Predicting ADMET Properties from Molecule SMILE: A Bottom-Up Approach Using Attention-Based Graph Neural Networks

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in pharmacology

Background:

• Accurate prediction of Absorption, Distribution, Metabolism, Excretion, and Toxicity (ADMET) properties is vital for successful drug development.
• Early assessment of ADMET properties minimizes late-stage failures and reduces development costs.
• Current methods for ADMET prediction can be computationally intensive and require manual feature engineering.

Purpose of the Study:

• To develop and validate an innovative attention-based graph neural network (GNN) model for predicting ADMET properties.
• To demonstrate the model's ability to directly utilize molecular structures from Simplified Molecular Input Line Entry System (SMILE) notation.
• To provide a computationally efficient alternative to traditional methods for ADMET property prediction.

Main Methods:

• Utilized an attention-based graph neural network (GNN) architecture.
• Employed a bottom-up approach, processing molecular information from substructures to the entire molecule.
• Represented molecules directly from SMILE notation, bypassing the need for explicit molecular descriptors.
• Validated the model on six benchmark datasets for both regression (lipophilicity, aqueous solubility) and classification (CYP inhibition) tasks.

Main Results:

• The attention-based GNN model demonstrated significant effectiveness in predicting various ADMET properties.
• The model successfully performed regression tasks, including lipophilicity and aqueous solubility prediction.
• The model achieved strong performance in classification tasks, predicting inhibition of key cytochrome P450 enzymes (CYP2C9, CYP2C19, CYP2D6, CYP3A4).
• The approach eliminated the need for computationally expensive retrieval and selection of molecular descriptors.

Conclusions:

• The developed attention-based GNN offers a powerful and efficient tool for predicting ADMET properties.
• This method facilitates high-throughput screening and early-stage assessment in drug discovery.
• The model enhances the likelihood of success for candidate drugs by enabling proactive identification of potential liabilities.