The Variant That Only Showed Its Hand in RNA

Bioinformatics Insight Series

The Variant That Only Showed Its Hand in RNA

Whole-genome sequencing routinely surfaces deep-intronic and synonymous variants that prediction tools can't confidently call spliceogenic. RNA-seq is the functional readout that turns one of those variants of uncertain significance into a diagnosis — or clears it. Here is how the two assays actually work together, and where the pairing quietly fails.

Kanna Nandakumar, PhD · Zetobit · 11 min read

A whole-genome sequencing report lands, the panel of known disease genes comes back clean, and the case stalls. Then someone notices a variant sitting 150 bases deep in an intron — nowhere near a canonical splice site, silent by every coding metric, flagged by an in silico tool with a middling score that could mean anything. On DNA alone, that variant is unresolvable. It is a coin toss dressed up as a probability. The only way to know whether it does anything is to look at what the gene actually transcribes.

This is the specific, high-value niche where pairing WGS with RNA-seq earns its cost. Not as a general fishing expedition, but as a targeted functional assay for a class of variants that WGS is uniquely good at finding and uniquely bad at interpreting: the ones that break splicing without touching the protein-coding grammar we know how to read.

Why WGS finds them and can't call them

Exome sequencing, still the standard-of-care first test in many settings, has a diagnostic rate around 31%,¹ and genome sequencing adds perhaps another 10–15% on top of that.² A meaningful share of what remains unsolved is not missing from the data — it is present but uninterpretable. WGS reads across the whole intron, so it sees the deep-intronic variant that exome capture would have missed entirely. It sees the synonymous change that alters an exonic splicing enhancer. It sees the variant twenty bases into an intron that creates a brand-new cryptic donor. The DNA evidence is there.

What is missing is a reliable way to predict consequence. Splicing prediction has improved sharply — SpliceAI is genuinely useful — but on real diagnostic cohorts its ceiling is modest. In one clinical implementation, SpliceAI at its recommended threshold recovered 57% of confirmed pathogenic aberrant-splicing variants, and MMSplice recovered 38%,³ with both performing worst precisely on the deep-intronic and coding-region variants where you most need help. A more recent benchmark across millions of rare variants in 49 tissues found DNA-based models capped near 12% precision at 20% recall for tissue-specific aberrant splicing.⁴ These are prioritization tools, not adjudicators. They tell you where to look. They do not tell you what happened.

A splice prediction narrows the suspect list. RNA-seq is the eyewitness that puts the variant at the scene.

What RNA-seq actually measures

RNA-seq resolves splice variants through three orthogonal signals, and it is worth being precise about which one carries the weight in a given case, because they fail in different ways.

Signal	What it detects	Typical WGS variant it explains
Aberrant splicing	Exon skipping, intron retention, cryptic splice-site use — reads that cross junctions the reference transcript never uses	Deep-intronic cryptic-site creators; canonical-site disruptors; synonymous ESE variants
Allelic expression imbalance	One allele expressed far below the other, often the signature of nonsense-mediated decay eliminating a defective transcript	Variants triggering NMD; regulatory and promoter variants that silence one copy
Expression outliers	A gene whose overall level sits far outside the reference distribution across a control cohort	Deep regulatory variants; structural events that abolish or amplify expression

For splice-variant resolution specifically, the first signal is the headline and the second is the confirmation. A deep-intronic variant that creates a cryptic donor will show up as a novel junction in the split-read data, and — if that junction introduces a premature stop — as allelic imbalance from the mutant transcript being degraded. When both fire on the same allele, you are no longer interpreting a prediction. You are reading a measurement.

How a silent-on-DNA variant becomes visible in RNA. A deep-intronic variant creates a cryptic splice donor, splicing in a pseudoexon that the reference transcript never contains. WGS sees the variant but cannot confidently predict its effect; RNA-seq shows the novel junctions directly as split reads, and — where the pseudoexon introduces a premature stop — nonsense-mediated decay drops the mutant allele's expression, giving a second, independent line of evidence on the same allele.

The interpretation loop, in practice

The two assays are not run in parallel and compared at the end. They work in a loop, and the order matters for both yield and cost.

1. WGS proposes; RNA-seq disposes

The most efficient workflow uses WGS variants — filtered by rarity and by splice-prediction score — to define a short list of candidate positions, then interrogates the RNA-seq data at exactly those loci for aberrant junctions and allelic imbalance. This candidate-directed approach is where the pairing is strongest. In cohorts with a prior candidate splice VUS, RNA-seq resolution rates are high: one clinical study reported a 60% diagnostic uplift among cases carrying a candidate VUS, versus under 3% in cases with no prior candidate.⁵ The DNA gives the RNA-seq somewhere to look.

2. RNA-seq proposes; WGS explains

The loop also runs the other way. A transcriptome-wide scan for splicing and expression outliers — using tools such as FRASER for aberrant splicing and OUTRIDER for expression — can surface an aberrant event with no prior DNA candidate at all. But an unexplained aberrant junction is not yet a diagnosis; you then return to the WGS to find the causal variant underneath it, which is often the deep-intronic change that no coding-focused filter would ever have surfaced. In the Baylor Undiagnosed Diseases Network implementation, genome sequencing was specifically required alongside RNA-seq to identify the genomic event responsible for the transcriptomic finding in several cases.⁶

Reclassification, not just discovery

The clinically underappreciated payoff is bidirectional reclassification. RNA-seq does not only upgrade VUS to pathogenic — it also clears them. A variant with a scary splice prediction that shows a completely normal transcript can be moved toward benign with functional confidence.

