When a virus or a gene-therapy vector inserts itself into human DNA, everything about the consequences depends on exactly where the insertion landed — which gene it disrupted, which promoter it activated. Genomic integration site mapping finds those single-nucleotide junctions, usually by sequencing the tumor or the therapy product and looking for "split reads": sequencing reads whose two halves align to two different genomes, one viral and one human. The human half of that read must align somewhere, and today it aligns to a single linear reference assembly, GRCh37 or GRCh38 — an assembly built overwhelmingly from a small number of genomes of largely European origin. That creates a blind spot with a population structure: recent long-read sequencing of 539 individuals across five continental populations found 45,284 unique stretches of human sequence that simply do not exist in the standard reference, and 35.6% of these are carried only by specific populations (Human pangenome analysis of sequences missing from the reference genome). Any integration event landing in that sequence is invisible to every standard pipeline.
This matters clinically, not just conceptually. HPV integration status is now a cervical-cancer triage test with 92.8% specificity for high-grade lesions, cutting colposcopy referrals from 27.3% to 10.7% in a 4,086-woman prospective cohort — but that cohort was drawn from a single county in China, and the assay's silence about an integration could mean either "no integration" or "integration in sequence our reference cannot see" (Risk assessment and triage strategy of cervical cancer primary screening on HPV integration status: 5-year follow-up of a prospective cohort study). The same logic applies to CAR T-cell product release testing, where vector integration profiles are read against the linear reference before infusion (Genome-wide profiling of retroviral DNA integration and its effect on clinical pre-infusion CAR T-cell products). The parallel to polygenic risk scores is direct: absolute measures computed on a European-shaped reference do not transfer across ancestries, and the failure mode is systematic, not random (On cross-ancestry cancer polygenic risk scores).
Core Claim
I propose that integration site mapping carries an unmeasured, ancestry-dependent detection bias, and that this bias can be quantified and corrected by re-implementing junction detection against a population-annotated graph reference. The concrete claim has three parts. First, a measurable fraction of the human genome — population-specific nonreference sequence — is a detection dead zone for every deployed integration caller, in the same way the Retrovirus Integration Database already documented assembly-gap "dead zones" like the first 14 million base pairs of chromosome 22 being technical artifacts rather than true biological resistance to integration (Retrovirus Integration Database (RID): a public database for retroviral insertion sites into host genomes). Second, this dead zone is not ancestry-neutral: it is concentrated in populations least represented in the reference, so apparent differences in integration rate or hotspot usage between ancestries confound biology with reference artifact. Third, a graph-aware caller recovers these events and changes ancestry-stratified conclusions.
It is useful to write the bias as one quantity. Let $R$ be the linear reference, $N$ the union of all nonreference sequence, and $N_a$ the subset of nonreference sequence carried by ancestry group $a$. A caller restricted to $R$ has a detection ceiling for a uniformly placed breakpoint of roughly
$$V(a) \;=\; \frac{|R|}{|R| \,\cup\, N_a},$$
the fraction of ancestry-$a$ genome space in which a junction can be found at all. The proposal's first deliverable is estimating $V(a)$ for each of the five pangenome populations, and its central methodological deliverable is a caller that pushes the ceiling toward 1 by searching $R \cup N$.
Research Program
The program has three components: a resource, a method, and a reanalysis.
The resource is a population-annotated augmented reference: the standard linear assembly plus all 45,284 unique nonreference sequences from the 539-genome long-read pangenome, each contig tagged with the populations in which it was observed, and cross-referenced against that study's functional annotations — 565 expression quantitative trait loci and 154 sequences in strong linkage disequilibrium with 258 disease-associated variants (Human pangenome analysis of sequences missing from the reference genome). The annotation matters because nonreference sequence is not junk: it contains regulatory and disease-linked sequence, and HPV integration is already known to prefer transcriptionally active, open-chromatin regions — 66% of 119 mapped integration sites fell within genes, with a 1.78-fold enrichment at DNase hypersensitive sites (Transcriptionally Active Regions Are the Preferred Targets for Chromosomal HPV Integration in Cervical Carcinogenesis). There is no biological reason to expect nonreference chromatin to be spared; the targeting logic predicts the opposite.
The method is a junction caller — graph-VISA, in effect — that aligns candidate chimeric reads to the augmented graph rather than the linear reference, calls breakpoints at single-nucleotide resolution as VERSE does within the PCAWG pipeline (The landscape of viral associations in human cancers), and reports for every call whether the host breakpoint lies in reference sequence, shared nonreference sequence, or population-specific sequence. Junction detection itself does not need to be reinvented; the contribution is the reference representation and the population-stratified bookkeeping around it.
