SIEVE: Sparse Interpretable Exome Variant Explainer.

Deep-learning framework reading every observed exonic variant without a frequency filter, representing genomic position through self-attention and calibrating attributions against a permuted-label null.

Whole-exome case-control studies contain both rare and common variation, yet analytical methods usually partition the frequency spectrum, discard positional context or depend on fixed annotations. SIEVE is a deep-learning framework for interpretable variant and gene prioritisation: it reads every observed exonic variant without a frequency filter, represents genomic position through self-attention and calibrates attributions against a permuted-label null. Evaluated on coronary artery disease, early-onset myocardial infarction and Crohn's disease, it achieves discrimination matching the liability-threshold expectation for each trait, with recovery of catalogued associations rising with annotation depth. Compared with burden testing, single-variant association and polygenic scoring, SIEVE recovers overlapping but largely distinct candidates.

Synthesis written by Geno'X. For the full original abstract, please refer to the source publication.

The methodological interest is real — dispensing with a frequency threshold and calibrating against a permuted null avoids two classic biases — but the headline result is ambiguous: candidates largely distinct from established approaches may reflect a gain in sensitivity just as much as an excess of false positives, and nothing reported here allows a decision. The three traits tested belong to common disease genetics, not Mendelian diagnostics: there is no demonstration of usefulness in clinical interpretation. Worth revisiting after peer review and validation on a reference gene set.

Analysis by Dr Thibaut Benquey