DeepMAP: Deep CNN Classifiers Applied to Optical Mapping for Fast and Precise Species-Level Metagenomic Analysis

Abstract

DNA optical mapping is a powerful technique commonly used for structural variant calling and genome assembly verification. Despite being inherently high-throughput, the method has not yet been applied to highly complex settings such as species identification in microbiome analysis due to the lack of alignment algorithms that can both assign large numbers of reads in minutes and handle large database size. In this work, we present a novel genomic classification pipeline based on deep convolutional neural networks for optical mapping data (DeepMAP), which can perform fast and accurate assignment of individual optical maps to their respective genomes. We furthermore achieve a superior performance of DeepMAP in the presence of evolutionary divergent sequences, making it robust to the presence of unknown strains within metagenomic samples. We evaluate DeepMAP on genomic DNA extracted from bacterial mixtures, reaching species-level resolution with true positive rates of around 75% and a false positive rate of less than 1%, with measured classification speeds significantly outpacing those of previously developed approaches for high-density optical mapping data alignment.