and RDKit fingerprints/descriptors.
• Molecular SMILES are converted into feature vectors using these six featurizers to capture structural and physicochemical properties.
• XGBoost is trained on the combined feature set, using a regularized objective function to minimize overfitting.
• The objective function combines a loss term and a regularization term Ω(f) = γT + (λ/2)Σω², controlling tree complexity.
• Hyperparameters are tuned via randomized grid search cross-validation to optimize performance.
• Trained models are deployed on a web server using SMU's high-performance computing resources.

Experimental results

Research questions

• RQ1 Can an ensemble of diverse molecular fingerprints and descriptors improve ADMET prediction accuracy?
• RQ2 How does XGBoost compare to other machine learning models on the standardized TDC ADMET benchmark?
• RQ3 Which molecular features contribute most significantly to ADMET prediction performance?
• RQ4 Can a single unified model achieve top-tier performance across diverse ADMET tasks?
• RQ5 How does the performance of XGBoost compare when using only Mordred descriptors versus all features?

Key findings

• The XGBoost model achieved the highest rank in 11 out of 22 ADMET tasks on the TDC benchmark.
• The model ranked in the top 3 for 21 out of 22 tasks, demonstrating consistent high performance.
• Mordred descriptors were the most important feature type across all ADMET tasks, followed by Mol2Vec and circular fingerprints.
• Using only Mordred descriptors, the model outperformed the full-feature model in three tasks (HIA, AqSol, PPBR), but underperformed in metabolism and excretion tasks.
• The web server ADMETboost, hosted at https://ai-druglab.smu.edu/admet, delivers predictions in seconds with visualized 2D and 3D molecular structures.
• The study demonstrates that XGBoost effectively learns from a diverse set of molecular features, enabling superior ADMET prediction.