CI-SpliceAI—Improving machine learning predictions of disease causing splicing variants using curated alternative splice sites

Spectrum and genotyping strategies of "dark" genetic matter in germline susceptibility genes of tumor syndromes

PURPOSE

Despite the widespread use of high-throughput genotyping strategies, certain mutation types remain understudied. We provide an overview of these often overlooked mutation types, with representative examples from common hereditary cancer syndromes.

METHODS

We conducted a comprehensive review of the literature and locus-specific variant databases to summarize the germline pathogenic variants discovered through non-routine genotyping methods. We evaluated appropriate detection and analysis methods tailored for these specific genetic aberrations. Additionally, we performed in silico splice predictions on deep intronic variants registered in the ClinVar database.

RESULTS

Our study suggests that, aside from founder mutations, most cases are sporadic. However, we anticipate a relatively high likelihood of splice effects for deep intronic variants. The findings underscore the significant clinical utility of genome sequencing techniques and the importance of applying relevant analysis methods.

Transformers significantly improve splice site prediction

Mutations that affect RNA splicing significantly impact human diversity and disease. Here we present a method using transformers, a type of machine learning model, to detect splicing from raw 45,000-nucleotide sequences. We generate embeddings with residual neural networks and apply hard attention to select splice site candidates, enabling efficient training on long sequences. Our method surpasses the leading tool, SpliceAI, in detecting splice sites in GENCODE and ENSEMBL annotations. Using extensive RNA sequencing data from an Icelandic cohort of 17,848 individuals and the Genotype-Tissue Expression (GTEx) project, our method demonstrates superior performance in detecting splice junctions compared to SpliceAI-10k (PR-AUC = 0.834 vs. PR-AUC = 0.820) and is more effective at identifying disease-related splice variants in ClinVar (PR-AUC = 0.997 vs. PR-AUC = 0.996). These advancements hold promise for improving genetic research and clinical diagnostics, potentially leading to better understanding and treatment of splicing-related diseases.

Homozygous variants in WDR83OS lead to a neurodevelopmental disorder with hypercholanemia

WD repeat domain 83 opposite strand (WDR83OS) encodes the 106-aa (amino acid) protein Asterix, which heterodimerizes with CCDC47 to form the PAT (protein associated with ER translocon) complex. This complex functions as a chaperone for large proteins containing transmembrane domains to ensure proper folding. Until recently, little was known about the role of WDR83OS or CCDC47 in human disease traits. However, biallelic variants in CCDC47 were identified in four unrelated families with trichohepatoneurodevelopmental syndrome, characterized by a neurodevelopmental disorder (NDD) with liver dysfunction. Three affected siblings in an additional family share a homozygous truncating WDR83OS variant and a phenotype of NDD, dysmorphic features, and liver dysfunction. Using family-based rare variant analyses of exome sequencing (ES) data and case matching through GeneMatcher, we describe the clinical phenotypes of 11 additional individuals in eight unrelated families (nine unrelated families, 14 individuals in total) with biallelic putative truncating variants in WDR83OS. Consistent clinical features include NDD (14/14), facial dysmorphism (13/14), intractable itching (9/14), and elevated bile acids (5/6). Whereas bile acids were significantly elevated in 5/6 of individuals tested, bilirubin was normal and liver enzymes were normal to mildly elevated in all 14 individuals. In three of six individuals for whom longitudinal data were available, we observed a progressive reduction in relative head circumference. A zebrafish model lacking Wdr83os function further supports its role in the nervous system, craniofacial development, and lipid absorption. Taken together, our data support a disease-gene association between biallelic loss-of-function of WDR83OS and a neurological disease trait with hypercholanemia.

