SpliceVault predicts the precise nature of variant-associated mis-splicing

Spatial enrichment and genomic analyses reveal the link of NOMO1 with amyotrophic lateral sclerosis

Amyotrophic lateral sclerosis (ALS) is a severe motor neuron disease with uncertain genetic predisposition in most sporadic cases. The spatial architecture of cell types and gene expression are the basis of cell-cell interactions, biological function and disease pathology, but are not well investigated in the human motor cortex, a key ALS-relevant brain region. Recent studies indicated single nucleus transcriptomic features of motor neuron vulnerability in ALS motor cortex. However, the brain regional vulnerability of ALS-associated genes and the genetic link between region-specific genes and ALS risk remain largely unclear. Here, we developed an entropy-weighted differential gene expression matrix-based tool (SpatialE) to identify the spatial enrichment of gene sets in spatial transcriptomics. We benchmarked SpatialE against another enrichment tool (multimodal intersection analysis) using spatial transcriptomics data from both human and mouse brain tissues. To investigate regional vulnerability, we analysed three human motor cortex and two dorsolateral prefrontal cortex tissues for spatial enrichment of ALS-associated genes. We also used Cell2location to estimate the abundance of cell types in ALS-related cortex layers. To dissect the link of regionally expressed genes and ALS risk, we performed burden analyses of rare loss-of-function variants detected by whole-genome sequencing in ALS patients and controls, then analysed differential gene expression in the TargetALS RNA-sequencing dataset. SpatialE showed more accurate and specific spatial enrichment of regional cell type markers than multimodal intersection analysis in both mouse brain and human dorsolateral prefrontal cortex. Spatial transcriptomic analyses of human motor cortex showed heterogeneous cell types and spatial gene expression profiles. We found that 260 manually curated ALS-associated genes are significantly enriched in layer 5 of the motor cortex, with abundant expression of upper motor neurons and layer 5 excitatory neurons. Burden analyses of rare loss-of-function variants in Layer 5-associated genes nominated NOMO1 as a novel ALS-associated gene in a combined sample set of 6814 ALS patients and 3324 controls (P = 0.029). Gene expression analyses in CNS tissues revealed downregulation of NOMO1 in ALS, which is consistent with a loss-of-function disease mechanism. In conclusion, our integrated spatial transcriptomics and genomic analyses identified regional brain vulnerability in ALS and the association of a layer 5 gene (NOMO1) with ALS risk.

RNA variant assessment using transactivation and transdifferentiation

Understanding the impact of splicing and nonsense variants on RNA is crucial for the resolution of variant classification as well as their suitability for precision medicine interventions. This is primarily enabled through RNA studies involving transcriptomics followed by targeted assays using RNA isolated from clinically accessible tissues (CATs) such as blood or skin of affected individuals. Insufficient disease gene expression in CATs does however pose a major barrier to RNA based investigations, which we show is relevant to 1,436 Mendelian disease genes. We term these "silent" Mendelian genes (SMGs), the largest portion (36%) of which are associated with neurological disorders. We developed two approaches to induce SMG expression in human dermal fibroblasts (HDFs) to overcome this limitation, including CRISPR-activation-based gene transactivation and fibroblast-to-neuron transdifferentiation. Initial transactivation screens involving 40 SMGs stimulated our development of a highly multiplexed transactivation system culminating in the 6- to 90,000-fold induction of expression of 20/20 (100%) SMGs tested in HDFs. Transdifferentiation of HDFs directly to neurons led to expression of 193/516 (37.4%) of SMGs implicated in neurological disease. The magnitude and isoform diversity of SMG expression following either transactivation or transdifferentiation was comparable to clinically relevant tissues. We apply transdifferentiation and/or gene transactivation combined with short- and long-read RNA sequencing to investigate the impact that variants in USH2A, SCN1A, DMD, and PAK3 have on RNA using HDFs derived from affected individuals. Transactivation and transdifferentiation represent rapid, scalable functional genomic solutions to investigate variants impacting SMGs in the patient cell and genomic context.

