Hertzberg, J. | bioRxiv Science

bioRxiv2025

Accurate detection of pathogenic structural variants guided by multi-platform comparison

Structural variants (SVs) are a common cause of human diseases and greatly contribute to inter-individual variability. Their detection represents a significant challenge due to their diversity, size, and enrichment in repetitive regions. With the use of high quality long-read technologies, the majority of these challenges can now be mitigated. However, many downstream applications ranging from clinical diagnostics and genome-wide association studies to large-scale aggregation of population variants continue to rely on short-read-sequencing data due to its high through-put and cost-effectiveness. Thus, the challenges of short-read sequencing SV detection remain a constant and relevant obstacle. We created dicast, a machine-learning method to improve the identification of true-positive SVs from short-read sequencing using genomic context and alignment features. Dicast is driven by a novel and comprehensive benchmark call-set created through the combination of several sequencing technologies and rigorous manual curation. This benchmark set served as the basis for a systematic evaluation of five sequencing platforms and fifteen SV detection methods across different SV classes, sizes, and genomic contexts, enabling us to quantify the strengths and weaknesses of each and inform the training of our model. Leveraging these insights, dicast outperforms state-of-the-art short-read-based callers and consensus approaches, identifying considerably more true-positive variants while maintaining a high precision. We also demonstrate the methods applicability in diagnostic scenarios using putative pathogenic candidates in a limb malformation cohort, as well as known pathogenic variants in atrial fibrillation and neuromuscular disease cohorts. Dicast identifies all known pathogenic variants and 20% more manually confirmed candidate deletions than a consensus approach, and therefore can be used to reliably reduce time-consuming manual inspection in diagnostics. Graphical Abstract O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=153 SRC="FIGDIR/small/655285v1_ufig1.gif" ALT="Figure 1"> View larger version (50K): org.highwire.dtl.DTLVardef@1c1702aorg.highwire.dtl.DTLVardef@aaecddorg.highwire.dtl.DTLVardef@1a4de08org.highwire.dtl.DTLVardef@ed0811_HPS_FORMAT_FIGEXP M_FIG C_FIG

bioinformatics↗

bioRxiv2020

TADA - a Machine Learning Tool for Functional Annotation based Prioritisation of Putative Pathogenic CNVs

The computational prediction of disease-associated genetic variation is of fundamental importance for the genomics, genetics and clinical research communities. Whereas the mechanisms and disease impact underlying coding single nucleotide polymorphisms (SNPs) and small Insertions/Deletions (InDels) have been the focus of intense study, little is known about the corresponding impact of structural variants (SVs), which are challenging to detect, phase and interpret. Few methods have been developed to prioritise larger chromosomal alterations such as Copy Number Variants (CNVs) based on their pathogenicity. We address this issue with TADA, a method to prioritise pathogenic CNVs through manual filtering and automated classification, based on an extensive catalogue of functional annotation supported by rigorous enrichment analysis. We demonstrate that our machine-learning classifiers for deletions and duplications are able to accurately predict pathogenic CNVs (AUC: 0.8042 and 0.7869, respectively) and produce a well-calibrated pathogenicity score. The combination of enrichment analysis and classifications suggests that prioritisation of pathogenic CNVs based on functional annotation is a promising approach to support clinical diagnostic and to further the understanding of mechanisms that control the disease impact of larger genomic alterations.View Full Text

genetics↗