This analysis was conducted using 6 publicly available human gut microbiome datasets, each of which contains counts of operational taxonomic units (OTUs) derived from 16s rRNA gene sequences for both lean and obese people ( Table 1 ). Samples were analyzed from a total of 1148 lean individuals and 1287 obese individuals. The selection of these datasets was based on their comprehensive representation of gut microbial profiles across different populations and the presence of lean and obese individuals in each, making it possible to carry out meaningful cross-comparisons. These datasets were obtained utilizing a variety of sequencing systems, including Illumina and 454 pyrosequencing. To reduce platform-specific technical heterogeneity, we aggregated data at the phylum level and used the same preprocessing strategy (Walters et al., 2014) for all datasets. This degree of taxonomic precision has been shown to be robust in cross-platform comparisons (Walters et al., 2014). Inclusion requirements also required the availability of BMI metadata, a sufficient sample size, and public accessibility. The six datasets provide a broad overview of the gut microbiome in lean, overweight, and obese individuals, encompassing diverse age ranges, phenotypic categories, and some sex distribution data. Out of the six datasets, four (American Gut, Ross, HMP, and Goodrich) provided gender distribution data, revealing a total of 4,165 female and 3,899 male participants across all groups. The remaining two datasets (Turnbaugh and Gordon) did not include gender distribution, representing 7.0% of the total samples. In terms of phenotypes, the datasets include a total of 5,157 samples classified as H (Healthy) and 5,909 samples classified as OB (Obese), summing lean and overweight individuals into the obese category. Age data were available for most datasets, covering a broad range from 0 to 93 years, with mean ages varying between datasets: 25.9 to 61.8 years for healthy individuals and 28.0 to 61.3 years for obese individuals. Obesity is defined as abnormal or excessive fat accumulation that presents a risk to health. An individual is generally considered overweight if they have a body mass index (BMI) value exceeding 25 kg m^-2 and obese if they have a value exceeding 30 kg m^-2. Due to data limitations, for the purposes of this study, we classified an individual as “lean” if they were identified as such or had a reported BMI less than or equal to 25 kg m^-2, and “obese” if they were reported as being overweight or obese or had a reporting BMI exceeding 25 kg m^-2. However, it also proves that diet affects the microbiota and, thus, while this analysis is confined to the evaluation of microbiota composition and interactions, the differences in dietary habits among the populations studied might influence the patterns of microbial dynamics. While we aimed to control for potential confounding factors such as dietary habits and geographic differences, we recognize that these factors may still influence our findings. Future research should strive for more balanced sample sizes and include detailed dietary assessments. Dietary components such as fiber, lipids, and proteins have been proven to have a significant effect on gut microbiota composition. High-fiber diets promote the growth of beneficial bacteria such as Bacteroidetes and Firmicutes, which have been associated with improved metabolic health (Zeevi et al., 2015). Saturated fat-rich meals, on the other hand, may promote pathogenic bacteria, contributing to dysbiosis and obesity (Castaner et al., 2018). Clinical studies have demonstrated that dietary interventions can lead to measurable changes in microbiota composition, reinforcing the notion that diet is a critical factor in shaping microbial dynamics (Walters et al., 2014; Tseng and Wu, 2019).

Counts in each database were aggregated from the OTU level to the phylum level in R Studio Version 1.4.1717 using the R package phyloseq (McMurdie and Holmes, 2013). The model was built at all taxonomic levels for each individual database, with the phylum level showing the highest coefficient of determination (R²). Given this, the phylum level was chosen as the baseline for analysis, offering a balance between identifying microbial interactions and maintaining model performance and interpretability. Taxa within phyla appear to show differential responses to obesity status, but we focused on the phylum level to facilitate comparisons. Phyla were only included in the analysis if they were found in at least 33% of the samples for that database ( Table 1 ), which is somewhat more stringent than the often-used criterion of 25% prevalence (Li et al., 2021). Using the 25% criterion would have resulted in the inclusion of 1 additional phylum (Verrucomicrobia) in the analysis of HMP dataset, and of 1 additional phylum (Proteobacteria) in the analysis of the Gordon dataset. Because we had a limited number of samples and BEEM-static had been tested on over 4,000 microbiome data points, we decided to use 37% prevalence as a criterion. The same OTUs are still significant at 25% and 33% prevalence, and because they are the most abundant overall, any pruning variant in between these criteria would result in the same outcome. BEEM-static detected microbiomes that were not in equilibrium and automatically deleted them from further investigation.