A systematic assessment of the impact of rare canonical splice site variants on splicing using functional and in silico methods

Canonical splice site variants (CSSVs) are often presumed to cause loss-of-function (LoF) and are assigned very strong evidence of pathogenicity (according to American College of Medical Genetics/Association for Molecular Pathology criterion PVS1). The exact nature and predictability of splicing effects of unselected rare CSSVs in blood-expressed genes are poorly understood. We identified 168 rare CSSVs in blood-expressed genes in 112 individuals using genome sequencing, and studied their impact on splicing using RNA sequencing (RNA-seq). There was no evidence of a frameshift, nor of reduced expression consistent with nonsense-mediated decay, for 25.6% of CSSVs: 17.9% had wildtype splicing only and normal junction depths, 3.6% resulted in cryptic splice site usage and in-frame insertions or deletions, 3.6% resulted in full exon skipping (in frame), and 0.6% resulted in full intron inclusion (in frame). Blind to these RNA-seq data, we attempted to predict the precise impact of CSSVs by applying in silico tools and the ClinGen Sequence Variant Interpretation Working Group 2018 guidelines for applying PVS1 criterion. The predicted impact on splicing using (1) SpliceAI, (2) MaxEntScan, and (3) AutoPVS1, an automatic classification tool for PVS1 interpretation of null variants that utilizes Ensembl Variant Effect Predictor and MaxEntScan, was concordant with RNA-seq analyses for 65%, 63%, and 61% of CSSVs, respectively. In summary, approximately one in four rare CSSVs did not show evidence for LoF based on analysis of RNA-seq data. Predictions from in silico methods were often discordant with findings from RNA-seq. More caution may be warranted in applying PVS1-level evidence to CSSVs in the absence of functional data.

Long-read RNA-seq demarcates cis- and trans-directed alternative RNA splicing

Genetic regulation of alternative splicing constitutes an important link between genetic variation and disease. Nonetheless, RNA splicing is regulated by both cis-acting elements and trans-acting splicing factors. Determining splicing events that are directed primarily by the cis- or trans-acting mechanisms will greatly inform our understanding of the genetic basis of disease. Here, we show that long-read RNA-seq, combined with our new method isoLASER, enables a clear segregation of cis- and trans-directed splicing events for individual samples. The genetic linkage of splicing is largely individual-specific, in stark contrast to the tissue-specific pattern of splicing profiles. Analysis of long-read RNA-seq data from human and mouse revealed thousands of cis-directed splicing events susceptible to genetic regulation. We highlight such events in the HLA genes whose analysis was challenging with short-read data. We also highlight novel cis-directed splicing events in Alzheimer's disease-relevant genes such as MAPT and BIN1. Together, the clear demarcation of cis- and trans-directed splicing paves ways for future studies of the genetic basis of disease.

Splicing analysis of STAT3 tandem donor suggests non-canonical binding registers for U1 and U6 snRNAs

Tandem donor splice sites (5'ss) are unique regions with at least two GU dinucleotides serving as splicing cleavage sites. The Δ3 tandem 5'ss are a specific subclass of 5'ss separated by 3 nucleotides which can affect protein function by inserting/deleting a single amino acid. One 5'ss is typically preferred, yet factors governing particular 5'ss choice are not fully understood. A highly conserved exon 21 of the STAT3 gene was chosen as a model to study Δ3 tandem 5'ss splicing mechanisms. Based on multiple lines of experimental evidence, endogenous U1 snRNA most likely binds only to the upstream 5'ss. However, the downstream 5'ss is used preferentially, and the splice site choice is not dependent on the exact U1 snRNA binding position. Downstream 5'ss usage was sensitive to exact nucleotide composition and dependent on the presence of downstream regulatory region. The downstream 5'ss usage could be best explained by two novel interactions with endogenous U6 snRNA. U6 snRNA enables the downstream 5'ss usage in STAT3 exon 21 by two mechanisms: (i) binding in a novel non-canonical register and (ii) establishing extended Watson-Crick base pairing with the downstream regulatory region. This study suggests that U6:5'ss interaction is more flexible than previously thought.

