Classification models for KCNQ1 variants distinguish functional and trafficking effects to enhance pathogenicity interpretation.

Random forest classifiers predicting seven experimental KCNQ1 metrics — four electrophysiology and three trafficking measurements — from large machine learning model predictions combined with protein-specific biophysical values.

Missense variants in KCNQ1 underlie most cases of congenital long QT syndrome, one of the most common genetic arrhythmias, by affecting protein stability, trafficking and function — three measurable properties that support variant interpretation. Leveraging the experimental data generated by their own laboratories, the authors built random forest classifiers predicting seven metrics: four electrophysiology and three trafficking measurements. The features combine predictions from large machine learning models with protein-specific biophysical values, outperforming either feature set alone. Applied to ClinVar variants of uncertain significance and AlphaMissense-ambiguous variants, the classifiers yield global dysfunction and mistrafficking scores that separate benign from pathogenic variants and complement AlphaMissense by linking variants to long QT-causing mechanisms. The authors present the approach as generalisable to other ion channels and recommend the kind of systematic benchmarking performed here to assess future variant effect predictors.

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

The model remains gene-specific, which limits its immediate reach, but KCNQ1 is precisely a gene where missense variants of uncertain significance accumulate and where reference electrophysiology is too demanding to be requested case by case. The practical contribution is twofold: an in silico surrogate calibrated on functional data generated in-house, and above all a mechanistic read-out — channel dysfunction versus mistrafficking — that AlphaMissense does not provide and that shapes the interpretation argument, provided one remembers that a predicted score is not an actual functional measurement. The announced transferability to other ion channels and the agreement between predictions and experiment on genuinely unseen variants still need to be demonstrated.

Analysis by Dr Thibaut Benquey