PhenoGenX: A Dual-Engine, Data-Driven Platform for HIV-1 Drug Resistance Interpretation Integrating Ensemble Machine Learning and Rule-Based Algorithms

Background : HIV-1 drug resistance (HIVDR) interpretation relies on expert rule-based algorithms that translate mutation–drug relationships into clinical categories but do not directly model phenotypic susceptibility and may have limited sensitivity to complex mutational patterns. We developed PhenoGenX (PGX), a dual-engine platform combining a phenotype-trained machine learning (ML) model with an extended rule-based system to integrate data-driven inference with expert knowledge for resistance interpretation in LMICs. Methods : From 45,039 HIV-1 clinical isolates, we curated 42,587 genotype–phenotype pairs with phenotypic fold-change (FC) measurements across 22 antiretroviral drugs. PGX integrates two independent engines: an ensemble ML model trained on mutation-level features and a rule-based interpreter derived from curated mutation knowledge bases. Model selection was guided by a Composite Resistance Performance Score (CRPS) incorporating predictive fit, error magnitude, rank correlation, categorical accuracy, and cross-validation stability. Ensemble predictions were calibrated to the PhenoSense assay scale and mapped to clinical resistance categories using safety-oriented cutoffs prioritizing minimization of very major errors. The ML engine was evaluated using an independent phenotypic dataset of 11,769 clinical isolates. The rule-based engine was benchmarked against Stanford HIVDB using 1,945 HIV-1 pol sequences (23,329 drug–sequence pairs) for NRTIs, NNRTIs, and PIs, with an additional 2,539 integrase sequences for INSTI validation. Findings : Ensemble ML models showed consistent predictive performance across drugs (R² range 0.50–0.95). Calibration improved agreement with measured phenotypes (mean log-scale correlation r=0.78), and optimized cutoffs achieved high diagnostic accuracy with low very major error rates. Most drugs achieved AUC values ≥0.80. The rule-based engine demonstrated high concordance with Stanford HIVDB (overall agreement 85.6%, weighted κ=0.72), with exact agreement exceeding 92% for integrase inhibitors. Interpretation : By integrating phenotype-calibrated ensemble ML with an extended rule-based interpreter, PhenoGenX provides a standardized framework for HIVDR interpretation that preserves biological plausibility and concordance with expert systems while maintaining a safety-weighted error profile. This approach may support HIV drug resistance surveillance and treatment decision-making where interpretation relies primarily on genotypic data in the next-generation sequencing era.