A Deep Intronic PKHD1 Variant Identified by SpliceAI in a Deceased Neonate With Autosomal Recessive Polycystic Kidney Disease

The etiologies of newborn deaths in neonatal intensive care units usually remain unknown, even after genetic testing. Whole-genome sequencing, combined with artificial intelligence-based methods for predicting the effects of non-coding variants, provide an avenue for resolving these deaths. Using one such method, SpliceAI, we identified a maternally inherited deep intronic PKHD1 splice variant (chr6:52030169T>C), in trans with a pathogenic missense variant (p.Thr36Met), in a newborn who died of autosomal recessive polycystic kidney disease at age 2 days. We validated the deep intronic variant's impact in maternal urine-derived cells expressing PKHD1. Reverse transcription polymerase chain reaction followed by Sanger sequencing showed that the variant causes inclusion of 147bp of the canonical intron between exons 29 and 30 of PKHD1 into the mRNA, including a premature stop codon. Allele-specific expression analysis at a heterozygous site in the mother showed that the mutant allele completely suppresses canonical splicing. In an unrelated healthy control, there was no evidence of transcripts including the novel splice junction. We returned a diagnostic report to the parents, who underwent in vitro embryo selection.

Reference-informed prediction of alternative splicing and splicing-altering mutations from sequences

Alternative splicing plays a crucial role in protein diversity and gene expression regulation in higher eukaryotes and mutations causing dysregulated splicing underlie a range of genetic diseases. Computational prediction of alternative splicing from genomic sequences not only provides insight into gene-regulatory mechanisms but also helps identify disease-causing mutations and drug targets. However, the current methods for the quantitative prediction of splice site usage still have limited accuracy. Here, we present DeltaSplice, a deep neural network model optimized to learn the impact of mutations on quantitative changes in alternative splicing from the comparative analysis of homologous genes. The model architecture enables DeltaSplice to perform "reference-informed prediction" by incorporating the known splice site usage of a reference gene sequence to improve its prediction on splicing-altering mutations. We benchmarked DeltaSplice and several other state-of-the-art methods on various prediction tasks, including evolutionary sequence divergence on lineage-specific splicing and splicing-altering mutations in human populations and neurodevelopmental disorders, and demonstrated that DeltaSplice outperformed consistently. DeltaSplice predicted ~15% of splicing quantitative trait loci (sQTLs) in the human brain as causal splicing-altering variants. It also predicted splicing-altering de novo mutations outside the splice sites in a subset of patients affected by autism and other neurodevelopmental disorders, including 19 genes with recurrent splicing-altering mutations. Among the new candidate disease risk genes, MFN1 is involved in mitochondria fusion, which is frequently disrupted in autism patients. Our work expanded the capacity of in silico splicing models with potential applications in genetic diagnosis and the development of splicing-based precision medicine.

Deciphering the impact of coding and non-coding SCN1A gene variants on RNA splicing

Variants that disrupt normal pre-mRNA splicing are increasingly being recognized as a major cause of monogenic disorders. The SCN1A gene, a key epilepsy gene that is linked to various epilepsy phenotypes, is no exception. Approximately 10% of all reported variants in the SCN1A gene are designated as splicing variants, with many located outside of the canonical donor and acceptor splice sites, and most have not been functionally investigated. However, given its restricted expression pattern, functional analysis of splicing variants in the SCN1A gene could not be routinely performed. In this study, we conducted a comprehensive analysis of all reported SCN1A variants and their potential to impact SCN1A splicing and conclude that splicing variants are substantially misannotated and under-represented. We created a splicing reporter system consisting of 18 splicing vectors covering all 26 protein-coding exons with different genomic contexts and several promoters of varying strengths in order to reproduce the wild-type splicing pattern of the SCN1A gene, revealing cis-regulatory elements essential for proper recognition of SCN1A exons. Functional analysis of 95 SCN1A variants was carried out, including all 68 intronic variants reported in the literature, located outside of the splice sites canonical dinucleotides; 21 exonic variants of different classes (synonymous, missense, nonsense and in-frame deletion) and six variants observed in patients with epilepsy. Interestingly, almost 20% of tested intronic variants had no influence on SCN1A splicing, despite being reported as causative in the literature. Moreover, we confirmed that the majority of predicted exonic variants affect splicing unravelling their true molecular mechanism. We used functional data to perform genotype-phenotype correlation, revealing distinct distribution patterns for missense and splice-affecting 'missense' variants and observed no difference in the phenotype severity of variants leading to in-frame and out-of-frame isoforms, indicating that the Nav1.1 protein is highly intolerant to structural variations. Our work demonstrates the importance of functional analysis in proper variant annotation and provides a tool for high-throughput delineation of splice-affecting variants in SCN1A in a whole-gene manner.

