Prediction of the Phenotype of Human Missense Mutations with an Ensemble of Deep Neural Networks

Rapid progress in the area of LLM and ML has increased the likelihood of clinical use for methods that predict the effects of genetic variations. Methods that can quickly obtain such predictions and enable rapid evaluation of the effects of residue changes in the genome will enable clinicians to better understand and treat the phenotypes presented in the clinic. Here, we present a new method to rapidly predict the effects of single-residue mutations in proteins using a single sequence-based approach by using rapid sequence alignments against a limited dataset. Non-local input features were also used. These integrate residue features such as accessible surface area over the sequence with discrete periodic weights. The predictor is built using a multi-layered neural network and has been trained on both definite phenotype assignment and definite and likely phenotype assignments. Human data for training was obtained from ClinVar. The proposed method gives an AUC of $0.88\pm 0.01$ and an MCC of $0.62\pm 0.02$ . This accuracy level is comparable to the one obtained from multiple sequence alignment-based methods. The Shoni prediction servers are available at http://www.mamiris.com/Shoni/ .