Comparative analysis of metagenomes can be used to detect sub-metagenomes (species or gene sets) that are associated with specific phenotypes (e.g., host status). The typical workflow is to assemble and annotate metagenomic datasets individually or as a whole, followed by statistical tests to identify differentially abundant species/genes. We previously developed subtractive assembly (SA), a de novo assembly approach for comparative metagenomics that first detects differential reads that distinguish between two groups of metagenomes and then only assembles these reads. Application of SA to type 2 diabetes (T2D) microbiomes revealed new microbial genes associated with T2D. Here we further developed a Concurrent Subtractive Assembly (CoSA) approach, which uses a Wilcoxon rank-sum (WRS) test to detect k-mers that are differentially abundant between two groups of microbiomes (by contrast, SA only checks ratios of k-mer counts in one pooled sample versus the other). It then uses identified differential k-mers to extract reads that are likely sequenced from the sub-metagenome with consistent abundance differences between the groups of microbiomes. Further, CoSA attempts to reduce the redundancy of reads (from abundant common species) by excluding reads containing abundant k-mers. Using simulated microbiome datasets and T2D datasets, we show that CoSA achieves strikingly better performance in detecting consistent changes than SA does, and it enables the detection and assembly of genomes and genes with minor abundance difference. A SVM classifier built upon the microbial genes detected by CoSA from the T2D datasets can accurately discriminates patients from healthy controls, with an AUC of 0.94 (10-fold cross-validation), and therefore these differential genes (207 genes) may serve as potential microbial marker genes for T2D.

Microbial community members exhibit various forms of interactions. Taking advantage of the increasing availability of microbiome data, many computational approaches have been developed to infer bacterial interactions from the co-occurrence of microbes across diverse microbial communities. Additionally, the introduction of genome-scale metabolic models have also enabled the inference of cooperative and competitive metabolic interactions between bacterial species. By nature, phylogenetically similar microbial species are more likely to share common functional profiles or biological pathways due to their genomic similarity. Without properly factoring out the phylogenetic relationship, any estimation of the competition and cooperation between species based on functional/pathway profiles may bias downstream applications. To address these challenges, we developed a novel approach for estimating the competition and complementarity indices for a pair of microbial species, adjusted by their phylogenetic distance. An automated pipeline, PhyloMint, was implemented to construct competition and complementarity indices from genome scale metabolic models derived from microbial genomes. Application of our pipeline to 2,815 human-gut associated bacteria showed high correlation between phylogenetic distance and metabolic competition/cooperation indices among bacteria. Using a discretization approach, we were able to detect pairs of bacterial species with cooperation scores significantly higher than the average pairs of bacterial species with similar phylogenetic distances. A network community analysis of high metabolic cooperation but low competition reveals distinct modules of bacterial interactions. Our results suggest that niche differentiation plays a dominant role in microbial interactions, while habitat filtering also plays a role among certain clades of bacterial species.

Recent years have witnessed unprecedented accumulation of DNA sequences and therefore protein sequences (predicted from DNA sequences), due to the advances of sequencing technology. One of the major sources of the hypothetical proteins is the metagenomics research. Current annotation of metagenomes (collections of short metagenomic sequences or assemblies) relies on similarity searches against known gene/protein families, based on which functional profiles of microbial communities can be built. This practice, however, leaves out the hypothetical proteins, which may outnumber the known proteins for many microbial communities. On the other hand, we may ask: what can we gain from the large number of metagenomes made available by the metagenomic studies, for the annotation of metagenomic sequences as well as functional annotation of hypothetical proteins in general? Here we propose a community profiling approach for predicting functional associations between proteins: two proteins are predicted to be associated if they share similar presence and absence profiles (called community profiles) across microbial communities. Community profiling is conceptually similar to the phylogenetic profiling approach to functional prediction, however with fundamental differences. We tested different profile construction methods, the selection of reference metagenomes, and correlation metrics, among others, to optimize the performance of this new approach. We demonstrated that the community profiling approach alone slightly outperforms the phylogenetic profiling approach for associating proteins in species that are well represented by sequenced genomes, and combining phylogenetic and community profiling further improves (though only marginally) the prediction of functional association. Further we showed that community profiling method significantly outperforms phylogenetic profiling, revealing more functional associations, when applied to a more recently sequenced bacterial genome.

The microbiome research is going through an evolutionary transition from focusing on the characterization of reference microbiomes associated with different environments/hosts to the translational applications, including using microbiome for disease diagnosis, improving the efficacy of cancer treatments, and prevention of diseases (e.g., using probiotics). Microbial markers have been identified from microbiome data derived from cohorts of patients with different diseases, treatment responsiveness, etc, and often predictors based on these markers were built for predicting host phenotype given a microbiome dataset (e.g., to predict if a person has type 2 diabetes given his or her microbiome data). Unfortunately, these microbial markers and predictors are often not published so are not reusable by others. In this paper, we report the curation of a repository of microbial marker genes and predictors built from these markers for microbiomebased prediction of host phenotype, and a computational pipeline called Mi2P (from Microbiome to Phenotype) for using the repository. As an initial effort, we focus on microbial marker genes related to two diseases, type 2 diabetes and liver cirrhosis, and immunotherapy efficacy for two types of cancer, non-small-cell lung cancer (NSCLC) and renal cell carcinoma (RCC). We characterized the marker genes from metagenomic data using our recently developed subtractive assembly approach. We showed that predictors built from these microbial marker genes can provide fast and reasonably accurate prediction of host phenotype given microbiome data. As understanding and making use of microbiome data (our second genome) is becoming vital as we move forward in this age of precision health and precision medicine, we believe that such a repository will be useful for enabling translational applications of microbiome data.

Microbial organisms play important roles in many aspects of human health and diseases. Encouraged by the numerous studies that show the association between microbiomes and human diseases, computational and machine learning methods have been recently developed to generate and utilize microbiome features for prediction of host phenotypes such as disease versus healthy cancer immunotherapy responder versus nonresponder. We have previously developed a subtractive assembly approach, which focuses on extraction and assembly of differential reads from metagenomic data sets that are likely sampled from differential genomes or genes between two groups of microbiome data sets (e.g., healthy vs. disease). In this article, we further improved our subtractive assembly approach by utilizing groups of k-mers with similar abundance profiles across multiple samples. We implemented a locality-sensitive hashing (LSH)-enabled approach (called kmerLSHSA) to group billions of k-mers into k-mer coabundance groups (kCAGs), which were subsequently used for the retrieval of differential kCAGs for subtractive assembly. Testing of the kmerLSHSA approach on simulated data sets and real microbiome data sets showed that, compared with the conventional approach that utilizes all genes, our approach can quickly identify differential genes that can be used for building promising predictive models for microbiome-based host phenotype prediction. We also discussed other potential applications of LSH-enabled clustering of k-mers according to their abundance profiles across multiple microbiome samples.