A genetic variant is not a verdict: a better path through uncertainty

One of precision medicine’s least glamorous but most consequential problems is the variant of uncertain significance. A patient may receive a genetic result associated with a disease gene without knowing whether the exact substitution changes how the protein works. Such ambiguity can obstruct diagnosis, counseling and eligibility for emerging therapies.

Investigators developed EpiPred, a gene-specific machine-learning method for STXBP1 variants associated with developmental and epileptic encephalopathies. Rather than judging every variant using a generic genome-wide score, the model was trained on curated pathogenic and benign variants and its predictions were compared with biochemical assays of protein abundance, solubility, stability and interaction with syntaxin 1. The paper reports better performance than broad predictors and identifies candidate variants for re-evaluation.

The transferable idea is not that an algorithm can replace molecular diagnosis. It is that predictions should be tied to experimentally measured effects, curated clinical evidence and the possibility of reclassification. Functional assay agreement does not automatically demonstrate clinical penetrance or predict treatment benefit in an individual.