BEEM-Static takes as input an OTU table of counts comprised of samples (in columns) and taxa (in rows). While BEEM-Static can provide useful insights by inferring ecological interactions from cross-sectional data, it is essential to acknowledge its limitations. The gut ecosystem is extraordinarily dynamic and responds to various environmental pressures, making it susceptible to complex microbial interactions that can change over time. Although our goal was to identify ‘key players’ that distinguish between obese and lean populations, we recognize that BEEM-Static cannot fully capture the temporal dynamics inherent in microbiome changes, particularly in response to dietary interventions aimed at preventing or treating obesity. Longitudinal studies using dynamical models are better suited for understanding these complexities because they can track changes over time and provide insights into how interventions impact microbiome dynamics. For instance, our previous work demonstrated the importance of longitudinal analysis in understanding urinary microbiota dynamics (Ceprnja et al., 2021). However, such models require substantial collection of data and may be more susceptible to individual response variability. As a result, while our approach offers a practical method of analyzing cross-sectional data, it should be viewed as complementary to longitudinal studies rather than a replacement.

For each dataset, samples for lean and obese individuals were separately analyzed using BEEM-Static at the phylum level with the aim of first inferring the bacterial interaction within the community for each population (Méndez-Salazar et al., 2018), and then comparing the estimated carrying capacities and interaction coefficients between microbiota of lean and obese people.

Estimated carrying capacities for the lean and obese datasets were compared using pooled two-sample t-tests. Interaction coefficients that were identified as significant in both datasets were also compared in this way. The interaction strengths estimated by BEEM-Static represent the effect of one microbial phylum on another. A negative number indicates a competitive or inhibitory interaction, in which a rise in one phylum’s abundance suppresses the other. A positive score, on the other hand, indicates mutualistic or cooperative relationships, in which a rise in one phylum’s abundance benefits the other. Values close to zero, such as -0.1 or +0.1, indicate weak or negligible interactions, but values between -0.2 and -0.5, or +0.2 to +0.5, are considered moderate interactions. Values less than -0.5 or greater than +0.5 indicate significant interactions, with values approaching -1 or +1 indicating highly significant relationships. For instance, an interaction strength of -0.2 between Firmicutes and Bacteroidetes suggests moderate competition, whereas a value of -0.7 would signify a stronger competitive interaction. These thresholds provide a clearer framework for interpreting the significance and impact of interaction strengths within the gut microbiome. BEEM-Static also yields a coefficient of determination (R²) for each phylum that provides an indication of how well the model performed in predicted the sample relative abundances. The BEEM-Static package was downloaded from GitHub (Li et al., 2021), built in RStudio (version 1.4.1717), and compiled on a Windows 10 machine running R-4.0.2 (code available on https://github.com/enmelvan/Chapter4). To promote transparency and reproducibility, all datasets and analysis codes used in this study are publicly available on GitHub.

The total number of phyla retained for analysis ranged from 4 to 7 depending on the database, including Actinobacteria, Bacteroidetes, Firmicutes, Lentisphaerae, Proteobacteria and Verrucomicrobia ( Table 2 ). These phyla were selected due to their known relevance in obesity-related gut microbiota research and their consistent presence across datasets.

A total of 12 BEEM-Static analyses were performed, including one lean-population analysis and one obese-population analysis, for each of the 6 datasets. For each analysis, BEEM-Static yielded a set of estimates for carrying capacities and interaction coefficients that together can be used to predict the relative abundances of each phylum in each sample. Phylum-level coefficients of determination between the observed and predicted relative abundances consistently yielded R2 values greater than or equal to 0.55 ( Supplementary Table 1 ), indicating that the fitted models provided robust descriptions of the data. The consistent R² values of ≥0.55 across datasets confirm the robustness of the BEEM-Static model in accurately predicting phylum-level relative abundances, underscoring its utility in analyzing gut microbiota dynamics in lean and obese individuals.

