(T1030-10-68) Machine Learning-Based Prediction of Aldehyde Oxidase Activity in Multiple Species

Purpose: Aldehyde oxidase (AO) plays a critical role in drug metabolism and clearance. It is well known that there is a large interspecies variation in AO activities which often leads to complexity during drug development. Moreover, due to diverse factors affecting AO activities, experimental determination of its activity could be very challenging. To address these issues, we employed machine learning techniques to predict AO activity in different species.

Methods: We selectively curated a dataset of ~1400 compounds from the NCATS ADME 5K library, focusing on those with the highest likelihood of being metabolized by AO. We tested the metabolic stability of these compounds in human, mouse, rat and monkey cytosol fractions. In order to account for the similarity in substrate specificity between Aldehyde Oxidase (AO) and xanthine oxidase (XO), an XO-inhibitor (allopurinol) was introduced to the incubations. Utilizing the compound half-life data, we categorized them into unstable (half-life < 30 min) and stable (half-life > 30 min), converting the information into binary data representing AO activity. From the SMILES representation of each compound, we computed 2048 fingerprints and 1623 molecular descriptors, incorporating them into the training dataset. We employed the classifier and regressor of three machine learning models: Decision tree, Random Forest, and XGBoost, with random split 5-fold cross-validation.

Results: Among these three models, the regressor model, using real-number data for prediction, consistently demonstrated low accuracies below 20% in the test set. In contrast, the classifier model with binary data showed a high accuracy of 88%, effectively recognizing compound stability against AO.

Conclusion: In summary, our study shows that machine learning-based classifiers can accurately predict AO-stability based on its molecular properties. This approach helps identify compounds that may be quickly cleared by AO, making it a valuable tool for optimizing drug candidates and lead selection. Predicting AO activity early in drug discovery can reduce the chances of discontinuing promising drug candidates during drug development and clinical trials. Ultimately, it will aid drug development process more efficient.

References: 1. Nenad Manevski et al, Metabolism by Aldehyde Oxidase: Drug Design and Complementary Approaches to Challenges in Drug Discovery, Journal of Medicinal Chemistry 2019 62 (24), 10955-10994
2. Coelho C etal., Structural insights into xenobiotic and inhibitor binding to human aldehyde oxidase. Nat Chem Biol. 2015.11(10):779-83

Structure of aldehyde oxidase. There is an active site where the substrate binds and a site where the cofactor binds. The structures of the homology models of the four species are nearly identical.

Aldehyde oxidase active site and loop structure according to species. We created a missing loop, primate has a long loop and rodent has a relatively short loop.

Comparison of machine learning models. SMILES fingerprint and descriptor were used as features. In the results between the classifier and the regressor model, the result of the classifier is significantly better.