Microbiome Multi-Omics Integration: Combining Metagenomics with Metabolomics and Proteomics

Microbiome multi-omics integration is the practice of combining metagenomic sequence data with metabolite profiles and protein measurements to explain how microbial communities function instead of only what taxa they contain. This article covers the data inputs, workflow choices, quality controls, interpretation limits, and reporting standards that researchers and analysts need when planning an integration project. The focus is on practical decisions: which omics layers to collect, how to preprocess each data type, which integration strategies fit which biological questions, and how to avoid common failure patterns that produce misleading correlations.

At a Glance

Multi-omics integration does not mean collecting every possible data type and running a single algorithm. The choice of integration strategy depends on the biological question, the sample size, the expected effect size, and the resources available for validation. The table below summarizes the main integration approaches, their data requirements, and their typical outputs.

The practical implication is that no single method works for every project. A study with 500 samples and a clear clinical outcome may support supervised machine learning approaches. A study with 30 samples and no outcome labels may only support unsupervised module detection or pathway-based interpretation. The evaluation literature supports this view: benchmarking studies across cancer datasets show that adding more omics data types does not always improve clustering accuracy, and some combinations reduce performance [17]. Researchers should therefore decide on the integration strategy before data collection, not after the results come back.

Why Integrate Metagenomics with Metabolomics and Proteomics

Metagenomic sequencing describes which microbial taxa are present and which functional genes they carry. This information is necessary but not sufficient for understanding microbial activity. A gene can be present in the genome but not expressed under the conditions of the sample. A metabolite can be produced by multiple microbial species or by the host. A protein can be translated but not enzymatically active. Each omics layer captures a different point in the biological flow from genetic potential to functional output.

Metabolomics measures the small molecules present in a sample, which represent the downstream products of microbial and host metabolism. Proteomics measures the proteins present, which represent the expressed functional machinery. Combining these layers with metagenomics allows researchers to ask whether the presence of a microbial gene corresponds to the presence of its protein product and whether that protein corresponds to the expected metabolite. This type of question is central to understanding host-microbiome interactions in conditions such as inflammatory bowel disease, where perturbations of the gut microbiota and metabolites are closely associated with disease progression [5].

The value of integration is visible in studies that use multiple omics layers to characterize disease. In irritable bowel syndrome, a multi-omics assessment incorporating 16S rRNA gene sequencing, metatranscriptomics, and untargeted metabolomics revealed that the disease is associated with global alterations in microbiome composition and function, including differential abundance of bacterial taxa, increased tyramine, decreased gentisate and hydrocinnamate, and transcripts related to fructooligosaccharide and polyol utilization [8]. A composition-only analysis would have missed the functional and metabolic dimensions of the disease.

The same logic applies outside human health. In food fermentation research, integrated metagenomic and targeted metabolomic analyses of traditional watermelon soybean paste identified the microbial synthesis pathways of flavor-related metabolites and showed that lactic acid bacteria and Aspergillus were associated with the synthesis of flavor compounds through amino acid metabolism and isoflavone biosynthesis pathways [21]. In vinegar production, enzyme-assisted ultrasonication pretreatment combined with co-fermentation was studied using volatile metabolomics and metagenomic pathway integration [20]. These examples show that the integration approach transfers across sample types and research questions.

Data Inputs and Experimental Design

Sample Collection and Metadata

The quality of any multi-omics integration project depends on the quality of the samples and the completeness of the metadata. Samples for metagenomics, metabolomics, and proteomics are often collected from the same biological specimen, such as feces, soil, or fermentation broth. The collection protocol must be standardized across all samples and all time points. For fecal samples, this means controlling for collection time, storage temperature, and extraction method. For fermentation samples, this means controlling for sampling time relative to the start of the process and for the physical location within the vessel.

Metadata should include all variables that could influence the omics profiles. In human microbiome studies, these variables include age, sex, body mass index, diet, medication use, and comorbid conditions. The irritable bowel syndrome study adjusted for age, sex, race and ethnicity, body mass index, diet, and anxiety in multivariate models [8]. In fermentation studies, the equivalent variables are substrate composition, temperature, pH, inoculum source, and fermentation duration. Missing metadata is a common cause of spurious associations in integration analyses.

