Research Use Only (RUO): This article discusses research study design and evidence interpretation for AAV integration site analysis. It is not intended for clinical diagnosis, treatment decisions, or individual health assessment.

Low-frequency AAV integration candidates are where integration site analysis becomes genuinely difficult—not because the concept of a vector–host junction is mysterious, but because the noise floor of sequencing, library construction, and alignment begins to overlap the signal you care about.

At high abundance, evidence tends to be redundant: multiple read types converge on the same breakpoint, mapping is stable, and technical replicates reproduce the site. At low abundance, you often have a small number of junction-like reads competing with multiple plausible alternative explanations: ligation artifacts, cross-sample contamination, ambiguous mapping in repeats, or over-amplified duplicates.

This guide is written for gene therapy and viral vector teams who already have (or expect) low-level candidates and need a defensible way to decide:

• Is the candidate technically credible as a candidate?
• What additional computational review is needed before ordering primers?
• Which orthogonal confirmation is worth the time and sample?
• How should low-frequency candidates be reported without implying confirmation?

Key takeaway: Treat low-frequency candidates as hypotheses. Advance them to validation only when read-level evidence, mapping context, duplicate handling, and controls together support a coherent vector–host junction model.

1. Introduction: Why low-frequency integration calls require an evidence framework

A low-frequency AAV integration call is not inherently "weak." It may represent a real event in a small subpopulation, a transient DNA repair intermediate, or a rare integration in a complex tissue sample. The problem is that, below some dataset-specific abundance, the call quality is dominated by process-specific artifacts.

A practical evidence framework does three things:

1. Separates detection from confirmation. A computational call is a candidate until it is supported by independent molecular evidence.
2. Forces explicit assumptions. What alignment reference was used? How were duplicates handled? What negative controls define your background?
3. Optimizes validation effort. You cannot validate every weak call; you need a rational triage scheme.

A comparative study of rAAV insertion site detection methods showed that even vehicle controls and spike-in controls can yield apparent insertion sites, and that library preparation steps (notably adapter ligation) can contribute to artifactual junction-like signals; switching to tagmentation reduced these artifacts in that setting (a 2023 comparative analysis). The takeaway for low-frequency work is straightforward: the absence of an evidence framework is not neutral—it increases the probability you will validate artifacts.

2. What counts as a low-frequency AAV integration site candidate?

For the purpose of validation, an "AAV integration site candidate" should mean:

• There is sequence evidence consistent with a junction between host genomic DNA and vector-derived sequence.
• The evidence localizes to a specific host breakpoint interval (even if not single-base precise).
• The candidate can be described by a breakpoint model: host coordinate(s), host strand/orientation, vector breakpoint region (e.g., within the transgene cassette, within/near ITR-adjacent sequence, or within an internal fragment), and the implied junction orientation.

Crucially, a candidate is not defined by a single software flag or a single read mapping to the vector. Candidates can arise from multiple computational paradigms, including:

• Split-read breakpoint detection on a hybrid host+vector reference
• Discordant paired-end clustering where mates map to different references
• Local assembly around soft-clipped sequences to reconstruct junction contigs
• Capture-based approaches where vector-targeted reads are enriched and junctions are inferred

In all cases, a candidate should be reducible to a reviewable set of supporting reads (names/IDs) and a reviewable alignment view (e.g., genome browser snapshot or exported alignment segment). If the original analysis cannot provide those, the call is not practically validate-able.

3. Types of sequencing evidence

Different read signals answer different questions. Low-frequency validation requires you to treat these signals as complementary, not interchangeable.

Split reads

Split reads (often represented as soft-clipped or supplementary alignments) provide the most direct evidence for a breakpoint because a single molecule spans both sides of the junction.

What split reads can support:

• Base-level localization of the host breakpoint (within alignment uncertainty)
• Orientation of the junction (host-to-vector vs vector-to-host)
• Whether the junction includes microhomology, small insertions, or small deletions at the breakpoint (when read length and quality permit)

What to check:

• Does the read have a clear host-aligned segment and a clear vector-aligned segment?
• Are the clipped bases high quality, or are they dominated by low-quality tails?
• Do multiple split reads agree on the same breakpoint window and orientation?

