#31 · Genomics

Computational methods resolve hard-to-call and hard-to-interpret variants in human genomes

The common thread is computational resolution of variant-level ambiguity in human genomes, where variants are difficult either to detect or to interpret. A preprint uses cohort summary statistics from gnomAD to identify non-reference processed pseudogene insertion artifacts, flagging 1498 processed pseudogenes, 1268 of them novel. CanDrivR-CS trains 50 cancer-specific gradient boosting models on ICGC data to separate recurrent from rare somatic missense variants using DNA shape features. SyMetrics integrates splicing, stability, conservation and allele-frequency predictors to score the pathogenicity of synonymous variants. SDrecall improves variant detection within segmental duplications by realigning reads via homolog networks and phasing haplotypes. What is new is the extension of computational frameworks, including machine learning and cohort-scale analysis, to silent coding changes, somatic missense variants and duplicated sequence.