3Cnet: Pathogenicity prediction of human variants using knowledge transfer with deep recurrent neural networks

Abstract

Thanks to the improvement of Next Generation Sequencing (NGS), genome-based diagnosis for rare disease patients become possible. However, accurate interpretation of human variants requires massive amount of knowledge gathered from previous researches and clinical cases. Also, manual analysis for each variant in the genome of patients takes enormous time and effort of clinical experts and medical doctors. Therefore, to reduce the cost of diagnosis, various computational tools have been developed for the pathogenicity prediction of human variants. Nevertheless, there has been the circularity problem of conventional tools, which leads to the overlap of training data and eventually causes overfitting of algorithms. In this research, we developed a pathogenicity predictor, named as 3Cnet, using deep recurrent neural networks which analyzes the amino-acid context of a missense mutation. 3Cnet utilizes knowledge transfer of evolutionary conservation to train insufficient clinical data without overfitting. The performance comparison clearly shows that 3Cnet can find the true disease-causing variant from a large number of missense variants in the genome of a patient with higher sensitivity (recall = 13.9 %) compared to other prediction tools such as REVEL (recall = 7.5 %) or PrimateAI (recall = 6.4 %). Consequently, 3Cnet can improve the diagnostic rate for patients and discover novel pathogenic variants with high probability.