ETD

• fine-mapping
• genetics
• machine learning
• sequence-to-function

Non-coding genetic variation accounts for the majority of trait- and disease-associated loci, yet statistical fine-mapping, though effective at narrowing association signals to small sets of candidate variants, cannot establish which of these variants is functionally causal, nor can it test whether multiple variants jointly shape transcription. This thesis asks whether sequence-to-function deep learning models can supply this missing functional evidence in the simplest possible multivariate setting: pairs of linked non-coding variants. We evaluate two complementary S2F architectures, AlphaGenome, a large multitask transformer trained on genome-wide functional tracks, and PARM, a compact convolutional model trained on massively parallel reporter assay data, on two tasks: recovering the correct direction of fine-mapped GTEx breast-tissue eQTLs when a linkage-disequilibrium partner is introduced alongside the prioritized variant, and detecting experimentally validated cis-regulatory epistasis between variant pairs. AlphaGenome, in a substantial subset of diplotypes, recovered the correct effect direction through a linked variant when the fine-mapped variant itself showed negligible predicted perturbation of gene expression; one such case, at the RAVER2 locus, was independently corroborated by PARM in-silico mutagenesis and by orthogonal chromatin and promoter annotations, converging on a reassignment of functional weight within the credible set. For genuine non-additive interactions, PARM recovered a meaningful fraction of experimentally measured epistatic pairs, with cell-type-dependent accuracy, and its base-pair-resolution output could mechanistically explain the recovered interactions. AlphaGenome, predicted near-complete additivity in both tested configurations and failed to reproduce the experimental interactions, aside from partial agreement on the single strongest experimental signal. Together, these results indicate that S2F models can provide orthogonal, mechanistically grounded functional evidence to complement statistical fine-mapping, despite both models remaining biased toward additivity, underscoring that closing the gap toward reliable multivariant prediction will require training data explicitly enriched for genuine variant interactions rather than larger models built on the same reference-genome assumptions.