Machine Learning Identifies Potential Accessory Resistance-Associated Mutations in HIV-1 Integrase

Although integrase strand transfer inhibitors (INSTIs) have a high genetic barrier to resistance, cases of virological failure continue to emerge, sometimes in the absence of major resistance-associated mutations. Conventional genotypic and phenotypic resistance testing is costly, time-intensive, and remains limited in its ability to identify novel resistance pathways in resource-limited settings. Machine learning offers a scalable approach to uncover previously unrecognized resistance-associated patterns in HIV-1 genomic data.

We analyzed 41,247 publicly available HIV-1 integrase sequences from ART-naïve and ART-experienced individuals using interpretable machine learning algorithms. Random Forests (RF), Support Vector Machines (SVM), Logistic Regression (LR), and Gradient Boosting Machines (GBM) classifiers were trained to distinguish treatment status based solely on HIV-1 integrase mutation profiles. RF outperformed other classifiers, with an accuracy of 0.94 and an AUC of 0.98 when including known INSTI resistance mutations. Top-ranking mutations identified by the RF classifier, including S283G, T112V, D278A, K136Q, T125A, V201I, V31I, T124A, I72V, K14R, A265V, G134N, D167E, and I135V, were significantly more prevalent in ART-experienced sequences. Structural analysis showed that most mutations potentially destabilize the three-dimensional structure of HIV-1 integrase. Relative risk (RR) analysis identified nine significant co-occurring mutation pairs with major INSTI resistance mutations, including G118R-D278A (RR = 1.9), G140C-T124A (RR = 2.2), and I135V-Y143A (RR = 2.3). These associations clustered within established resistance pathways (G118R, Q148/G140, Y143, and N155).

Interpretable machine learning effectively identifies potential accessory resistance-associated mutations in HIV-1 integrase. These mutations may contribute to INSTI resistance via epistatic interactions with known major resistance mutations. Experimental validation and longitudinal studies are needed to clarify their impact on treatment outcomes and on the evolution of ART resistance.