Comprehensive splicing analysis of the alternatively spliced CHEK2 exons 8 and 10 reveals three enhancer/silencer-rich regions and 38 spliceogenic variants

Splicing is controlled by a large set of regulatory elements (SREs) including splicing enhancers and silencers, which are involved in exon recognition. Variants at these motifs may dysregulate splicing and trigger loss-of-function transcripts associated with disease. Our goal here was to study the alternatively spliced exons 8 and 10 of the breast cancer susceptibility gene CHEK2. For this purpose, we used a previously published minigene with exons 6-10 that produced the expected minigene full-length transcript and replicated the naturally occurring events of exon 8 [Δ(E8)] and exon 10 [Δ(E10)] skipping. We then introduced 12 internal microdeletions of exons 8 and 10 by mutagenesis in order to map SRE-rich intervals by splicing assays in MCF-7 cells. We identified three minimal (10-, 11-, 15-nt) regions essential for exon recognition: c.863_877del [ex8, Δ(E8): 75%] and c.1073_1083del and c.1083_1092del [ex10, Δ(E10): 97% and 62%, respectively]. Then 87 variants found within these intervals were introduced into the wild-type minigene and tested functionally. Thirty-eight of them (44%) impaired splicing, four of which (c.883G>A, c.883G>T, c.884A>T, and c.1080G>T) induced negligible amounts (<5%) of the minigene full-length transcript. Another six variants (c.886G>A, c.886G>T, c.1075G>A, c.1075G>T, c.1076A>T, and c.1078G>T) showed significantly strong impacts (20-50% of the minigene full-length transcript). Thirty-three of the 38 spliceogenic variants were annotated as missense, three as nonsense, and two as synonymous, underlying the fact that any exonic change is capable of disrupting splicing. Moreover, c.883G>A, c.883G>T, and c.884A>T were classified as pathogenic/likely pathogenic variants according to ACMG/AMP (American College of Medical Genetics and Genomics/Association for Molecular Pathology)-based criteria. © 2024 The Authors. The Journal of Pathology published by John Wiley & Sons Ltd on behalf of The Pathological Society of Great Britain and Ireland.

Combining full-length gene assay and SpliceAI to interpret the splicing impact of all possible SPINK1 coding variants

BACKGROUND

Single-nucleotide variants (SNVs) within gene coding sequences can significantly impact pre-mRNA splicing, bearing profound implications for pathogenic mechanisms and precision medicine. In this study, we aim to harness the well-established full-length gene splicing assay (FLGSA) in conjunction with SpliceAI to prospectively interpret the splicing effects of all potential coding SNVs within the four-exon SPINK1 gene, a gene associated with chronic pancreatitis.

RESULTS

Our study began with a retrospective analysis of 27 SPINK1 coding SNVs previously assessed using FLGSA, proceeded with a prospective analysis of 35 new FLGSA-tested SPINK1 coding SNVs, followed by data extrapolation, and ended with further validation. In total, we analyzed 67 SPINK1 coding SNVs, which account for 9.3% of the 720 possible coding SNVs. Among these 67 FLGSA-analyzed SNVs, 12 were found to impact splicing. Through detailed comparison of FLGSA results and SpliceAI predictions, we inferred that the remaining 653 untested coding SNVs in the SPINK1 gene are unlikely to significantly affect splicing. Of the 12 splice-altering events, nine produced both normally spliced and aberrantly spliced transcripts, while the remaining three only generated aberrantly spliced transcripts. These splice-impacting SNVs were found solely in exons 1 and 2, notably at the first and/or last coding nucleotides of these exons. Among the 12 splice-altering events, 11 were missense variants (2.17% of 506 potential missense variants), and one was synonymous (0.61% of 164 potential synonymous variants). Notably, adjusting the SpliceAI cut-off to 0.30 instead of the conventional 0.20 would improve specificity without reducing sensitivity.