Investigation of cryptic JAG1 splice variants as a cause of Alagille syndrome and performance evaluation of splice predictor tools

Haploinsufficiency of JAG1 is the primary cause of Alagille syndrome (ALGS), a rare, multisystem disorder. The identification of JAG1 intronic variants outside of the canonical splice region as well as missense variants, both of which lead to uncertain associations with disease, confuses diagnostics. Strategies to determine whether these variants affect splicing include the study of patient RNA or minigene constructs, which are not always available or can be laborious to design, as well as the utilization of computational splice prediction tools. These tools, including SpliceAI and Pangolin, use algorithms to calculate the probability that a variant results in a splice alteration, expressed as a Δ score, with higher Δ scores (>0.2 on a 0-1 scale) positively correlated with aberrant splicing. We studied the consequence of 10 putative splice variants in ALGS patient samples through RNA analysis and compared this to SpliceAI and Pangolin predictions. We identified eight variants with aberrant splicing, seven of which had not been previously validated. Combining these data with non-canonical and missense splice variants reported in the literature, we identified a predictive threshold for SpliceAI and Pangolin with high sensitivity (Δ score >0.6). Moreover, we showed reduced specificity for variants with low Δ scores (<0.2), highlighting a limitation of these tools that results in the misidentification of true splice variants. These results improve genomic diagnostics for ALGS by confirming splice effects for seven variants and suggest that the integration of splice prediction tools with RNA analysis is important to ensure accurate clinical variant classifications.

Germline Variants in Patients Affected by Both Uveal and Cutaneous Melanoma

Uveal melanoma (UM) and nonacral cutaneous melanoma (CM) are distinct entities with varied genetic landscapes despite both arising from melanocytes. There are, however, similarities in that they most frequently affect people of European ancestry, and high penetrance germline variants in BAP1, POT1 and CDKN2A have been shown to predispose to both UM and CM. This study aims to further explore germline variants in patients affected by both UM and CM, shedding light on the underlying genetic mechanism causing these diseases. Using exome sequencing we analysed germline DNA samples from a cohort of 83 Australian patients diagnosed with both UM and CM. Eight (10%) patients were identified that carried pathogenic mutations in known melanoma predisposition genes POT1, MITF, OCA2, SLC45A2 and TYR. Three (4%) patients carried pathogenic variants in genes previously linked with other cancer syndromes (ATR, BRIP1 and MSH6) and another three cases carried monoallelic pathogenic variants in recessive cancer genes (xeroderma pigmentosum and Fanconi anaemia), indicating that reduced penetrance of phenotype in these individuals may contribute to the development of both UM and CM. These findings highlight the need for further studies characterising the role of these genes in melanoma susceptibility.

Functional characterization of 2,832 JAG1 variants supports reclassification for Alagille syndrome and improves guidance for clinical variant interpretation

Pathogenic variants in the JAG1 gene are a primary cause of the multi-system disorder Alagille syndrome. Although variant detection rates are high for this disease, there is uncertainty associated with the classification of missense variants that leads to reduced diagnostic yield. Consequently, up to 85% of reported JAG1 missense variants have uncertain or conflicting classifications. We generated a library of 2,832 JAG1 nucleotide variants within exons 1-7, a region with a high number of reported missense variants, and designed a high-throughput assay to measure JAG1 membrane expression, a requirement for normal function. After calibration using a set of 175 known or predicted pathogenic and benign variants included within the variant library, 486 variants were characterized as functionally abnormal (n = 277 abnormal and n = 209 likely abnormal), of which 439 (90.3%) were missense. We identified divergent membrane expression occurring at specific residues, indicating that loss of the wild-type residue itself does not drive pathogenicity, a finding supported by structural modeling data and with broad implications for clinical variant classification both for Alagille syndrome and globally across other disease genes. Of 144 uncertain variants reported in patients undergoing clinical or research testing, 27 had functionally abnormal membrane expression, and inclusion of our data resulted in the reclassification of 26 to likely pathogenic. Functional evidence augments the classification of genomic variants, reducing uncertainty and improving diagnostics. Inclusion of this repository of functional evidence during JAG1 variant reclassification will significantly affect resolution of variant pathogenicity, making a critical impact on the molecular diagnosis of Alagille syndrome.

