A Guide to Dietary Pattern–Microbiome Data Integration

ABSTRACT

The human gut microbiome is linked to metabolic and cardiovascular disease risk. Dietary modulation of the human gut microbiome offers an attractive pathway to manipulate the microbiome to prevent microbiome-related disease. However, this promise has not been realized. The complex system of diet and microbiome interactions is poorly understood. Integrating observational human diet and microbiome data can help researchers and clinicians untangle the complex systems of interactions that predict how the microbiome will change in response to foods. The use of dietary patterns to assess diet–microbiome relations holds promise to identify interesting associations and result in findings that can directly translate into actionable dietary intake recommendations and eating plans. In this article, we first highlight the complexity inherent in both dietary and microbiome data and introduce the approaches generally used to explore diet and microbiome simultaneously in observational studies. Second, we review the food group and dietary pattern–microbiome literature focusing on dietary complexity—moving beyond nutrients. Our review identified a substantial and growing body of literature that explores links between the microbiome and dietary patterns. However, there was very little standardization of dietary collection and assessment methods across studies. The 54 studies identified in this review used ≥7 different methods to assess diet. Coupled with the variation in final dietary parameters calculated from dietary data (e.g., dietary indices, dietary patterns, food groups, etc.), few studies with shared methods and assessment techniques were available for comparison. Third, we highlight the similarities between dietary and microbiome data structures and present the possibility that multivariate and compositional methods, developed initially for microbiome data, could have utility when applied to dietary data. Finally, we summarize the current state of the art for diet–microbiome data integration and highlight ways dietary data could be paired with microbiome data in future studies to improve the detection of diet–microbiome signals.

Introduction

Dietary modulation of the human gut microbiome is an enticing practical pathway to realize the promise of the microbiome to prevent disease (1). The integration of human diet and gut microbiome data has the potential to reveal new relations between microbes and foods (2). When including dietary data collection in microbiome studies, researchers face numerous decisions. First, researchers must decide between interventional feeding trials or observational study designs. If researchers choose an observational study, they must select between dietary assessment methods, such as the choice between 24-h records or FFQs (3–5), as these are the most commonly used and validated methods for dietary collection. Data integration of dietary patterns with microbiome data adds another layer of complexity after data collection. Rather than focus on study design or tool selection [reviewed previously (3)], this review summarizes the current state of the art for diet–microbiome data integration (6, 7). While much has been written about microbiome analysis methods, there is less information about operationalizing dietary patterns (8) for incorporation into microbiome analysis.

Recently, dietary patterns have risen to the forefront of nutritional analysis in microbiome studies—moving away from conventional nutrient analysis of macro- and micronutrients. As such, in this review we primarily explore the food group and dietary pattern–microbiome literature, focusing on the methods used to incorporate dietary complexity. The use of dietary patterns to assess diet–microbiome relations holds promise to identify interesting associations and to result in findings that can directly translate into actionable dietary intake recommendations. Dietary patterns and food groups associate more strongly with microbiome composition than conventional nutrients (9–11), further reinforcing this approach as viable for exploring diet–microbiome relations. Humans do not consume individual nutrients in isolation; instead, they eat complex diets. Food-based dietary pattern analysis avoids some limitations inherent in nutrient databases in which the representation of food-derived biomolecules is lacking (12). Analysis using dietary patterns also inherently incorporates the concept of food synergy (i.e., that foods behave differently when consumed in combination rather than individually); allowing for complexity and nuance that is lacking from single-nutrient or single-food analysis (13–15).

Overview of Diet and Microbiome Data Structures

Dietary data

It is essential to understand dietary and microbiome data structures to appreciate the complexity involved in diet–microbiome datasets. The 2 most common forms of dietary data in human observational studies are FFQs and 24-h recalls or records. FFQs are validated tools that aim to estimate usual intake over a defined period, typically from 1 to 3 mo to ≥1 y, and 24-h recalls are usually collected by telephone or electronically by asking subjects to recall what they ate in the previous 24 h. Repeated 24-h recalls can be used to estimate habitual dietary intake. In contrast, dietary records are collected by logging or recording foods and drinks in real time when consumed throughout the day on paper or electronically for a defined number of days. For this review, we collectively refer to 24-h recalls and dietary records as records for simplicity because the downstream analyses of these methods are generally similar.

