SVFX: a machine-learning framework to quantify the pathogenicity of structural variants

Abstract

A rapid decline in sequencing cost has made large-scale genome sequencing studies feasible. One of the fundamental goals of these studies is to catalog all pathogenic variants. Numerous methods and tools have been developed to interpret point mutations and small insertions and deletions. However, there is a lack of approaches for identifying pathogenic genomic structural variations (SVs). That said, SVs are known to play a crucial role in many diseases by altering the sequence and three-dimensional structure of the genome. Previous studies have suggested a complex interplay of genomic and epigenomic features in the emergence and distribution of SVs. However, the exact mechanism of pathogenesis for SVs in different diseases is not straightforward to decipher. Thus, we built an agnostic machine-learning-based workflow, called SVFX, to assign a "pathogenicity score" to somatic and germline SVs in various diseases. In particular, we generated somatic and germline training models, which included genomic, epigenomic, and conservation-based features for SV call sets in diseased and healthy individuals. We then applied SVFX to SVs in six different cancer cohorts and a cardiovascular disease (CVD) cohort. Overall, SVFX achieved high accuracy in identifying pathogenic SVs. Moreover, we found that predicted pathogenic SVs in cancer cohorts were enriched among known cancer genes and many cancer-related pathways (including Wnt signaling, Ras signaling, DNA repair, and ubiquitin-mediated proteolysis). Finally, we note that SVFX is flexible and can be easily extended to identify pathogenic SVs in additional disease cohorts.

StrVCTVRE: A supervised learning method to predict the pathogenicity of human genome structural variants

“…Surprisingly, several exonic features had relatively low importance, which may have been caused by the sparsity of SVs in our dataset that affect just a single exon. The low importance of TAD boundaries is counter to findings from a recent cancer SV impact predictor 15 and may reflect StrVCTVRE's focus on exonic SVs. Additionally, the low importance of deletion/duplication status suggests that on average, for exonic deletions and duplications, the region affected by an SV is more important than whether there was a gain or loss of genome content.…”

“…Indeed, 92% of SVs identified by sequencing are rare (AF < 1%), so the salient challenge is to distinguish rare pathogenic variants from rare benign variants 11 . Existing SV predictors have been trained and assessed on common benign variants 14,15 , which may cause them to instead rely on features that separate common from rare SVs and result in lower accuracy in clinical use 16 .…”

“…A grid search was performed to find the optimal hyperparameters by using a leave-one-chromosome-out cross validation strategy and validation only on ClinVar data, as described previously. The hyperparameters searched included: the max depth of a tree (5,10,15, No limit), max features considered at each split (1, 2, 3, 4), the minimum samples at each leaf node (1, 2, 4), the minimum samples required to split a node (2,4), the number of trees generated (500, 1000, 3000), and whether to use out-of-bag samples to estimate accuracy (True, False). Several combinations of features performed similarly well, and we chose one that performed well while unlikely to over-fit to the training data-max depth: 10, max features considered at each split: 1, minimum samples at each leaf node: 2, minimum samples required to split a node: 4, number of trees: 1,000.…”

“…SVScore then applies an aggregating operation across the CADD scores (the default is the mean of the top 10%), and this approach has shown promise in identifying SVs under purifying selection 14 . Another standalone predictor, SVFX, integrates multiple features, but focuses on somatic SVs in cancer and germline SVs in common diseases 15 . The majority of SVs identified by sequencing are rare (AF < 1%), so the salient challenge in resolving undiagnosed cases is to distinguish rare pathogenic variants from rare benign variants 11 .…”

“…The majority of SVs identified by sequencing are rare (AF < 1%), so the salient challenge in resolving undiagnosed cases is to distinguish rare pathogenic variants from rare benign variants 11 . Since these existing SV predictors have been trained and assessed on common benign variants 14,15 , they may rely on features that separate common from rare SVs and be less clinically useful 16 . In this manuscript, we introduce StrVCTVRE, the first method to provide a score based on a machine learning approach trained on sets of rare germline SVs from disease cases.…”