Optimal Transport Based Graph Kernels for Drug Property Prediction

Abstract

Objective: The development of pharmaceutical agents relies heavily on optimizing their pharmacodynamics, pharmacokinetics, and toxicological properties, collectively known as ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity). Accurate assessment of these properties during the early stages of drug development is challenging due to resource-intensive experimental evaluation and limited comprehensive data availability. To overcome these obstacles, there has been a growing reliance on computational and predictive tools, leveraging recent advancements in machine learning and graph-based methodologies. This study presents an innovative approach that harnesses the power of optimal transport (OT) theory to construct three graph kernels for predicting drug ADMET properties. This approach involves the use of graph matching to create a similarity matrix, which is subsequently integrated into a predictive model. Results: Through extensive evaluations on 19 distinct ADMET datasets, the potential of this methodology becomes evident. The OT-based graph kernels exhibits exceptional performance, outperforming state-of-the-art graph deep learning models in 9 out of 19 datasets, even surpassing the most impactful Graph Neural Network (GNN) that excels in 4 datasets. Furthermore, they are very competitive in 2 additional datasets. Conclusion: Our proposed novel class of OT-based graph kernels not only demonstrates a high degree of effectiveness and competitiveness but also, in contrast to graph neural networks, offers interpretability, adaptability and generalizability across multiple datasets.

Keywords: ADMET properties; Optimal tranpsort; graph kernels; graph matching; wasserstein distance.

Figure 1.

This figure depicts the method ranks across 19 datasets, with the top-performing methods positioned on the left (ranked one) and the lower-performing ones gradually appearing towards the right (ranked ten).