Accurate and Robust Characterization of Structural Variants at Low Coverages with ARCLID

Structural variants (SVs) shape genome diversity and underlie human diseases, yet accurate detection remains difficult, especially at low sequencing coverages. Most existing callers are optimized for deep sequencing and lose sensitivity as coverage falls, limiting their use in studies that sequence genomes at shallow depth. We present ARCLID, a deep learning-based SV caller that reframes SV detection as an object detection task. We benchmarked ARCLID against six well-established callers on four samples and challenging medically relevant genes, across coverages from 28x down to 5x, multiple SV sizes, and both relaxed and strict breakpoint matching. At 5x, ARCLID achieved the highest F1 score of all evaluated callers on every sample, exceeding the second-best tool by up to 8%. This advantage widened under strict breakpoint matching, indicating genuine variant recovery with accurate breakpoint placement. At higher coverages, ARCLID remains competitive with the strongest current tools. By preserving accuracy at low depth, it makes reliable SV discovery feasible from shallow data, enabling lower sequencing cost, time, and storage for population-scale and clinical genomics. We additionally curated SV benchmarking regions for HG00733 and NA19240 that complement the widely used HG002 GIAB benchmark regions, supporting evaluation of SV callers across ancestrally distinct genomes.