Sequencing Depth and Coverage

Metagenomic sequencing depth determines the sensitivity of taxonomic and functional profiling. Shallow sequencing may miss low-abundance taxa and rare functional genes. The choice of depth depends on the expected diversity of the community and the question being asked. A study focused on dominant taxa in a low-diversity fermentation community requires less depth than a study of a high-diversity soil or gut community.

The same principle applies to metabolomics and proteomics. Untargeted metabolomics detects a broader range of molecules but with lower quantitative accuracy than targeted approaches. Targeted metabolomics measures a predefined set of metabolites with higher precision. The choice between these approaches depends on whether the research question is discovery-oriented or hypothesis-driven. The watermelon soybean paste study used targeted metabolomics to focus on flavor-related metabolites [21]. The irritable bowel syndrome study used untargeted metabolomics to capture a broader metabolic profile [8].

Replication and Power

Multi-omics studies are expensive, and replication is often limited by budget. However, the number of samples determines the statistical power for detecting associations and the ability to validate findings. Cross-cohort validation is a common strategy to address the limitations of single-cohort studies. The inflammatory bowel disease study analyzed 9 metagenomic and 4 metabolomics cohorts from different populations and used cross-cohort integrative analysis to identify consistent characteristics of commensal gut microbiota [5]. This design allowed the authors to distinguish reproducible signals from cohort-specific noise.

For machine learning approaches, sample size is critical. Models trained on small datasets are prone to overfitting, and the reported performance may not generalize. The atherosclerosis study used 456 metagenomic samples and 111 16S rRNA gene sequencing samples for microbial profiling, alongside 118 RNA-Seq samples and 302 microarray samples, and validated diagnostic performance with 5-fold cross-validation, study-to-study transfer validation, and leave-one-study-out validation [6]. This level of validation is the standard that integration projects should aim for when the goal is biomarker discovery.

Data Preprocessing and Quality Control

Metagenomic Data Preprocessing

Metagenomic data preprocessing begins with raw sequence reads. Quality trimming removes low-quality bases and adapter sequences. Host DNA removal is necessary for host-associated samples to avoid wasting sequencing depth on host reads. Taxonomic profiling assigns reads to taxa using reference databases or marker gene approaches. Functional profiling assigns reads to genes and pathways using databases such as eggNOG or KEGG. The watermelon soybean paste study used both eggNOG and KEGG databases for functional composition analysis [21].

The output of metagenomic preprocessing is a feature table with samples as rows and taxa or functional genes as columns. These tables are sparse, meaning most entries are zero. Normalization is required to account for differences in sequencing depth across samples. Common approaches include relative abundance scaling, rarefaction, and variance stabilizing transformations. The choice of normalization method affects downstream analyses, and the method should be reported clearly.

Metabolomic Data Preprocessing

Metabolomic data preprocessing depends on the platform. Mass spectrometry data requires peak detection, alignment, and annotation. Nuclear magnetic resonance data requires spectral binning and metabolite identification. The output is a feature table with samples as rows and metabolite features as columns. Many features remain unidentified, and the annotation rate varies by sample type and platform.

Normalization for metabolomics includes correction for batch effects, internal standards, and sample concentration. Missing values are common and require imputation or filtering. The choice of imputation method affects downstream analyses, and the proportion of missing values should be reported.

Proteomic Data Preprocessing

Proteomic data preprocessing involves peptide identification, protein inference, and quantification. Label-free quantification compares peptide intensities across samples. Tandem mass tag or isobaric labeling allows multiplexed quantification. The output is a protein abundance table with samples as rows and proteins as columns.

Proteomic data are often incomplete, with many proteins detected in only a subset of samples. Filtering criteria for missing values must be established before analysis. Normalization methods include total protein normalization, median normalization, and variance stabilizing normalization. Batch effects are a major concern in proteomic studies, especially when samples are processed in multiple batches.

Cross-Omics Quality Checks

Before integration, each omics layer should pass its own quality checks. The integration step adds new quality concerns. Sample identifiers must match across omics layers. The number of samples should be consistent, and any mismatches should be resolved before analysis. Batch effects should be assessed within each omics layer and, if present, corrected before integration.

