A practical introduction to holo-omics

Introduction

The realization of the importance of microorganisms for animal and plant biology,¹ along with the increased capacity to generate and process molecular data,² has given rise to a new holistic way to study biological systems.³ Holo-omics refers to the technical approach to jointly (hence the prefix holo-) analyze multiple non-targeted molecular data layers (known as multi-omics) from both hosts and their associated microorganisms,^4,5 aimed at unraveling their intricate relationships.³ The generation, analysis, and integration of holo-omic datasets requires deep knowledge of the myriad of conceptual and technical steps involved in the process.⁴ Here, we provide an overview of the main steps researchers must undergo while highlighting the challenges that should be overcome to obtain meaningful results that enable them to address complex biological questions.

While acknowledging the relevance and value of other methods for studying host-microbiota interactions (e.g., 16S rRNA amplicon sequencing, spatial transcriptomics), in this review we primarily consider seven omic layers characterized by a non-targeted data generation approach without spatial resolution. These methods require specific data generation and analysis strategies before integrating them into multi-omic statistical models (Figure 1). Four of them are based on nucleic acid sequencing, namely host genomics (HG), host transcriptomics (HT), microbial metagenomics (MG), and microbial metatranscriptomics (MT). The three remaining layers are based on molecular spectroscopy, namely host proteomics (HP), microbial metaproteomics (MP), and (meta)metabolomics (ME), which is not split between hosts and microorganisms for the difficulties in assigning an origin to a given metabolite. Each omic layer contains fundamentally different information, at different levels of biological organization. A host contains a single genome (HG) that encodes thousands of genes (HT) which are expressed to produce an even larger number of proteins (HP). A host might have multiple (often hundreds or thousands) associated bacterial species (n), each with a distinct genome (MG_n) containing a few thousand genes that can be expressed (MT_n), thus potentially encoding microbial proteins (MP_n). The biological activity enabled by those proteins builds the metabolomic (ME) landscape that not only shapes host phenotype but also conditions the environment in which host-associated microbes live (e.g., the gut).

HG—which in this manuscript is solely discussed in terms of nucleotide sequences without considering other levels such as epigenetic and DNA folding that may also shape the hologenome^3,6,7—can inform about population structure and the genetic features of individuals but is invariable to treatments or environmental disturbances. In contrast, MG contains the genetic information of microbial communities that are likely to change over very short timescales and reflects the immediate responses of hosts to treatments or disturbances. Although HG and MG inform about the potential of hosts and microorganisms to perform biological functions, HT and MT provide snapshots of the actual functional activity of the system, which can be validated using HP, MP, and ME that result from the activity of the expressed genes. Acknowledging the distinct biological characteristics of the omic layers is therefore essential to design experiments and analytical pipelines for better solving the complex puzzle of host-microbiota relationships.

In light of this, we navigate from sample acquisition to data interpretation through six sections: (1) sample collection and preservation, (2) laboratory sample processing, (3) bioinformatic analysis of raw data, (4) data filtering, imputation, and distillation, (5) initial quantitative exploration of omic layers, and (6) multi-omic data integration. Each of these sections deals with key methodological aspects, focusing on the commonalities and main differences to generate and analyze the various types of omics, with a special focus on optimizing data generation and integration. We stress the importance of considering each step before collecting and processing data, to ensure every scientific question is aligned with the appropriate study design and the technical tools needed to address your question at hand.³

Sample collection and preservation

Decision-making starts with the choice of sample collection and preservation procedures, a non-trivial election that can affect the quality and accuracy of any omic datasets to the risk of masking the biological signal of interest.^8,9 Though the current “gold standard” of sample preservation is to snap-freeze and store samples at −80°C, this is not always feasible.¹⁰ Hence, multiple different preservatives have been developed, allowing samples to be stored at ambient temperature for extended periods. Studies have shown that preservatives can introduce considerable biases when generating different omic layers,^11,12 thus keeping consistency among preservatives is critical. In addition, the biological and chemical properties of the molecules (e.g., DNA, RNA, proteins, metabolites) used to generate omic data must also be acknowledged. HG is the less sensitive omic layer because host DNA is usually abundant and stable. In contrast, MG should be treated more cautiously because microbial communities can change (e.g., the abundance of saprophytic bacteria tends to increase after sampling) unless biochemical reactions are blocked. HT and MT require even more rapid preservation if representative gene expression patterns are to be captured, because RNA typically decays faster than DNA.¹³ In addition, one must bear in mind that gene expression can vary considerably across host tissues,¹⁴ whereas microbial physiology can vary rapidly across time and space.¹⁵ Host gene expression can vary drastically between, for instance, intestinal tissue, liver, and brain, whereas microbial gene expression can change between the mucosal layer, digesta, and feces, for example. Ultimately, selecting appropriate sample types for answering the biological question of interest in advance is essential to capture the desired signal. Lastly, metabolites have varying chemical properties, spanning large stable steroids to extremely volatile short-chain fatty acids. The choice of appropriate preservatives for the generation of multiple omic layers is therefore necessary, which entails deciding whether the omic data will be generated from a single (requires appropriate preservative for all omics) or multiple biological samples (each sample type can have a dedicated preservative). Lastly, due to the varying physicochemical properties of samples, collection and storage methods validated for one sample type cannot be assumed to be optimal for other types. Hence, preliminary optimization tests are always advisable,¹⁶ and methodological consistency is a requirement to generate reliable omic data.

