Optimal Transport Based Graph Kernels for Drug Property Prediction.

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Optimizing pharmaceutical agent properties (ADMET) is crucial but challenging due to experimental costs and data limitations.
• Computational and predictive tools, including machine learning and graph-based methods, are increasingly vital in early drug development.
• Existing methods face challenges in accurately and efficiently predicting complex ADMET profiles.

Purpose of the Study:

• To develop and evaluate novel graph kernels based on optimal transport (OT) theory for predicting drug ADMET properties.
• To assess the performance of OT-based graph kernels against state-of-the-art deep learning models.
• To highlight the interpretability, adaptability, and generalizability advantages of the proposed approach.

Main Methods:

• Utilized optimal transport (OT) theory to construct three graph kernels.
• Employed graph matching to generate a similarity matrix.
• Integrated the similarity matrix into a predictive modeling framework for ADMET properties.

Main Results:

• OT-based graph kernels demonstrated superior performance across 19 ADMET datasets.
• Outperformed state-of-the-art graph deep learning models in 9 out of 19 datasets.
• Showed competitive results in 2 additional datasets, surpassing even advanced Graph Neural Networks (GNNs) in some cases.

Conclusions:

• The novel OT-based graph kernels offer a highly effective and competitive approach for ADMET property prediction.
• These methods provide advantages over traditional graph neural networks, including enhanced interpretability, adaptability, and generalizability.
• The study validates the potential of OT theory in advancing computational drug discovery and development.