The data structures generated from records and FFQs are generally similar, with some critical naming conventions and format differences depending on the FFQ used or dietary data collection system. For both methods, all reported foods are matched to representative foods in food and nutrient databases to generate food and nutrient reports. The default dietary data outputs from most collection systems are tabular data with summary nutrient information for each subject. FFQ data typically include estimated nutrient intake over the FFQ-defined time period. Data are arranged so that rows are research subjects and columns are FFQ-estimated nutrients such as overall kilocalorie intake, macro- and micronutrient values, and food group serving size estimates. Similarly, summary record data typically consists of tabular data for which rows are records and columns are nutrients. Because multiple records generally are collected for each subject, there are usually summary data for each subject, with 1 row per subject, and longer formats with 1 row per intake day. The 2 most common collection systems for dietary records used in the United States [the Automated Self-Administered 24-hour (ASA24) Dietary Assessment Tool and the Nutrition Data System for Research (NDSR)] also provide data on a per food basis for each subject (16, 17). Therefore, depending on the level of aggregation, nutrient data from records can be summarized for a subject or population or provided with high resolution at the level of individual foods. In some cases, food level data from mixed dishes can be disaggregated to the ingredient level.

Regardless of the method for dietary data collection, there are potential issues for underreporting; all methods are subject to measurement error due to recall bias resulting from memory reliance. Missing items in dietary databases and software could potentially have ramifications specific to microbiome studies. For example, the primary purpose of dietary collection software in the past ∼50 y has been to quantify intake in terms of energy consumed by humans and measure intake of nutrients of concern, like sodium. Subsequently, many nonnutritive food items high in small molecules like polyphenols, such as spices and herbs, are underrepresented in most dietary databases and software systems.

Microbiome data

Microbiome data generated from human stool samples are uniquely complex. DNA is extracted from samples and prepped for either marker gene amplification (i.e., 16S rRNA sequencing) or metagenomic sequencing (i.e., shotgun sequencing). The methods for microbiome analysis are beyond the scope of this review [reviewed (18)]. Briefly, generated sequences are passed through computational pipelines and aligned to representative genome databases to create tabular data of operational taxonomic units, amplicon sequence variants, or taxonomic assignments depending on the program and algorithm employed. Microbiome data are organized in taxonomy tables with rows of bacteria and columns of samples.

Before analysis, microbiome taxonomy tables may be transformed to relative abundances or median normalized counts or transformed with sophisticated mathematical approaches. For example, transformations such as the centered log-ratio transformation may be applied to account for microbiome compositionality. Alternatively, relations between species based on evolutionary distances in terms of phylogenetic trees can be harnessed to transform taxonomy tables such as the phylogenetically informed isometric log-ratio (PhILR) (19). Phylogenetic relations and data structures are also crucial for analyzing overall microbiome composition similarity between subjects. Phylogenetic trees can be used to statistically account for the proximal relations between related microbial species when measuring overall microbiome composition similarity and differences with UniFrac (20).

Shared Features of Diet and Microbiome Data

While there are numerous differences between dietary and microbiome data, inherently these data structures share similarities (Figure 1). Like microbiome data, dietary record data are compositional, contain numerous zero values, are autocorrelated, and have non-normal distributions.

FIGURE 1.

Side-by-side workflow for data processing of diet and gut microbiome data. Dietary data and microbiome data each require multiple steps for data preparation and processing. After processing, the data structures for both diet and microbiome data share similarities in format, lending to the potential for similar statistical analysis and data visualization opportunities. ASV, amplicon sequence variant, MCMC, Markov chain Monte Carlo; OTU, operational taxonomic unit.

Compositionality

Compositional data are parts of a whole, so they are relative to each other and are therefore difficult to compare across samples or subjects (21). First, consider dietary intake. Intake is generally bounded by appetite and energy needs such that a person cannot eat an unlimited amount of food each day. Therefore, dietary data are constrained and have fundamental differences from unconstrained data. Consider dietary intake in terms of food groups; when intake of whole grains increases, then intake of other food groups must decrease to compensate. This concept is important for analysis because researchers must consider changes in specific foods or food groups and how those changes result in substitutions or reduction of intake of other foods, which in turn is reflected in overall diet quality (15, 22). Compositionality likely contributes to the often contradictory findings that are reported in the nutritional epidemiology literature.

Now consider microbiome data. Microbiome data are inherently compositional. Sequences are generated together on a sequencer that is bounded by the total number of sequences it can produce. This means that even when reported as counts, microbiome data are still compositional and should be viewed relative to all other sequences in a sample. When the number of sequences for one type of bacteria increases in a sample, the number of sequences generated for other bacteria in that sample must decrease. One solution to this problem is to adjust for compositionality (23) or only analyze relative changes (24).

