A-G.32: SIEVE: Sparse Interpretable Exome Variant Explainer
Deep learning methods for case-control variant discovery depend on functional annotations, treat variants as unordered sets, and usually restrict the analysis to either rare or common variants. These choices leave fundamental questions unaddressed: would the model discover different biology without annotations? Does genomic position carry signal that permutation-invariant architectures discard? What is missed when common and rare variants are not modelled jointly? We present SIEVE, a framework addressing five methodological gaps: (1) sinusoidal positional encoding enables learning of position-dependent relationships; (2) attribution regularisation encourages sparse, stable variant rankings during training; (3) an annotation-ablation protocol trains the same architecture at four different annotation levels comparing discoveries systematically; (4) frequency-agnostic processing analyses the full allele-frequency; (5) epistatic interactions are intrinsically estimated from the attention layer at both the individual-level and variant-level, as well as by collapsing variant attributions at gene-level. We applied SIEVE to a coronary artery disease whole-exome cohort from the Ottawa Heart Genomics Study. Classification performance is consistent with expectations from exome variants, and null-baseline permutation confirms that attributions exceed chance. Pairwise Jaccard overlap across ablation-annotation levels is low (0.03–0.07), confirming that each level discovers genuinely different biology. Annotation-free models identify genes implicated in TNF signalling, ubiquitin-proteasome regulation, and endothelial apoptosis, i.e. pathways with established cardiovascular role. Adding positional encoding alone shifts top-ranked genes toward HIF-1 signalling, FOXO pathway, and mitochondrial import. Top-ranked attributions span the full MAF spectrum, confirming joint modelling of rare and common variation. The workflow is implemented as a reproducible open-source Nextflow pipeline.
Co-authors: Davide Bagordo, Cezar Grigorean
Contact Attendee
Warning: Attempt to read property "user_email" on string in /home/1276969.cloudwaysapps.com/ydbgzhdjeq/public_html/wp-content/plugins/my-conference-now/functions.php on line 2826