CodonBench: separating gene context, exon position, and variant-specific component in synonymous-variant classification

Gene identity and exon position primarily predict synonymous-variant pathogenicity, with a simple distance-to-exon-junction feature outperforming neural models at 0.84 accuracy.

• Why it matters: Understanding what influences the classification of synonymous variants is crucial because these variants do not alter proteins but can still be pathogenic, highlighting gaps in current models and benchmarks.
• What they did: The authors assembled SynPath, a benchmark of ClinVar synonymous variants, and evaluated models under gene-held-out splits, comparing neural representations and a sequence-independent feature based on exon distance.
• The result: A simple feature based on exon distance achieved 0.84 accuracy, surpassing neural models, indicating that gene identity and position dominate pathogenicity signals, and that random-split benchmarks may overstate model performance.