Zero inflation

Zero inflation simply means that these data tend to be sparse, having a large number of zeros. This can complicate analysis by generating statistical artifacts and introducing errors. A high level of skewness can harm the model fit and performance, particularly in regression-based models. The tail region can behave as an outlier in statistical models, leading to erroneous results. Consider first dietary data. A common goal with dietary data is to estimate usual intake. When collecting records reported as foods consumed, there is an enormous amount of variation from person to person in terms of the foods that are reported. Within a person, the time period captured by the dietary record is short and conceivably could miss their intake of specific foods that are part of their episodic diet and may be necessary for estimating usual intake. In nutritional epidemiology, complex Markov chain Monte Carlo models have been developed to account for the error due to zero inflation when modeling usual intake from record data (25).

In microbiome data, especially when working with data at the species or strain level, there is high variation between people. Many microbiome species or strains are only present in a few subjects and absent in all others. Zero inflation also shows up in microbiome data when trying to model changes over time within a subject. When species or strains are in low abundance, they may look like they come and go, when really, they are just sitting at or around the level of detection of the sequencing protocol, so modeling their changes can be extremely difficult. Numerous approaches have been developed in the microbiome field to model and manage zero-inflated microbiome data (26, 27).

Autocorrelation

In both fields, data features can have a high amount of autocorrelation, meaning that ≥2 features are highly correlated with each other. Autocorrelation violates the independent observations assumption in traditional statistical analyses. In nutrition, dietary data can be correlated because foods are often consumed together. For example, burger patties are usually consumed with burger buns, lettuce, tomato, ketchup, and mustard. In the microbiome, bacteria can self-organize into guilds that increase and decrease in concert, so multiple species respond as a group to changing environmental pressures (28). The same may also be true for genome interaction groups which, like guilds, change collectively and correlate with disease outcomes independently of taxa (29).

Data distribution

Microbiome data at the feature level (either taxonomy assignments, operational taxonomic units, or amplicon sequence variants) are rarely normally distributed; statistical approaches that do not assume normality are essential for these data types. Microbiome alpha diversity metrics, which measure the diversity within each sample, generally approximate a normal distribution. Beta diversity metrics, which measure similarity between microbiome samples, are unlikely to be normally distributed, so nonparametric approaches are better suited to these analyses.

Summary data for dietary intake, such as macro- and micronutrient intake, may approximate a normal distribution, though some have log-normal distributions. Foods and food groups reported using records and even FFQs are unlikely to follow a normal distribution and are likely to violate assumptions of parametric statistical tests. Numerous data transformations may be necessary to analyze dietary data depending on the approach and the hypothesis tested. Food-level data are likewise unlikely to be normally distributed given the zero inflation expected with these data.

The Microbiome Complicates Concepts of Nutrients and Serving Sizes

Dietary data can take many forms. The level of detail available in the data is somewhat dependent on the assessment tool used. Data collected with food records and FFQs can typically be processed with nutrient databases to provide nutrient information. At the nutrient level, kilocalories and macro- and micronutrients are provided in output files in the units that best correspond with their interpretation relative to dietary guidelines and recommendations ( i.e., recommended dietary allowances, adequate intakes, or daily values) (30). There are also ways to explore dietary data in terms of food weights, such as the gram weight of foods consumed or normalized food weights that account for typically consumed serving sizes or the ratio of nutrients to energy (expressed as nutrient contents per 100 kcal or 1000 kcal). For the analysis of record data, these concepts can have important implications for interpreting, understanding, and communicating results. For example, using the gram weight of foods when studying overall dietary intake in terms of reported foods will unintentionally weight the intake of liquids and beverages more than solid foods and skew results. Using foods normalized for usual serving sizes can minimize this problem. The use of serving sizes can be challenging to implement as serving sizes are not always readily available in all types of dietary data output. Alternatively, subtracting water or moisture content from foods to generate dry food weights may imperfectly approximate normalization when serving sizes are not available.

When aiming to connect diet with the microbiome, how to quantify intake is further complicated beyond the usual context of estimating intake and consumption. Most work dealing with error adjustment for nutrient intake has focused on more accurately estimating the intake of essential and nonessential nutrients from a human-centric perspective. When it comes to connecting diet with the microbiome, this paradigm may need to shift, ultimately changing the way we consider dietary inputs to models. Connecting food servings to the quantity of food that reaches the distal colon where it can ultimately impact the composition of the gut microbiome will require new approaches to dietary assessment and analysis. Furthermore, quantifying the nutritional composition of foods by the time they reach the distal colon may be impossible. Absorption of nutrients along the gastrointestinal tract is likely to vary with meal composition. At the same time, there are likely also important differences introduced by food particle size that may be due to other components of eating context and environment, including meal preparation and the amount of chewing a person does. Subsequently, it may ultimately make sense to consider diet not in terms of absolute quantity or nutritional content, but rather in terms of presence or absence of a food because there are so many unknowns about the amount of each food product that will persist in a large enough quantity to impact the gut. On the other hand, there is also a possibility that digestion and absorption of nutrients in larger quantities could concentrate the relative amount of small biochemicals in foods, meaning that small contributors to foods, like polyphenols, flavors, and color molecules that make up a relatively small volume of food at the point of ingestion, could have an oversized impact on the microbiome after other components of foods have been digested and absorbed. Continued development of in vitro and ex vivo simulated gut environments remains an important experimental approach to explore the role of digestion and gut biogeography on the relations between foods and food components in the gut microbiome (31, 32).

