The Reads That Missed on Purpose

Zetobit · Bioinformatics Insight Series

The Reads That Missed on Purpose

A hybrid-capture panel is designed to enrich a few hundred kilobases and throw the rest of the genome away. But the rest of the genome shows up anyway — thinly, unevenly, and more usefully than most pipelines admit.

By Kanna Nandakumar, PhD · Zetobit, LLC

Every hybrid-capture assay has a target: the exons, hotspots, or gene set the probes are designed to pull down. And every hybrid-capture assay has a shadow — the reads that landed everywhere else. In a typical panel or exome run, that shadow is not a rounding error. A large fraction of sequenced reads fall outside the targeted regions, driven by imperfect probe specificity, near-target fragment overhang, and the simple physics of shearing genomic DNA into a library.¹

The instinct is to discard them. On-target rate is a headline QC metric, off-target reads are "wasted" sequencing dollars, and the whole point of a targeted assay is to not sequence the genome. All true. But treating off-target reads as pure loss throws away a low-density, genome-wide layer of information that is already paid for — and that, in specific situations, changes what a report can say.

Off-target reads are ultra-low-depth whole-genome sequencing you didn't ask for. The question is whether your pipeline knows what to do with a free half-genome at sub-1× depth.

What the off-target fraction is telling you

Before rescuing anything, the off-target fraction is a diagnostic about the run itself. The proportion of reads landing on target reflects capture efficiency, and it drifts for real reasons: probe lot changes, library input, hybridization time, and the fragment-size distribution of the input DNA.² A sudden rise in off-target rate across a batch is rarely random — it points at a chemistry or protocol shift that will also degrade on-target uniformity. Read as a control chart rather than a single pass/fail gate, the off-target fraction is one of the earliest warnings that a batch is drifting.

Off-target reads also feed contamination detection. Cross-sample contamination estimators work by modeling allele fractions at a panel of common SNPs against population frequencies; the more genome-wide sites you can pile up, the more stable the estimate.³ Because contamination has to sit well below your minimum allele fraction of interest before somatic calls are trustworthy — a sample with ~1% contamination can force you to disregard variants below roughly 5% VAF³ — the extra sites contributed by off-target coverage are not a curiosity. They tighten the number that decides whether low-VAF calls are reportable at all.

Two kinds of coverage from one assay. Over the target, depth is high enough for confident single-nucleotide and indel calling. Everywhere else, off-target reads produce a sparse, sub-1× layer — too thin for base-level calls, but sufficient, aggregated over wide bins, to carry copy-number and other genome-scale signal.⁴

Where off-target reads actually rescue a call

The rescue cases are specific, and it helps to be precise about them rather than promising that "free data" saves everything.

Copy-number variants beyond the target. This is the strongest use. Standard panel and exome CNV callers only see dosage inside the captured regions, so a clinically relevant deletion or duplication whose body — or breakpoints — fall outside the target is invisible to them. Off-target reads, binned into wide windows and normalized against a panel of samples, recover a genome-wide copy-number profile at effectively ultra-low-pass resolution. Tools built for this, such as SavvyCNV, call large deletions and duplications from off-target data across the genome and can localize breakpoints that sit outside targeted regions.^4,5 The tradeoff is honest: at sub-1×, evidence comes from read-depth shifts over broad regions, not from split reads or fine breakpoints, so this is a tool for large events, not focal single-exon changes.⁴

Genome-scale scores that don't need base resolution. Because homologous-recombination-deficiency (HRD) scoring depends on genome-wide allelic-imbalance and loss patterns rather than individual variants, off-target reads from a tumor panel can be repurposed to estimate HRD without a separate whole-exome or whole-genome assay.⁶ The same logic extends to tumor-fraction estimation in cell-free DNA panels, where genome-wide copy signal in the off-target fraction informs how much tumor is present.

Loci the panel never targeted. Mitochondrial DNA is a clean example. Most nuclear-targeted capture kits don't include mtDNA, yet its high copy number means off-target reads pile up there anyway — enough to genotype pharmacogenetically important variants such as the aminoglycoside-ototoxicity allele m.1555A>G in MT-RNR1 directly from off-target data, with concordance against orthogonal genotyping.⁷ A variant your assay was never designed to see becomes reportable because the reads were already in the BAM.

