Overview of microSLAM

MicroSLAM is a statistical framework designed to perform MWAS using microbiome population structure-aware generalized linear mixed effects models. It was inspired by the application of mixed models in human genetics and was specifically developed to address the unique characteristics of microbiome data. By modeling both strain-level and gene-level effects, microSLAM ensures that microbial contributions to host traits are accurately identified while mitigating confounding effects of microbial population structure on gene-trait associations. Relative abundance testing can also be incorporated, allowing users to test for its association with the trait while adjusting for it in the strain-level and gene-level tests. The examples in this study focus on the human gut microbiome and IBD case/control cohorts, but microSLAM can be applied to any trait or environment. The microSLAM method is implemented as an open-source R package at: https://github.com/pollardlab/microSLAM.

Adapting mixed modeling to microbiome genotypes

microSLAM is designed to analyze microbial gene presence/absence data (0/1), which differs from single nucleotide polymorphism (SNP) data from diploid organisms, typically represented as the number of copies of the derived allele (0/1/2). Gene presence/absence genotypes require a distinct approach to modeling genetic relatedness and performing association tests. To accomplish this, we extended the SAIGE mixed modeling approach [29] to (i) perform separate modeling and testing per microbiome species, (ii) compute species-specific genetic relatedness matrices (GRMs) using gene presence/absence with similarity metrics appropriate to binary data (e.g., pairwise Manhattan distance), and (iii) test each gene in each species’ pangenome for evidence that gene presence associates with a host trait using the same score-based approach as SAIGE. MicroSLAM modeling is performed downstream of running a pangenome profiling tool, such as PanPhlAn 3 [21], Roary [12], or MIDAS v3 [11], to call gene presence/absence across a set of metagenomic samples. By treating gene presence as the genotype, microSLAM is distinct from other extensions of SAIGE that aggregate SNPs within regions or genes [46,47].

microSLAM framework

microSLAM consists of three major components that are run for each microbial species and a host trait (Fig 1A, Methods): (1) construction of the GRM between samples; (2) estimation of variance components with a permutation test to determine whether the strain-level population structure is associated with the trait (τ test); (3) for each gene in the species’ pangenome, estimation of the association between the trait gene presence/absence given the population structure and other covariates with a score test for statistical significance (β test).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. MicroSLAM motivation and approach.

A) Flow chart of microSLAM modeling approach (diagram created in BioRender). B – G) Two bacterial species with different population structures. First row: High population structure species Phocaeicola dorei (260 IBD cases, 218 controls); Second row: Low population structure species Blautia massiliensis (73 IBD cases, 44 controls). B&E) Heatmap of gene-by-gene correlation matrix based on gene presence/absence across IBD samples, using the top 1000 most variable genes. Red: high positive correlation, Blue: high negative correlation. C&F) Heatmap of sample-by-sample GRM (1 minus Manhattan distance of gene presence/absence profiles). Dark green: high similarity, White: low similarity. D&G) Q-Q plot for p-values from tests of association between case/control status and presence/absence of individual genes in the pangenome. Tests are based on micoSLAM and standard logistic regression that does not adjust for population structure (glm). The diagonal line shows expected p-values under the null hypothesis of no association. Pangenome profiling for the metagenomes was done using MIDAS v3 [11].

https://doi.org/10.1371/journal.pcbi.1012277.g001

Step 1: Constructing the GRM: A species-level n-times-n GRM is computed using Manhattan-distance based on gene presence/absence data for n metagenomes.

Step 2: τ test (variance component estimation): The variance component τ represents the strain-level population structure that is associated with the host trait. microSLAM extends SAIGE’s generalized linear mixed modeling approach to adapt to gene presence/absence data rather than SNP genotypes. For a binary trait the model in step 1 is:

where:

: trait value for sample (e.g., disease/control)

: covariate matrix (e.g., age, sex, diet, species abundance)

: fixed-effect coefficients for the covariates in

: random effect for sample capturing strain-level population structure

: species-specific GRM

: residual variance

We then employ AI-REML (Average Information Restricted Maximum Likelihood) and PQL (Penalized Quasi-Likelihood) to estimate variance components and fit the generalized linear mixed model, following the same approach as SAIGE [29]. Instead of using a likelihood ratio test to assess the significance of population structure effects, microSLAM applies a non-parametric permutation test [48] to compute the p-value for , providing a more robust evaluation of how significantly within-species population structure is associated with the trait.

A large value and small p-value indicate that cases are enriched for particular strains or lineages compared to controls. Identifying strain-trait associations is important because this improves the precision of research and therapeutics development based on cultured strains beyond simply picking a random strain from a trait-associated species, which may in fact not be one of the strains driving the species-level association. The genes jointly defining trait-associated strains (i.e., those positively or negatively associated with ) also provide signatures that can be used for predictive modeling and potential diagnostics.

Step 3: β test (gene-trait association test): The β-test evaluates how significantly a specific gene’s presence/absence is associated with the trait while controlling for strain-level population structure and covariates. After fitting the model in the test once per species, it can be used as a null model to test if adding a gene’s presence/absence significantly increases the likelihood. For a binary trait, the model for gene g is:

where variables are defined as above with the addition of:

: the presence/absence of the gene g in sample i

: effect size (log odds ratio for the trait given gene presence/absence)

The null hypothesis H₀: β = 0 (no gene-trait association) is tested using a score test. To mitigate inflated significance due to unbalanced case-control ratios, microSLAM employs saddlepoint approximation (SPA), following SAIGE’s method to improve p-value accuracy [29].

For a quantitative trait, the logit link function in Step 2 and Step 3 is replaced with the identity link function to fit linear mixed effects models.

