Neural RNA splicing models rely on reading frame information unavailable to the splicing machinery

Common black-box neural splice-site predictors are strongly affected by deletions or mutations that change translational reading frame, including those creating in-frame stop codons, in human genome data.

• Why it matters: It is unclear whether sequence-based neural models simulate natural splicing or instead use cues unavailable to the splicing machinery, especially for clinically important poison exons.
• What they did: The study tests splice-site effects of mutations expected to have little direct impact, focusing on frame-alignment–altering variants, and compares commonly used neural tools with grey-box models that incorporate splicing-specific structure.
• The result: Frame dependence makes black-box models systematically miss many poison exons, while this dependence is much weaker in grey-box models, implying black-box predictions use inaccessible information.

AI summary of the abstract · may contain errors · not medical advice · read the paper.