Discordant read pairs

Discordant pairs (one mate maps to host, the other maps to vector; or abnormal insert size/orientation) are often more abundant than split reads, especially in short-read datasets.

What discordant pairs can support:

• A region-level candidate (often a window rather than a precise breakpoint)
• Evidence that the junction is not a single-read anomaly

Limitations:

• Discordant pairs rarely define the breakpoint precisely without additional split-read or assembly evidence.
• Discordant pairs are more sensitive to mapping ambiguity, especially if one mate is short, low complexity, or maps to repetitive sequence.

Vector-host junction reads

"Vector–host junction reads" is an umbrella term: any read (or read pair) that contains evidence bridging vector and host sequence.

To make this category useful for validation, subdivide junction reads by how they support the junction:

• Single-molecule spanning junction (split read): strongest
• Paired-end bridging (discordant pair): supportive but less precise
• Assembly-derived contig: can be strong if contig is well supported by multiple reads and includes sufficient unique host sequence

A useful internal rule is: a junction read should contribute either (a) breakpoint precision, or (b) breakpoint reproducibility across independent molecules. Reads that do neither are often noise.

Unique molecules

Read counts are not independent evidence when many reads derive from the same original template molecule.

Depending on library design, independent evidence can be approximated by:

• Distinct start/end positions after random fragmentation (shear-site diversity)
• Molecular barcode diversity (molecular barcodes assigned prior to amplification)
• Distinct fragment lengths supporting the same junction (especially when random shearing is used)

A key point from integration site quantification literature is that abundance can be distorted by uneven PCR amplification of junction fragments; coordinate-only deduplication can also remove real signal in complex regions, so "unique molecule" handling should be explicitly defined and audited rather than assumed (IS-Seq pipeline discussion, 2023).

Strand and breakpoint consistency

Low-frequency candidates become more credible when evidence is internally consistent:

• Host breakpoints cluster within a small interval rather than scattering across a kilobase
• The vector-side breakpoints point to a coherent region (e.g., consistent within the same vector segment)
• Read orientation is compatible with one or a small number of plausible junction structures
• Evidence is present on both strands (when library and mapping support this expectation)

Inconsistency does not prove artifact, but it raises the bar for additional evidence or orthogonal confirmation.

Figure 2. Genome-browser-style illustration of split reads, discordant read pairs, and vector–host junction evidence at a candidate low-frequency AAV integration site.

4. Why supporting-read count alone is insufficient

"Integration supporting reads" is a necessary metric, but it is not a sufficient credibility measure for low-frequency calls.

Reasons supporting-read count can mislead:

1. PCR amplification can inflate reads from a single molecule. Many reads may represent one original junction fragment.
2. Certain junctions amplify preferentially. Sequence composition, fragment length, and primer/probe behavior can bias which junctions dominate.
3. Mapping ambiguity can inflate counts. Multi-mappers can be counted multiple ways across pipelines.
4. Background artifacts produce junction-like reads at low levels. Vehicle controls can contain insertion-site-like signals depending on method and library preparation (see the 2023 comparative analysis of rAAV insertion site detection methods).

A better framing is:

• Read count is a starting clue.
• Independent molecules, mapping uniqueness, breakpoint coherence, and controls determine whether read count is informative.

5. PCR duplicates and amplification bias

Low-frequency AAV integration candidates are especially sensitive to duplicate handling because the absolute number of junction templates may be small.

Two failure modes to avoid

Failure mode A: Under-deduplication (counting PCR copies as evidence).

• Inflates apparent support and can push artifacts into "validate" territory.

Failure mode B: Over-deduplication (collapsing distinct molecules incorrectly).

• Collapses evidence in repetitive or low-complexity regions where mapping coordinates are not a stable proxy for molecular identity.

What you should request from the original analysis

To review duplicates responsibly, ask for:

• The deduplication method used (coordinate-based? barcode-based? both?)
• Whether deduplication was performed before or after extracting junction reads
• Duplicate rate metrics for the library (overall and within junction-supporting reads)
• For each candidate: counts of (a) raw reads, (b) deduplicated reads, and (c) estimated independent molecules

Practical review cues

• If a candidate is supported by many reads but only one or two independent molecules, it is typically a weak validation target unless other evidence is unusually strong (e.g., long host unique sequence in split reads).
• If independent molecules exist but all share identical breakpoint and identical read start positions, suspect over-amplification or a library bottleneck.