Population structure in inflammatory bowel disease gut microbiomes

We compiled 710 publicly available gut metagenomes from five inflammatory bowel disease (IBD) case/control studies (S1 Table) and performed pangenome profiling of them using MIDAS v3. There were 71 species with sufficient sequencing coverage to analyze within species genetic variation. After dropping gene families that are nearly always present or nearly always absent (Methods), we had an average of 2,254 gene families per species and a total of ~160 thousand across species. In some species, such as Phocaeicola dorei (P. dorei), many gene families are co-evolving and show a high correlation in their presence/absence across hosts (Fig 1B). In turn, we see two distinct subgroups of strains in the GRM (Fig 1C). This high level of structure might be the result of selection pressures, drift, or a recent population expansion. When we perform MWAS for all P. dorei gene families using logistic regression (glm), we observe that most genes are significantly associated with IBD case/control status (Fig 1D). This inflation is similar to the well-known problem in human genetics in which ancestry-associated variants are all highly significant when genetic ancestry differs between cases and controls [49]. In contrast, the gene-level test in microSLAM does not show inflation, because our model adjusts for population structure when testing individual gene families for disease associations. We therefore hypothesize that inflation is a consequence of high population structure resulting from a high correlation between gene families. Supporting this, Blautia massilensis does not have many genes that are correlated (Fig 1E) and shows less structure in its GRM (Fig 1F). Accordingly, the glm p-values do not show inflation, and the microSLAM output is very similar to that of glm.

These results suggest that if we wish to identify individual gene families with unexpectedly high associations with a host trait given the species’ GRM, the mixed modeling approach in microSLAM provides a way to adjust for population structure across hosts, just as mixed models have enabled human geneticists to account for confounding from genetic ancestry. However, population structure is not necessarily a confounder in microbiome research, and it may also be of interest to identify trait-associated strains, defined by the presence and absence of many gene families relative to other strains. These genes would not be significant in the microSLAM β test, because they are highly correlated with population structure. For this reason, microSLAM also includes a strain-level test, the τ test.

Consider, for example, Ruminococcus B gnavus (R. gnavus), a species that has long been associated with IBD [18,50,51]. The R. gnavus GRM shows two distinct groups when hosts are sorted based on their values (Fig 2A). One of these groups only contains IBD individuals, while the other is split between IBD and controls. The b_i estimates better separate cases and controls than do the first two principal coordinates of the gene presence/absence matrix (Fig 2B). Not surprisingly, when we apply the microSLAM test to R. gnavus, we obtain a large and statistically significant measure of association (τ = 3.92, permutation p-value = 0.0001; Fig 2C), and the resulting model can classify IBD cases with high accuracy (ROC AUC = 0.969; Fig 2D). Now, if we look at the genes that are most highly correlated to the estimated values of the samples (correlation to 0.5), we identify 186 out of the ~ 1200 non-core, non-rare genes used in the analysis that are all nearly equally associated with IBD (Fig 2E). None of these genes are significantly associated with IBD after adjusting for population structure. Thus, we were able to identify several hundreds highly correlated genes that form a predictive signature for the R. gnavus strains present in IBD patients versus controls. These observations illustrate the importance of including the τ test in microSLAM.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. MicroSLAM detects both strain and gene associations.

A) The GRM for Ruminococcus B gnavus with hosts sorted by their estimated b values and annotated by their disease status. B) PCoA from the R. gnavus gene presence/absence colored by host disease status (as in A). C) Histogram of permutation test statistics (-values) from the test for R. gnavus. The line denotes the observed value of t. D) ROC plot for the microSLAM test model for R. gnavus. The statistic τ quantifies population structure. E) Gene presence/absence plot for a subset of genes associated with the random effect b for R. gnavus. Samples are ordered by b and annotated by their disease status.

https://doi.org/10.1371/journal.pcbi.1012277.g002

MicroSLAM controls false positive rates and increases specificity in simulations

The examples in Figs 1 and 2 suggest that microSLAM’s τ test can detect strain-trait associations when species have a high degree of population structure and its β test may control false positive gene-disease associations better than a standard glm, albeit somewhat conservatively. But the ground truth is unknown in real data. Hence, we designed a series of simulations to assess the performance of both of microSLAM’s tests. Our simulation strategy leveraged the IBD compendium in order to capture the range of patterns observed in real data, while varying parameters such as effect size and sample size. For the β test, we compared microSLAM to glm in order to evaluate the effects of adjusting for population structure via the random effects b_i.

First, we evaluated the τ test. To the best of our knowledge, there is no other method that performs this type of association test, so we did not compare microSLAM to alternative approaches. To quantify type 1 error (i.e., false positive rate), we simulated gene presence/absence matrices with core and accessory genes, as well as a set of strain-specific genes, and for each host a binary trait was simulated independently of the gene presence/absence matrix so that the GRM is not associated with the trait (τ test simulation simulation 1; Methods). We investigated a sample size of 100 hosts, which is on the low end of what we observed for species in the IBD compendium, and repeated the simulation 1000 times, keeping track of how many iterations had a permutation p-value < 0.05. We observed a false positive rate of 0.054, which is very close to the expected value of 0.05. This indicates that the false positive rate of microSLAM’s τ test is approximately correct.

To evaluate the power of the τ test, we modified the prior simulation so that the trait depends on presence/absence of a particular strain (τ test simulation 2; Methods). We varied the strength of the strain-trait association (odds ratio) and explored sample sizes ranging from 60 to 250. As expected, power increases with the odds ratio and sample size (Fig 3B). MicroSLAM achieves ~80% power at an odds ratio of 1.5 with 250 samples, whereas an odds ratio greater than 2.0 is needed for similar power with only 60 samples. These results provide practical guidelines for the expected performance of the τ test.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Simulations show that microSLAM improves power and false positive rates.

