FrameRate: learning the coding potential of unassembled metagenomic reads

Abstract

Motivation: Metagenomic assembly is a slow and computationally intensive process and despite needing iterative rounds for improvement and completeness the resulting assembly often fails to incorporate many of the input sequencing reads. This is further complicated when there is reduced read-depth and/or artefacts which result in chimeric assemblies both of which are especially prominent in the assembly of metagenomic datasets. Many of these limitations could potentially be overcome by exploiting the information content stored in the reads directly and thus eliminating the need for assembly in a number of situations. Results: We explored the prediction of coding potential of DNA reads by training a machine learning model on existing protein sequences. Named 'FrameRate', this model can predict the coding frame(s) from unassembled DNA sequencing reads directly, thus greatly reducing the computational resources required for genome assembly and similarity-based inference to pre-computed databases. Using the eggNOG-mapper function annotation tool, the predicted coding frames from FrameRate were functionally verified by comparing to the results from full-length protein sequences reconstructed with an established metagenome assembly and gene prediction pipeline from the same metagenomic sample. FrameRate captured equivalent functional profiles from the coding frames while reducing the required storage and time resources significantly. FrameRate was also able to annotate reads that were not represented in the assembly, capturing this 'missing' information. As an ultra-fast read-level assembly-free coding profiler, FrameRate enables rapid characterisation of almost every sequencing read directly, whether it can be assembled or not, and thus circumvent many of the problems caused by contemporary assembly workflows.