Fig. 1: SVision-pro overview.

a, Overview of the sequence-to-image representation module in SVision-pro. SVision-pro sketches the structures of a candidate SV locus and renders ACTs (above) into the sparse image regions. The ACT is generated from mapped alignments by the three-channel RGB augmentation (below). Dup., duplicated-matching; Rev., reversed-matching; For., forward-matching. b, Overview of the comparative recognition module in SVision-pro. The neural-network-based instance segmentation framework outputs a segmentation mask, providing intuitive SV types (above). By comparative genotyping analysis of the colored regions in the upper and lower panels (below), we can determine the SV differences between case and control genomes. c, Neural-network model training and selection strategy of SVision-pro. SVision-pro was trained with five basic SV subcomponent types along with wild type (identical to reference genome) and was able to recognize CSVs with several internal subcomponents (above). To select an efficient instance segmentation models (red solid circle), we leveraged three factors: validation accuracy, parameter size and interpretability. d, Attribution maps of the Lite-Unet model. Pixels relevant for a certain prediction class are highlighted. DEL, deletion; DUP, duplication; INV, inversion; INS, insertion; invDUP, inverted-duplication; WT, wild type; R, red; G, green; B, blue; w₁, w₂ and w_n, parameter weights.