Estimated carrying capacities ( Supplementary Table 2 ) differed significantly between the lean and obese datasets for most phyla in at least 4 of the 6 datasets ( Table 3 ), with Firmicutes and Bacteroidetes generally having the highest values, as expected given their generally greater abundances. Interpretation of these differences are, however, challenging because the BEEM-Static algorithm estimated these parameters by setting the median total abundance estimate to an arbitrary default value of 1000, making these estimates sensitive to the average taxon richness of samples. The challenge in interpreting these differences arises from the arbitrary setting of the median total abundance estimate, which may affect the sensitivity of the carrying capacity estimates to variations in sample richness. We therefore will not discuss these results any further.

Among the 6 datasets, 57 significant interactions were identified for obese populations, 79% of which were negative, while only 37 significant interactions were identified for lean populations, 92% of which were negative ( Supplementary Table 3 ; see https://github.com/enmelvan/Chapter4 for the complete set of significant interactions and statistical comparisons among coefficients). Two interactions were identified as significant only in the lean-data analyses, whereas 5 interactions were identified as only significant in the obese-data analyses. All of the interaction coefficients that were identified as significant, and of the same sign, in 3 or more databases were negative ( Table 4 ), suggesting that the microbiota is predominantly structured by amensal and competitive interactions, at least at the phylum level. Negative effects of Firmicutes on the abundance Bacteroidetes, and Bacteroidetes on Firmicutes, were the most commonly observed interactions. The predominance of negative interactions in obese microbiotas suggests a competitive and possibly dysbiotic microbial environment, which could be a target for therapeutic interventions aimed at restoring a healthier balance.

The interaction coefficients identified as being significant by BEEM-Static can be represented as directed-network graphs ( Supplementary Figure 1 ). These network graphs are crucial for visualizing and interpreting the complex interactions within the microbiota, providing a clear representation of the differences between lean and obese individuals. Using analyses of HMP data as an example ( Figure 1 ), the BEEM-Static analyses indicate that increases in Actinobacteria abundance have a strong positive effect on Bacteroidetes abundance among lean people (denoted by a thick blue arrow in Figure 1 ), but a weak negative effect on Bacteroidetes abundance among obese people (denoted by a thin red arrow).

All notable differences in microbial interactions within lean and obese individuals can be found in Table 5 , and all significant interactions can be found in https://github.com/enmelvan/Chapter4.

Five or more of the 6 datasets showed a significant negative effect of Firmicutes on Bacteroidetes, as well as a significant negative effect of Bacteroidetes on Firmicutes ( Table 4 ), indicating that the two taxa are in competition with one another. Three of the 6 datasets yielded Bacteroidetes effects on Firmicutes that differed significantly between the lean and obese analyses. The average negative interaction strength from Bacteroidetes to Firmicutes in obese people was -0.4, compared to -0.26 in lean people.

In the lean population, there were just two significantly distinct interactions: the previously described Firmicutes to Bacteroidetes interaction and, not surprisingly, the Bacteroidetes to Firmicutes interaction. The observed competition between Bacteroidetes and Firmicutes suggests that dietary interventions targeting these phyla could be critical in managing obesity. It is important to note, however, that the predictive value of the Firmicutes-to-Bacteroidetes (F/B) ratio for obesity remains controversial. Several recent studies have reported inconsistent or non-significant associations between this ratio and obesity status. For instance, a 2023 study found no significant differences in the F/B ratio between obese and lean individuals, questioning its utility as a universal biomarker for obesity (Zhang et al., 2023). Similarly, a 2022 review highlighted that variations in diet, geography, and methodological approaches contribute to the heterogeneity of findings, and emphasized the need for more refined methods to assess microbiome composition and function (Cheng et al., 2022). Our findings support this perspective by shifting the focus from static abundance ratios to dynamic interaction patterns, which may offer more robust insights into dysbiosis and host-microbiome relationships. For example, high-fiber diets are known to favor Bacteroidetes, potentially offering a strategy to modulate this balance in favor of a healthy microbiome.

Another interaction that occurs in both lean and obese people is the negative interaction between Actinobacteria and Firmicutes, with the average strength of the interaction being much stronger in lean people than obese people (-1.4, versus -0.13, Table 4 ), which is consistent with the previous finding of higher Actinobacteria abundance in obese people’s gut microbiome (John and Mullin, 2016; Tseng and Wu, 2019; Castaner et al., 2018). Actinobacteria exhibited a negative interaction with Bacteroidetes in the lean population, with an interaction strength of -0.72; however, this interaction did not significantly differ between the lean and obese groups, which contradicts established findings in the literature (Turnbaugh et al., 2009; Clarke et al., 2012).