A) The false positive rates of the τ test of microSLAM were estimated using simulations with varying GRMs but no trait associations. We simulated gene presence/absence and GRMs for the 1000 iterations (τ test simulation 1; Methods). A histogram of p-values for the τ tests shows that the percentage of tests with a p-value < 0.05 is 5.4%. B) Power of the τ test for simulations with a range of values for the odds ratio of the simulated y compared to presence of the trait-associated strain (τ test simulation 2; Methods), repeated for different numbers of samples (N). C) False positive rates of the β tests for glm and microSLAM were estimated using simulations with varying levels of population structure (τ) but no trait associations. We simulated gene presence/absence using the GRMs for the 71 species in the IBD compendium (β test simulation 1; Methods). The false positive rate increases with for the glm and is generally above the targeted level (0.05; horizontal line), while it decreases and is generally below 0.05 for microSLAM. D) Power for 3 simulated species with different τ values and numbers of samples (N). For a subset of genes, presence/absence is simulated based on the trait using a range of odds ratios; other genes have presence probabilities that do not depend on the trait (β test simulation 2; Methods).

https://doi.org/10.1371/journal.pcbi.1012277.g003

Next, we investigated the type 1 error and power of microSLAM’s β test compared to glm, which is the standard approach used in the literature. We considered the case where gene presence/absence is not associated with a binary trait, but it is associated with population structure, and hence genes are correlated with each other. To do so, we simulated gene presence/absence using principal components of the observed GRM for each of the 71 species in the IBD compendium (β test simulation simulation 1; Methods). MicroSLAM controlled the false positive rate below 0.05 for all but two species where it is exactly 0.05 (Dorea A longicatena, Roseburia sp900552665). In contrast, the glm without a random effect adjusting for population structure failed to do so for all but 9 species (Faecalibacterium prausnitzii, Bifidobacterium adolescentis, Bariatricus comes, Blautia A faecis, Faecalibacillus intestinalis, Gemmiger qucibialis, Akkermansia muciniphila, Roseburia sp900552665, Acetatifactor sp900066565), a failure rate of 87.3% (62/71). In addition, as the estimated τ increased, the false positive rate of the glm dramatically increased while the false positive rate for microSLAM decreased slightly (Fig 3C).

To explore if this conservative control of the false positive rate affects the power of microSLAM’s β test, we performed simulations where 100 true positive genes are added to the previously stimulated genes, meaning that they have a presence/absence pattern that is associated with the simulated trait (β test simulation 2; Methods). We varied the strength of the association (odds ratio) and evaluated power at an empirical false positive rate of 0.05 (calculated using the non-trait-associated genes). These analyses show that microSLAM consistently has either the same or higher power than the glm at the same false positive rate (Fig 3D), with the difference between methods being most pronounced with a higher number of samples and a high degree of population structure (τ) (S1 Fig). In order to understand exactly what types of genes lead to the higher false positive rate we simulated a data set completely de novo without using the IBD compendium (β test simulation 3; Supplement). This showed that the strain-associated genes tended to lead to an increase in false positives for the glm, while microSLAM was able to differentiate the true positives from genes linked to the strain but not directly associated with the trait (S2 Fig). Thus, microSLAM’s β test has better specificity than does a standard glm.

MicroSLAM reveals IBD associations across 71 gut microbiome species

We next sought to examine associations in our IBD compendium using microSLAM’s population structure and gene tests. First, we performed the standard species-level analysis in which the relative abundance of each gut species (quantified using Kraken2 and Bracken [45–47]; Methods) is tested for association with IBD case/control status using logistic regression, adjusting for host age, which was consistently reported across studies. Medications and other clinical covariates are also important confounders but unfortunately were not provided in the publicly available datasets. For comparability, we used the same 71 species and species-specific number of samples that had sufficient coverage to perform microSLAM within-species association tests. We found that 3/71 species (4%) had significant relative abundance associations (localFDR < 10%; Fig 4A): Lachnospira rogosae_A, Flavonifractor plautii, and Alistipes putridinis. If the full set of 100 species and samples are utilized, the proportion with relative abundance associations rises to 18%.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. MicroSLAM identifies novel IBD associations.

We analyzed all 71 species in our IBD compendium for three types of associations with case/control status: relative abundance (Kraken2 + Bracken: amount of the species predicts disease), population structure (microSLAM τ: strain predicts disease), and gene family (microSLAM β: gene presence/absence predicts disease). A) Venn diagram showing the number of species with significant IBD associations of each type. For genes, we counted the species if it had at least one significant gene family; species varied in the number of hits (S2 Table). All tests are localFDR adjusted for multiple testing. B) Boxplots showing the AUC ROC from τ test models for all 71 species, stratified by bacterial class. C) UHGG species tree for all 71 species, colored by order. The τ value, p-value for τ test, number of significant genes, and number of samples for each species are plotted in the outer rings. D) Volcano plot for β tests with significant genes (localFDR < 0.2) colored by bacterial order. (S3 Table) E) Bar plot of COG categories for the 83 genes with significant β tests.

https://doi.org/10.1371/journal.pcbi.1012277.g004

Focusing on the 71 species subset, we next ran microSLAM using each species’ gene presence/absence matrix and corresponding GRM (quantified using MIDAS v3 [11]; Methods). Because we observed significant variations in species abundance across studies, we used the MIDAS v3 genes flow to compute the median coverage of 15 universal, single-copy genes (“marker depth”) in each sample (S3 Fig). This statistic is correlated with species relative abundance (by Kraken2 + Bracken), while also capturing cohort-to-cohort variation beyond abundance (e.g., sequencing depth, unmeasured confounders). Therefore, we added marker depth as a covariate in microSLAM as a way to indirectly account for cohort-specific batch effects without explicitly controlling for study name, thereby retaining statistical power (one parameter versus four parameters) while still addressing inter-study variation. Age was also included as a covariate. After including these covariates, the random effects () were not strongly associated with study (S3 Fig).

We used microSLAM’s τ test to identify species whose population structure is associated with IBD case/control status. At localFDR < 10%, 56/71 species (69%) were significant (Fig 4A), meaning that cases and controls tend to harbor distinct strains consistently across studies. These included two out of the three species with relative abundance associations, indicating that specific strains of these two species may be responsible for differential abundance between cases and controls and leaving only one species (Alistipes putridinis) with a significant species-wide IBD-abundance association.

