Improving Predictive\nEfficacy for Drug Resistance\nin Novel HIV‑1 Protease Inhibitors through Transfer Learning\nMechanisms

The human immunodeficiency virus presents a significant\nglobal\nhealth challenge due to its rapid mutation and the development of\nresistance mechanisms against antiretroviral drugs. Recent studies\ndemonstrate the impressive performance of machine learning (ML) and\ndeep learning (DL) models in predicting the drug resistance profile\nof specific FDA-approved inhibitors. However, generalizing ML and\nDL models to learn not only from isolates but also from inhibitor\nrepresentations remains challenging for HIV-1 infection. We propose\na novel drug-isolate-fold change (DIF) model framework that aims to\npredict drug resistance score directly from the protein sequence and\ninhibitor representation. Various ML and DL models, inhibitor representations,\nand protein representations were analyzed through realistic validation\nmechanisms. To enhance the molecular learning capacity of DIF models,\nwe employ a transfer learning approach by pretraining a graph neural\nnetwork (GNN) model for activity prediction on a data set of 4855\nHIV-1 protease inhibitors (PIs). By performing various realistic validation\nstrategies on internal and external genotype–phenotype data\nsets, we statistically show that the learned representations of inhibitors\nimprove the predictive ability of DIF-based ML and DL models. We achieved\nan accuracy of 0.802, AUROC of 0.874, and <i>r</i> of 0.727\nfor the unseen external PIs. By comparing the DIF-based models with\na null model consisting of isolate-fold change (IF) architecture,\nit is observed that the DIF models significantly benefit from molecular\nrepresentations. Combined results from various testing strategies\nand statistical tests confirm the effectiveness of DIF models in testing\nnovel PIs for drug resistance in the presence of an isolate.