NGS Variant Callers in Clinical Practice

What this research found

A 23-page review written for clinical geneticists and molecular pathologists, surveying the software that turns raw next-generation sequencing reads into reportable variant calls. It covers germline, somatic and structural variant detection, the ACMG/AMP and AMP/ASCO/CAP classification frameworks, and the validation requirements a diagnostic laboratory must satisfy under CAP and CLIA rules. This is a synthesis of published work — 45 citations spanning 2009 to 2025 — rather than a new benchmarking exercise.

• Deep learning has overtaken purely statistical methods for germline calling. A trio-sequencing comparison found DeepVariant produced a Mendelian error rate of 3.09% against 5.25% for GATK HaplotypeCaller, although GATK retains an advantage on rare variants, which are often the ones that matter in rare-disease diagnosis.
• Small-variant calling is near its ceiling on benchmark data. Current tools reach SNP F1-scores above 99.9% and indel F1-scores of roughly 99% to 99.5%, with insertions and deletions consistently harder than single-nucleotide variants.
• Somatic calling is a different problem because the signal sits much lower. Germline heterozygous variants appear near 50% variant allele frequency, subclonal tumor mutations can fall to 1–5%, and circulating tumor DNA in liquid biopsy may require detection at 0.1% or below — where VarScan2's default 20% threshold would miss clinically relevant mutations unless adjusted.
• Combining callers outperforms any single one. A 2024 benchmark identified MuSE, Mutect2 and Strelka2 as the best somatic ensemble for single-nucleotide variants and Mutect2, Strelka2 and VarScan2 for indels, with at least two tools required to agree; for structural variants, majority voting improved precision over the best individual caller but reduced recall.
• Structural variant detection remains the least mature area. Variants larger than 50 base pairs must be inferred indirectly from discordant read pairs, split reads and read-depth changes, and no standardized interpretation framework matches the 28 evidence criteria the ACMG/AMP guidelines provide for sequence variants.

How it was done

Forty-five papers published between 2009 and 2025, from journals including Nature Biotechnology, Genome Research, Nature Genetics and Genome Medicine, were synthesized into a review organized by variant class and by stage of the clinical workflow. Tools were grouped into germline callers (GATK HaplotypeCaller, DeepVariant, DRAGEN, Clair3, DNAscope), somatic callers (Mutect2, Strelka2, VarScan2, MuSE) and structural or copy-number callers (Manta, DELLY, LUMPY, GRIDSS, CNVnator), each assessed on algorithmic approach, reported accuracy and fitness for clinical use. Regulatory material was summarized alongside it — ACMG/AMP germline criteria, AMP/ASCO/CAP somatic tiers and AMP/CAP pipeline validation recommendations — together with practical guidance on quality metrics, reference genome version control and laboratory information system integration. The output is a 23-page paper with a graphical abstract and seven schematic figures.

Data sources

• 45 peer-reviewed papers, 2009–2025, including Nature Biotechnology, Genome Research, Nature Genetics and Genome Medicine
• Richards et al., Genetics in Medicine 17:405 (2015) — ACMG/AMP sequence variant interpretation standards
• Li et al., Journal of Molecular Diagnostics 19:4 (2017) — AMP/ASCO/CAP four-tier somatic variant guidelines
• Roy et al., Journal of Molecular Diagnostics 20:4 (2018) — AMP/CAP bioinformatics pipeline validation recommendations
• Genome in a Bottle consortium reference materials, including the NA12878 (HG001) cell line
• gnomAD — population variant frequencies from 141,456 humans
• precisionFDA Truth Challenge V2 — Olson et al., Cell Genomics 2:100129 (2022)