In addition to assessing the statistical significance of τ via a permutation test, we also report the area under the curve (AUC) for the receiver operating characteristic (ROC) from the τ test model. This shows how well the population structure component is able to separate the cases from the controls. The AUC is calculated within the same training data, because the random effects b, which are per-host parameters (i.e., on per subject) generated from the GRM, are unknown for new hosts and hence the fitted model does not generalize beyond the training set. Overall, the AUC from the τ tests was quite high; 55 species had AUC over 0.9. Class Clostridia tended to have the highest AUC values and a smaller variance in AUC values compared to Bacteroidia (Fig 4B). In addition to R. gnavus (Fig 2), species with significant τ tests included Agathobacter rectalis (previously found to be related to IBD under certain conditions [48]) and Phocaeicola coprocola (formally Bacteroides coprocola, which has been shown to have a relationship with ulcerative colitis [52]). In both of these species, there were no significant genes with the β test, but with the information from the τ test genes differentiating IBD-associated strains can be identified.

To investigate specific gene families associated with IBD case/control status, above and beyond the genes that define IBD-associated strains, we next applied microSLAM’s β test. Across the 71 species, 53 genes from 20 species showed significant associations after adjusting for population structure (localFDR < 20%, which is the threshold with optimal lift and somewhat more lenient than the 10% threshold used for the other two tests). Of these species with genes that had significant β tests, three did not have significant relative abundance or population structure associations (Fig 4A). Conversely, 38 species only had significant τ tests, and of those 25/38 (66%) are from class Clostridia. These results underscore the unique information captured by each of microSLAM’s tests.

Having analyzed IBD associations at the species, strain and gene level, we integrated these results across the 71 species to look for phylogenetic trends (Fig 4C). Out of the 71 analyzed gut species, 11 (spread across the phyla Firmicutes A, Firmicutes C, Bacteriodota, Actinobacteria, and Verrucomicrobiota) had no significant IBD associations, possibly due to a lower number of samples (N < 100 for ten species). As noted above, of the three species with relative abundance associations, two were detected with one or both of microSLAM’s tests (Fig 4A), suggesting that relative abundance differences are often accompanied by differences in gene content. Looking across the phylogenetic tree, Lactobacillus species tend to have the least IBD-associated population structure (low values of τ), although there is a subclade of two species with higher τ values. On the other hand, Oscillospirales tend to have high values of, and most species in this order do not have any significant genes. Finally, Bacteroidales stands out as the order with the most significant genes (32/53), consistent with species in this order having many mobile and accessory genes [53].

To further explore the functions of genes identified by microSLAM’s β test (Fig 4D), we looked at the functional annotation provided by MIDASDB [11]. As expected, many (47%) of the 53 significant genes had no functional annotation. For example, 25/53 are in the EggNOG COG category “Function Unknown”. The remaining annotated genes were too few in number to perform well-powered enrichment analyses, but we did note several interesting trends (Fig 4E). The most common COG, encompassing 7 genes from 4 species, was “Replication, recombination and repair”. Four genes were annotated as carbohydrate transport and metabolicm, 8 genes were annotated as putative plasmids by geNomad [54], among which 4 “unknown function” by EggNOG. Considering these annotations and the fact that these genes remain significant beyond the overall population structure of their species, we conclude that many of the significant genes identified by the β test are likely associated with mobile genetic elements.

Seven-Gene GRF operon is a structural variant in Faecalibacterium prausnitzii

One significant gene identified by microSLAM was annotated as “subunit D of the fructoselysine/glucoselysine phosphotransferase (PTS) system” by BlastKOALA [55] (S4 Table). It was negatively associated with IBD case status in Faecalibacterium prausnitzii D (UHGG species id 102272), a species whose relative abundance is positively associated with IBD in our compendium. This hit intrigued us, because F. prausnitzii is a well-studied bacteria with roles in short-chain fatty acid metabolism and inflammation [56,57]. Predicting a new molecular mechanism underlying this host-microbe interaction would enable future functional studies (e.g., in gnotobiotic mice) and potentially could be useful for developing diagnostics, dietary interventions, or other therapies.

To explore this gene family, we first compiled 85 high-quality and diverse F. prausnitzii genome assemblies from NCBI (S5 Table) and clustered them into eight clades (Fig 5A; Methods). We observed that seven genes (plus occasionally an eighth gene) were consistently found together, with a conserved order and orientation across 53% of the NCBI F. prausnitzii genomes (49/85) (S4 Fig). Annotations suggest that these genes encode a fructoselysine/glucoselysine PTS system operon. Having established that this operon is variably present across distantly related F. prausnitzii strains, we expanded our search to include all high-quality F. prausnitzii genomes available in the United Human Gut Microbiome Database (UHGG v2) [1]. This analysis indicated that the complete seven-gene operon is present in all nine F. prausnitzii clades, with between ~3% and ~24% of genomes per clade containing the operon (Table 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. A seven-gene operon present in all nine F. prausnitzii clades in UHGG v2.

https://doi.org/10.1371/journal.pcbi.1012277.t001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Investigation of F. prausnitzii fructoselysine PTS system operon.

A) 52 representative genomes selected from NCBI and colored by the dRep secondary cluster (Selection of Faecalibacterium prausnitzii genomes, Methods). B) Comparison of S. Typhimurium operon to operon in F. prausnitzii D. C) Graphic of the F. prausnitzii fructoselysine PTS system operon and its products (made in BioRender). D) P-values for F. prausnitzii fructoselysine PTS system operon genes in microSLAM β tests across the four F. prausnitzii species defined by UHGG. The flanking genes are much less significant than the genes within the operon. Subunit D (most significant gene in microSLAM analysis) is located at 0, and all other indices are relative to this gene.

https://doi.org/10.1371/journal.pcbi.1012277.g005

Since genes can be syntenic without being functionally related, we conducted further analysis to determine the relationship between the seven-gene operon in F. prausnitzii and the well-characterized GFR operon in Salmonella Typhimurium 14028s [58]. We successfully mapped five genes from the F. prausnitzii operon to the corresponding S. Typhimurium operon (gfrABCDF) (Fig 5B). Notably, the gfrE gene, encoding a deglycase that cleaves glucoselysine 6-phosphate, is absent in F. prausnitzii. In addition, the operon in F. prausnitzii includes a gene without homology to any gene in the GFR S. Typhimurium operon. The regulatory genes, which are located at the start of the operons, also differ between the two species. We hypothesize the seven-gene F. prausnitzii operon identified in our analysis functions solely as a fructoselysine PTS system (Fig 5C).

