Accurate and interpretable ADMET prediction: Integrating structural, geometric, and global molecular context...

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Accurate prediction of absorption, distribution, metabolism, excretion, and toxicity (ADMET) is crucial for successful drug discovery.
• Existing computational models often rely on single molecular representations, limiting their ability to capture complex structure-property relationships.

Purpose of the Study:

• To develop a comprehensive deep learning framework (LGSM) for integrative molecular property prediction.
• To improve the accuracy of ADMET profiling in early drug discovery by fusing diverse molecular representations.

Main Methods:

• LGSM integrates 2D topological, 3D conformational, and sequence-derived semantic molecular representations.
• A language model encodes SMILES sequences for global molecular context, complementing local functional group information.
• The framework fuses these multi-view representations to capture both local chemical environments and spatial relationships.

Main Results:

• LGSM demonstrated superior performance in predicting ADMET properties across four public datasets, even under strict scaffold splits simulating real-world drug design.
• The model consistently outperformed standard computational baselines.
• LGSM provided chemical interpretability by identifying key substructures associated with metabolic liabilities and transporter interactions.

Conclusions:

• LGSM offers a powerful and interpretable deep learning approach for accurate ADMET prediction.
• The framework's ability to integrate multi-view molecular data enhances its utility in guiding rational lead optimization and reducing late-stage drug discovery failures.