An interpretable model of pre-mRNA splicing for animal and plant genes

Pre-mRNA splicing is a fundamental step in gene expression, conserved across eukaryotes, in which the spliceosome recognizes motifs at the 3' and 5' splice sites (SSs), excises introns, and ligates exons. SS recognition and pairing is often influenced by protein splicing factors (SFs) that bind to splicing regulatory elements (SREs). Here, we describe SMsplice, a fully interpretable model of pre-mRNA splicing that combines models of core SS motifs, SREs, and exonic and intronic length preferences. We learn models that predict SS locations with 83 to 86% accuracy in fish, insects, and plants and about 70% in mammals. Learned SRE motifs include both known SF binding motifs and unfamiliar motifs, and both motif classes are supported by genetic analyses. Our comparisons across species highlight similarities between non-mammals, increased reliance on intronic SREs in plant splicing, and a greater reliance on SREs in mammalian splicing.

Predicting the impact of rare variants on RNA splicing in CAGI6

Variants which disrupt splicing are a frequent cause of rare disease that have been under-ascertained clinically. Accurate and efficient methods to predict a variant's impact on splicing are needed to interpret the growing number of variants of unknown significance (VUS) identified by exome and genome sequencing. Here, we present the results of the CAGI6 Splicing VUS challenge, which invited predictions of the splicing impact of 56 variants ascertained clinically and functionally validated to determine splicing impact. The performance of 12 prediction methods, along with SpliceAI and CADD, was compared on the 56 functionally validated variants. The maximum accuracy achieved was 82% from two different approaches, one weighting SpliceAI scores by minor allele frequency, and one applying the recently published Splicing Prediction Pipeline (SPiP). SPiP performed optimally in terms of sensitivity, while an ensemble method combining multiple prediction tools and information from databases exceeded all others for specificity. Several challenge methods equalled or exceeded the performance of SpliceAI, with ultimate choice of prediction method likely to depend on experimental or clinical aims. One quarter of the variants were incorrectly predicted by at least 50% of the methods, highlighting the need for further improvements to splicing prediction methods for successful clinical application.

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.

Use of genome sequencing to hunt for cryptic second-hit variants: analysis of 31 cases recruited to the 100 000 Genomes Project

BACKGROUND

Current clinical testing methods used to uncover the genetic basis of rare disease have inherent limitations, which can lead to causative pathogenic variants being missed. Within the rare disease arm of the 100 000 Genomes Project (100kGP), families were recruited under the clinical indication 'single autosomal recessive mutation in rare disease'. These participants presented with strong clinical suspicion for a specific autosomal recessive disorder, but only one suspected pathogenic variant had been identified through standard-of-care testing. Whole genome sequencing (WGS) aimed to identify cryptic 'second-hit' variants.

METHODS

To investigate the 31 families with available data that remained unsolved following formal review within the 100kGP, SVRare was used to aggregate structural variants present in <1% of 100kGP participants. Small variants were assessed using population allele frequency data and SpliceAI. Literature searches and publicly available online tools were used for further annotation of pathogenicity.

RESULTS

Using these strategies, 8/31 cases were solved, increasing the overall diagnostic yield of this cohort from 10/41 (24.4%) to 18/41 (43.9%). Exemplar cases include a patient with cystic fibrosis harbouring a novel exonic LINE1 insertion in CFTR and a patient with generalised arterial calcification of infancy with complex interlinked duplications involving exons 2-6 of ENPP1. Although ambiguous by short-read WGS, the ENPP1 variant structure was resolved using optical genome mapping and RNA analysis.

CONCLUSION

Systematic examination of cryptic variants across a multi-disease cohort successfully identifies additional pathogenic variants. WGS data analysis in autosomal recessive rare disease should consider complex structural and small intronic variants as potentially pathogenic second hits.

Structural and non-coding variants increase the diagnostic yield of clinical whole genome sequencing for rare diseases

BACKGROUND

Whole genome sequencing is increasingly being used for the diagnosis of patients with rare diseases. However, the diagnostic yields of many studies, particularly those conducted in a healthcare setting, are often disappointingly low, at 25-30%. This is in part because although entire genomes are sequenced, analysis is often confined to in silico gene panels or coding regions of the genome.

METHODS

We undertook WGS on a cohort of 122 unrelated rare disease patients and their relatives (300 genomes) who had been pre-screened by gene panels or arrays. Patients were recruited from a broad spectrum of clinical specialties. We applied a bioinformatics pipeline that would allow comprehensive analysis of all variant types. We combined established bioinformatics tools for phenotypic and genomic analysis with our novel algorithms (SVRare, ALTSPLICE and GREEN-DB) to detect and annotate structural, splice site and non-coding variants.