Pro tip: Always interpret "supporting reads" alongside "independent molecules." If the pipeline can't provide a molecule-level view, treat the call as lower confidence and consider confirmation by an alternative sequencing approach rather than PCR alone.

6. Mapping quality and ambiguous genomic regions

At low frequency, many candidates are either true events in difficult genomic contexts or artifacts created by those contexts. You must treat genomic mappability as a first-class variable.

Mapping quality is necessary but not sufficient

High mapping quality (MAPQ) on its own does not guarantee correctness, especially if:

• The reference is missing relevant vector variants (e.g., plasmid-derived sequence differences)
• The read is short and the "best" alignment is still weak in absolute terms
• The alignment is forced by aggressive mismatch penalties

Still, MAPQ remains a key discriminator. For each candidate, review:

• MAPQ distribution for all junction-supporting reads
• Whether supporting reads are uniquely mapped or multi-mapped
• Whether the host breakpoint overlaps known repeats, segmental duplications, satellite DNA, or low-complexity regions

Ambiguous regions that commonly degrade credibility

• Short tandem repeats and microsatellites
• Segmental duplications and paralogous gene families
• Centromeric/pericentromeric regions
• Regions containing high similarity to vector payload (e.g., if transgene contains human cDNA segments)

If the host-side alignment is ambiguous, PCR validation can fail or produce non-specific products. In these cases, long-read confirmation or capture-based resequencing can be a more realistic path than conventional PCR.

7. Vector-reference and host-reference alignment issues

Low-frequency candidates are often mis-scored because the reference model is wrong.

Vector reference issues

Ask explicitly:

• Which vector reference sequence was used? (production plasmid? packaged genome consensus? a generic serotype reference?)
• Does it include ITR-adjacent sequences in the correct orientation?
• Were known vector variants (mutations, deletions, truncations) considered?

If the vector reference is mismatched, true junction reads may align poorly, and artifact reads may align "better" than real ones.

Host reference issues

Similarly:

• Which genome build was used (e.g., GRCh38 vs GRCh37)?
• Were decoy contigs and alternative loci included?
• Was the host reference augmented with common contaminants or spike-ins?

Hybrid reference alignment: what to check

Many pipelines align reads to a hybrid host+vector reference. That approach is practical, but it introduces specific review needs:

• Ensure that vector and host segments are not being mis-assigned due to short overlaps.
• Verify that supplementary alignments are retained and not filtered out.
• Confirm that secondary alignments are handled consistently (some pipelines count them; others drop them).

8. Control samples and background artifacts

Controls are not optional for interpreting low-frequency calls. Without them, you cannot know whether your "rare events" are above your pipeline's artifact baseline.

Minimum control logic

Ideally, you want at least:

• Negative control processed through the same workflow (vehicle-treated, mock-transduced, or no-vector control)
• Process controls that capture index hopping/cross-contamination risk (e.g., unique dual indexing, blank libraries)
• Spike-in control or contrived positive control when feasible (to ensure the workflow can detect junctions when present)

The 2023 comparative analysis cited above noted apparent insertion sites in controls and attributed a substantial component of the signal to library preparation conditions. You should interpret that as permission to be strict: if a candidate's features also appear in negatives, it is not a priority validation target.

Control-aware candidate review questions

For each candidate, ask:

• Is the same host breakpoint window present in negative controls?
• Is the same vector breakpoint region present in negatives?
• Are similar candidates enriched in certain batches (batch effect)?
• Are candidates associated with specific index combinations (index hopping signal)?

If controls are absent, consider the analysis incomplete for decision-making. At minimum, request a re-analysis including control datasets if they exist.

9. Recurrent breakpoints versus recurrent technical artifacts

Recurrence is often treated as evidence of biological plausibility. At low frequency, recurrence can also indicate pipeline artifacts.