Fructoselysine is a spontaneous product of Amadori rearrangements, and its presence in the human gut environment can promote the growth of bacteria capable of importing and using this carbohydrate as an energy source [59,60]. In F. prausnitzii, this is performed by a seven-gene operon encoding proteins that phosphorylate the substrate while transporting it across the bacterial cell membrane, making it available as a source of carbon. While only subunit D of this operon was significant in microSLAM’s gene test after accounting for multiple comparisons, all of the genes in the operon had unadjusted p-values less than 0.05 and flanking genes in the F. prausnitzii reference genome did not (Fig 5D). Variability in gene detection from shotgun metagenomics data is a likely source of the difference in significance for subunit D versus the other genes. MicroSLAM analysis for three other F. prausnitzii species in the IBD compendium did not yield any significant genes, primarily due to inadequate sample sizes, which restricts the statistical power of microSLAM’s β test. Nonetheless, the gene presence/absence matrices for these species are consistent with the operon being variably present and depleted in IBD cases.

Altogether, these results suggest that the genes in this F. prausnitzii PTS operon are co-evolving in terms of their presence/absence across hosts, potentially independently of neighboring genes, and that the presence of this operon is more common in healthy hosts. Our data also suggest that the fructoselysine PTS system operon could be a mobile genetic element. Supporting this possibility, many NCBI and UHGG contigs carrying this operon are predicted by geNomad [54] to be plasmids (Table 1). We also observe sequences associated with mobile elements and horizontal gene transfer (HGT) in the genomic context surrounding the operon. These are computational predictions only, and no plasmids have been previously reported in F. prausnitzii. We therefore checked for the operon in the other 70 species in our microSLAM analysis, detecting it in strains from two other phyla: Gemmiger qucibialis (species_id 103937) and Faecalibacterium sp90053945 (species_id 103899). It is also known to be present in Escherichia coli, Bacillus subtilis and Agrobacterium tumefaciens Ti plasmid [60]. While not conclusive, these data are consistent with HGT. Regardless of the mechanism of acquisition or loss, the variable presence/absence of the operon across F. prausnitzii strains indicates that certain lineages of this species can acquire and utilize fructoselysine, thereby enhancing their adaptability and competitiveness in the dynamic gut ecosystem relative to strains without the operon.

Compendium of IBD/healthy case/control metagenomic studies

We compiled a total of 2625 publicly available paired-end shotgun metagenomic samples, sourced from five studies related to either inflammatory bowel disease (IBD) or the Human Microbiome Project (HMP2) and having an average read count greater than 20 million (accession numbers: PRJNA400072, PRJNA398089, PRJEB15371, PRJEB5224 and PRJEB1220). A stringent sample selection process was implemented to ensure (1) all samples included comprehensive metadata, such as disease status, age, and antibiotic usage; and (2) only one sample was selected per subject, considering that multiple time points could have been sequenced from the same subject. Specifically, for multiple time point samples from HMP2, we adopted the same selection criterion used by Lloyd-Price et al [68], selecting week 20 or greater for all subjects, the maximum read count for healthy subjects and the time point with the highest dysbiosis score for IBD patients. The first time point was chosen for the MetaHit project [1,69,70]. Ultimately, a total of 710 samples met these criteria and were included for downstream analysis (S1 and S6 Tables).

Bioinformatics analysis

Preprocessing of the downloaded metagenomic sequencing libraries was performed using a QC pipeline that includes the following steps: (1) Adapter removal and quality trimming: adapter sequences were removed and low-quality reads were trimmed using Trimmomatic (v 0.39) [71]. (2) Human contamination removal: reads were aligned against the complete human reference genome (CHM13v2.0) [72] and a collection of 2250 genomes known to be contaminated by human sequences [73], using Bowtie2 (v 2.5.1) [74] to identify and remove human contamination. (3) Low-complexity read filtering: low-complexity reads were filtered out using BBduk (v 37.62) [75]. This step involved removing reads with an average entropy less than 0.5, with entropy k-mer length of 5 and a sliding window of 50 (parameters: entropy = 0.5, entropywindow = 0.5, entropyk = 4). Additionally, reads shorter than 50 base pairs (bp) post-filtering were removed (parameters: minlen = 50). (4) Quality reporting: a quality report of the cleaned-up reads was generated with FastQC (v 0.12.1) [76]. After preprocessing, samples with read counts lower than 1 million were removed, resulting in 710 high-quality samples retained for further analysis.

Pangenome profiling using MIDAS v3

To determine which species within the 710 samples are both prevalent and sufficiently abundant for pangenome profiling, we implemented a two-step analysis using MIDAS v3 [11]. In the first step, we quickly scanned each sample to detect the presence of species by assessing the vertical coverage of 15 universal single-copy marker genes across 3956 distinguishable species in the UHGG v2 database [1,77]. In the second step, we adopted a whole-genome read alignment-based methodology to quantify the abundance of each species. This involved running MIDAS’s single-nucleotide variant (SNV) pipeline for species that meet specific criteria from the first step: a median marker coverage of at least 2X and at least 50% of the marker genes uniquely covered. We further classify a species as present based on whole-genome based species abundance estimation, applying the criteria of horizontal coverage > 0.4 and vertical coverage > 5. We further excluded sample-species pairs where the ratio of genome-wide vertical coverage to single-copy marker gene coverage exceeded 4, which helps us to eliminate potential false positives caused by cross-mapping of reads among closely related species and conserved gene families. This stringent criterion also improves computational efficiency [67]. After implementing the aforementioned filtering, 71 species that were present in more than 60 samples and met the abundance criteria were selected for subsequent pangenome profiling analysis. There were 619 samples with at least one species present.