Laboratory sample processing

Bioinformatic analysis of raw data

Data filtering, imputation, and distillation

Bioinformatic processing of raw sequencing and spectrometry data yields a massive amount of information organized in structurally diverse datasets. Host genomes often contain thousands of genes, with hundreds of thousands of nucleotide variants.²³ It is also common to generate catalogs of hundreds of bacterial genomes, with millions of genes, as well as metabolomic profiles containing thousands of metabolites. One of the first considerations is to filter these data, to reduce the information to only that required for answering the biological question of interest. Following qualitative criteria, one may decide to exclude entire omic layers to address specific questions or to only consider a subset of an omic dataset, such as a specific family of genes. Following quantitative criteria, it is necessary to establish thresholds to discard data points under certain threshold values to offset the biases and limitations of data generation techniques. Insufficient sequencing effort, genome completeness, or coverage of genomes are some of the reasons researchers need to filter some of the data out, to avoid technical aspects introducing noise that potentially shadows the true biological signal in the statistical analyses.

As many statistical approaches do not deal adequately with missing data, imputation is another strategy to be considered in the initial pre-processing of the data. Imputation refers to the process of replacing missing data with the most probable substituted values based on the context. Genotype imputation is broadly employed in the context of HG,²³ functional imputation has also been recently proposed as an approach to minimize bias from low genome completeness in functional MG analyses,^64,65 and imputation strategies are also being developed for HT and ME.^66,67

In some cases, integration and interpretation of these multivariable data can be intractable, which might necessitate the conversion of original datasets into a reduced set of biologically meaningful data through a process known as data distillation.⁴⁵ The main advantage of data distillation is to gain the capacity to interpret results within a given biological context. Moreover, distillation also reduces the number of features, thus decreasing computational requirements and gaining statistical power for downstream analyses. One example of biological data distillation is to convert occurrence data of dozens of genes present in a bacterial genome into a genome-inferred functional trait, which proxies the capacity of that genome to perform a given function.⁶⁸

Initial quantitative exploration of omic layers

The analysis of multi-omic data should begin with an independent analysis of each omic layer to learn about its structure and variability before jumping to multi-omic data integration. Despite the fundamental differences between the seven omic layers considered in this article, all share the characteristic of being multivariate, i.e., they contain multiple features (genomic variants, genes, metabolic pathways, proteins, or metabolites) collected across multiple observations. In the following, we provide a summary of some of the most relevant statistical techniques suitable for multi-omic studies, which have been covered more in depth in the literature.^69,70,71,72

Multi-omic data integration

When it comes to multi-omic data integration, researchers often need to prioritize between predicting a given outcome (e.g., phenotype, disease state, microbial community) based on the multi-omic data (predictive modeling) or using multi-omic data for understanding the biological processes leading to that outcome (explanatory modeling or causal inference). A good predictive model includes a set of variables, the observation of which is systematically associated with changes in an outcome, thus maximizing predictive capacity. By contrast, a good causal model includes a set of variables, the intervention of which would provoke a change in an outcome.⁹⁸ Both concepts are related in that a good causal model should have predictive capacity; however, the best predictive model need not have a causal meaning.⁹⁹

Acknowledging these priorities is important for choosing the best strategy for multi-omic data integration, which can be broadly categorized into two types: multi-staged analysis and meta-dimensional analysis.^100,101 In the multi-staged approximation, data analysis is divided into multiple steps, relating two omic layers at a time, and the final step links the relevant omic layers with the outcome of interest.¹⁰⁰ This approach leverages the central dogma of molecular biology to assume that the variation in omic datasets is hierarchical, such that variation in DNA leads to variation in RNA and so on. By contrast, in meta-dimensional analysis, all omic datasets are analyzed in a single analysis, spanning the whole landscape of features simultaneously and enabling to account for inter-omic interactions.¹⁰⁰

Until recently, the multi-staged approach dominated the field of integrated analysis of biological data,¹⁰¹ mostly relying on traditional statistical tools and hypothesis testing approximations mentioned in the previous section. Multi-staged analysis has the advantage of relating multi-omic datasets in an organized, stepwise manner, which enables building knowledge for later use to test causally oriented hypotheses. In addition, multi-staged approaches are better suited to accommodate the biological asymmetries between multi-omic datasets. For instance, although both HG and MG data comprise DNA sequences that could be treated similarly, the process that generated their compositions profoundly differs. While HG belongs to a single individual (i.e., the host), the MG belongs to a microbial community of multiple individuals. Species communities are assembled following the basic processes of selection, drift, dispersal, and speciation,¹⁰² the combination of which determines the community composition, functional profile, and biodiversity of the microbiome. Thus, inferences made from MG may benefit from specific statistical frameworks such as joint species distribution modeling,^87,103 whereas this might not be the case for HG data. As a downside, such a stepwise approach assumes that omic layers linearly influence each other to determine a complex trait of interest, which implies that complex inter-omic interactions are likely to be overlooked.¹⁰⁰