One analysis concluded that up to 31% of splicing VUS could reach a likely-pathogenic or likely-benign classification once RNA-seq evidence was applied.⁷ Removing a false lead is worth nearly as much as finding the true one, and it is a benefit pure DNA reanalysis cannot deliver.

Where the pairing quietly fails

This is the part that separates a workable clinical service from a research demo. RNA-seq resolves splice variants only when the transcript is present to be read, and that condition fails more often than the headline yield figures suggest.

The tissue problem

You can only sequence RNA from what you can sample — overwhelmingly blood and cultured fibroblasts, the clinically accessible tissues. If the disease gene isn't adequately expressed there, the assay is blind. The numbers are sobering: roughly 37% of coding genes in blood and 48% in fibroblasts show low expression (TPM < 1),⁸ and even among expressed genes, tissue-specific splicing means a canonical transcript can be present but spliced differently than in the affected tissue. One systematic analysis found that in inaccessible tissues, about 40% of genes have splicing inadequately represented by at least one accessible tissue, and roughly 6% are inadequately represented by all of them.⁹ A normal blood RNA-seq does not exonerate a variant in a gene that blood cannot report on.

The NMD trap

Nonsense-mediated decay is both a signal and a saboteur. It gives you the allelic-imbalance evidence that confirms a loss-of-function transcript — but it does so by degrading that transcript, which means the aberrant splice product you most want to characterize can be nearly gone from the pool before you sequence it. Definitive characterization sometimes requires culturing patient cells with an NMD inhibitor to stabilize the transcript long enough to see what it actually is, which pushes the work out of a standard blood-draw workflow entirely.

The reference and depth problem

Novel junction detection depends on split-read alignment, and splice-aware aligners inherit a bias toward annotated junctions. A truly novel cryptic junction has fewer supporting reads and less annotation help, so calling it reliably needs adequate depth at the locus and careful outlier statistics against a control cohort — not a single-sample eyeball. Underpowered RNA-seq produces the worst outcome: a confident-looking normal result that is really just insufficient coverage.

What to ask before ordering the pair

The pairing is powerful and specific, not universal. Before adding RNA-seq to a WGS workflow for splice resolution, the questions that actually determine success are: Is the candidate gene expressed in an accessible tissue? (If not, the assay may be blind before it starts.) Is there a prior DNA candidate to direct the analysis? (Candidate-directed yields dwarf undirected ones.) Will the aberrant transcript survive to be sequenced, or will NMD erase it? And is the RNA-seq powered — depth, controls, and outlier statistics — to call a novel junction rather than just confirm annotated ones?

Answer those honestly and the pairing does something no amount of DNA reanalysis can: it replaces a splice prediction with a splice observation. The variant that only showed its hand in RNA was always doing its damage. WGS found where it was hiding. RNA-seq is what finally caught it in the act.

---

References

1. Retterer K, Juusola J, Cho MT, et al. Clinical application of whole-exome sequencing across clinical indications. Genetics in Medicine. 2016;18(7):696–704. doi:10.1038/gim.2015.148
2. Cummings BB, Marshall JL, Tukiainen T, et al. Improving genetic diagnosis in Mendelian disease with transcriptome sequencing. Science Translational Medicine. 2017;9(386):eaal5209. doi:10.1126/scitranslmed.aal5209
3. Yépez VA, Gusic M, Kopajtich R, et al. Clinical implementation of RNA sequencing for Mendelian disease diagnostics. Genome Medicine. 2022;14:38. doi:10.1186/s13073-022-01019-9
4. Wagner N, Çelik MH, Hölzlwimmer FR, et al. Aberrant splicing prediction across human tissues (AbSplice). Nature Genetics. 2023;55:861–870. doi:10.1038/s41588-023-01373-3
5. Zhang Y, et al. Blood RNA-seq in rare disease diagnostics: a comparative study of cases with and without candidate variants. Journal of Translational Medicine. 2025;23:552. doi:10.1186/s12967-025-06609-w
6. Murdock DR, Dai H, Burrage LC, et al. Transcriptome-directed analysis for Mendelian disease diagnosis overcomes limitations of conventional genomic testing. Journal of Clinical Investigation. 2021;131(1):e141500. doi:10.1172/JCI141500
7. Riepe TV, Khan M, Roosing S, Cremers FPM, 't Hoen PAC. Benchmarking deep learning splice prediction tools using functional splice assays. Human Mutation. 2021;42(7):799–810. doi:10.1002/humu.24225
8. Bournazos AM, Riley LG, Bommireddipalli S, et al. Standardized practices for RNA diagnostics using clinically accessible specimens reclassifies 75% of putative splicing variants. Genetics in Medicine. 2022;24(1):130–145. doi:10.1016/j.gim.2021.09.001
9. Aicher JK, Jewell P, Vaquero-Garcia J, Barash Y, Bhoj EJ. Mapping RNA splicing variations in clinically accessible and nonaccessible tissues to facilitate Mendelian disease diagnosis using RNA-seq. Genetics in Medicine. 2020;22(7):1181–1190. doi:10.1038/s41436-020-0780-y

Zetobit Bioinformatics Insight Series · Written by Kanna Nandakumar, PhD. Zetobit is a boutique bioinformatics consulting firm specializing in CAP/CLIA-compliant NGS pipelines and multi-omics analysis. zetobit.com