Bioinformatics Pipelines for Rare Variant Detection

Abstract

Rare genetic variants with minor allele frequencies below one percent account for a substantial fraction of disease heritability yet are notoriously difficult to detect reliably from short-read sequencing data due to their resemblance to sequencing errors. We evaluated 212 bioinformatics pipelines for rare variant detection deployed across genomics centres in Austria, Switzerland, and France between 2016 and 2021, spanning five pipeline architectures: conventional GATK-based germline callers, deep-learning variant callers, long-read-aware hybrid pipelines, population-informed joint calling frameworks, and somatic rare variant detectors for tumour sequencing. A Rare Variant Detection Pipeline Index (RVDPI) was constructed from five sub-scores -- sensitivity for true rare variants, precision against false positives, performance on difficult genomic regions, scalability to cohort-level calling, and clinical-grade validation depth -- with weights from regression against sustained adoption into diagnostic or research sequencing workflows. RVDPI correlated with adoption at r = +0.84 and discriminated adopted from non-adopted pipelines with an AUC of 0.884. Deep-learning callers scored highest (mean RVDPI 0.826), while somatic detectors trailed at 0.598. Only 35.8 percent of pipelines exceeded the 0.75 threshold. Sensitivity carried the largest regression weight (beta = +0.278), followed by precision (beta = +0.232).