Advances in the field of artificial intelligence (AI) and ML have shifted the balance toward meta-dimensional approaches where the whole complexity of multi-omic datasets are pooled into a single analysis.⁹⁷ Meta-dimensional approaches can be classified into concatenation-based, transformation-based, and model-based integration methods.^97,100 Concatenation-based integration combines multiple omic datasets, raw or pre-processed, into a single large matrix; then, the concatenated table can be used for unsupervised learning with methods such as multi-omics factor analysis (MOFA),¹⁰⁴ or for supervised learning with any of the above-mentioned ML algorithms. In transformation-based integration, omic datasets are first transformed into an intermediate representation, typically a graph or a kernel matrix, and they are then merged before building the final model. Graph-based analyses have the advantage of easier interpretability and lower computational requirements whereas, overall, kernel-based methods provide higher predictive performance.¹⁰⁵ Model-based integration builds intermediate models from each omic layer and then builds a final model combining all intermediate models. An advantage of the model-based approach is that it allows the merging of multiple omic types that have been collected in different sets of sampling units if the outcome of interest is the same across datasets (e.g., specific disease).¹⁰⁰ Multi-omic data integration efforts such as analysis tool for heritable and environmental network associations (ATHENA)¹⁰⁶ or multi-omics supervised autoencoder (MOSAE)¹⁰⁷ use model-based integration for disease prediction by combining a variety of modeling frameworks and algorithms. Unquestionably, meta-dimensional integration of multi-omic data has improved predictive accuracy in many applications, including disease diagnostics and prognosis or biomarker discovery.^108,109 However, these applications usually respond to rather “simple” problems, such as classifying individuals into health statuses.⁹⁷ Such simple settings are the exception rather than the rule in studies on fundamental biology, which often deal with intricate hierarchical, spatial, temporal, or phylogenetic structures derived from complex experimental designs and the heterogeneity of nature. Although approaches are being developed to account for those biases in ML algorithms,^110,111 currently generalized linear mixed models and Bayesian hierarchical models represent the most established and accessible statistical frameworks to account for them by means of random effects.

Lastly, one must bear in mind that predictive modeling is not suited for causal inference, because non-causal associations may increase predictive accuracy, but undermine our understanding of the system under study.^112,113 Unlike the multi-staged approximation, meta-dimensional analysis favors not using previous knowledge to independently reduce omic datasets, arguing that domain-knowledge-oriented procedures may introduce bias by ignoring previously undiscovered biology.¹⁰⁰ However, blindly pooling hundreds of variables into a model may introduce biases as well, such as masking true causal associations through overcontrol bias or generating spurious (i.e., non-causal) correlations through collider bias.¹¹⁴ Hence, while maximizing predictive accuracy is highly valuable for many applications of multi-omic data, we argue it should not be the final goal when trying to understand complex host-microbe and microbe-microbe interactions in basic holo-omic research. In those situations, the application of meta-dimensional approaches should ideally be an intermediate step to generate hypotheses on causal relationships that later will need to be tested using randomized experiments or causally oriented analyses, which have recently started to flourish in the context of multi-omic data integration.^115,116

Conclusions and future perspectives

The number of elements involved, the asymmetry between host and microbial organisms, the dynamism of the interactions, and the multi-scale nature of the effects render host-microbiota systems, or holobionts, as one of the most challenging biological entities to be studied. Holo-omics tackles that complexity by leveraging the full potential of laboratory, bioinformatic, and statistical methodologies that continue advancing at a vertiginous pace. Although access to great data and technology comes with great power, with great power comes great responsibility. In certain scenarios, employing simpler targeted methodologies, such as characterizing taxonomic compositions of microbial communities through 16S rRNA sequencing or quantifying specific metabolites such as short-chain fatty acids, could prove more cost effective than adopting non-targeted approaches. However, for more complex inquiries, higher resolution strategies such as epigenomics, single-cell genomics, Hi-C genomics, spatial transcriptomics, or spatial metabolomics might be necessary to effectively address research questions demanding a complete representation of the environment. In any case, generating millions of data points does not entail that all of them need to be used without any filtering or distillation. Instead, we argue that the workflow delineated in this article, with careful generation and distillation of data followed by independent exploration and analyses of the single omic layers, will aid researchers in having a better understanding of the study system before the integration of multiple omic layers in a final model is attempted. We, therefore, advocate for an integral approximation in which study design (acknowledging the experimental setup that best allows addressing the problem of interest), data generation (acknowledging the pitfalls and biases of employed techniques), and data analysis (acknowledging the properties of the biological elements under study) are considered as a whole. Only then will we be able to maximize the power of holo-omics and address some of the most challenging questions in modern biology.