Food Groups, Dietary Patterns, and Dietary Quality Indices

Food groups are another important concept for exploring dietary data. Food grouping structures and food ontologies largely stem from an original need in food and nutrient databases to facilitate dietary data entry and database maintenance. For example, the Food and Nutrient Database for Dietary Studies is organized in a hierarchical structure (33). These grouping structures do not necessarily perfectly match the food groups that are defined by the What We Eat In America (WWEIA) database (34). They also do not perfectly correspond to food groups as they are identified in the USDA's Dietary Guidelines for Americans (30) and used in subsequent scoring tools to assess dietary quality like the Healthy Eating Index (HEI) (35). Nevertheless, these structures may provide analogous tree-like, lattice, or hierarchical structures that organize food data and can inform analysis of dietary patterns using approaches analogous to UniFrac (9). The Use of WWEIA food groups has also been demonstrated to improve the ability to find agreement between diet and microbiome composition and features in large observational studies (11).

Dietary indices allow for assessment of dietary pattern and quality that typically reduce multivariate dietary data to a single continuous variable. Approximately 60 dietary indices have been developed, defined, and exist in the published literature (36). All are intended to determine adherence to a dietary pattern and are defined based on their association with improved health outcomes. Dietary indicies are considered a priori because they are based on prior scientific evidence, current expert nutrition knowledge, and consensus around defined dietary components. Dietary indices capture multiple dimensions of diet, including adequate intake of essential nutrients and moderation of intake across food groups considered harmful to health. These scores typically also incorporate balance and variety. For example, the HEI (35) is a common index used in US populations to assess how well eating patterns of a given study population align with the USDA's Dietary Guidelines for Americans (30). The benefit of the wide adoption of this approach to assess dietary patterns is that it allows for comparison across studies and with data collected from representative populations. However, as with other indices, there are also shortcomings. All a priori indices depend on expert recommendations, which have fluctuated and changed as nutritional epidemiology has changed course on the definition of a healthy diet.

The alternative approach to dietary indices is to use a data-driven or a posteriori derivation of dietary patterns. Data-driven derivations of dietary patterns are specific to a given study population and are difficult to generalize to another population unless the study population is adequately representative. These derivations are also not necessarily likely to correspond with expert consensus and agreed upon nutritional knowledge, as perfect adherence to dietary guidelines is generally not observed in study populations. There are multiple ways to derive dietary patterns from a dataset. When data from FFQs are available, factor analysis can be used to derive a 2-factor solution, which is typically used to identify 2 patterns, termed “prudent” and “Western”, based on the variables that contribute most to each pattern. Additionally, clustering approaches can be used to determine dietary patterns from food group–based principal components scores, where K-means clusters of subjects with similar eating patterns are identified from the dataset (11). Other approaches incorporating different methods to define the distance metrics between subjects (e.g., tree-based, weighted, unweighted, or Jaccard) can be derived for diet, building off multivariate approaches previously defined for use in ecological assessment (9, 37).

Dietary Assessment is Dominated by FFQ

To explore the associations of dietary patterns and food groups with the gut microbiome, we conducted a literature search to identify studies in which dietary patterns were compared with the microbiome (Supplemental Table 1). Identification of relevant research studies was conducted in PubMed using the keywords (“food” OR “diet” OR “dietary”) AND (“gut microbiota” OR “gut microbiome”) AND (“prospective” OR “nested case-control” OR “case-cohort” OR “cross-sectional”). The search was limited to human studies and English language papers published until October 2021. Healthy adults (aged ≥18 y) and either >50 subjects or >100 stool samples were also required to be included in the review. Reference lists and the relevant review articles were manually reviewed, and Google Scholar was also screened for related citations.

Of the studies identified, the majority were cross-sectional studies, and only a small number were prospective cohort studies. The sample sizes covered a wide range, from small longitudinal studies with high numbers of repeated samples and low numbers of subjects (9, 38) to large studies of almost 5000 subjects with a single microbiome collection timepoint (39). Most cross-sectional studies identified had sample sizes under 500 subjects (Figure 2A). These studies use many methods to collect dietary data, including records (9, 38, 40–43), short intake questionnaires (44–49), and self-identified dietary types such as vegetarian or vegan dietary patterns (50–56), but the majority relied on FFQs. The dietary methods used in the identified studies appear to correlate with the study design. Studies enrolling more subjects preferentially used faster and cheaper dietary assessment methods, like FFQs, than smaller studies (Figure 2B). The reliance on FFQs to assess diet was also evident in prospective studies. While sample sizes were generally larger in these studies, most used FFQs during earlier phases of the prospective study to compare with later measured microbiome composition.

