Background: Missense variants in genes encoding voltage-gated sodium channels are associated with a spectrum of severe diseases affecting neuronal and muscle cells, the so-called sodium channelopathies. Variant effects on the biophysical function of the channel correlate with clinical features and can in most cases be categorized as an overall gain- or loss-of-function. This information enables a timely diagnosis, facilitates precision therapy, and guides prognosis. Machine learning models may be able to rapidly generate supporting evidence by predicting variant functional effects.
Methods: Here, we describe a novel multi-task multi-kernel learning framework capable of harmonizing functional results and structural information with clinical phenotypes. We included 62 sequence- and structure-based features such as amino acid physiochemical properties, substitution radicality, conservation, protein-protein interaction sites, expert annotation, and others. We harmonized phenotypes as human phenotype ontology (HPO) terms, and compared different measures of phenotypic similarity under simulated sparsity or noise. The final model was trained on whole-cell patch-clamp recordings of 375 unique non-synonymous missense variants each expressed in mammalian cells.
Results: Our gain- or loss-of-function classifier outperformed both conventional baseline and state-of-the-art methods on internal validation (mean accuracy 0.837 SD 0.035, mean AU-ROC 0.890 SD 0.023) and on an independent set of recently described variants (n = 30, accuracy 0.967, AU-ROC 1.000). Model performance was robust across different phenotypic similarity measures and largely insensitive to phenotypic noise or sparsity. Localized multi-kernel learning offered biological insight and interpretability by highlighting channels with implicit genotype-phenotype correlations or latent task similarity for downstream analysis.
Conclusions: Learning with phenotypic similarity makes efficient use of clinical information to enable accurate and robust prediction of variant functional effects. Our framework extends the use of human phenotype ontology terms towards kernel-based methods in machine learning. Training data, pre-trained models, and a web-based graphical user interface for the model are publicly available.