RESULTS

Our diagnostic yield was 43/122 cases (35%), although 47/122 cases (39%) were considered solved when considering novel candidate genes with supporting functional data into account. Structural, splice site and deep intronic variants contributed to 20/47 (43%) of our solved cases. Five genes that are novel, or were novel at the time of discovery, were identified, whilst a further three genes are putative novel disease genes with evidence of causality. We identified variants of uncertain significance in a further fourteen candidate genes. The phenotypic spectrum associated with RMND1 was expanded to include polymicrogyria. Two patients with secondary findings in FBN1 and KCNQ1 were confirmed to have previously unidentified Marfan and long QT syndromes, respectively, and were referred for further clinical interventions. Clinical diagnoses were changed in six patients and treatment adjustments made for eight individuals, which for five patients was considered life-saving.

CONCLUSIONS

Genome sequencing is increasingly being considered as a first-line genetic test in routine clinical settings and can make a substantial contribution to rapidly identifying a causal aetiology for many patients, shortening their diagnostic odyssey. We have demonstrated that structural, splice site and intronic variants make a significant contribution to diagnostic yield and that comprehensive analysis of the entire genome is essential to maximise the value of clinical genome sequencing.

Advances in alternative splicing identification: deep learning and pantranscriptome

In plants, alternative splicing is a crucial mechanism for regulating gene expression at the post-transcriptional level, which leads to diverse proteins by generating multiple mature mRNA isoforms and diversify the gene regulation. Due to the complexity and variability of this process, accurate identification of splicing events is a vital step in studying alternative splicing. This article presents the application of alternative splicing algorithms with or without reference genomes in plants, as well as the integration of advanced deep learning techniques for improved detection accuracy. In addition, we also discuss alternative splicing studies in the pan-genomic background and the usefulness of integrated strategies for fully profiling alternative splicing.

FexSplice: A LightGBM-Based Model for Predicting the Splicing Effect of a Single Nucleotide Variant Affecting the First Nucleotide G of an Exon

Single nucleotide variants (SNVs) affecting the first nucleotide G of an exon (Fex-SNVs) identified in various diseases are mostly recognized as missense or nonsense variants. Their effect on pre-mRNA splicing has been seldom analyzed, and no curated database is available. We previously reported that Fex-SNVs affect splicing when the length of the polypyrimidine tract is short or degenerate. However, we cannot readily predict the splicing effects of Fex-SNVs. We here scrutinized the available literature and identified 106 splicing-affecting Fex-SNVs based on experimental evidence. We similarly identified 106 neutral Fex-SNVs in the dbSNP database with a global minor allele frequency (MAF) of more than 0.01 and less than 0.50. We extracted 115 features representing the strength of splicing cis-elements and developed machine-learning models with support vector machine, random forest, and gradient boosting to discriminate splicing-affecting and neutral Fex-SNVs. Gradient boosting-based LightGBM outperformed the other two models, and the length and nucleotide compositions of the polypyrimidine tract played critical roles in the discrimination. Recursive feature elimination showed that the LightGBM model using 15 features achieved the best performance with an accuracy of 0.80 ± 0.12 (mean and SD), a Matthews Correlation Coefficient (MCC) of 0.57 ± 0.15, an area under the curve of the receiver operating characteristics curve (AUROC) of 0.86 ± 0.08, and an area under the curve of the precision-recall curve (AUPRC) of 0.87 ± 0.09 using a 10-fold cross-validation. We developed a web service program, named FexSplice that accepts a genomic coordinate either on GRCh37/hg19 or GRCh38/hg38 and returns a predicted probability of aberrant splicing of A, C, and T variants.

Identified eleven exon variants in PKD1 and PKD2 genes that altered RNA splicing by minigene assay

BACKGROUND

Autosomal dominant polycystic kidney disease (ADPKD) is a common monogenic multisystem disease caused primarily by mutations in the PKD1 gene or PKD2 gene. There is increasing evidence that some of these variants, which are described as missense, synonymous or nonsense mutations in the literature or databases, may be deleterious by affecting the pre-mRNA splicing process.

