Benchmarking splice variant prediction algorithms using massively parallel splicing assays.

Area of Science:

• Genomics
• Computational Biology
• Molecular Genetics

Background:

• Identifying splice-disruptive variants (SDVs) is critical for understanding genetic disorders, but current computational tools face challenges in accuracy and generalization.
• Existing variant interpretation methods are often validated on biased clinical datasets, limiting their effectiveness for non-canonical splice site mutations.

Approach:

• Benchmarked eight splicing effect prediction algorithms using experimentally validated ground-truth data from massively parallel splicing assays (MPSAs).
• Assessed 3,616 variants across five genes, comparing bioinformatic predictions with experimentally measured splicing outcomes.
• Evaluated algorithm performance, focusing on concordance with MPSA measurements and identifying challenges in predicting exonic variants.

Key Points:

• Deep learning-based predictors, particularly those trained on gene model annotations, demonstrated superior performance in distinguishing disruptive from neutral variants.
• SpliceAI and Pangolin exhibited high sensitivity for identifying SDVs, outperforming other algorithms when controlling for genome-wide call rates.
• Algorithm concordance was lower for exonic variants compared to intronic variants, highlighting the difficulty in predicting missense or synonymous SDVs.

Conclusions:

• SpliceAI and Pangolin represent the current state-of-the-art in splice effect prediction but require further refinement, especially for variants within exons.
• Practical considerations such as optimal score cutoff selection and variability in gene model annotations significantly impact genome-wide variant scoring.
• Strategies for optimizing splice effect prediction must address these challenges to improve variant interpretation in genetic research.