Multimodal Convolutional Neural Network for Pathogenicity Classification of Missense Variants

Abstract

Interpreting genetic missense variants is critical for diagnosing inherited disease. In the ClinVar missense data used here, 80% of variants were classified as Variants of Uncertain Significance (VUS), which leaves patients without clear diagnostic guidance. This thesis combines ClinVar with DescribePROT, a database of per-residue structural and functional descriptors, to classify missense variants by their clinical significance. These descriptors have not been applied to variant classification before. Each variant is represented by a 31 × 15 matrix of the protein window around the mutation, holding 14 structural and functional amino acid features, together with an eight-feature vector of population allele frequencies and physicochemical substitution scores. Five models were compared on 117,843 variants: three main architectures (VGG, VGG CNN + MLP, and DeepInsight with DenseNet121) and two single-input ablations. Once VUS are removed, ClinVar labels the remaining variants in four categories, and the models were trained on those. The VGG CNN + MLP model performed best, with a Macro F1-score of 0.6179, an accuracy of 66.77%, and a Macro ROC-AUC of 0.8910, and it held a Macro ROC-AUC of 0.8811 on an independent external ClinVar validation set of 34,146 variants. Grouped into the binary benign against pathogenic split that clinical practice acts on, the same architecture reached a ROC-AUC of 0.972. The ablations located the predictive signal: the eight variant-level features alone recovered 92% of the best model's Macro F1, while the DeepInsight image alone performed near chance. Feature integration and variant-level annotation, rather than added network complexity, drive performance on this task.