The FAIR Guiding Principles provide a framework for data management that supports reproducibility and reuse [4]. These principles state that data should be findable, accessible, interoperable, and reusable. For multi-omics projects, this means depositing raw data in public repositories, providing complete metadata, using standard file formats, and documenting all preprocessing steps. The NCBI Data Resources provide repositories for sequence data, and the EMBL-EBI Training program offers guidance on data management practices [2][1].

Integration Strategies and Their Tradeoffs

Early Integration

Early integration, also called concatenation, combines all omics feature tables into a single matrix before analysis. This approach is simple to implement and allows any downstream method to be applied to the combined matrix. The main problem is that the omics layers often have very different numbers of features. A metagenomic feature table may contain thousands of taxa and genes, while a metabolomic table may contain hundreds of metabolites. The larger layer can dominate the analysis, and the smaller layers may contribute little to the result.

Early integration is most appropriate when the feature sets are balanced across omics layers and when the goal is classification or clustering. The evaluation of integration methods for cancer subtyping showed that the choice of data types and combinations affects performance, and that adding more omics data does not always improve results [17]. Researchers should test whether the inclusion of each omics layer improves or degrades the performance metric of interest.

Intermediate Integration

Intermediate integration reduces each omics layer to a lower-dimensional representation before combining them. Methods include canonical correlation analysis and its extensions, which find linear combinations of features from two datasets that are maximally correlated. The MintTea approach combines canonical correlation analysis extensions, consensus analysis, and an evaluation protocol to identify disease-associated multi-omic modules comprising features from multiple omics that shift in concord [12]. Applied to a metabolic syndrome study, MintTea identified a module with serum glutamate and TCA cycle-related metabolites along with bacterial species linked to insulin resistance [12].

Intermediate integration is well suited for identifying coordinated shifts across omics layers. The output is a set of modules or latent factors that can be tested for association with outcomes. The main challenge is parameter tuning. The number of components or factors must be chosen, and the results can be sensitive to this choice. Validation is essential to ensure that the identified modules are reproducible.

Late Integration

Late integration analyzes each omics layer independently and then combines the results. This approach is common in practice because it allows each layer to be analyzed with its optimal method. The results are combined by looking for overlapping features, shared pathways, or concordant direction of effects. The inflammatory bowel disease study used this type of approach by first identifying differentially abundant microbes and metabolites separately and then constructing multi-omics biological correlation maps to highlight gut microbial biotransformation deficiencies and alterations in aminoacyl-tRNA synthetases [5].

Late integration is appropriate for hypothesis generation and for studies where the omics layers have very different data structures. The main limitation is that it does not capture cross-omics interactions directly. A metabolite may be correlated with a microbial taxon even if neither shows a significant difference in single-omics analysis. Late integration would miss this relationship.

Pathway-Based Integration

Pathway-based integration maps features from each omics layer to biological pathways and tests for enrichment. This approach uses external knowledge from pathway databases to provide biological context. The watermelon soybean paste study identified valine, leucine, and isoleucine biosynthesis, starch and sucrose metabolism, glycolysis and gluconeogenesis, and pyruvate metabolism as the predominant metabolic pathways through functional composition analysis [21].

Pathway-based integration is useful for mechanistic interpretation. The main limitation is that pathway databases are incomplete, and many microbial genes and metabolites are not annotated. The results depend on the quality and completeness of the databases used.

Machine Learning Integration

Machine learning methods can integrate multi-omics data for classification, prediction, and biomarker discovery. These methods include random forests, support vector machines, and deep learning approaches. The application of artificial intelligence and machine learning to multi-omics datasets has potential for clinical implementation, including discovery of microbial biomarkers for disease classification or prediction, prediction of response to specific treatments, and fine-tuning of microbiome-modulating therapies [11].

The irritable bowel syndrome study used differential features to construct random forests classifiers [8]. The inflammatory bowel disease study identified multi-omics biomarkers for diagnosis that were validated across multiple global cohorts with area under the receiver operating characteristic curve values ranging from 0.92 to 0.98 [5]. These examples show the potential of machine learning for diagnostic applications.

