Using an integrative machine learning approach utilising homology modelling to clinically interpret genetic variants: CACNA1F as an exemplar

Abstract

Advances in DNA sequencing technologies have revolutionised rare disease diagnostics and have led to a dramatic increase in the volume of available genomic data. A key challenge that needs to be overcome to realise the full potential of these technologies is that of precisely predicting the effect of genetic variants on molecular and organismal phenotypes. Notably, despite recent progress, there is still a lack of robust in silico tools that accurately assign clinical significance to variants. Genetic alterations in the CACNA1F gene are the commonest cause of X-linked incomplete Congenital Stationary Night Blindness (iCSNB), a condition associated with non-progressive visual impairment. We combined genetic and homology modelling data to produce CACNA1F-vp, an in silico model that differentiates disease-implicated from benign missense CACNA1F changes. CACNA1F-vp predicts variant effects on the structure of the CACNA1F encoded protein (a calcium channel) using parameters based upon changes in amino acid properties; these include size, charge, hydrophobicity, and position. The model produces an overall score for each variant that can be used to predict its pathogenicity. CACNA1F-vp outperformed four other tools in identifying disease-implicated variants (area under receiver operating characteristic and precision recall curves = 0.84; Matthews correlation coefficient = 0.52) using a tenfold cross-validation technique. We consider this protein-specific model to be a robust stand-alone diagnostic classifier that could be replicated in other proteins and could enable precise and timely diagnosis.

Improving the clinical interpretation of missense variants in X linked genes using structural analysis

“…In a previous study, we integrated genetic and structural biology data to predict variant–disease association with high accuracy in the X linked gene CACNA1F (MIM: 300110); the area under the receiver operating characteristic (ROC) and precision–recall (PR) curves was 0.84; Matthews correlation coefficient (MCC) was 0.52. 6 Here, we replicate the accuracy and robustness of this approach in several other disease-implicated X linked genes. Furthermore, we evaluate seven prediction tools and show that the meta-predictors REVEL (rare exome variant ensemble learner), 7 VEST4 (variant effect scoring tool 4.0), 8 and ClinPred 9 are generally the most accurate in predicting the impact of missense variants in this group of disorders.…”

“…We previously reported a protein-specific or gene-specific approach to variant pathogenicity prediction in the X linked CACNA1F gene. 6 To assess the generalisability of this approach, we again selected X linked disease-causing genes as our test case. From a total of 482 X linked genes from HGMD, no missense variants were reported for 329 genes.…”

“… 7 We limited our analyses to X linked genes from HGMD that contained a minimum of 70 pathogenic missense variants as informed by earlier findings. 6 …”

“…As in our previously described methodology, 6 a set of sequence-based and structure-based features were defined to integrate clinical and structural data. The ‘colour’ column obtained from ConSurf server, 25 using default parameters, was used to measure conservation at the variant site instead of using a protein-specific multiple sequence alignment.…”

Predicting the functional impact of KCNQ1 variants with artificial neural networks

“…To overcome these difficulties, specific machine learning models tailored to predict the functional effects of KCNQ1 variants were developed [ 23 , 24 ]. Similar approaches have also been applied to other cardiac ion channels [ 31 , 32 ].…”