Multi-Model Consensus Learning With Metaheuristic Optimization and Transformer Encoding for HIV-1 Biomarker Identification
Abstrak
The heterogeneity of patient responses to antiretroviral therapy (ART) remains a critical challenge in HIV-1 management, yet robust transcriptomic biomarkers distinguishing ART responders from non-responders have not been established. This study presents a novel multi-model consensus learning framework integrating metaheuristic feature selection, transformer-based encoding, and ensemble classification to identify high-confidence gene expression biomarkers. Gene expression data from 406 HIV-1-positive patients across five GEO datasets, comprising 19,033 features, were analyzed. Three metaheuristic algorithms—Genetic Algorithm (GA), Particle Swarm Optimization (PSO), and Ant Colony Optimization (ACO)—were evaluated for feature selection, followed by a custom GeneTransformer encoder and five diverse classifiers (Random Forest, SVM, Logistic Regression, XGBoost, MLP) with consensus-based biomarker ranking. ACO achieved a 99.74% reduction in feature space, yielding a 50-gene signature that attained 95.1% accuracy and 0.995 AUC on internal testing, while external validation on an independent cohort (GSE18233) produced an F1-score of 0.771, confirming generalizability. Consensus analysis identified six biomarkers (EIF4G1, ANKHD1, CNGB1, BTF3, GGA3, and PPM1B) with published mechanistic roles in HIV-1 pathogenesis such as host translational shutoff, apoptosis, tissue degeneration, and viral particle release. The framework delivers a compact, externally validated gene panel and a reproducible pipeline, offering a quantitative foundation for developing prognostic assays to guide personalized HIV-1 management, pending further prospective validation.
Identifier
- DOI: https://doi.org/10.1109/access.2026.3717647
- OpenAlex: https://openalex.org/W7171500551