When recurrence increases confidence

Recurrence is more convincing when:

• The same site appears across biological replicates processed independently
• Independent molecules support the site in each replicate
• Junction structure is consistent (orientation, breakpoint window)
• The site is absent in negative controls

When recurrence is a red flag

Recurrence may be a technical artifact when:

• The site clusters in difficult regions (repeats/low mappability) across many samples
• The same vector breakpoint "hotspot" appears in many unrelated samples without other supportive structure
• Candidates track with batch, lane, index, or library-prep condition

If you observe "recurrent" candidates, treat recurrence as a hypothesis generator: it tells you where to look harder at controls, mapping context, and sample processing—not a shortcut to confirmation.

10. An evidence grading framework

The goal of grading is not to label events as true/false. It is to standardize what kind of evidence you have and what validation is justified.

Evidence-grade table

Evidence grade	Typical read-level pattern	Mapping context	Duplicate / independence signals	Control behavior	Recommended next action
Low-confidence candidate	Only discordant pairs or 1–2 ambiguous split reads; breakpoints scattered; inconsistent orientation	Host breakpoint overlaps repeats/low mappability or MAPQ mostly low; vector alignment short/weak	Support largely collapses to one molecule or cannot be assessed	Similar signals seen in negatives or batch-specific	Do not advance to PCR by default; prioritize re-alignment, stricter filtering, alternative aligner check, or re-sequencing / alternative library prep
Moderate-confidence candidate	At least one clean split read with coherent breakpoint window; discordant pairs support the same locus; orientation mostly consistent	Host segment maps uniquely enough to place locus; vector segment aligns to a coherent region	More than one independent molecule or evidence suggests independence (distinct fragment starts/lengths)	Absent in negatives; not tracking with batch	Consider targeted junction PCR / nested PCR + Sanger; if locus is hard (repeats), consider capture re-seq
High-confidence candidate	Multiple split reads with concordant breakpoint and orientation, supported by discordant pairs; junction sequence reconstructable	High-confidence mapping on host side; breakpoint not solely in low-complexity; vector mapping stable	Multiple independent molecules; not dominated by duplicates; reproducible across replicate libraries when available	Clean separation from controls; consistent across runs	Advance to orthogonal confirmation; consider alternative sequencing (long read) if junction complexity or ITR-adjacent structure is suspected

⚠️ Warning: These grades are qualitative by design. Avoid universal read-count thresholds—acceptable evidence depends on sequencing depth, enrichment design, duplicate handling, and genomic context.

11. A decision table for whether a candidate should advance to validation

Use this as a practical triage table during review meetings.

Question	If YES	If NO
Does the candidate have at least one junction-spanning split read with a clear host segment and a clear vector segment?	Continue	Treat as low-confidence unless other evidence is unusually strong; prioritize deeper review / re-analysis
Do supporting reads cluster tightly around one host breakpoint window (rather than scattered)?	Continue	Suspect artifact, mapping ambiguity, or mixed events; do not validate yet
Are supporting reads largely independent (not all duplicates of one molecule)?	Continue	Revisit deduplication; request molecule-level evidence; consider alternative sequencing confirmation
Is the host breakpoint in a reasonably mappable region (or does the read include enough unique host sequence)?	Continue	PCR/Sanger may fail; prefer capture re-seq or long-read validation
Is the candidate absent in negative controls and not explained by batch/index effects?	Continue	Deprioritize; likely integration site false positive or contamination
Is the junction structure coherent (orientation consistent; vector breakpoint region plausible)?	Continue	Re-align with alternative parameters/reference; consider local assembly

A candidate should usually advance to validation only when most answers are YES. If the decision is ambiguous, choose a validation method that maximizes interpretability (e.g., capture-based resequencing or long reads) rather than a single PCR.

12. Orthogonal confirmation options (AAV insertion site confirmation)

Orthogonal confirmation is where you move from "candidate" to "confirmed integration junction sequence in this sample." Each method has strengths and predictable failure modes.

Targeted PCR

Targeted PCR uses primers flanking the host breakpoint region with an expectation of amplifying across the junction.

When it works best:

• Host side is unique and primerable
• Expected junction is not extremely long or structurally complex
• Candidate is moderate/high confidence

Common issues:

• Non-specific amplification in repeat-rich loci
• Failure due to low template abundance (candidate below assay detection)
• Primer mismatch if the local host sequence differs from reference (polymorphisms) or if vector reference is wrong

Junction-specific PCR

Junction-specific PCR explicitly designs one primer on the vector side and one on the host side (or nested sets), attempting to amplify only when the specific vector–host junction exists.

