IUPAC Consensus References Improve Short-Read Variant Detection in Clinically Challenging Regions: A Stratified Benchmarking Study with BurdenBench
MotivationReference bias depresses variant detection in low-mappability regions, segmental duplications and the major histocompatibility complex (MHC) -- precisely the regions of greatest clinical relevance. Existing benchmarks rely on aggregate precision, recall and F1 metrics that obscure the absolute true-positive and false-positive counts that determine laboratory workload. No study has systematically evaluated IUPAC consensus references for short-read whole-genome sequencing (WGS) variant calling across Genome in a Bottle (GIAB) stratifications, multiple allele-frequency thresholds and multiple variant callers. ResultsWe aligned 30x WGS from three GIAB samples to IUPAC consensus references (allele frequency [≥]10% and [≥]30%) using the ambiguity-aware aligner novoAlign, benchmarking against BWA-MEM/GRCh38 and novoAlign/GRCh38 baselines across BCFtools, FreeBayes and GATK HaplotypeCaller. SNV recall increased by 3.1-3.9 percentage points (pp) in low-mappability regions and 1.8-3.1 pp in segmental duplications; INDEL recall rose by 4.5-5.8 pp and 2.4-3.8 pp, respectively, with similar gains in the MHC and challenging medically relevant genes (CMRG). Decomposition analysis showed that the aligner change drove most INDEL gains, while IUPAC encoding contributed additional SNV-specific improvement. We introduce BurdenBench, an open-source framework that computes net benefit and region-size-normalised metrics directly from standard hap.py outputs, revealing divergent caller-specific trade-off profiles that are invisible to aggregate F1: FreeBayes showed the most favourable precision-recall balance in low-mappability regions, while GATK achieved positive net benefit in the MHC. A controlled comparison using an identical variant set showed severe recall and precision losses for SALT (a published SNP-aware dual-index aligner) across all three callers, supporting the value of preserving linear reference structure. Pan-human and population-specific consensuses performed within 0.2 pp of one another. All findings are descriptive and hypothesis-generating from three samples. Availability and implementationTo mitigate potential bias associated with software developed by an authors employer, primary hap.py outputs and derived burden metrics were independently verified by co-authors with no affiliation to that employer. BurdenBench (v1.0.0) is implemented in Python (pandas, numpy; Python [≥]3.7) and freely available under the MIT licence at https://github.com/akzam/BurdenBench, including raw hap.py outputs and an audit trail enabling independent recomputation without a novoAlign licence. novoAlign and novoUtil (version 4, Novocraft Technologies) are commercial software with no-cost academic trial licences. Contactleanne.dibbens@adelaide.edu.au Supplementary informationSupplementary tables, figures and methods are available online.