CONCLUSIONS

By integrating FLGSA with SpliceAI, we have determined that less than 2% (1.67%) of all possible coding SNVs in SPINK1 significantly influence splicing outcomes. Our findings emphasize the critical importance of conducting splicing analysis within the broader genomic sequence context of the study gene and highlight the inherent uncertainties associated with intermediate SpliceAI scores (0.20 to 0.80). This study contributes to the field by being the first to prospectively interpret all potential coding SNVs in a disease-associated gene with a high degree of accuracy, representing a meaningful attempt at shifting from retrospective to prospective variant analysis in the era of exome and genome sequencing.

A metabolic signature for NADSYN1-dependent congenital NAD deficiency disorder

Nicotinamide adenine dinucleotide (NAD) is essential for embryonic development. To date, biallelic loss-of-function variants in 3 genes encoding nonredundant enzymes of the NAD de novo synthesis pathway - KYNU, HAAO, and NADSYN1 - have been identified in humans with congenital malformations defined as congenital NAD deficiency disorder (CNDD). Here, we identified 13 further individuals with biallelic NADSYN1 variants predicted to be damaging, and phenotypes ranging from multiple severe malformations to the complete absence of malformation. Enzymatic assessment of variant deleteriousness in vitro revealed protein domain-specific perturbation, complemented by protein structure modeling in silico. We reproduced NADSYN1-dependent CNDD in mice and assessed various maternal NAD precursor supplementation strategies to prevent adverse pregnancy outcomes. While for Nadsyn1+/- mothers, any B3 vitamer was suitable to raise NAD, preventing embryo loss and malformation, Nadsyn1-/- mothers required supplementation with amidated NAD precursors (nicotinamide or nicotinamide mononucleotide) bypassing their metabolic block. The circulatory NAD metabolome in mice and humans before and after NAD precursor supplementation revealed a consistent metabolic signature with utility for patient identification. Our data collectively improve clinical diagnostics of NADSYN1-dependent CNDD, provide guidance for the therapeutic prevention of CNDD, and suggest an ongoing need to maintain NAD levels via amidated NAD precursor supplementation after birth.

Benchmarking splice variant prediction algorithms using massively parallel splicing assays

BACKGROUND

Variants that disrupt mRNA splicing account for a sizable fraction of the pathogenic burden in many genetic disorders, but identifying splice-disruptive variants (SDVs) beyond the essential splice site dinucleotides remains difficult. Computational predictors are often discordant, compounding the challenge of variant interpretation. Because they are primarily validated using clinical variant sets heavily biased to known canonical splice site mutations, it remains unclear how well their performance generalizes.

RESULTS

We benchmark eight widely used splicing effect prediction algorithms, leveraging massively parallel splicing assays (MPSAs) as a source of experimentally determined ground-truth. MPSAs simultaneously assay many variants to nominate candidate SDVs. We compare experimentally measured splicing outcomes with bioinformatic predictions for 3,616 variants in five genes. Algorithms' concordance with MPSA measurements, and with each other, is lower for exonic than intronic variants, underscoring the difficulty of identifying missense or synonymous SDVs. Deep learning-based predictors trained on gene model annotations achieve the best overall performance at distinguishing disruptive and neutral variants, and controlling for overall call rate genome-wide, SpliceAI and Pangolin have superior sensitivity. Finally, our results highlight two practical considerations when scoring variants genome-wide: finding an optimal score cutoff, and the substantial variability introduced by differences in gene model annotation, and we suggest strategies for optimal splice effect prediction in the face of these issues.

CONCLUSION

SpliceAI and Pangolin show the best overall performance among predictors tested, however, improvements in splice effect prediction are still needed especially within exons.