Given the strong interaction between Actinobacteria and other phyla, dietary components known to promote Actinobacteria, such as prebiotics, could be leveraged to influence these microbial dynamics positively. Our findings confirm the dual nature of Actinobacteria’s impact, wherein their abundance and interactions can shift from beneficial to detrimental based on the host’s metabolic state, emphasizing the importance of precision in microbiota-targeted interventions. Personalized dietary strategies based on microbiome profiling could mitigate the negative effects of Actinobacteria in obese individuals, such as promoting strains that enhance metabolic health, and simultaneously leveraging their beneficial potential in lean individuals through tailored prebiotic or probiotic formulations.

Our study has shown that the Actinobacteria and Proteobacteria interactions both with each other and with other phyla are both characteristic and significant within obese individuals, and both phyla appear as the network hub. This is also in accordance with the current literature, as both the Proteobacteria and Actinobacteria phyla are more prevalent in obese people (Castaner et al., 2018; Méndez-Salazar et al., 2018

The negative interaction from Actinobacteria to Proteobacteria was significantly different for the lean and obese populations, confirming a hypothesis of this meta-analysis. The interaction strength of Proteobacteria to Actinobacteria also differed significantly between the lean and obese populations.

The carrying capacity of Actinobacteria is also significantly higher in obese populations than lean populations in 5 datasets, and the strength of the coefficient was twice as large in the obese population, which is in line with the current research of Actinobacteria and obesity. The relative differences between groups are still relevant despite BEEM-Static’s arbitrary total abundance baseline (set at 1000). All carrying capacities are proportionately rescaled by the model, thus comparisons between groups within the same taxon have biological significance even when absolute values cannot be interpreted in isolation. For instance, persistently higher carrying capacities for Firmicutes and Proteobacteria in obese people across several datasets imply that these phyla might have a larger ecological niche in microbiomes linked to obesity, possibly as a result of changes in nutrient availability or microbial shifts brought on by inflammation. We acknowledge that differences in sample sizes across datasets may influence the estimation of carrying capacities. Although BEEM-Static is designed to be robust to varying sample numbers, smaller datasets may yield less stable estimates. Ideally, achieving reliable and generalizable conclusions at the level of individual datasets would require large, homogeneous cohorts—potentially exceeding 10,000 samples per study. However, due to practical constraints, our study relied on existing publicly available data. Therefore, while trends observed consistently across multiple datasets increase our confidence in the results, findings at the level of individual datasets should be interpreted with caution.

The contrasting roles of Actinobacteria underscore the need for personalized dietary interventions (Castaner et al., 2018). Tailoring nutrition based on individual microbiome profiles could mitigate the adverse effects associated with certain Actinobacteria in obese individuals while promoting beneficial strains in lean individuals

Interestingly, Firmicutes and Actinobacteria displayed negative interaction with the Proteobacteria phylum only in the obese data, consistent with recent evidence indicating that inflammatory host response promotes Proteobacteria growth (Shin et al., 2015). A negative interaction from Proteobacteria to Firmicutes was also observed only in the obese individuals.

The Firmicutes to Proteobacteria negative interaction and Proteobacteria to Firmicutes negative interaction are significantly different interactions among the obese populations, with the latter being much greater in absolute magnitude than the former (average of -1 compared to -0.12).

The prominence of Proteobacteria in obese individuals suggests that diets aimed at reducing inflammation, such as those rich in polyphenols, could help suppress this phylum’s overgrowth and improve gut health.

Some significant interactions found in only one database are also worthy of note. In the HMP database, Bacteroidetes had a positive interaction with Actinobacteria in lean individuals, while in obese individuals, this interaction was negative. In the American Gut database, Actinobacteria had a negative interaction with Firmicutes in lean and obese individuals. In the Goodrich database, Verrucomicrobia and Lentisphaerae were also identified as having significant interactions that ranged from negative (obese population) to positive (lean population).

These findings underscore the importance of considering geographic and dietary differences when developing personalized nutrition strategies. For instance, dietary interventions effective in Western populations may need to be adjusted for Eastern diets to achieve similar microbiome modulation.

