Ecologically informed microbial biomarkers and accurate classification of mixed and unmixed samples in an extensive cross-study of human body sites

Abstract

BackgroundThe identification of body site-specific microbial biomarkers and their use for classification tasks have promising applications in medicine, microbial ecology, and forensics. Previous studies have characterized site-specific microbiota and shown that sample origin can be accurately predicted by microbial content. However, these studies were usually restricted to single datasets with consistent experimental methods and conditions, as well as comparatively small sample numbers. The effects of study-specific biases and statistical power on classification performance and biomarker identification thus remain poorly understood. Furthermore, reliable detection in mixtures of different body sites or with noise from environmental contamination has rarely been investigated thus far. Finally, the impact of ecological associations between microbes on biomarker discovery was usually not considered in previous work.ResultsHere we present the analysis of one of the largest cross-study sequencing datasets of microbial communities from human body sites (15,082 samples from 57 publicly available studies). We show that training a Random Forest Classifier on this aggregated dataset increases prediction performance for body sites by 35% compared to a single-study classifier. Using simulated datasets, we further demonstrate that the source of different microbial contributions in mixtures of different body sites or with soil can be detected starting at 1% of the total microbial community. We apply a biomarker selection method that excludes indirect environmental associations driven by microbe-microbe associations, yielding a parsimonious set of highly predictive taxa including novel biomarkers and excluding many previously reported taxa. We find a considerable fraction of unclassified biomarkers (“microbial dark matter”) and observe that negatively associated taxa have a surprisingly high impact on classification performance. We further detect a significant enrichment of rod-shaped, motile, and sporulating taxa for feces biomarkers, consistent with a highly competitive environment.ConclusionsOur machine learning model shows strong body site classification performance, both in single-source samples and mixtures, making it promising for tasks requiring high accuracy, such as forensic applications. We report a core set of ecologically informed biomarkers, inferred across a wide range of experimental protocols and conditions, providing the most concise, general, and least biased overview of body site-associated microbes to date.Electronic supplementary materialThe online version of this article (10.1186/s40168-018-0565-6) contains supplementary material, which is available to authorized users.

Rapid Inference of Direct Interactions in Large-Scale Ecological Networks from Heterogeneous Microbial Sequencing Data

“…We also found MVs to be interesting in their own right: in our HMP analysis, they constituted central nodes in the association network with many directly associated OTUs, in line with the expected habitat preference of many microbes (The Human Microbiome Project Consortium 2012) and known primer biases (Tremblay et al, 2015). Consistent with these results, closely related approaches have previously identified parsimonious sets of predictive microbial biomarkers for human body sites and a skin disease (Tackmann et al, 2018;Statnikov et al, 2013).…”

Assessing time dependent changes in microbial composition of biological crime scene traces using microbial RNA markers

“…For instance, Hanssen et al [14] showed that a combination of principal component analysis (PCA) and linear discriminant analysis (LDA) can successfully be used to differentiate samples from saliva being deposited on skin and samples of skin only. In addition, Tackmann et al [43] have shown that machine learning algorithms trained on large heterogeneous datasets provide high accuracy when predicting body fluids, not only from single source samples but also from mixtures generated in-silico. In another study, Hanssen et al [15] developed a prediction model using partial least squares (PLS) in combination with LDA using data from the Human Microbiome Project for the identification of samples originating from the oral, nasal and vaginal cavity as well as skin and feces.…”

Issues and current standards of controls in microbiome research

“…2017) (although the authors report a clean negative control) and sometimes pure in-silico studies might identify contaminations as biologically relevant (Tackmann et al . 2018). These examples come in addition to further publications mentioned in (de Goffau et al .…”