Abstract:
Machine learning methods are increasingly used across a wide range of scientific disciplines,
including health-related and biomedical research, where data are often high dimensional,
heterogeneous, and limited in sample size. In health-related virology, clinical datasets often
combine patient-level information with high-dimensional viral sequence data, while the number of available samples is constrained by ethical and privacy considerations, clinical practice, and the cost of data collection. These characteristics lead to machine learning problems in high-dimensional, low-sample-size settings, in which model selection, data handling, and evaluation require particular care to avoid overfitting and misleading conclusions. This thesis examines the application of machine learning methods to three classification problems in health-related virology under such data constraints, with a focus on Hepatitis C Virus (HCV) and Human Immunodeficiency Virus (HIV). All three problems are characterized by low-sample-size settings, and the first two additionally involve high-dimensional data.
The first part of the thesis focuses on developing machine learning models to distinguish the
HCV 2k/1b recombinant variant from subtype 1b using sequence data from three non-structural proteins. This distinction is challenging for standard genotyping assays because the recombination breakpoint lies outside of the commonly analyzed genomic regions. The resulting models are integrated into the open access geno2phenoHCV web tool, enabling genotype prediction based on sequence data that is routinely measured for drug resistance determination instead of more expensive whole-genome sequencing data, and supporting molecular epidemiological analyses.
The second part of the thesis addresses the early prediction of multidrug class resistance in
HIV using viral sequence data. The prediction problem is formulated under multiple settings
that differ with respect to both the level of pre-existing drug class resistance and the time
interval between sample collection and the onset of multidrug class resistance. To evaluate
model behavior across these settings, feature importance analysis is used to identify amino
acid positions and mutations that contribute most strongly to the predictions, including both
previously reported and newly identified ones.
The third part of the thesis examines HLA footprint analysis by applying binomial generalized linear mixed models to next-generation viral sequencing data from HIV-infected patients. Several model formulations are considered, differing in their inclusion of demographic vari-
ables, individual-level HLA allele indicators, and HLA escape mutation pair information, while
the group structure capturing phylogenetic relatedness is kept fixed across models. In the
absence of a ground-truth set, model behavior is evaluated using a subsampling strategy that
assesses the effects of limited sample size, incomplete HLA representation, and variable read depth.
Virology