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

Abstract

Background It is estimated that up to 50% of all disease causing variants disrupt splicing. Due to its complexity, our ability to predict which variants disrupt splicing is limited, meaning missed diagnoses for patients. The emergence of machine learning for targeted medicine holds great potential to improve prediction of splice disrupting variants. The recently published SpliceAI algorithm utilises deep neural networks and has been reported to have a greater accuracy than other commonly used methods. Methods and findings The original SpliceAI was trained on splice sites included in primary isoforms combined with novel junctions observed in GTEx data, which might introduce noise and de-correlate the machine learning input with its output. Limiting the data to only validated and manual annotated primary and alternatively spliced GENCODE sites in training may improve predictive abilities. All of these gene isoforms were collapsed (aggregated into one pseudo-isoform) and the SpliceAI architecture was retrained (CI-SpliceAI). Predictive performance on a newly curated dataset of 1,316 functionally validated variants from the literature was compared with the original SpliceAI, alongside MMSplice, MaxEntScan, and SQUIRLS. Both SpliceAI algorithms outperformed the other methods, with the original SpliceAI achieving an accuracy of ∼91%, and CI-SpliceAI showing an improvement at ∼92% overall. Predictive accuracy increased in the majority of curated variants. Conclusions We show that including only manually annotated alternatively spliced sites in training data improves prediction of clinically relevant variants, and highlight avenues for further performance improvements.

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

“…However, most SpliceAI current public implementations or pre-computed whole genome VCFs currently do not process complex delins variations (i.e., other than deletion, insertion, or substitution), nor does Pangolin. Of note, those complex variations are handled by CI-SpliceAI but with numerical results [12]. The functional study of this variant by a minigene assay has shown the skipping of an entire out-of-frame exon [30].…”

“…This de Sainte Agathe et al Human Genomics (2023) 17:7 ability to focus not only on the nearby site (destruction or creation) but at the whole transcript level is a unique feature of these deep-learning-based next-generation splicing predictors, such as SpliceAI or Pangolin [11]. In a recent improvement, the SpliceAI neural network has been retrained with a curated and manually validated isoforms dataset [12]. Still, the standard version of SpliceAI (currently v1.3.1) has some limitations.…”

“…broad insti tute. org/) or pre-computed whole genome VCFs only annotate simple variants (i.e., substitutions, insertions, deletions), prohibiting the interpretation of more complex deletions-insertions or inversions, with the notable exception of the recent CI-SpliceAI [12].…”

An Interpretable Model of pre-mRNA Splicing for Animal and Plant Genes

“…In every case, the organism-specific SMsplice model had the best performance, as expected. As a point of comparison, we also applied the neural net model CI-SpliceAI to the same test sets (Strauch et al, 2022). This black box model was trained on human sequences and so the most direct comparison is to the SMsplice model with fully human parameters.…”

Computational prediction of human deep intronic variation

“…For example, Pangolin [26] uses splicing quantifications from multiple species and tissues to not only predict whether a position is a splice site (as SpliceAI does) but also to predict splice site usage (e.g., how much a splice site is being used in a given tissue). In contrast, CI-SpliceAI [27] uses different training labels for true and false splice site positions based on a collapsed transcript structure derived from GENCODE [28] annotations.…”