Benchmarking splice variant prediction algorithms using massively parallel splicing assays.

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Variants affecting mRNA splicing contribute significantly to genetic disorders.
• Identifying splice-disruptive variants (SDVs) beyond canonical sites is difficult.
• Existing computational predictors often show discordance and are validated on biased clinical data.

Purpose of the Study:

• To benchmark eight widely used splicing effect prediction algorithms.
• To evaluate predictor performance against experimental ground-truth data from massively parallel splicing assays (MPSAs).
• To assess the generalizability of computational predictors for variant interpretation.

Main Methods:

• Benchmarking eight splicing effect prediction algorithms.
• Utilizing massively parallel splicing assays (MPSAs) to generate experimental ground-truth data.
• Comparing bioinformatic predictions with experimentally measured splicing outcomes for 3,616 variants across five genes.

Main Results:

• Algorithm concordance was lower for exonic variants compared to intronic variants.
• Deep learning predictors, particularly SpliceAI and Pangolin, demonstrated superior sensitivity in distinguishing disruptive from neutral variants.
• Variability in gene model annotation significantly impacts variant scoring and prediction accuracy.

Conclusions:

• SpliceAI and Pangolin exhibit the best performance among the tested predictors.
• Further improvements in splice effect prediction are necessary, especially for variants located within exons.
• Optimizing score cutoffs and accounting for gene model annotation variability are crucial for accurate genome-wide variant scoring.