A molecular modeling and machine learning-based approach for the identification of potential human ALOX15 inhibitors.

CONTEXT: Arachidonate 15-lipoxygenase (ALOX15) metabolises polyunsaturated fatty acids into lipid mediators that regulate inflammation. Depending on the cell type and substrate, ALOX15 influences both pro-inflammatory and pro-resolving pathways. Its dysregulation is associated with asthma, atherosclerosis, neurodegeneration, ferroptosis, inflammation, oxidative stress and disrupted lipid homeostasis. Therefore, targeting ALOX15 offers a promising strategy for modulating pathological lipid oxidation.

METHOD: In this study, a hybrid ligand-based and structure-based machine learning (ML) framework was developed that combines physicochemical and 3D ligand features with interacting features (including SIFts) derived from protein-ligand interactions for the discovery of putative ALOX15 inhibitors. To identify potent human ALOX15 inhibitors, predictive machine learning (ML) models were developed, a dataset was obtained from BindingDB and multiple molecular representations (1-D, 2-D, 3-D descriptors, fingerprints and post-docking features) were computed. Two classes of ML models, Type A and Type B, were developed using 8 regression algorithms. The best Type A model, XGBoost, achieved training R2 of 0.915 (RMSE = 0.299; MAE = 0.224) and test R2 of 0.688 (RMSE = 0.581; MAE = 0.456), whereas the best Type B model, Random Forest, attained training R2 of 0.938 (RMSE = 0.254; MAE = 0.195) and test R2 of 0.659 (RMSE = 0.607; MAE = 0.473). Over 5 million compounds from various databases were screened, and five compounds were identified. ADME predictions and molecular dynamics simulations suggest favourable pharmacokinetic properties and stable binding within the active site of human ALOX15. These compounds are proposed as putative inhibitors and require further experimental validation.