To perform pangenome profiling, we utilized the Genes module from MIDAS v3 [11], which features careful curation of the pangenome database and comprehensive functional annotation. Specifically, a single Bowtie2 index was built for all 71 species, and QC-ed paired-end reads for each sample were aligned to this index. Our analysis included only genes covered by at least 4 reads (--read_depth 4). Genes with an estimated copy number greater than 0.4 were classified as present (--min_copy 0.4). This threshold was selected based on exploratory data analysis and simulations previously performed by our lab. The resulting sample-by-gene presence/absence binary matrix was then used for subsequent association analysis with microSLAM, excluding core genes (absent in less than 10 samples) and rare genes (present in less than 30 samples).

microSLAM runtime checklist

1. 1. High-Quality Metagenomic Data

Because proper QC is critical for reliable read-mapping based pangenome profiling, we performed bioinformatics quality control to the raw metagenomic sequencing data, including removing low-quality reads, filtering out low-complexity reads, and removing host reads by aligning to an updated version of the reference genome (e.g., CHM13v2.0). We also removed samples with low post-processed read count (< 1 M reads). Users may follow our QC procedure (details in the Bioinformatics section under the Methods section) or use their preferred QC pipeline.

1. Pangenome Profiling using MIDAS v3.

The following steps outline how to produce the sample-by-gene presence/absence matrix from raw metagenomic shotgun reads using MIDAS v3. However, other genotyping pipelines may also be used as alternatives, such as PanPhlAn 3 [21] and Roary [12]. The required input for microSLAM consists of two tables: (1) a metadata table specifying traits and covariates for each sample, and (2) a sample-by-gene presence/absence matrix that includes accessory genes with sufficient variation across samples, as well as samples with sufficient abundance for the given species.

Step 1: Sample Selection via Species Profiling

Purpose: To identify abundant species suitable for downstream ogene profiling.

Use run_species from MIDAS v3 to estimate species abundance using 15 universal single-copy marker genes. Retain species/sample pairs that meet these criteria: median_marker_depth > 2X and unique_fraction_covered > 0.5.

In this study, we also ran the SNPs workflow of MIDAS v3, which introduced an additional sample selection filter requiring whole-genome vertical coverage ≥ 5X and horizontal coverage ≥ 0.4. While this enhances data quality, it also increases computational demands, making it an optional step for users based on their available resources.

Step 2: Pangenome Profiling

Purpose: To quantify gene coverage.

Use run_genes from MIDAS v3 to profile the abundance of genes on a species level with the following parameters: --read_depth 0.4 --min_copy 0.4. For non-gut metagenomes, use the database `--midasdb_name gtdb`. We included species present in at least 60 samples to ensure statistical power.

Step 3: Prepare Sample-by-Gene Presence/Absence Matrix

Purpose: To convert gene coverage to presence/absence.

Run merge_genes.py from MIDAS v3 with minimal filtering (--genome_depth 0.01, --sample_counts 60). Define genes as present in a sample using the following criteria: minimum read depth of 4 (--read_depth 4); copy number threshold ≥ 0.4 (--min_copy 0.4). Filter genes with too little variation: exclude core genes (absent in less than 10 samples) or rare genes (present in <30 samples). Functional annotations for all the genes can be accessed through MIDASDB v3.

Step 4. Covariates & Trait Data

Purpose: To prepare sample-level covariates and trait data.

To control for confounding, we recommend including appropriate covariates, such as: age, species abundance. For binary traits, be cautious with severely unbalanced classes to minimize bias. For quantitative traits, confirm that normalization (e.g., inverse normal transformation) produces a near-Gaussian distribution.

Step 5. Format input data for microSLAM

Purpose: To prepare data for input to microSLAM.

The first input is a binary sample-by-gene presence/absence matrix (), where represents samples (rows), and represents gene families (columns). Each entry is either 0 (absence) or 1 (presence) for a gene family in a given sample. The second input is a phenotype metadata file with rows. The first column must be a sample identifier (sample_name). This is followed by the trait data and covariates. Consistency check: Ensure that sample names match across both input files.

Step 6. Running microSLAM

Purpose: To perform MWAS using the microSLAM R package.

Use the wrapper function run_microSLAM() to execute the full pipeline for a given species: First, compute the genetic relatedness matrix (GRM); Then perform the τ test once per species to evaluate whether population structure (if present) is associated with the traits of interest. This step identifies strain-level associations. Finally, run the β test for each gene to assess whether individual genes are associated with the trait while adjusting for the GRM. For step-by-step analysis and visualization, refer to the stepwise code provided in the microSLAM GitHub repository (https://github.com/pollardlab/microSLAM).

Statistical model for strain-trait and gene-trait associations while accounting for population structure in metagenomes

We present microSLAM, a 3-step modeling procedure for detecting within-species genetic variation associated with host biology (Fig 1A). The inputs to microSLAM, for a given species, are a p x n binary matrix of gene family presence/absence values for n host samples and p gene families, a 1 x n vector of trait values for each sample (binary or quantitative), and optionally a q x n matrix X of data for q covariates. The outputs are a measure of population structure (τ) with a permutation p-value and, for each gene family, a coefficient (β) measuring the gene’s association with the trait (e.g., log odds ratio for binary traits and logistic regression) with its local false discovery rate (localFDR) adjusted p-value [78]. Results from different species can be interpreted jointly to identify shared trends in trait-associations, such as enriched pathways (see below).