FIGURE 2.

Overview of the dietary pattern-gut microbiome research landscape. (A) Bar chart showing the majority of studies in this space have sample sizes of <500 subjects, with only 36% of the studies identified enrolling larger sample sizes. (B) Bar chart showing the proportion of dietary methods used to collect dietary data for studies identified by sample size, with higher proportions of larger studies identified using FFQs to collect dietary data. (C) Stream plot showing the different types of research completed in this space with the multiple types of dietary data collection and how those data ultimately map onto the final dietary assessment modality used to explore relations between diet and the microbiome. Dietary patterns and or food groups were used to explore diet–microbiome relations in most studies identified, with only a handful of studies using mixed exposures (both dietary patterns and food groups) and only a single study identified using nutrient biomarkers.

The underlying assumption of numerous diet–microbiome studies is that dietary intake history going back years, if not decades, is predictive of microbiome composition. While it appears that historical habitual diet may influence some microbial features, it is also apparent that more proximal diet measures are valuable when exploring temporal microbiome variation (9, 22, 38). Proximal measures of diet may also outperform historical dietary data collection, though the latter is still helpful for predicting some markers of microbiome community structure (57). This context is essential when considering the relation between dietary patterns and the microbiome through the lens of precision nutrition. If we intend to shift the microbiome to a more “healthful” state, then short-term and long-term dietary pattern–microbiome relations likely need to be considered in tandem.

When considering all studies identified by this review, what is most apparent is simply the diversity of methodology employed to capture dietary intake. This heterogeneity ultimately influences researchers’ ability to identify consistent signals and similarities across studies. The 54 studies identified used at least 7 different methods to assess diet. Coupled with the variation in final dietary parameters stemming from these methods, few studies with shared methods and assessment techniques are available for comparison (Figure 2C). Only a few studies used a mixed cross-sectional and prospective study design (37, 57). Nutrient biomarkers allow for objective measurement of nutrient and food intake, enabling stronger associations to refine and better understand the relations between specific metabolites and microbiome features and functions. However, nutrient biomarkers were rarely used as the primary dietary exposure of interest in identified studies (43, 58).

Microbiome Collection in Dietary Pattern Studies

In most of the identified studies, stool samples for microbiome analysis were collected from fewer than 100 subjects at a single time point (42, 48, 50, 51, 59–64). Only a few studies collected repeated stool samples (9, 37, 38, 52, 55, 65, 66), with the largest number of collections being 17 times on consecutive days in 34 subjects (9) and 479 stool samples from 2 subjects over a 1-y period (38). The time frame between dietary assessment and stool sample collection varied across studies. While some studies collected dietary data during the same period as stool sample collection, others reported several years between diet data collection and stool sample collection. The time intervals between dietary data collection and stool sample collection were not documented clearly in the study methods in a majority of studies. Gut microbial composition was primarily profiled using 16S rRNA sequencing, with a minority of more recent studies using shotgun metagenomic sequencing approaches (9, 37, 39, 65, 67–70).

Microbiome Alpha Diversity and Dietary Patterns

The majority of studies identified in this review examined dietary patterns and food groups. Others incorporated combinations of food groups, dietary patterns, and nutrient biomarkers. Alpha diversity was the most common microbiome outcome of interest. Linear regression models were most frequently used to compare dietary patterns and food groups with alpha diversity, followed by ANOVA and correlation analysis. Numerous studies have found relations between alpha diversity and features of diet, including food intake amounts, dietary pattern scores, or dietary types (vegetarians, vegans, or omnivores) (11, 37, 39, 52, 56–58, 67, 70–76, 76–78). HEI and the Mediterranean diet score (MDS) are the most commonly used a priori indices studied in relation to gut microbial diversity. For the HEI, results were inconsistent; some authors found a positive association between HEI and alpha diversity indices (57, 77), while others found no association (79). Similarly, the results for the MDS score were heterogeneous across studies. Three studies reported no significant association between MDS and alpha diversity indices (49, 66, 80), while 1 study found a positive association (77). Interestingly, almost half of the studies identified from the literature review did not report testing for relations between diet and alpha diversity, bringing into question if the tests were not run, or if there simply was no significant association between alpha diversity and diet, which could indicate less consensus overall between dietary patterns and alpha diversity. While many studies did not study or find an association of dietary indices and foods with alpha diversity, they found compositional differences in microbiome (38, 40, 44–46, 59–61, 63, 65, 68, 79, 81–86).

