TOWARDS A CLINICAL PATHOGEN IDENTIFICATION SYSTEM USING DEEP LEARNING AND BAYESIAN INFERENCE ON HIGH PERFORMANCE COMPUTING PLATFORMS

Abstract

The advent of next generation sequencing (NGS) has provided a powerful approach to identify pathogens through bioinformatics pipelines. However, there are several limitations of current bioinformatics tools: (i) the scalability is usually not optimized (ii) the accuracy on divergent species is limited and (iii) the interpretation of a pathogen report may not be intuitive for clinicians.

To address the scalability of bioinformatics pipelines, we applied profiling and different optimization strategies for high performance computing (HPC) systems on SURPI pipeline and successfully accelerated it and demonstrated its scalability. To address the accuracy of taxonomic assignment, we developed a multi-task convolutional neural network (MT-CNN) to perform both pathogen identification and genomic coverage estimation for human viruses which outperformed current state-of-the-art tools. Finally, we explored how the interpretability of pathogen reports could be enhanced by incorporating a Bayesian approach to assigning the likelihood of a pathogen based on relevant factors.