MicroSLAM fits generalized linear mixed effects models that account for the genetic relatedness of strains of a given species across hosts. In Step 1, an n x n sample genetic relatedness matrix (GRM) is computed from the gene presence/absence matrix. To do so, we create an n x n Manhattan distance matrix and then transform this into relatedness using 1-distance. The GRM is used in Step 2 to test if the species’ population structure is associated with the trait, which would indicate that hosts with similar trait values tend to have similar strains. For example, for a case/control study, this step aims to detect species where a subset of related strains confers risk. We call Step 2 the τ test, because population structure is modeled using a parameter τ. In Step 3, random effects estimated from the GRM are used to adjust for population structure in a model that is used to test gene families for associations with the trait beyond simply being present in trait-associated strains. We call Step 3 the β test, because a parameter denoted β is used to quantify gene-trait associations.

In Step 2 (τ test), microSLAM fits a generalized linear mixed model. The trait is modeled as a function of any covariates X (with coefficients α) and random effects that are estimated from the GRM. The link function f() is the identity function for normalized quantitative traits (linear regression) or the logit function for binary traits (logistic regression):

(1)

One key component of fitting Model 1 is estimating τ, the variance on the random effects, which depends on the association of the trait to the GRM. This is done iteratively using the average information restricted maximum likelihood (AI-REML) algorithm from the GMMAT [28] method. From this, we obtain a point estimate of τ, a point estimate of the random effects , and a statistic, , that measures how associated the species’ population structure is with the trait. This T statistic is derived from [48] and computed using the linear setup from [79]. To assess the statistical significance of T, we randomly permute the trait values B times (e.g., B = 1000), repeat model fitting, compute a T statistic for each permutation, and use these as an empirical null distribution to estimate a p-value based on how many of the permuted T statistics exceed the observed T statistic. Species with a significant T statistic have population structure that associates with the trait.

In Step 3 (β test), microSLAM fits a second model using the random effects (b) estimated in Step 2 and the presence/absence vector for each gene family, denoted g (with coefficients ):

(2)

Model 2 is fit separately for each gene family within each species. measures the gene’s association with the trait given the species’ population structure and the covariates. To test the null hypothesis that a given gene’s , we follow the strategy employed in SAIGE [29], which uses a score statistic and does not re-estimate or b. By only fitting Model 1 once per species, this approach is computationally efficient. Microbiome case/control studies are often unbalanced, for example, when a bacterial species is detected in many more controls than cases. To obtain accurate p-values in this scenario, we approximate the score statistics for testing the null hypothesis that is zero using the Saddle Point Approximation (SPA) of the true distribution, as implemented in SAIGE.

To adjust the resulting p-values for multiple testing, we use localFDR [78], which accounts for the high correlation between gene families (i.e., when genes co-occur across strains) that invalidates methods such as Benjamini-Hochberg FDR [80] or Storey’s q-value [81]. We transform SPA p-values into Z-values by dividing by two, multiplying times the sign of the estimated coefficient, and converting the resulting numbers to quantiles. Then, localFDR uses maximum likelihood estimation to approximate the null Z-value distribution and identify Z statistics that deviate from this distribution. We implement this using the locfdr v1.1-8 package in R, fitting the null distribution to the Z-values between the 10th and 90th percentiles across all species (S5 Fig).

Generalized linear model

A standard generalized linear model (glm) [82] was fit for all genes for all species that were analyzed with microSLAM. This is a logistic regression model using glm in R with case/control status as the outcome, gene presence/absence as a predictor, and age as a covariate.

τ test simulations

We performed simulations to assess the false positive rate and power of microSLAM’s τ test. To assess the false positive rate (simulation 1), we set up a simulation where a binary trait was generated independently of GRMs. For each of 1000 iterations and n = 100 samples, the trait y was simulated using a binomial distribution with a success probability of 0.5, and a covariate was simulated with a normal distribution centered at 45 with a standard deviation of 15, similar to the age distribution in our IBD compendium. Next, for each iteration, a gene presence/absence matrix was simulated with p = 1000 genes. This included 400 “core” genes simulated from a binomial distribution with a success probability of 0.8, 400 “accessory” genes simulated from a binomial distribution with a success probability of 0.2, and 200 genes simulated based on presence of a strain unrealted to the trait y, as follows. The strain’s presence/absene across samples was simulated using a binomial distribution with a success probability of 0.5, and then presence/absence for each of the 200 genes was set to absent if the strain was absent and simulated from a binomial distribution if the strain was present, where the success probabilities were chosen such that the average odds ratio of a given gene being present if the strain is present is 1.8. After the genes are simulated, the GRM is calculated and the population structure test is run with 100 permutations. The p-value is calculated for each iteration as the number of permutations with a more extreme T statistic than the observed T statistic. The false positive rate is calculated as the number of iterations with a p-value <0.05.

The power test (simulation 2) is carried out in a similar fashion, with two key changes. First, we explored a range of sample sizes (n = 60, 100, 250) to assess the relationship between sample size and power. Second, we simulated the trait y based on presence/absence of the simulated strain. Specifically, we explored a range of effect sizes, quantified with an odds ratio ranging from 1.0 to 2.5. For a given odds ratio, we set and then generated the trait using a binomial distribution with success probability equal to . This creates a stronger relationship between the trait and presence of the strain as the odds ratio increases. For each odds ratio and sample size, 125 iterations were run and power was calculated as the proportion of iterations with a significant τ test divided by 125.

β test simulations

We performed simulations to assess the false positive rate and power of microSLAM’s β test versus a standard glm. To assess false positives (simulation 1), we generated data in which no genes were associated with the trait, so that all genes are false positives. We computed a p-value for each gene using each modeling approach and tracked the proportion of genes with p < 0.05. In order to simulate real population structure, while introducing some random variation, we used the observed GRM for each of the 71 species in our metagenomic compendium to help generate simulated gene presence/absence matrices. Specifically, we first decomposed the observed GRM for each species into its first 10 principal components (PCs). We then standardized each PC by dividing each value by the PC’s standard deviation and computed the standard normal probability for each sample’s loading on each standardized PC (one per sample i per PC dimension j): = . The probabilities retain relationships between samples across the 10 dimensions. For each of the 10 PCs, we simulated the presence/absence of 90 genes using a binomial distribution with a success probability equal to the sample’s for that PC, for a total of 900 genes correlated with one dimension of the population structure. We also simulated 100 uncorrelated genes using a binomial distribution with a success probability chosen from a uniform distribution between 0.2 and 0.8. From the resulting 1000 x n gene presence/absence matrix, we simulated a binary trait (y) using the first two PCs (PC1 and PC2), as follows. We set y equal to one in a given sample if its loadings on PC1 and PC2 had opposite signs (either PC1 > 0 and PC2 < 0 or PC1 < 0 and PC2 > 0). This created a nonlinear relationship between y and 180 of the simulated genes (S6 Fig).