Microbiome Beta Diversity and Dietary Patterns

Compared to alpha diversity, the methods used for the analysis for beta diversity were relatively consistent across studies; beta diversity was commonly described using principal coordinates analysis (PCoA) based on either Bray-Curtis distance or UniFrac, with PERMANOVA being the most common approach to test group differences statistically. Relations between dietary patterns and beta diversity were reported in multiple studies. Acute changes to the diet have been shown to affect the beta diversity on short timescales (9), with approximately 1% of the variation in beta diversity that can be attributed to specific biomarkers of vegetable intake in diverse populations (58). Similar regional diets such as Western, Japanese, and Mediterranean patterns have signals that can be detected in beta diversity space (37, 49, 60, 75), as do dietary patterns that focus on the exclusion of specific food groups such as vegetarian or vegan diets (51–53, 55), or exclusion of sources of carbohydrate foods (11). The consumption of specific food groups like dairy products and sugar or sugar-sweetened beverages may also contribute to patterns seen in beta diversity space between subjects (43, 72). In addition, fruits and vegetables (58, 71), nuts (87, 88), and fish (74) have been reported to be related to beta diversity. Generally, intake of high-quality (39) dietary patterns with significant fiber content are associated with microbial beta diversity differences across populations (39, 77). A recent retrospective analysis found that a posteriori dietary patterns based on a variety of food categories were more likely to be related to beta diversity than a priori dietary patterns based on dietary fibers, proteins, and 12 HEI-2010 subcomponents (11). These data suggest that when connecting diet with the gut microbiome, it is more important to consider the overall diet rather than focusing on single nutrients or food categories, corroborating earlier results that gut microbiota had a stronger correlation with food choices than with nutrients (9).

Differential Abundance and Dietary Patterns

A number of differences in taxa abundances at each taxonomic level were reported across the studies. Analytical models used for differential bacteria abundance varied widely across studies; linear regression-based models were most common, followed by Spearman's or Pearson's correlation analyses, contributing to the variability in findings (89). In addition, ANOVA or t-tests were used in several studies. There is a wide difference in types of food and dietary patterns studied, as well as the analytical approaches applied, which makes the direct comparison of results difficult. Briefly, at the phylum level, Bacteroidetes and Firmicutes are the 2 most prevalent phyla-level taxa to be investigated, but with substantially contradictory results. At the genus level, higher abundance of Faecalibacterium (51, 75, 82, 90),Lachnospira(41, 66, 90), Prevotella (54, 66, 75, 91), and Roseburia (41, 66, 75, 82) were observed with plant-based foods, vegetarian or vegan diets, and higher values of diet quality indices, while at the species level Escherichia coli (50, 53, 55, 86) was more prevalent in the nonvegetarian or omnivore groups and subjects with lower MDS. The mixed results reported in the literature to date highlight the need for refined diet–microbiome integration protocols and a deep understanding of dietary data to improve our ability to explore relations between these 2 complex data sets. It is important to note that differences in bacterial taxa do not necessarily represent differences in microbial activity and the most highly differentially abundant taxa may not represent meaningful differences in function, while low-abundant taxa (which are routinely excluded from analysis if <1% relative abundance) may also influence function (92).

Opportunities to Expand Diet–Microbiome Integration Approaches

Applying computational tools designed for the microbiome to dietary datasets

Microbiome analysis provides a fertile space for the development of computational approaches to understand the temporal changes in complex systems. Microbiome data provide a unique analogous data type to nutritional data that can inform the possibilities for longitudinal dietary assessment. The similarities between dietary data and microbiome data in terms of compositionality and zero inflation should allow for the direct application of many computational approaches from the microbiome space to the dietary assessment space. The microbiome literature draws heavily from ecology and mathematical approaches originally defined by ecologists for measuring variation and stability. It follows that these ecological approaches (not to be confused with ecological studies in epidemiology) to assess variation and stability should have utility when applied to our understanding of dietary variation and stability beyond what is currently captured in existing dietary diversity metrics.

Both dietary data and microbiome data can be analyzed using univariate and multivariate techniques (Figure 3). Our review of the literature demonstrates that a majority of studies to date have applied univariate approaches to dietary data while using multivariate approaches for microbiome data. The use of multivariate approaches for diet may improve integration of diet and microbiome data to reveal stronger diet–microbiome signatures, and importantly to develop methods to control for the variation in microbiome composition introduced by diet in studies with microbiome-focused outcomes.

FIGURE 3.