The main risk with machine learning is overfitting. Models with many features and few samples can achieve perfect performance on training data but fail on new data. External validation is essential. The atherosclerosis study used multiple validation strategies, including 5-fold cross-validation, study-to-study transfer validation, and leave-one-study-out validation [6]. Researchers should adopt similar validation protocols and report performance metrics for both training and validation sets.

Workflow Design for Integration Projects

Step 1: Define the Biological Question

The biological question determines the integration strategy. A question about disease classification supports supervised machine learning. A question about coordinated shifts across omics layers supports intermediate integration. A question about specific metabolic pathways supports pathway-based integration. The question should be written down before data collection and used to guide all subsequent decisions.

Step 2: Inventory Available Data

List all available omics layers and their sample sizes. Check whether the same samples have data for all layers. If not, decide whether to analyze the complete-case subset or to use methods that handle missing data. The choice affects statistical power and the generalizability of the results.

Step 3: Preprocess Each Omics Layer

Apply the preprocessing steps appropriate for each omics layer. Document all parameters and software versions. Save the normalized feature tables in a standard format. The preprocessing steps should be reproducible by another researcher.

Step 4: Assess Batch Effects and Confounders

Test each omics layer for batch effects and associations with known confounders. Correct for batch effects if present. Include confounders as covariates in downstream models. The irritable bowel syndrome study adjusted for age, sex, race and ethnicity, body mass index, diet, and anxiety in multivariate models [8].

Step 5: Select the Integration Strategy

Choose the integration strategy based on the biological question, sample size, and data structure. Test multiple strategies if resources allow. The kidney disease study used two complementary integration methods, one unsupervised and one supervised, on the same data and identified shared enriched pathways including the complement and coagulation cascades, cytokine-cytokine receptor interaction pathway, and the JAK/STAT signaling pathway [13]. Using multiple methods on the same data can identify robust signals.

Step 6: Validate the Results

Validate the integration results using internal and external approaches. Internal validation includes cross-validation and permutation testing. External validation includes testing in independent cohorts or in independent samples from the same cohort. The inflammatory bowel disease study validated multi-omics biomarkers across multiple global cohorts [5]. The atherosclerosis study used study-to-study transfer validation and leave-one-study-out validation [6].

Step 7: Interpret and Report

Interpret the results in the context of the biological question and the limitations of the data. Report all preprocessing steps, integration parameters, and validation results. Deposit raw data and code in public repositories. The FAIR Guiding Principles provide a framework for data management that supports reproducibility [4].

Records and Measurements for Integration Projects

Documentation Standards

Every integration project should maintain a record of the following items: sample collection protocols, DNA extraction and sequencing parameters, metabolite extraction and mass spectrometry parameters, protein extraction and mass spectrometry parameters, bioinformatics software versions, preprocessing parameters, normalization methods, integration method parameters, and validation results. This documentation should be sufficient for another researcher to reproduce the analysis.

Quality Metrics

Each omics layer has quality metrics that should be recorded. For metagenomics, these include sequencing depth, number of reads passing quality filters, proportion of host reads, and number of taxa detected. For metabolomics, these include number of features detected, proportion of features annotated, and proportion of missing values. For proteomics, these include number of proteins detected, proportion of missing values, and coefficient of variation for technical replicates.

Reproducibility Checks

Reproducibility checks include running the same analysis pipeline on the same data twice and confirming identical results, testing the sensitivity of results to preprocessing parameters, and testing the sensitivity of results to integration method parameters. The evaluation of integration methods for cancer subtyping showed that the choice of data types and combinations affects performance [17]. Sensitivity analyses should be reported alongside the main results.

Common Failure Patterns in Multi-Omics Integration

Overfitting and Circular Analysis

Overfitting occurs when a model is tuned to perform well on the training data but fails on new data. This is a particular risk in machine learning approaches with many features and few samples. Circular analysis occurs when the same data are used for feature selection and model evaluation. The solution is to use nested cross-validation or to evaluate the model in an independent cohort.

Batch Effects Confounded with Biological Signal

Batch effects occur when samples are processed in groups and the processing differences create systematic variation in the data. If batch is confounded with the biological variable of interest, the batch effect cannot be distinguished from the biological signal. The solution is to randomize samples across batches and to include batch as a covariate in the analysis.