The reanalysis applies the caller to the three richest published integration datasets: the PCAWG pan-cancer callset (2,658 tumors, 38 cancer types, 356 virus-positive patients), the pre-infusion CAR T and TCR product dataset (75 and 6 products), and the Longyou HPV screening cohort. For each, the questions are how many previously unassigned or discarded reads now resolve into junction calls, where those calls fall relative to population-annotated sequence, and whether ancestry-stratified integration statistics — rates, hotspot enrichments such as the recurrent TERT promoter integrations — shift once the blind spot is filled. This is deliberately an analysis-first project: no new wet-lab cohort is required to establish the bias, though a small spike-in validation closes the loop (below).
Proposed Experiments
The detection-bias experiment is a simulation benchmark built from real data. Starting from the PCAWG framework, I will simulate chimeric reads whose host-side breakpoints are placed uniformly across the augmented genome, with realistic read lengths and error profiles, and then run the standard linear-reference pipeline and the graph caller side by side. The primary metric is recall stratified by breakpoint location: reference sequence, shared nonreference, population-specific nonreference. The expected signature of the bias is a recall cliff exactly at the population-specific stratum for the linear pipeline and no cliff for the graph caller; the size of the cliff is the first quantitative estimate of $1 - V(a)$ per population. No reported result found on Lacuna for integration-caller recall against nonreference sequence — that absence is the gap this fills.
The reanalysis experiment then tests whether the bias is material in real cohorts. For PCAWG, the published numbers anchor the comparison: viruses detected in 13% (356/2,658) of patients, with HBV DNA integrated in 59% (36/61) of HBV-positive liver cancers (The landscape of viral associations in human cancers). I will re-map the non-human-aligning and discordant read fraction through the graph caller and ask whether the virus-positive fraction or the integration-positive fraction within HBV-positive tumors rises — an integration-call recovery rate is the headline number. For the CAR T dataset, the analogous anchor is the vector-genomic-element distribution — for example, gamma-retroviral promoters at 14% versus 4% for lentiviral — and the test is whether integrations into nonreference sequence alter that distribution or the identified hotspots (Genome-wide profiling of retroviral DNA integration and its effect on clinical pre-infusion CAR T-cell products). For the Longyou cohort, the stakes are sharpest: the assay's negative-predictive value rests on the 1.6% immediate CIN3+ risk among integration-negative women, and the experiment asks what fraction of "integration-negative" calls could plausibly be reference-invisible rather than truly negative, given the population-specific sequence carried by the screened population (Risk assessment and triage strategy of cervical cancer primary screening on HPV integration status: 5-year follow-up of a prospective cohort study).
The spike-in validation experiment grounds the simulation in physical reality: construct a small panel of sequencing libraries containing synthetic junctions at known coordinates, including several inside nonreference contigs, sequence them, and confirm the graph caller recovers the planted breakpoints while the linear pipeline misses exactly the nonreference ones. This costs little and rules out the objection that simulated chimeras behave unlike real ones.
Finally, an interpretation experiment connects to the ancestry-bias literature directly. The cross-ancestry PRS study taught a specific lesson: absolute score distributions shift badly across ancestries, but within-ancestry relative stratification stays predictive (breast-cancer top-decile odds ratios of 2.81 in Europeans, 2.88 in Africans, 2.60 in East Asians, 2.33 in South Asians) (On cross-ancestry cancer polygenic risk scores). I will test the analogous proposition for integration mapping: cross-ancestry comparisons of absolute integration rates are invalid until the detection ceiling is equalized, but within-ancestry hotspot enrichments and integration-based triage remain usable if — and only if — the reference used covers that ancestry's sequence. This yields a concrete reporting recommendation for clinical integration assays.
Baselines
The naive reference point is the deployed standard: split-read junction calling against the unmodified linear reference, which is what the PCAWG pipeline family (CaPSID, P-DiP, SEPATH, with VERSE for breakpoints) did across 2,658 tumors and nearly 196 billion non-human-aligning reads (The landscape of viral associations in human cancers). This is the "do nothing" comparator because it is what every published integration callset, including the CAR T VISA profiles, was produced with. It will be run on exactly the same simulated and real read sets as the graph caller, with the same read filters and the same consensus-count thresholds, so any recovered-call difference is attributable to the reference alone.
The strongest existing competitor is that full PCAWG consensus architecture rather than any single tool: three independent discovery pipelines with high-confidence calls requiring agreement of at least two. It is the strongest because consensus across callers controls pipeline-specific false positives, and its published performance — 13% viral prevalence, 59% HBV integration in HBV-positive liver cancers — is the field's best-powered result. I will run the consensus setup unchanged against the linear reference and against the augmented reference, so the comparison isolates the one variable under study. Beating it means recovering junction calls in population-specific nonreference sequence at a controlled false-positive rate (measured on the spike-in panel) while reproducing its linear-reference calls to near-perfect overlap; any graph-caller gain that came at the cost of losing published calls would be a failure.