Compilation of Genotype and Phenotype Data in GCDH-LOVD for Variant Classification and Further Application

Glutaric aciduria type 1 (GA-1) is a rare but treatable autosomal-recessive neurometabolic disorder of lysin metabolism caused by biallelic pathogenic variants in glutaryl-CoA dehydrogenase gene (GCDH) that lead to deficiency of GCDH protein. Without treatment, this enzyme defect causes a neurological phenotype characterized by movement disorder and cognitive impairment. Based on a comprehensive literature search, we established a large dataset of GCDH variants using the Leiden Open Variation Database (LOVD) to summarize the known genotypes and the clinical and biochemical phenotypes associated with GA-1. With these data, we developed a GCDH-specific variation classification framework based on American College of Medical Genetics and Genomics and the Association for Molecular Pathology guidelines. We used this framework to reclassify published variants and to describe their geographic distribution, both of which have practical implications for the molecular genetic diagnosis of GA-1. The freely available GCDH-specific LOVD dataset provides a basis for diagnostic laboratories and researchers to further optimize their knowledge and molecular diagnosis of this rare disease.

Genome-wide detection of human intronic AG-gain variants located between splicing branchpoints and canonical splice acceptor sites

Human genetic variants that introduce an AG into the intronic region between the branchpoint (BP) and the canonical splice acceptor site (ACC) of protein-coding genes can disrupt pre-mRNA splicing. Using our genome-wide BP database, we delineated the BP-ACC segments of all human introns and found extreme depletion of AG/YAG in the [BP+8, ACC-4] high-risk region. We developed AGAIN as a genome-wide computational approach to systematically and precisely pinpoint intronic AG-gain variants within the BP-ACC regions. AGAIN identified 350 AG-gain variants from the Human Gene Mutation Database, all of which alter splicing and cause disease. Among them, 74% created new acceptor sites, whereas 31% resulted in complete exon skipping. AGAIN also predicts the protein-level products resulting from these two consequences. We performed AGAIN on our exome/genomes database of patients with severe infectious diseases but without known genetic etiology and identified a private homozygous intronic AG-gain variant in the antimycobacterial gene SPPL2A in a patient with mycobacterial disease. AGAIN also predicts a retention of six intronic nucleotides that encode an in-frame stop codon, turning AG-gain into stop-gain. This allele was then confirmed experimentally to lead to loss of function by disrupting splicing. We further showed that AG-gain variants inside the high-risk region led to misspliced products, while those outside the region did not, by two case studies in genes STAT1 and IRF7. We finally evaluated AGAIN on our 14 paired exome-RNAseq samples and found that 82% of AG-gain variants in high-risk regions showed evidence of missplicing. AGAIN is publicly available from https://hgidsoft.rockefeller.edu/AGAIN and https://github.com/casanova-lab/AGAIN.

Alternative Splicing Variation: Accessing and Exploiting in Crop Improvement Programs

Alternative splicing (AS) is a gene regulatory mechanism modulating gene expression in multiple ways. AS is prevalent in all eukaryotes including plants. AS generates two or more mRNAs from the precursor mRNA (pre-mRNA) to regulate transcriptome complexity and proteome diversity. Advances in next-generation sequencing, omics technology, bioinformatics tools, and computational methods provide new opportunities to quantify and visualize AS-based quantitative trait variation associated with plant growth, development, reproduction, and stress tolerance. Domestication, polyploidization, and environmental perturbation may evolve novel splicing variants associated with agronomically beneficial traits. To date, pre-mRNAs from many genes are spliced into multiple transcripts that cause phenotypic variation for complex traits, both in model plant Arabidopsis and field crops. Cataloguing and exploiting such variation may provide new paths to enhance climate resilience, resource-use efficiency, productivity, and nutritional quality of staple food crops. This review provides insights into AS variation alongside a gene expression analysis to select for novel phenotypic diversity for use in breeding programs. AS contributes to heterosis, enhances plant symbiosis (mycorrhiza and rhizobium), and provides a mechanistic link between the core clock genes and diverse environmental clues.