Feature Imbalance Across Omics Layers

Feature imbalance occurs when one omics layer has many more features than another. In early integration, the larger layer can dominate the analysis. The solution is to use intermediate integration methods that reduce each layer to a comparable number of features before combining them.

Ignoring Missing Data

Missing data are common in metabolomics and proteomics. Ignoring missing values by dropping samples or features can introduce bias. The solution is to use appropriate imputation methods or to use integration methods that handle missing data explicitly.

Overinterpreting Correlations

Correlation does not imply causation. A correlation between a microbial taxon and a metabolite may reflect a direct metabolic relationship, an indirect relationship through a third variable, or a spurious association. The inflammatory bowel disease study constructed multi-omics biological correlation maps to highlight gut microbial biotransformation deficiencies and alterations in aminoacyl-tRNA synthetases [5]. These maps are hypotheses, not proof. Experimental validation is required to confirm causal relationships.

Adding More Data Without a Plan

The intuition that more omics data always produces better results is not supported by evidence. Benchmarking studies show that there are situations where integrating more omics data negatively impacts the performance of integration methods [17]. Researchers should test whether each additional omics layer improves the performance metric of interest.

Interpretation Limits and Validation Requirements

Limits of Observational Data

Multi-omics integration studies are typically observational. They can identify associations between microbial taxa, metabolites, proteins, and outcomes, but they cannot establish causation. The inflammatory bowel disease study identified metabolites with significant differences whose roles in the disease are still unknown [5]. This uncertainty is inherent to observational research.

Mendelian Randomization for Causal Inference

Mendelian randomization is a method that uses genetic variants as instrumental variables to test for causal effects. A multi-omics Mendelian randomization study of Crohn's disease integrated oxidative stress gene expression, DNA methylation, and gut microbiota interaction to identify putative causal effects and underlying mechanisms [9]. This approach can strengthen causal claims but requires appropriate genetic instruments and careful sensitivity analysis.

External Validation as the Standard

External validation in independent cohorts is the standard for biomarker discovery. The inflammatory bowel disease study validated multi-omics biomarkers across multiple global cohorts [5]. The atherosclerosis study used multiple validation strategies to confirm diagnostic performance [6]. Researchers should plan for external validation from the start of the project.

The Role of Experimental Validation

Experimental validation in vitro or in vivo can confirm the expression patterns of key genes and the activity of key pathways. The idiopathic pulmonary fibrosis study used quantitative PCR and enzyme-linked immunosorbent assay in vitro, in vivo, and in patient plasma to confirm the expression patterns of key genes [10]. This type of validation is essential for translating computational findings into biological understanding.

Safety and Regulatory Context

Data Sharing Policies

Multi-omics data often include human genetic and clinical information. The NIH Genomic Data Sharing Policy sets expectations for data sharing, including informed consent, data de-identification, and data use limitations [3]. Researchers should review the applicable policies before collecting data and should plan for data deposition in appropriate repositories.

Data Repositories

The NCBI Data Resources provide repositories for sequence data, including the Sequence Read Archive and the database of Genotypes and Phenotypes [2]. The EMBL-EBI Training program offers guidance on data management and deposition [1]. Researchers should identify the appropriate repositories for each omics data type and should deposit data at the time of publication.

Ethical Considerations

Multi-omics studies involving human participants require ethical approval and informed consent. The consent should cover the intended uses of the data, including future research uses. The FAIR Guiding Principles support responsible data sharing by making data findable, accessible, interoperable, and reusable [4].

Professional Escalation Criteria

Researchers should escalate to a statistician or bioinformatics specialist when any of the following situations arise: the sample size is too small for the planned analysis, the batch effects cannot be corrected, the integration results are highly sensitive to preprocessing parameters, the machine learning models show signs of overfitting, or the interpretation requires causal claims that the data cannot support. The evaluation literature provides guidance on method selection and performance assessment [16][18]. A statistician can help with study design, power calculations, and validation planning.

Frequently Asked Questions

What is the difference between metagenomics and metabolomics?