Opportunities and approaches for diet–microbiome data integration. Analysis of microbiome data using alpha and beta diversity metrics lend themselves well to integration with conventional nutrition analysis in terms of single nutrient analysis and dietary indices. Modern approaches to dietary pattern assessment using hierarchical tree-of-foods classification, factor analysis, or complex clustering algorithms from dietary data can be used to define dietary patterns in PCoA space. The resulting analysis outputs lend themselves well to integration with the microbiome and other omics data through correlation or regression (shown with an example heatmap) where dietary features including dietary indices, single nutrients, food groups, and dietary PCoA, or clusters can be compared with differentially abundant taxonomic groups, alpha diversity metrics, or microbiome PCoAs (with or without adjustment for covariates). Procrustes analysis or other multivariate metrics can be used to compare microbiome ordination directly with dietary ordination. Finally, constrained ordination of microbiome space by dietary features, which can include single nutrients, dietary patterns, or simply dietary PCoAs as inputs, can reveal microbiome gradients driven by diet, or be used to infer the explanatory power of diet on microbiome composition with or without the addition of other explanatory variables (i.e., age, sex, BMI, etc.). Figure created using BioRender. PCoA, principal coordinate analysis.

The application of ecological metrics to diet, such as alpha diversity, can reveal dimensions of diet variation that have previously been unexplored in relation to microbiome composition (93). However, this raises important questions about the meaning of alpha diversity when applied to diet: What does it mean to have highly diverse dietary scores on a daily basis? Is there an optimal timeframe that should be considered when calculating dietary diversity? Much like when estimating usual intake, is it better to estimate dietary diversity from multiple days of records or from FFQ data? Or is it acceptable to estimate dietary diversity from individual days of intake? What is the best alpha diversity metric to use to capture dietary diversity, and should caloric content be controlled for, given that as diversity increases energy intake may also increase?

The application of beta diversity to diet is somewhat easier to interpret, as factor analysis of diet has already been operationalized to define dietary patterns (e.g. prudent v. Western). In the dietary beta diversity multi-dimensional space, data points that are close relative to each other represent days or subjects with similar overall dietary patterns. The use of tree-based analysis metrics, similar to phylogenetic trees used in microbiome science, may also have utility in the dietary space when integrating diet and microbiome data (9). Using a tree-structure has the potential to allow for easy analysis of multiple levels of dietary patterns, foods, and nutrients along with food synergy somewhat simultaneously (13). Work to define the optimal tree structure for this purpose is still needed and could incorporate actual phylogenetic distances for related plants (i.e., Linnaean taxonomy), though defining distances between foods is challenging. The development of a metabolite-informed tree structure for foods that accurately represents the diversity of food components in a hierarchical way without dependence on expert-defined food grouping structures or hierarchies could further improve this approach. Regardless of the approach, multivariate analysis of diet can allow for use the of Procrustes analysis or the Mantel test for direct comparison between microbiome communities and dietary patterns in their equivalent beta diversity spaces (Figure 3).

Integration approaches for diet data with omics data

Constrained ordination/redundancy analysis

Canonical correspondence analysis (CCA) and redundancy analysis (RA) approaches are frequently applied to microbiome data. These approaches allow for the incorporation of additional variables to explore their impact on microbiome composition, and these types of approaches are routinely used to identify gradients in microbiome data (94). Once dietary patterns have been generated from food or food-group level data, then they can be passed to CCA and RA to explore if the constraint of microbiome beta diversity distances by diet explains any amount of the variation in microbiome composition (Figure 3).

Multiomics where diet is an “ome”

Co-occurrence networks and co abundance network analysis using both diet data at the food or food group level and microbiome features at the species or genera level with enough data density has the potential to reveal relations between dietary intake and specific members of the microbial community. These approaches are helpful for hypothesis generation and also lend themselves well to the addition of more omics data such as fungal data, functional gene content predictions, or protein expression data. These methods are likely to be improved by Bayesian approaches and causal inference that can incorporate time lag into the models and networks developed.

Compositional approaches for differential abundance

Approaches for differential abundance analysis have been developed for microbiome account for compositionality [i.e., Songbird's “reference frames” and ANCOMBC (analysis of compositions of microbiomes with bias correction) (23, 24)]. Diet can be assessed as a feature in these models to control for habitual diet, or the diet itself could be assessed using these methods to identify differences in diet between study groups that can then be explored for their impact on the microbiome space.

