Improved KCNQ2 gene missense variant interpretation with artificial intelligence.

Advances in DNA sequencing technologies have revolutionized rare disease diagnosis, resulting in increasing genomic data availability. There is a wide variety of genome-wide computational tools aiming at predicting the pathogenicity of missense variants. Despite this wealth of information and improved procedures to combine data from various sources, identifying the pathogenic causal variants and distinguishing between severe and tolerated variants remains a key challenge. Mutations in the Kv7.2 voltage-gated potassium channel gene (KCNQ2) have been linked to different subtypes of epilepsies, such as benign familial neonatal epilepsy (BFNE) and epileptic encephalopathy (EE). Previous reports suggest that these genome-wide tools have limited applicability to KCNQ2 gene related diseases due to overestimation of deleterious mutations and failure to identify tolerated variants correctly, being, therefore, of limited use in clinical practice. Here, we evaluate different machine learning protocols. Despite the similarity in performance, logistic regression, support vector machine, random forest and gradient boosting algorithms resulted in significantly improved specificity and sensitivity values. We found that combining primate genome information, readily available features, such as AlphaFold structural information, allele frequency index, residue conservation and other commonly used protein descriptors, provides foundations to build reliable gene-specific machine learning models. We present transferable methodology able to accurately classify KCNQ2 missense variants with sensitivity and specificity scores above 97%.