Genetic testing can identify mutations linked to disease, but discovering a variant doesn’t always reveal what it means for a patient. Many genetic changes are classified as variants of uncertain significance, and even variants predicted to be harmful can disrupt proteins in very different ways.
A by Ana Chang-Gonzalez and colleagues shows how machine learning could help move genetic medicine from simply predicting whether a variant is dangerous to explaining how it causes disease.
The researchers focused on KCNQ1, a potassium channel that helps the heart maintain a normal electrical rhythm. Mutations in KCNQ1 are a major cause of congenital long QT syndrome (LQTS), an inherited condition that can lead t0聽dangerous cardiac arrhythmias.

Importantly, KCNQ1 mutations do not all cause disease through the same mechanism. Some prevent enough channel protein from reaching the cell surface, a process called mistrafficking, while others reach the membrane normally but fail to conduct electrical current or open and close properly.
To distinguish these mechanisms, the researchers trained machine-learning models using experimental measurements from 201 KCNQ1 variants with functional data and 145 with trafficking data. Rather than producing a single prediction of pathogenicity, the models predict seven properties of the channel, including electrical current, channel opening and closing, and protein trafficking to the cell surface. The key was combining broad machine learning predictions with detailed knowledge of KCNQ1. The models incorporated information from tools such as AlphaMissense, ESM-2 and ProteinMPNN alongside features describing the channel鈥檚 three-dimensional structure, evolution and biophysical properties. Together, these sources generally performed better than either approach alone. This added detail proved particularly useful for difficult variants.
When tested on 23 ClinVar variants of uncertain significance that had been excluded from training, the models predicted their functional effects better than AlphaMissense and an earlier KCNQ1-specific predictor. The researchers also developed overall dysfunction and mistrafficking scores that helped separate variants annotated as benign from those considered pathogenic.

The broader promise is precision medicine. A patient whose mutation prevents KCNQ1 from reaching the cell surface might benefit from a drug that restores trafficking, whereas someone whose channel reaches the membrane but functions poorly could require a treatment that increases channel activity. The study notes that a small-molecule trafficking corrector has already been identified for certain KCNQ1 variants, suggesting that such mechanism-specific therapies are becoming plausible.
Important challenges remain. The training dataset is relatively small and imbalanced, and the current models cannot reliably distinguish loss-of-function from gain-of-function effects, a distinction that will be essential for clinical use. Still, the significance extends well beyond KCNQ1. The study provides a framework for combining general machine learning models with targeted biological experiments to predict not just whether a mutation is harmful but to explain why.
The same strategy could potentially be adapted to other ion channels and disease-associated proteins. As genetic sequencing makes variants easier to discover, approaches like this could help solve the harder problem, turning genetic information into biological understanding, and ultimately into more personalized treatments.