Why it's preferred for low-frequency candidates:

• It is much more specific than host-only strategies
• It directly tests the hypothesized junction model

In a 2026 study, AAV integration sites identified using CRISPR-Cas9-based enrichment and long-read nanopore sequencing were compared with hybridization-capture short-read sequencing and further confirmed by PCR. Although the experimental context involved clonal cell models rather than low-frequency tissue samples, the study illustrates the value of using an independent method to confirm candidate junctions.

Sanger sequencing

Sanger sequencing is the typical readout for PCR-based confirmation.

What Sanger can confirm:

• The existence of a specific junction amplicon
• Base-level breakpoint sequence across the vector–host junction

What Sanger cannot guarantee:

• That the junction represents the dominant structure in the sample
• That the integration is the only vector-derived structure present

Sanger also struggles when PCR products are mixed (multiple similar junctions), often producing ambiguous chromatograms.

Alternative sequencing confirmation

When PCR/Sanger is unlikely to succeed or interpret cleanly, consider sequencing-based confirmation:

• Capture-based resequencing targeting vector sequence, optimized for junction recovery
• Long-read confirmation (e.g., targeted enrichment + long reads) when junction structure is complex or when repeats make short reads ambiguous

Long reads can be especially valuable when you need contiguous context: they can place the junction within a longer host sequence and reveal whether the vector-side segment is truncated or rearranged.

Figure 3. Validation workflow for low-frequency AAV integration site candidates, from computational detection through evidence review, control comparison, and orthogonal confirmation.

13. Situations in which validation may fail

Validation failure does not necessarily mean the original candidate was false, but it often reflects predictable limitations.

Common reasons PCR/Sanger validation fails

1. Candidate abundance below wet-lab detection. Low-frequency events can be real but still not amplifiable from available input DNA.
2. Ambiguous host locus. Repeats, segmental duplications, or low complexity prevent primer specificity.
3. Wrong breakpoint model. Misplaced vector or host breakpoint due to alignment ambiguity; primers flank the wrong coordinates.
4. Junction complexity. Micro-insertions, rearrangements, concatemer-like structures, or multiple junctions generate mixed PCR products.
5. Template composition issues. Fragmentation patterns, inhibitors, or damaged DNA reduce effective amplifiable templates.

Common reasons sequencing-based confirmation fails

• Insufficient on-target enrichment (capture probes not covering the relevant vector segment)
• Bioinformatic filters that remove genuine low-support reads as "noise"
• Index hopping or contamination that produces low-level junctions that are real sequencing artifacts rather than sample-derived biology

A good practice is to treat validation as an iterative process: if PCR fails, do not automatically discard the candidate—reassess mapping context and consider whether the confirmation method was mismatched to the locus.

14. Reporting low-frequency candidates without overinterpretation

Low-frequency candidates should be reported in a way that preserves their potential importance without implying confirmation.

Recommended reporting language:

• Use "candidate" or "putative junction" terminology.
• Explicitly state the evidence type(s): split reads, discordant pairs, assembled contigs.
• Report duplicate handling and independent molecule estimates.
• State whether the candidate was observed in controls.
• State whether orthogonal confirmation was attempted and the outcome.

Avoid:

• Calling candidates "confirmed integration" without junction confirmation.
• Overstating implications (no clinical safety conclusions).
• Presenting a single read count as "frequency" without explaining assumptions.

If you need a structured format, consider including an evidence grade for each candidate plus a brief validation recommendation.

15. Information needed from the original analysis

If you are reviewing a report (especially from a CRO), request the minimal information needed to audit low-frequency calls:

1. Reference definitions

   • Host genome build and decoy/alt handling
   • Vector reference sequence and any known variants
   • Whether alignment used a hybrid host+vector reference

2. Read-level evidence package (per candidate)

   • Read IDs for supporting reads
   • Alignment excerpts or genome browser snapshots showing split reads and discordant pairs
   • Breakpoint coordinates (host + vector) and orientation

3. Duplicate and independence metrics

   • Raw supporting reads vs deduplicated reads
   • Independent molecule estimation approach
   • Global duplication rate and junction-read duplication rate

4. QC and artifact context

   • Negative control results and whether candidates appear in controls
   • Batch/lane/index information
   • Any steps taken to mitigate library-prep artifacts

5. Pipeline parameters

   • Mapping tool(s), key filters, multi-mapper handling
   • Breakpoint clustering strategy
   • Whether alternative aligner or local assembly was used for ambiguous candidates

Without these items, you cannot meaningfully separate "weak true signal" from "weak artifact."

