Self-Supervised Learning with Masked Images for Structural Variant Analysis in Short-Read Genome Sequencing

Structural variants (SVs), including deletions and insertions, are a major source of human genetic variation and play important roles in both rare and common disease. Despite their biological significance, accurate SV genotyping from short-read genome sequencing (SRS) data remains challenging due to limited read lengths and reliance on indirect sequencing signals. To address these limitations, this work presents BASILISC (BEiT-based Analysis with Self-supervised Image Learning In Structural Variant Classification), a transformer-based framework for detecting structural variants in SRS data using self-supervised learning. Building on prior image-based SV genotyping approaches, this work introduces a novel pipeline that converts aligned short-read sequencing data into multi-channel genomic pileup images capturing biologically meaningful signals, including read depth, split reads, discordant read pairs, mapping quality, strand orientation, and allele support. These images are compressed into discrete token representations using a discrete variational autoencoder (dVAE), enabling efficient downstream modeling with transformer architectures. The core methodological contribution is the application of self-supervised masked image modeling (MIM), adapted from the BEiT vision transformer framework, to large collections of unlabeled genomic images. Through this pretraining strategy, the model learns generalizable representations of genomic structure without requiring extensive manually labeled SV datasets. The pretrained transformer is then fine-tuned with a classification head to predict the presence or absence of structural variants at candidate genomic loci. The BASILISC framework is evaluated using short-read data from the 1000 Genomes Project, with high-confidence SV labels derived from long-read sequencing assemblies provided by the Human Genome Structural Variation Consortium (HGSVC2). Results demonstrate that BASILISC achieves competitive or superior performance relative to established SV genotyping tools across deletions, insertions, and mixed variant regions, with particularly strong precision. Qualitative analyses using genome browser visualizations further show that the model captures biologically interpretable SV signals, even in repetitive or noisy genomic contexts. Overall, this work demonstrates that self-supervised, transformer-based representation learning is a powerful approach for structural variant analysis in short-read sequencing data. By leveraging abundant unlabeled genomic data and modeling long-range dependencies through attention mechanisms, BASILISC establishes a scalable foundation for future SV genotyping systems, with potential extensions to full genotype prediction, additional variant classes, and integration with simulation-based pipelines.