Using drug exposure for predicting drug resistance - A data-driven genotypic interpretation tool.

Area of Science:

• Virology
• Computational Biology
• Epidemiology

Background:

• Antiretroviral therapy (ART) success relies on predicting treatment outcomes.
• Patient treatment history and HIV-1 genotypes are key predictors.
• Treatment history is often incomplete or inaccurate in patients with complex care needs.

Purpose of the Study:

• To develop and validate statistical models for predicting past antiretroviral drug exposure based on current HIV-1 genotypes.
• To assess the models' performance in predicting drug resistance and therapy success.
• To provide a freely accessible tool for optimizing ART selection.

Main Methods:

• Training statistical models on a large dataset of 63,742 HIV-1 nucleotide sequences with known treatment histories.
• Utilizing 6,836 genotype-phenotype pairs (GPPs) to correlate genotypic predictions with phenotypic resistance.
• Evaluating model performance using ROC-AUC for drug exposure and therapy success prediction.

Main Results:

• Models achieved mean ROC-AUC of 0.78 and 0.76 for predicting drug exposure.
• Correlation to phenotypic resistance was 0.51 (PhenoSense) and 0.46 (Antivirogram).
• Therapy success prediction yielded ROC-AUC of 0.71 and 0.63, comparable or superior to existing methods.

Conclusions:

• Data-driven models accurately infer past drug exposure and predict resistance from current HIV-1 genotypes.
• These models offer a valuable alternative when patient treatment history is unavailable.
• The freely available tool (www.geno2pheno.org) aids in selecting optimal antiretroviral therapy and predicting treatment outcomes.