RESULTS

This study aimed to determine the effect of these PKD1 and PKD2 variants on exon splicing combined with predictive bioinformatics tools and minigene assay. As a result, among the 19 candidate single nucleotide alterations, 11 variants distributed in PKD1 (c.7866C > A, c.7960A > G, c.7979A > T, c.7987C > T, c.11248C > G, c.11251C > T, c.11257C > G, c.11257C > T, c.11346C > T, and c.11393C > G) and PKD2 (c.1480G > T) were identified to result in exon skipping.

CONCLUSIONS

We confirmed that 11 variants in the gene of PKD1 and PKD2 affect normal splicing by interfering the recognition of classical splicing sites or by disrupting exon splicing enhancers and generating exon splicing silencers. This is the most comprehensive study to date on pre-mRNA splicing of exonic variants in ADPKD-associated disease-causing genes in consideration of the increasing number of identified variants in PKD1 and PKD2 gene in recent years. These results emphasize the significance of assessing the effect of exon single nucleotide variants in ADPKD at the mRNA level.

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.

A deep intronic TCTN2 variant activating a cryptic exon predicted by SpliceRover in a patient with Joubert syndrome

The recent introduction of genome sequencing in genetic analysis has led to the identification of pathogenic variants located in deep introns. Recently, several new tools have emerged to predict the impact of variants on splicing. Here, we present a Japanese boy of Joubert syndrome with biallelic TCTN2 variants. Exome sequencing identified only a heterozygous maternal nonsense TCTN2 variant (NM_024809.5:c.916C >T, p.(Gln306Ter)). Subsequent genome sequencing identified a deep intronic variant (c.1033+423G>A) inherited from his father. The machine learning algorithms SpliceAI, Squirls, and Pangolin were unable to predict alterations in splicing by the c.1033+423G>A variant. SpliceRover, a tool for splice site prediction using FASTA sequence, was able to detect a cryptic exon which was 85-bp away from the variant and within the inverted Alu sequence while SpliceRover scores for these splice sites showed slight increase (donor) or decrease (acceptor) between the reference and mutant sequences. RNA sequencing and RT-PCR using urinary cells confirmed inclusion of the cryptic exon. The patient showed major symptoms of TCTN2-related disorders such as developmental delay, dysmorphic facial features and polydactyly. He also showed uncommon features such as retinal dystrophy, exotropia, abnormal pattern of respiration, and periventricular heterotopia, confirming these as one of features of TCTN2-related disorders. Our study highlights usefulness of genome sequencing and RNA sequencing using urinary cells for molecular diagnosis of genetic disorders and suggests that database of cryptic splice sites predicted in introns by SpliceRover using the reference sequences can be helpful in extracting candidate variants from large numbers of intronic variants in genome sequencing.

SpliceAI-visual: a free online tool to improve SpliceAI splicing variant interpretation

SpliceAI is an open-source deep learning splicing prediction algorithm that has demonstrated in the past few years its high ability to predict splicing defects caused by DNA variations. However, its outputs present several drawbacks: (1) although the numerical values are very convenient for batch filtering, their precise interpretation can be difficult, (2) the outputs are delta scores which can sometimes mask a severe consequence, and (3) complex delins are most often not handled. We present here SpliceAI-visual, a free online tool based on the SpliceAI algorithm, and show how it complements the traditional SpliceAI analysis. First, SpliceAI-visual manipulates raw scores and not delta scores, as the latter can be misleading in certain circumstances. Second, the outcome of SpliceAI-visual is user-friendly thanks to the graphical presentation. Third, SpliceAI-visual is currently one of the only SpliceAI-derived implementations able to annotate complex variants (e.g., complex delins). We report here the benefits of using SpliceAI-visual and demonstrate its relevance in the assessment/modulation of the PVS1 classification criteria. We also show how SpliceAI-visual can elucidate several complex splicing defects taken from the literature but also from unpublished cases. SpliceAI-visual is available as a Google Colab notebook and has also been fully integrated in a free online variant interpretation tool, MobiDetails ( https://mobidetails.iurc.montp.inserm.fr/MD ).

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.