Also, only one interaction with Lentisphaerae was identified in the lean population, an unassigned bacterium, and this taxon had a positive effect on it. In obese individuals, a total of 10 different interactions were identified with Verrucomicrobia and Lentisphaerae. There is still not much research on these two phyla, except one on Chinese children and adolescents with obesity, which found that Verrucomicrobia and Lentisphaerae were both significantly lower in obese group than those in the control group (Hou et al., 2017). It should be noted that all the databases used in this meta-analysis involved people from Western countries, which have a very different diet than the Chinese population. Assessing potential differences between eastern and western countries is a topic worthy of further investigation due to geographically dietary differences.

Most of the interactions identified were negative (96%) and all the extracted interactions can be found in https://github.com/enmelvan/Chapter4. While most ecological interaction networks consist of positive and negative interactions in approximately the same ratio, results here indicate that interactions in the gut microbiome were mostly negative. While most ecological interaction networks include a balance of positive and negative relationships, our results indicate that gut microbiome interactions, particularly in obese individuals, are predominantly negative. This may be partially explained by the fact that the number of gut bacteria is relatively constrained, such that an increase in one taxon’s abundance often influence a decrease in another’s. Beyond this, such patterns may reflect a disrupted microbial ecosystem, marked by reduced diversity and intensified competition. In obese individuals, dominant taxa can suppress others more aggressively, leading to competitive exclusion and instability. Diet likely plays a role as well: low-fiber, high-fat diets, which are more common in obesity, reduce ecological niches and shared resources, potentially amplifying these antagonistic interactions (Shin et al., 2015; Castaner et al., 2018).

As noted in the Introduction, the largest meta-analysis of human microbiome links to obesity was conducted in 2014 by Walters et al, who identified few consistent relationships between obesity and species composition. In particular, the Firmicutes: Bacteroidetes ratio yielded no consistent patterns. This is why it is critical to investigate the dynamics of microbial communities. Quantifying interactions among taxa would help us better understand the ecology of the gut microbiome and better predict how perturbations will affect it. Removing a taxon that engages in mutualistic interactions, for example, may reduce the number of other taxa that rely on it for survival. Given their importance in understanding community ecology, there is a significant amount of interest in developing techniques to infer taxa interactions from metagenomic data.

One of the most serious restrictions of the presented meta-analysis is that network modeling could not be done on a merged dataset composed of all six different databases. Initially, we attempted to merge these datasets and applied XGBoost to identify patterns, but this approach did not yield meaningful results, as no clear pattern related to BMI was observed (Chen and Guestrin, 2016) (see Supplementary Data for XGBoost results). Consequently, we analyzed each database individually because cluster analyses performed on the combined dataset revealed no pattern related to BMI (see Supplementary Data ; Table 2 ). Instead, data clustered strongly by study, with no consistent BMI pattern after controlling for study, consistent with prior work (Walters et al., 2014). These findings suggest that per-study effects are much larger than biological effects differentiating lean from obese people. Differences between studies are typically caused by technical and/or clinical factors. This finding emphasizes the need to control for factors inducing heterogeneity among studies because a taxonomic signature will need to be consistent across populations in order to be useful in a clinical setting. Even though inferring interactions at this taxonomic level is challenging due to the significant levels of heterogeneity among phyla in terms of metabolic and other features of microorganisms, the findings presented here demonstrate its utility. While BEEM-Static was chosen for its ability to infer ecological interactions from cross-sectional data, we acknowledge that alternative methods such as SPIEC-EASI, CoNet, and machine learning classifiers (e.g., random forest) could offer complementary insights. These models differ in their assumptions, with some focusing on co-occurrence networks and others on predictive power. In contrast, BEEM-Static provides an ecological modeling framework grounded in generalized Lotka-Volterra dynamics, enabling estimation of both interaction strengths and carrying capacities. Future research should consider benchmarking BEEM-Static against these approaches on harmonized datasets to assess performance and interpretability.

In our research, we did data pruning in order to keep only taxa that appeared in at least one-third of samples, it is important to mention that the low-abundant species also deserve attention. The precise investigation of low-abundant species is a difficult task since low-abundance bacterial species’ exact quantities are notoriously difficult to determine and use to draw conclusions about underlying mutualistic relationships (Jousset et al., 2017; dos Santos, 2012).