The pattern across every rescue case is the same: off-target reads carry genome-scale signal, not base-level calls. Match the question to the resolution and they deliver; ask them for a single confident SNV outside the target and they won't.

The limits, stated plainly

Off-target coverage is uneven in a structured way. Some off-target loci attract reads far above background — comparable to real targets — because certain sequences are repeatedly, reproducibly pulled down; a minority of loci contribute the majority of off-target reads.⁸ That non-uniformity is exactly why off-target CNV methods lean on across-sample normalization and wide bins, and why large tandem-repeat regions produce noisy, high-variance bins that have to be masked. It is also why mixing on-target and off-target reads for genome-wide dosage can introduce batch effects from variable enrichment efficiency, so some workflows deliberately use off-target reads alone.⁶

None of this makes off-target reads a substitute for the right assay. If you need confident single-base calls genome-wide, order genome sequencing. What off-target reads offer is a way to extract genome-scale answers — copy number, HRD, tumor fraction, a handful of off-panel loci — from data you have already generated and paid to sequence, at the cost of some pipeline engineering and a clear-eyed sense of the resolution ceiling.

The practical takeaway

Three moves turn the shadow into an asset. First, monitor the off-target fraction as a trend, not just a threshold — it is a leading indicator of capture and protocol drift, and it feeds contamination estimates that gate low-VAF reporting. Second, when a CNV, an HRD score, or an off-panel locus like mtDNA is clinically in scope, ask whether off-target reads can answer it before commissioning a second assay. Third, respect the resolution: wide-window, genome-scale questions only. Read that way, the reads that missed on purpose stop being waste and start being the cheapest whole-genome layer you'll ever collect. ▪

---

References

1. Talkowski lab / capture methodology overviews; SavvyCNV report that up to ~70% of reads from exome and targeted sequencing can fall outside targeted regions. Laver TW, et al. SavvyCNV: Genome-wide CNV calling from off-target reads. PLOS Computational Biology (2022). journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1009940
2. Genoox Franklin Help Center. Understanding QC Metrics in Franklin — on-target rate, duplication, insert-size, and contamination metrics for targeted assays. help.genoox.com/en/articles/12645109
3. Illumina DRAGEN Documentation. QC Metrics and Coverage/Callability Reports — cross-sample contamination modeling and the requirement that contamination sit well below the minimum allele fraction of interest for safe somatic calling. support-docs.illumina.com/SW/dragen_v42
4. Laver TW, et al. SavvyCNV: Genome-wide CNV calling from off-target reads. PLOS Comput Biol 18(3):e1009940 (2022) — off-target reads as ultra-low-depth WGS; wide-window read-depth calling; breakpoint localization outside targets. pmc.ncbi.nlm.nih.gov/articles/PMC8959187
5. SavvyCNV benchmarking against MLPA and WGS-derived truth sets for germline CNV recall from panel and exome off-target reads (same work, benchmarking sections). PLOS Comput Biol, full text PDF
6. Deriving genome-wide copy-number profiles and HRD estimates from off-target reads of hybrid-capture panel sequencing; note that combining on-target and off-target reads can introduce enrichment-driven batch effects. Leveraging Off-Target Reads in Panel Sequencing for Homologous Recombination Repair Deficiency Screening. PMC12178388. pmc.ncbi.nlm.nih.gov/articles/PMC12178388
7. Extraction of MT-RNR1 pharmacogenetic variation, including m.1555A>G, from off-target reads of capture-based panels and WES, validated against fluorescence-based genotyping. A Novel Approach for the Identification of Pharmacogenetic Variants in MT-RNR1 through NGS Off-Target Data. PMC7408883. ncbi.nlm.nih.gov/pmc/articles/PMC7408883
8. Structured non-uniformity of off-target coverage: a minority of high-prevalence off-target loci contribute the majority of off-target reads. Off-target capture reduction in sequencing techniques (USPTO 11,624,084). image-ppubs.uspto.gov/.../11624084

ZETOBIT.COM · BIOINFORMATICS INSIGHT SERIES