A girl of three with a recessive neuromuscular condition has one clearly pathogenic variant in the expected gene and, on the other allele, a change four bases into an intron. The splice predictors disagree: one scores it as likely to weaken the donor site, another as negligible. The laboratory reports it as a variant of uncertain significance, which is correct, and the case is closed as unresolved, which is a choice.
A muscle biopsy had been taken a year earlier for histology. Nobody ordered RNA from it. The question the predictors could not settle had an answer in a freezer.
What the field knows
Variants that alter splicing are a substantial and under-called share of pathogenic variation. Canonical splice-site changes, at the two bases either side of an exon, are treated as loss of function under the interpretation framework (Richards et al., 2015), with the caveats the ClinGen sequence variant working group later set out for when that assumption fails (Abou Tayoun et al., 2018). Beyond those two bases, in the extended splice region, in exonic splice enhancers and deep in introns, variants can create or destroy splice signals, and these are the ones the exome either does not capture or captures and cannot score.
Prediction has improved and remains prediction. SpliceAI (Jaganathan et al., 2019) trained a deep network on pre-mRNA sequence to predict splice junctions and showed substantially better agreement with observed splicing effects than earlier position-weight and motif tools, and it has become the default in many pipelines. The ClinGen recommendations for applying splicing evidence (Walker et al., 2023) set out how a predictor's output and RNA data each map onto the evidence codes, and are explicit that a computational prediction alone reaches at most a supporting or moderate weight, while an RNA result demonstrating aberrant splicing can carry strong weight.
RNA sequencing finds what prediction cannot. Cummings and colleagues (2017), sequencing muscle RNA from patients with undiagnosed muscle disease, made diagnoses in a notable share by showing aberrant splicing, including from deep intronic variants that no exome had called and from variants of uncertain significance that no predictor had scored convincingly. Frésard and colleagues (2019) extended the approach to blood RNA across a broader range of conditions, with the limitation that the gene has to be expressed in the tissue sampled.
The two kinds of evidence answer different questions. A predictor says how much a sequence change resembles the changes that have altered splicing elsewhere. An RNA result says what this variant did to this transcript in this tissue. The second is a measurement; the first is an informed guess about what the measurement would show, and the guidance weights them accordingly.
Where practice breaks
The prediction is read as a verdict. A high splice score moves a variant to likely pathogenic in the scientist's mind before the evidence framework has been applied, and a low score dismisses one. The framework allows neither; the number is a prior for an experiment, and the experiment is what is missing.
The predictors disagree and the disagreement is resolved by preference. One laboratory trusts one tool; another laboratory trusts another; a variant is uncertain in one and supporting in the other, and the classification archive records the disagreement as a conflict with no indication that both laboratories were looking at the same unanswered question.
RNA is treated as a research assay. Even where the gene is expressed in blood, a request for RNA from the patient is unusual, slow and often routed through a research laboratory, so it is not made. Where a tissue sample already exists, as in the opening case, nobody connects the stored biopsy with the open splicing question, because the sample is recorded in pathology and the question in genetics.
Tissue is the quiet constraint. A gene expressed only in brain or retina cannot be tested from blood, and the predictor is then the only evidence there will be. That is a reason to say so plainly, with the weight the guidance allows, rather than a reason to let the prediction do the work of a measurement it cannot replace.
The case closes on 'uncertain significance' as if the uncertainty were a property of the variant. It is a property of the evidence collected so far. A report that says uncertain and stops is different from a report that says uncertain, names the assay that would resolve it, and says whether the tissue exists. The first ends the case. The second hands the next person a plan.
How Tulparus handles it
For every variant in or near a splice region, the workbench records the predictions, each with its tool and version, as what they are: computational evidence at the weight the ClinGen guidance allows. Where predictors disagree, both scores are shown with their thresholds, and the disagreement is named on the hypothesis rather than averaged away.
The splicing question becomes a missing-evidence item with a route to an answer. The workbench knows from the knowledge store whether the gene is expressed in blood, fibroblasts or muscle, and from the case which samples exist. For the girl above, the item would read: splicing effect of this variant unresolved; gene expressed in muscle; muscle biopsy recorded in this case; RNA analysis of that sample proposed, with the evidence weight a positive or negative result would carry.
The hypothesis that the gene explains the case is scored with the splicing variant as an unresolved second allele, so that the case stays open with a reason, and ranks the RNA step above further sequencing by what it would settle. When an RNA result arrives, it is attached as evidence with its own provenance, the variant's classification is revisited through the framework, and the report is regenerated from the updated case.
No prediction is ever displayed as a probability that the variant is pathogenic or that the gene is the diagnosis. The scores are labelled by tool and read against the published thresholds, and the page says what they can and cannot establish. A reader sees a splice score for what it is, one input among several to a hypothesis a person is still weighing.
What you can do tomorrow
The habit to build is to treat a splice prediction as the start of a plan.
- For every splice-region variant of uncertain significance in an open case, write down whether the gene is expressed in an accessible tissue and whether a sample exists. That alone will resolve some cases.
- Report splice predictions with the tool and version, and with the evidence weight the ClinGen guidance allows, so that a reader knows how much the number is worth.
- Agree a route to RNA analysis for clinical cases, even if it is a research laboratory with a service agreement, so that the request is routine rather than exceptional.
- When a case closes as uncertain because of a splicing question, say so in the report, and name the assay that would answer it.
Sources
- Richards S, Aziz N, Bale S, et al. Standards and guidelines for the interpretation of sequence variants: a joint consensus recommendation of the American College of Medical Genetics and Genomics and the Association for Molecular Pathology. Genetics in Medicine, 2015.
- Abou Tayoun AN, Pesaran T, DiStefano MT, et al. Recommendations for interpreting the loss of function PVS1 ACMG/AMP variant criterion. Human Mutation, 2018.
- Jaganathan K, Kyriazopoulou Panagiotopoulou S, McRae JF, et al. Predicting splicing from primary sequence with deep learning. Cell, 2019.
- Walker LC, Hoya M, Wiggins GAR, et al. Using the ACMG/AMP framework to capture evidence related to predicted and observed impact on splicing: recommendations from the ClinGen SVI Splicing Subgroup. American Journal of Human Genetics, 2023.
- Cummings BB, Marshall JL, Tukiainen T, et al. Improving genetic diagnosis in Mendelian disease with transcriptome sequencing. Science Translational Medicine, 2017.
- Frésard L, Smail C, Ferraro NM, et al. Identification of rare-disease genes using blood transcriptome sequencing and large control cohorts. Nature Medicine, 2019.