For teams designing or outsourcing ISA, a useful related reference is the CD Genomics overview on AAV integration site analysis, which can help define expected deliverables and evidence outputs.

16. Conclusion

Validating low-frequency AAV integration sites is fundamentally an exercise in evidence integration. Supporting reads matter, but they only become meaningful when interpreted through duplicate handling, mapping context, junction coherence, and controls.

If you do one thing differently after reading this guide, make it this: require a candidate-level evidence packet (reads, mapping context, independence, control comparison) before committing to wet-lab validation. That small discipline prevents most integration site false positive validations and makes true low-frequency events easier to confirm.

If you'd like an independent review of your candidate list (including evidence grading and a validation plan), CD Genomics can support integration site evidence review and escalation to alternative sequencing when PCR is unlikely to resolve junction structure.

17. FAQ

1) What is the minimum evidence needed to call something a candidate integration site?

A credible candidate usually requires at least one read or read pair that supports a vector–host junction model and localizes to a specific host breakpoint window. In practice, split-read evidence is the most informative because it directly spans the junction; discordant pairs alone may be sufficient for a "watch list" candidate but often aren't enough to justify PCR without additional review.

2) Why do low-frequency candidates show up in negative controls?

Low-level junction-like reads can arise from library preparation artifacts, low-level cross-sample contamination, or alignment ambiguity. A comparative rAAV insertion site methods study observed apparent insertion sites in vehicle and spike-in controls and tied a substantial component to library preparation conditions (see the 2023 comparative analysis). That's why control-aware filtering is essential.

3) Can I treat supporting read count as an estimate of integration frequency?

Not reliably. Read counts are affected by PCR amplification bias, enrichment efficiency, mapping filters, and duplicate handling. Unless you have a validated method to estimate independent molecules and normalize appropriately, supporting reads should be treated as an evidence measure—not a direct frequency readout.

4) What should I look for in a genome browser when reviewing a candidate?

Focus on (a) whether split reads have clean, high-quality clipped bases that map to vector sequence; (b) whether discordant pairs cluster around the same host locus; (c) whether mapping qualities are stable; and (d) whether breakpoint positions are coherent rather than scattered. Also inspect repeat annotations or local mappability, because those frequently explain low-support artifacts.

5) When should I skip PCR and go directly to alternative sequencing confirmation?

Consider escalating when the host locus is repeat-rich, when split reads are too short/ambiguous to place uniquely, when the junction appears structurally complex (multiple breakpoints or rearrangements), or when the expected amplicon would be large or difficult to amplify. In these cases, capture-based resequencing or long-read approaches may produce clearer evidence than iterative PCR.

6) If PCR validation fails, does that prove the candidate was a false positive?

No. PCR can fail because the event is below detection for available input DNA, primers are not specific due to repeats, the breakpoint model was off by enough bases to break primer design, or the junction is structurally complex and produces mixed products. A failed PCR should trigger re-review of mapping context and method choice, not an automatic dismissal.

7) How should I report low-frequency candidates without overclaiming?

Report them as candidates with an explicit evidence summary: read types, mapping context, duplicate/independence handling, and whether controls show similar signals. Use an evidence grade (low/moderate/high) and state whether orthogonal confirmation was attempted. Avoid labeling anything "confirmed integration" without junction sequence confirmation.

8) What should I request from a CRO or bioinformatics team to evaluate low-frequency calls?

Request the full evidence packet: reference definitions (host build and vector reference), candidate-level read IDs and alignment views, breakpoint coordinates and orientation, raw vs deduplicated support plus independent molecule estimates, and control comparisons. Without those, you are forced to trust a black-box call—exactly the situation where low-frequency artifacts get mistaken for biology.