interpretable-ADMET

interpretable-ADMET predicts and interprets Absorption, Distribution, Metabolism, Excretion, and Toxicity (ADMET) properties for small molecules to support lead optimization in drug discovery.

Key Features:

• Predictive Capability: 90 qualitative classification models and 28 quantitative regression models built with graph convolutional neural networks (GCNNs) and graph attention networks (GATs) to predict 59 ADMET-associated properties.
• Extensive Database: over 250,729 entries covering 59 types of ADMET properties for approximately 80,167 chemical compounds.
• Interpretability: uses gradient-weighted class activation maps (Grad-CAM) to identify molecular substructures that influence specific ADMET predictions.
• Optimization Module: generates virtual candidate molecules using matched molecular pair (MMP) rules to propose chemical modifications that alter predicted ADMET properties.

Scientific Applications:

• Lead optimization: supports lead optimization by predicting ADMET profiles and recommending matched molecular pair (MMP)-based chemical modifications.
• Rational compound design: enables attribution of ADMET effects to specific substructures via Grad-CAM to guide modification of pharmacokinetic profiles.

Methodology:

Computational methods comprise 90 qualitative classification models and 28 quantitative regression models implemented with graph convolutional neural networks (GCNNs) and graph attention networks (GATs); interpretability via gradient-weighted class activation maps (Grad-CAM); and candidate generation via matched molecular pair (MMP) rules.