Two further comparisons complete the picture. A decoy-augmented linear pipeline — same standard caller, but with all nonreference contigs appended as flat decoy sequences — is the serious cheap alternative and is treated as a true competitor, not a strawman; no reported result found on Lacuna for integration calling against an NRS-augmented decoy set, though the pangenome study reports only a 1.59% improvement in overall DNA read mapping rates from graph-based alignment, which does not tell us anything about junction resolution at breakpoints (Human pangenome analysis of sequences missing from the reference genome). The key ablation runs the graph caller with population-specific nonreference sequence removed, keeping only the cross-population shared subset (the 38.7% of nonreference sequences common to all five populations). If ancestry-stratified recovery differences persist in this ablation, they are driven by shared sequence and the population-specific claim is wrong; if they vanish, the ablation isolates exactly the ingredient the proposal is about.
Why not just…?
Why not just add the nonreference contigs as decoys to the existing pipeline? This is the alternative most likely to work, and I treat it as a co-equal arm rather than an argument to dismiss. Appending contigs as flat decoys recovers the raw reads that currently go unassigned, and for presence/absence detection it may match the graph caller. Where it plausibly falls short is breakpoint localization and ancestry bookkeeping: a decoy hit gives an unplaced contig name, not a population-resolved genomic coordinate, and linear callers handle multiply placed reads poorly. The experiments will settle this empirically — the decoy arm and the graph arm run on identical data, and if the decoy arm matches the graph caller on recall, localization accuracy, and ancestry-stratified reporting, the honest conclusion is that the graph machinery is unnecessary and the cheap fix suffices. That would still validate the proposal's central claim about the bias itself.
Why not just long-read sequence a diverse cohort? Long reads genuinely resolve complex integrations — nanopore assembly has been used to reconstruct full HPV genome structures (Whole Genome Assembly of Human Papillomavirus by Nanopore Long-Read Sequencing) — and I will use a small long-read spike-in set as the ground-truth layer. But long reads still must be placed against some reference to become coordinates, so they relocate rather than remove the bias, and cohort-scale long-read sequencing of thousands of tumors or screening samples is orders of magnitude more expensive than re-analyzing existing short-read data against a better reference. The right division of labor is short-read graph reanalysis at scale, with long reads validating a subset.
Why not just recruit more diverse cohorts and run the standard pipelines on them? More diverse cohorts with an unchanged reference would manufacture the exact false result this proposal warns against: integration rates and hotspot maps that differ across ancestries because the detection ceiling $V(a)$ differs, not because the biology does. The monogenic diabetes experience is the cautionary precedent — screening criteria derived in White populations failed in ethnic minorities and required ethnicity-specific cutoffs (Precision diagnosis for monogenic diabetes requires ethnicity specific criteria for genetic testing) — and the PRS experience shows the failure is systematic, not noisy. Diverse recruitment is necessary eventually, but only after the measurement instrument is ancestry-equalized; otherwise it hard-codes the bias into larger datasets.
Decisive Checks
The proposal stands or falls on three numbers. The simulation must show a recall cliff for the linear pipeline that tracks the population-specific stratum, of a magnitude large enough to matter — if the ceiling loss is, say, below half a percent of accessible genome for every population, the bias is real but clinically negligible and I would say so. The PCAWG reanalysis must recover a nonzero number of high-confidence junction calls in nonreference sequence, with a false-positive rate bounded by the spike-in panel; zero recovered calls would mean the bias is theoretical. And the ablation must show that removing population-specific sequence eliminates the ancestry stratification of recovered calls — if it does not, my proposed mechanism is not the operative one, and the honest next step is to look for the real source of any residual ancestry differences in the published callsets.
References
- Human pangenome analysis of sequences missing from the reference genome reveals their widespread evolutionary, phenotypic, and functional roles
- Risk assessment and triage strategy of cervical cancer primary screening on HPV integration status: 5-year follow-up of a prospective cohort study
- Genome-wide profiling of retroviral DNA integration and its effect on clinical pre-infusion CAR T-cell products
- On cross-ancestry cancer polygenic risk scores
- Retrovirus Integration Database (RID): a public database for retroviral insertion sites into host genomes
- Transcriptionally Active Regions Are the Preferred Targets for Chromosomal HPV Integration in Cervical Carcinogenesis
- The landscape of viral associations in human cancers
- Whole Genome Assembly of Human Papillomavirus by Nanopore Long-Read Sequencing
- Precision diagnosis for monogenic diabetes requires ethnicity specific criteria for genetic testing