Machine learning based prediction of functional capabilities in metagenomically assembled microbial genomes

Abstract

The increasing popularity of genome resolved meta genomics -the binning of genomes of potentially uncultured organisms direct from the environmental DNA -has resulted in a deluge of draft genomes. There is a pressing need to develop methods to interpret this data. Here, we used machine learning to predict functional and metabolic traits of microbes from their genomes. We collated an extensive database of 84 phenotypic traits associated with 9407 prokaryotic genomes and trained different machine learning models on this data. We found that a lasso logistic regression based on the frequency of gene orthologs had the best combination of functional prediction performance and interpretability. This model was able to classify 65 phenotypic traits with greater than 90 7 Microbiome Project [3], the Earth Microbiome Project [4] and the Tara Oceans 8 Project [5] have systematically sequenced the microbial communities in a huge variety of 9 environments at great depth. Amplicon sequencing, such as of the 16S rRNA gene, 10 allows detailed study of the taxonomic makeup of these communities, while shotgun 11 metagenomic sequencing allows characterisation of all genes present in an environment. 12 Increasing depth of coverage and improvements in genome binning algorithms for 13 clustering contigs into genomes, in particular the use of differential coverage across 14 different samples [6,7], are allowing more and more full and partial genomes to be 15 assembled from shotgun metagenomic studies. Many of these genomes are novel and 16 belong to uncultured organisms that are never studied in the laboratory. A recent 17 metagenomic study on aquifer systems [8], for example, reconstructed 2540 separate 18 high-quality, near-complete genomes, and claimed to have discovered an astonishing 47 19 new phylum-level lineages among them. 20Converting this exponentially growing sequence data into functional understanding 21 of microbial communities requires us to determine physiological functions from it [9]. 22This would allow the inference of key functions in microbial communities, and how these 23 PLOS1/22 functions change over ecological conditions and with time [9]. In turn, this ability, could 24 allow us to discern ecological adaptations in environmental microbial communities, as 25 well as to achieve functional mechanistic models of stability and function [10]. 26Efforts to achieve phenotype-genotype mapping from environmental sequence data 27 has so far mostly focussed on phylogenetic assignments using the 16S rRNA gene. This 28 highly conserved gene can provide a phylogenetic assignment at the species (or higher) 29 level, which can then be used to infer general functional traits. While this approach has 30 been commonly used to study ecological distribution of microbial functions e.g. [11][12][13], 31 its premise of a direct association of function with phylogenetic assignment (i.e. 32'functional coherence of microbial taxa') is questionable (e.g. [14]). The level of 33 taxonomic coherence of function is not clear even for strains o...

Machine Learning Approaches for Epidemiological Investigations of Food-Borne Disease Outbreaks

“…Moreover, such methods are rather time-consuming and computationally demanding, thus representing a bottleneck for efficient sequence data analysis (Sharma et al, 2015). ML algorithms could potentially increase the accuracy and speed of clinically and epidemiologically relevant predictions (Farrell et al, 2018). However, to yield accurate predictions, besides the choice of the most appropriate algorithm and a set of well-defined inputs and outputs of interest, ML-based strategies generally require large amounts of high-quality training data (Baker et al, 2018).…”

“…This presents a limitation, as currently microbial genome databases are known to be biased toward cultivable pathogenic bacteria. The current lack of large and comprehensive databases can be considered as the key bottleneck for the application of ML methods (Farrell et al, 2018). Hence, future improvements can be expected to come from better data curation and collection, in addition to development of new and improved classification algorithms (Farrell et al, 2018).…”

“…The current lack of large and comprehensive databases can be considered as the key bottleneck for the application of ML methods (Farrell et al, 2018). Hence, future improvements can be expected to come from better data curation and collection, in addition to development of new and improved classification algorithms (Farrell et al, 2018). Therefore, WGS data collection must be done in parallel with comprehensive and standartized metadata collection such as phenotypic profiling using traditional microbiology methods for isolate characterization (e.g., phenotypic profiling of antimicrobial resistance) (Maurer et al, 2017).…”

“…However, for large-scale read mapping, BLAST may be too slow (Deneke et al, 2017). Generally, this approach may fail to detect novel species in cases when closely related genomes are not found in the reference databases (Deneke et al, 2017), which are known to be biased toward cultivable pathogenic bacteria (Farrell et al, 2018). Average Nucleotide Identity (ANI) (Clingenpeel et al, 2015) has been recently proposed as an alternative metrics for the identification and classification of bacterial species, calculated by performing several pair-wise comparisons of all sequences shared between two given strains.…”

“…ML strategies have already been used for microbial diagnostics in diverse contexts, including (i) taxonomic grouping of metagenomics data (Sedlar et al, 2017; Afify and Al-Masni, 2018); (ii) classification of L. monocytogenes persistence in retail delicatessen environments (Vangay et al, 2014); (iii) phenotype prediction of bacterial strains based on presence/absence of particular genes (i.e., gene-trait matching) (Dutilh et al, 2013; Alkema et al, 2016; Farrell et al, 2018); (iv) to identify strains that demonstrate a higher probability to cause severe diseases (Wheeler et al, 2018); (v) to predict the host range of pathogens (Lupolova et al, 2017), e.g., identifying their signatures of host adaptation (Wheeler et al, 2018); and (vi) to predict the antimicrobial resistance potential of different E. coli strains (Her and Wu, 2018) or from different sources (Li et al, 2018).…”