RNA-Based Strategies for Cancer Therapy: In Silico Design and Evaluation of ASOs for Targeted Exon Skipping

Precision medicine in oncology has made significant progress in recent years by approving drugs that target specific genetic mutations. However, many cancer driver genes remain challenging to pharmacologically target ("undruggable"). To tackle this issue, RNA-based methods like antisense oligonucleotides (ASOs) that induce targeted exon skipping (ES) could provide a promising alternative. In this work, a comprehensive computational procedure is presented, focused on the development of ES-based cancer treatments. The procedure aims to produce specific protein variants, including inactive oncogenes and partially restored tumor suppressors. This novel computational procedure encompasses target-exon selection, in silico prediction of ES products, and identification of the best candidate ASOs for further experimental validation. The method was effectively employed on extensively mutated cancer genes, prioritized according to their suitability for ES-based interventions. Notable genes, such as NRAS and VHL, exhibited potential for this therapeutic approach, as specific target exons were identified and optimal ASO sequences were devised to induce their skipping. To the best of our knowledge, this is the first computational procedure that encompasses all necessary steps for designing ASO sequences tailored for targeted ES, contributing with a versatile and innovative approach to addressing the challenges posed by undruggable cancer driver genes and beyond.

Benchmarking splice variant prediction algorithms using massively parallel splicing assays

Background

Variants that disrupt mRNA splicing account for a sizable fraction of the pathogenic burden in many genetic disorders, but identifying splice-disruptive variants (SDVs) beyond the essential splice site dinucleotides remains difficult. Computational predictors are often discordant, compounding the challenge of variant interpretation. Because they are primarily validated using clinical variant sets heavily biased to known canonical splice site mutations, it remains unclear how well their performance generalizes.

Results

We benchmarked eight widely used splicing effect prediction algorithms, leveraging massively parallel splicing assays (MPSAs) as a source of experimentally determined ground-truth. MPSAs simultaneously assay many variants to nominate candidate SDVs. We compared experimentally measured splicing outcomes with bioinformatic predictions for 3,616 variants in five genes. Algorithms' concordance with MPSA measurements, and with each other, was lower for exonic than intronic variants, underscoring the difficulty of identifying missense or synonymous SDVs. Deep learning-based predictors trained on gene model annotations achieved the best overall performance at distinguishing disruptive and neutral variants. Controlling for overall call rate genome-wide, SpliceAI and Pangolin also showed superior overall sensitivity for identifying SDVs. Finally, our results highlight two practical considerations when scoring variants genome-wide: finding an optimal score cutoff, and the substantial variability introduced by differences in gene model annotation, and we suggest strategies for optimal splice effect prediction in the face of these issues.

Conclusion

SpliceAI and Pangolin showed the best overall performance among predictors tested, however, improvements in splice effect prediction are still needed especially within exons.

Computational prediction of human deep intronic variation

BACKGROUND

The adoption of whole-genome sequencing in genetic screens has facilitated the detection of genetic variation in the intronic regions of genes, far from annotated splice sites. However, selecting an appropriate computational tool to discriminate functionally relevant genetic variants from those with no effect is challenging, particularly for deep intronic regions where independent benchmarks are scarce.

RESULTS

In this study, we have provided an overview of the computational methods available and the extent to which they can be used to analyze deep intronic variation. We leveraged diverse datasets to extensively evaluate tool performance across different intronic regions, distinguishing between variants that are expected to disrupt splicing through different molecular mechanisms. Notably, we compared the performance of SpliceAI, a widely used sequence-based deep learning model, with that of more recent methods that extend its original implementation. We observed considerable differences in tool performance depending on the region considered, with variants generating cryptic splice sites being better predicted than those that potentially affect splicing regulatory elements. Finally, we devised a novel quantitative assessment of tool interpretability and found that tools providing mechanistic explanations of their predictions are often correct with respect to the ground - information, but the use of these tools results in decreased predictive power when compared to black box methods.

CONCLUSIONS

Our findings translate into practical recommendations for tool usage and provide a reference framework for applying prediction tools in deep intronic regions, enabling more informed decision-making by practitioners.