Procrustes analysis to pair diet and microbiome

Procrustes analysis, named after a character from Greek mythology known for forcing his victims to fit the size of a bed by chopping off their feet or legs, is nowhere near as brutal as its namesake. Procrustes analysis provides an exciting opportunity to compare diet and microbiome in ordination spaces directly, as has been demonstrated for microbiome–metabolite integration (95). Limited application of Procrustes analysis to diet and microbiome datasets has shown relations between diet and microbiome (9, 96). Researchers can generate ordination spaces for both diet and microbiome and simply ask if there is an agreement between the 2 sets of points using this permutation-based approach (Figure 3). When there is significant agreement between diet and microbiome, it suggests that the linked data contain some shared information, though caution should be taken when interpreting results as methods to adjust for covariates have not been defined for Procrustes.

Machine learning approaches

The field of deep learning continues to expand. The problem of understanding diet and microbiome relations potentially lends itself to deep learning and may benefit from neural network approaches to reveal interactions between the gut microbiome and diet. Using a neural network to understand these complex relations may be beneficial. Researchers can use a machine learning approach to uncover various overall dietary patterns obtained from repeated evaluations from a large population, construct models, and make microbiome composition predictions by learning from training data (97, 98).

Assessment of diet directly from stool samples

In the future, it may be possible to directly measure dietary intake or makers of dietary intake from stool samples. Preliminary work applying trnL DNA metabarcoding to human samples demonstrates that signals from plant foods can be detected directly in stool samples from subjects eating a high-plant diet (99). In settings where habitual diet is more reflective of a Western diet (i.e., lower in fiber and higher in meat and dairy products) it is probable that this methodology will be less successful. However, it is possible that plant metabarcodes could be used to approximate dietary quality indices given their ability to capture markers of plant intake. Alternatively, since dietary patterns are repeatedly shown to correlate with overall microbiome composition, we may reach a point where dietary intake or habitual dietary patterns can be inferred simply through exploration of microbiome community membership or functional genomic potential. Early exploration of this approach shows promising results when determining if a sample came from a subject who consumed specific foods (i.e., broccoli, almonds, etc.) as part of a series of controlled feeding intervention trials (100).

Connection to other data endpoints

The application of multivariate dietary pattern assessment in connection with microbiome data raises a number of questions for how to connect diet and microbiome datasets to other study end-points and health-related outcomes. Specifically, using these approaches to identify causal relations between diet, microbes, and health outcomes will require the development of statistical techniques for mediation analysis that can incorporate >1 highly multivariate model input. Methods specific to the microbiome space are under development and may have utility when applied to diet (101). Given that microbiome composition in terms of taxonomic composition may not accurately measure functional activity of the microbiome, there is a lack of research connecting dietary data to other microbiome outputs that better capture microbial function such as metatranscriptomics, metaproteomics, and metabolomics. Metatranscriptomics in particular, which gives information about gene expression could improve identification of important diet–microbiome-disease relations (102).

Conclusions

Our review of the literature revealed that dietary collection and assessment methods vary widely across studies. This methodological variation may contribute to inconsistency and limited agreement in the literature with respect to findings specific to how diet contributes to features of the microbiome, including the differential abundance of particular species or genera and alpha and beta diversity. Future research may benefit from considering diet as a multivariate, highly dimensional data type. Adopting multivariate approaches for diet will allow diet and dietary patterns to be treated as an additional “ome” for integration with other multi-omic data sets like microbiome sequencing data and fecal metabolomics. There remain opportunities to improve dietary analysis by using or adapting tools that are currently applied to microbiome datasets such as approaches incorporating dietary tree-of-foods approaches, the use of diversity metrics from ecology for diet, and data adjustment for compositionality. The adaptation of these approaches and data visualizations may help to integrate dietary patterns into existing microbiome analysis pipelines. Ultimately these approaches may help to reveal new relations, especially when applied to large datasets with dense sampling of both diet and microbiome.

Notes

AJJ is supported, in part, by the National Institutes of Health (P30CA077598 and R21DK125933).

Author disclosures: AJJ serves as the Associate Director of the Nutrition Coordinating Center at the University of Minnesota, which maintains the Nutrition Data System for Research (NDSR). All other authors report no conflicts of interest.

Supplemental Table 1 is available from the “Supplementary data” link in the online posting of the article and from the same link in the online table of contents at https://academic.oup.com/jn/.

Abbreviations used: ASA24, Automated Self-Administered Dietary Assessment Tool; HEI, Healthy Eating Index; MDS, Mediterranean diet score; NDSR, Nutrition Data System for Research; PCoA, principal coordinates analysis; PhILR, phylogenetic isometric log-ratio transform; RA, redundancy analysis; WWEIA, What We Eat In America.

Data Availability

All data presented are avaliable in the supplementary data file provided with this article. Code and summary data used to generate Figure 2 is avaliable on GitHub: https://github.com/computational-nutrition-lab/Diet-microbiome-integration-review/.