A New Classification Method Based on Dynamic Ensemble Selection and its Application to Predict Variance Patterns in HIV-1 Env

ABSTRACT Therapeutics that target the envelope glycoproteins (Envs) of human immunodeficiency virus type 1 (HIV-1) effectively reduce virus levels in patients. However, due to mutations, new Env variants are frequently generated, which may be resistant to the treatments. The appearance of such sequence variance at any Env position is seemingly random. A better understanding of the spatiotemporal patterns of variance across Env may lead to the development of new therapeutic strategies. We hypothesized that, at any time point in a patient, positions with sequence variance are clustered on the three-dimensional structure of Env. To test this hypothesis, we examined whether variance at any Env position can be predicted by the variance measured at adjacent positions. Sequences from 300 HIV-infected patients were applied to a new algorithm we developed. The k-best classifiers (KBC) method is a dynamic ensemble selection technique that identifies the best classifier(s) within the neighborhood of a new observation. It applies bootstrap resampling to generate out-of-bag samples that are used with the resampled set to evaluate each classifier. For many positions of Env, primarily in the CD4-binding site, KBC accurately predicted variance based on the variance at their adjacent positions. KBC improved performance compared to the initial learners, static ensemble, and other baseline models. KBC also outperformed other algorithms for predicting variance at multi-position footprints of therapeutics on Env. These understandings can be applied to refine models that predict future changes in HIV-1 Env. More generally, we propose KBC as a new high-performance dynamic ensemble selection technique.