For each simulated y and gene presence/absence matrix, we ran microSLAM to compute a GRM and estimate population structure (τ). These new τ values were different from the species’ observed τ values and greatly varied across species, as in the observed data (S7 Fig). After the GRM was calculated, and the τ test was run, both glm and microSLAM’s β test were run to test for gene-trait associations. The false positive rate was determined by summing the number of genes with p-values < 0.05 divided by the total number of genes, excluding genes simulated from PC1 or PC2 (p = 820 genes).

To assess power for the β test (simulation 2), we start with the data from simulation 1 and set for an odds ratio of . Then, we generated presence/absence for 100 additional genes from a binomial distribution with success probability . These genes, which we denote as G_y, are positives (associated with y) and all other genes are negatives (independent of y). At (i.e., an odds ratio of one), the generated genes G_y will not be associated with the trait. At , the average odds ratio will be 1.2. We investigated values between 0.52 and 0.78. We checked, and the odds ratios across simulations with the same value did not deviate more than 0.1 from the expected values.

We ran glm and microSLAM on each simulated dataset. As expected, the population structure test yielded estimated τ values that increase notably with the simulated odds ratio (i.e., as the association between y and the genes G_y increases). In order to assess power, we used the negative genes to establish a significance threshold for each modeling approach such that the empirical false positive rate would be no more than 0.05. Applying these thresholds to the positive genes G_y, we computed power as the proportion of positive genes detected. Power was compared between glm and microSLAM across odds ratios and species (each with different sample sizes and GRMs).

Relative abundance test

We calculated the relative abundance of each species by downloading the UHGG v2 Kraken database from MGnify [22] and running Kraken2 [43,45] with options --paired --minimum-hit-groups 3 and then Bracken [44,45] with options -l S -t 1000. We computed relative abundance as a given species’ bracken coverage divided by the total coverage, and we removed species with less than 0.05% relative abundance. We then performed logistic regression, using the case/control label as the dependent variable (y) and relative abundance as the independent variable, with age as a covariate. The estimated log odds ratios and p-values from these logistic regression analyses were compared with outputs from the microSLAM τ test, after using localFDR to adjust p-values at a 0.1 level.

Identification of seven-gene GFR operon in F. prausnitzii

The gene grfD from F. prausnitzii (UHGG species id 102272) reported as significant by microSLAM’s β test corresponds to the EIID component and is part of a putative GFR operon that encodes the Fructoselysine PTS system. A similar gene in S. Typhimurium 14028s has been identified as responsible for the utilization of fructoselysine [58,83]. To determine whether this EIID gene is part of a gene cluster that forms an operon - that is, genes that are sequentially arranged on the chromosome and co-regulated - we conducted the following analysis. First, we retrieved the neighboring genes upstream and downstream of this EIID gene in the UHGG reference genome for this F. prausnitzii species, considering up to five genes in each direction. We then used blastn [84] to identify homologous regions in three different sets of genomes: (1) 85 NCBI F. prausnitzii genomes, (2) all genes in the nine UHGG F. prausnitzii species clusters including metagenome-assembled-genomes (MAGs), and (3) all MAGs in the 71 species that were investigated in this study. These five genes from F. prausnitzii were also aligned to the corresponding S. Typhimurium operon (gfrABCDF) using Blastp [84]. All genomes were annotated using Prokka [85]. Genes were annotated with BlastKOALA [55] and eggNOG-mapper [86].

Selection of Faecalibacterium prausnitzii genomes

To avoid incorrectly assessing the seven-gene GFR operon as incomplete in an assembly simply due to fragmented contigs, we only selected high-quality F. prausnitzii genomes with assembly levels of scaffold, chromosome, or complete genome, and we specifically excluded atypical genomes. In total, we downloaded 105 genomes of F. prausnitzii from NCBI (using the taxon identifier 853). We assessed the genome quality using CheckM [87], and retained only genomes that met the following criteria: completeness >= 90, contamination <=5, and strain heterogeneity <= 10. After this filtering, 85 F. prausnitzii NCBI genomes were retained for the GFR operon screening analysis (S4 Fig). Next, we used dRep [88] to perform pairwise genome comparisons based on Average Nucleotide Identity (ANI). This dRep analysis involved first clustering all the genomes using the Mash heuristic for ANI [89] and subsequently using MUMMER [90] to compute ANI on sets of genomes that have at least 90% Mash ANI before performing a secondary clustering. As a result, 52 secondary clusters were formed at 98% MUMMER ANI (-comp 90 -con 5 -pa 0.95 -sa 0.98 -nc 0.8). Hierarchical clustering of the 52 representative genomes using average linkage was performed using the pairwise MASH similarity matrix (‘scipy.cluster.hierarchy’ package).

In addition to NCBI Genomes, we also collected all nine F. prausnitzii species clusters from UHGG v2 [1], using the same selection criteria to ensure assembly quality. We calculated the pairwise genome similarity between the resulting 52 F. prausnitzii genomes and the nine representative genomes from UHGG F. prausnitzii species using fastANI [91]. We also compared each NCBI F. prausnitzii genome to the nine representative genomes from UHGG, and each NCBI genome was assigned to the UHGG species cluster with the highest ANI (ANI >= 95%). If this similarity level was not reached, the NCBI genome remained unassigned. Eight F. prausnitzii species in UHGG were represented by the 52 NCBI F. prausnitzii genomes. The information for the 85 genomes is available in S5 Table.