The recognition of the microbiome’s importance in human health and disease has been constantly growing in recent decades. Aberrant microbial states have been extensively characterized, yet remain poorly understood. In most cases we still do not know what drives, sustains, or modulates these states - let alone how to reverse them. Causal effects of the microbiome on its host have been established in multiple clinical settings, yet this evidence is often at the ecosystem level, rarely pinpointing specific mechanisms or treatment targets. We presume this is because these phenotypes are being driven by emergent phenomena at the system level and not by specific microbes or metabolic pathways.
The major challenge in studying the microbiome lies in our limited ability to characterize, sample or manipulate it. In many cases, such as in the gut or lungs, it is difficult or impossible to sample these communities directly, forcing us to rely on dirty proxies such as stool; the intricate interactions with the host limit the applicability of animal models (and in some cases, no good models exist); and most microbes are unculturable, preventing detailed characterization in vitro.
We tackle this challenge directly, and combine the noisy outputs we can measure from the microbiome (metagenomics, metabolomics, etc.) into a model of microbial metabolism, studying how different microbes combine their metabolic activities to produce metabolites that have a systemic effect on the host.
Human microbiome analysis
We are developing novel ways to analyze microbiome related data (e.g., metagenomics), moving beyond counting microbe and gene abundances to inferring microbial growth, niche specialization, metabolic activity and host-specific effects.
Data-driven analysis of host-microbe interaction
We are devising structure learning and causal discovery methods that provide a data-driven view of the interaction between the microbiome and its host while accounting for environmental effects. This will enable us to detect putative causal and mechanistic effects, complementing knowledge-driven approaches.
Knowledge-driven metabolic analysis
We are adapting knowledge driven metabolic modelling approaches to the complex setting of the microbiome in order to predict the profile of microbially generated- and modulated- metabolites in a given interface with the host, identifying putative mechanistic pathways for host-microbiome interaction and potential intervention points.
We apply our research in diverse clinical settings, with a focus on adverse pregnancy outcomes and antibiotic resistance. We are searching for way to utilize the microbiome in personalizing medical care, and are seeking putative mechanisms through which microbes affect disease that could be manipulated by therapeutics.
Modeling contamination sources facilitates improved microbiome-based prediction.
Contamination in microbiome studies has been a cause of major controversies in the field, for example with respect to the placental microbiome. We developed a SCRuB, new method that detects and remove contamination from microbiome data using process (negative) controls. The key idea behind our method it is that instead of trying to identify whether a specific microbe is a contaminant, we model the entire contamination source – a mixture of several microbes with different abundances. Modeling contamination sources also allows us to handle the leakage of biological material from biological samples to adjacent negative controls during processing - a critical phenomenon which severely reduces the accuracy of contemporary decontamination methods. We demonstrated that our method facilitates the detection of signals that were so-far left hidden in contamination: we can obtain accurate diagnosis of melanomas based on circulating microbial DNA; and, separately, we are able to predict response to immune checkpoint inhibitor treatment in melanoma patients based on the microbiome profile of the tumor itself (Austin et al., Nature Biotechnology 2023).
SCRuB iteratively uses the shared information across samples and controls to estimate well-to-well leakage and the composition of the contamination source. It then uses estimates of the contamination sources to infer the underlying composition of the samples, and so on until convergence.
Studying the interaction between the vaginal microbiome, metabolome, and adverse pregnancy outcomes.
Despite many advances in diagnostics and precision medicine, most adverse pregnancy outcomes (APOs) are still a "surprise" - they are typically detected late, when symptoms are overt; it is then often also too late to administer effective treatment. The vaginal microbiome is associated with most APOs, and believed to have a causal role in many of them. We are studying how host-microbiome interactions affect subsequent APOs. In parallel, we are identifying biomarkers that can help us diagnose these conditions early in pregnancy.
We have recently conducted a large-scale analysis of the early-pregnancy vaginal microbiome and metabolome in a case-control study of preterm birth, the leading cause of neonatal morbidity and mortality. We found intriguing associations between several xenobiotics and subsequent preterm delivery, suggesting a role for exogenous exposures in its pathogenesis. We built supervised machine learning models able to predict subsequent prematurity, weeks-to-months in advance, with high accuracy, using measurements of the vaginal metabolome (Kindschuh et al., Nature Microbiology 2023). In another recent investigation, we found that strain diversity in the vaginal microbiome is also associated with preterm birth (Liao et al., bioRxiv 2023).
We accurately predict the risk for subsequent preterm birth in held-out samples from metabolomics data collected early in pregnancy.
A data-driven approach reveals robust associations between serum metabolites and host, microbial, and environmental factors.
The serum metabolome contains multiple biomarkers and causal agents that are important for a variety of human diseases. But even if these are identified, designing interventions that affect their levels remains a significant challenge. We took a data-driven approach to this problem, designing and contrasting machine-learning models that predict metabolite levels in held out data using different feature types - microbiome, diet, clinical data, etc. Our results show a strong contribution by microbiome and diet, with distinct contributions for each - identifying starting points for hundreds of potential interventions (Bar* & Korem* et al., Nature 2020). Our lab is currently combining this data-driven approach with knowledge-driven mechanistic methods to identify potential mechanisms by which the microbiome produces different metabolites.
The relative power of each feature type in predicting the serum metabolome.
Systematic examination of microbial structural variations facilitates mechanistic investigations
The same microbial strain could be very different in different people, and a difference in very few genes could have significant phenotypic effects. We analyzed sequencing coverage to systematically detect genomic structural variations across multiple microbes in two large cohorts. We found that these duplicated or deleted regions exhibit multiple novel associations with several disease risk factors. Better yet, examining genes coded in these regions allows us to raise hypotheses regarding potential mechanisms underlying these host-microbe interactions (Zeevi* & Korem* et al., Nature 2019).
Structural variation in Anaerostipes hadrus whose deletion is associated with higher disease risk. Genes in this region code a composite inositol catabolism - butyrate production pathway, potentially supplying the microbe with additional energy while supplying the host with beneficial butyrate.
Inferring dynamic microbial growth rates from a single, static snapshot
Most microbiome analysis approaches concentrate on compositions - which and how many bacteria or genes are present in a sample. We were able to infer the growth dynamics of multiple bacteria from just a single sample, by extracting and examining the DNA copy-number signal produced by bacterial DNA replication. We showed that these growth rates can potentially serve as biomarkers for disease, antibiotic treatment and pathogenic colonization (Korem et al., Science 2015, Joseph et al., Genome Research 2022).
Bacterial DNA replication generates a DNA copy number signal, which is discernible from coverage analysis (high and low coverage near the origin and terminus of replication).
Microbiome analysis informs personally tailored diets
The way our blood sugar levels respond to food is an important risk factor for diabetes, obesity, and other metabolic diseases. We profiled a large 800-participants cohort and showed that these sugar responses vary significantly between different individuals (even when they consume identical meals), and that using meal, lifestyle, and microbiome data allows us to accurately predict these responses personally for each individual and for any complex meal. We demonstrated the practical use of this approach, by demonstrating normalization of blood sugar levels in a prospective dietary intervention trial (Zeevi* & Korem* et al., Cell 2015). We later showed that for some dietary decisions, such as between white and whole-wheat bread, examining just the microbiome is probably enough to make an accurate prediction (Korem et al., Cell Metab. 2017)
Postprandial glycemic responses (PPGRs; y-axis) are reduced in a personally tailored glucose-lowering diet (green) compared to a glucose-increasing diet (red).