Random forest classifiers integrating machine learning predictions with biophysical values successfully predicted KCNQ1 dysfunction and mistrafficking, distinguishing benign from pathogenic variants.
A novel random forest classifier integrating machine learning and biophysical features improves the prediction of pathogenicity for KCNQ1 variants in long QT syndrome.
Missense variants in the potassium channel KCNQ1 underlie most cases of congenital long QT syndrome (LQTS), one of the most common genetic arrhythmias. Variants affect protein stability, trafficking, and function, which are measurable properties that support variant interpretation. Leveraging the extensive experimental data generated by our laboratories, we developed random forest classifiers that predict seven KCNQ1 metrics: four electrophysiology and three trafficking measurements. The features for our classifiers integrate predictions from large machine learning models with protein-specific biophysical values, outperforming using either set of features alone. We applied our classifiers to interpret ClinVar variants of uncertain significance and AlphaMissense-ambiguous variants and developed global dysfunction and mistrafficking scores which distinguished benign from pathogenic variants. Global scores complemented AlphaMissense predictions, linking variants with LQTS-causing mechanisms. While effective for KCNQ1, our approach to variant prediction is generalizable to other ion channels and we recommend systematic benchmarking as done in this work to fully assess performance of future variant effect predictors.
Chang-Gonzalez et al. (Tue,) conducted a other in congenital long QT syndrome (LQTS). Random forest classifiers was evaluated on Prediction of seven KCNQ1 metrics (four electrophysiology and three trafficking measurements). Random forest classifiers integrating machine learning predictions with biophysical values successfully predicted KCNQ1 dysfunction and mistrafficking, distinguishing benign from pathogenic variants.