Metagenomics measures the genetic material of the microbial community, providing information about which taxa are present and which functional genes they carry. Metabolomics measures the small molecules present in the sample, providing information about the metabolic activity of the community and the host. Metagenomics describes potential, while metabolomics describes activity. The two layers are complementary because a gene can be present without being expressed, and a metabolite can be produced by multiple species.

Which integration method should I use for my study?

The choice of integration method depends on the biological question, the sample size, and the data structure. For classification or prediction with a clear outcome, supervised machine learning methods are appropriate. For identifying coordinated shifts across omics layers, intermediate integration methods such as canonical correlation analysis are appropriate. For mechanistic interpretation, pathway-based integration is appropriate. Benchmarking studies show that no single method works for every project and that adding more omics data does not always improve performance [17].

How many samples do I need for a multi-omics integration study?

The required sample size depends on the expected effect size, the number of features, and the analysis method. Machine learning approaches require larger sample sizes to avoid overfitting. Cross-cohort validation is a common strategy to address the limitations of single-cohort studies. The inflammatory bowel disease study analyzed 9 metagenomic and 4 metabolomics cohorts [5]. The atherosclerosis study used 456 metagenomic samples and 111 16S rRNA gene sequencing samples [6]. Researchers should consult a statistician during study design.

How do I handle missing data in metabolomics and proteomics?

Missing data are common in metabolomics and proteomics. The first step is to determine whether the missingness is random or systematic. Systematic missingness can indicate technical problems that require correction. For random missingness, imputation methods can be used. The choice of imputation method affects downstream analyses, and the proportion of missing values should be reported.

How do I validate the results of a multi-omics integration analysis?

Validation can be internal or external. Internal validation includes cross-validation and permutation testing. External validation includes testing in independent cohorts or in independent samples from the same cohort. The inflammatory bowel disease study validated multi-omics biomarkers across multiple global cohorts [5]. The atherosclerosis study used 5-fold cross-validation, study-to-study transfer validation, and leave-one-study-out validation [6]. External validation is the standard for biomarker discovery.

What are the common causes of false findings in multi-omics integration?

Common causes include overfitting, batch effects confounded with biological signal, feature imbalance across omics layers, ignoring missing data, and overinterpreting correlations. Benchmarking studies show that adding more omics data does not always improve performance [17]. Researchers should test the sensitivity of results to preprocessing parameters and integration method parameters.

Can multi-omics integration establish causal relationships?

Multi-omics integration studies are typically observational and can identify associations but cannot establish causation. Mendelian randomization is a method that can strengthen causal claims by using genetic variants as instrumental variables. A multi-omics Mendelian randomization study of Crohn's disease used this approach to identify putative causal effects of oxidative stress genes [9]. Experimental validation in vitro or in vivo is required to confirm causal relationships.

What should I report in a multi-omics integration study?

Report all preprocessing steps, normalization methods, integration parameters, and validation results. Deposit raw data and code in public repositories. The FAIR Guiding Principles provide a framework for data management that supports reproducibility [4]. The NIH Genomic Data Sharing Policy sets expectations for data sharing for human data [3]. Complete documentation allows another researcher to reproduce the analysis.

Related Bioinformatics Guides

• Multi-Omics Integration Strategies
• The Rise of Omics: Genomics, Proteomics, and Metabolomics
• The KEGG Database and Pathway Analysis
• Metagenomics and Environmental DNA (eDNA) Analysis
• Alternative Splicing Analysis from RNA-Seq Data

References and Further Reading

• EMBL-EBI Training. European Bioinformatics Institute.
• NCBI Data Resources. National Center for Biotechnology Information.
• Genomic Data Sharing Policy. National Institutes of Health.
• The FAIR Guiding Principles. Scientific Data.
• Microbiome and metabolome features in inflammatory bowel disease via multi-omics integration analyses across cohorts.. Nature communications, 2023.
• Multi-omics integration reveals functional signatures of gut microbiome in atherosclerosis.. Gut microbes, 2025.
• Integrating multi-omics to unravel host-microbiome interactions in inflammatory bowel disease.. Cell reports. Medicine, 2024.
• Multi-omics profiles of the intestinal microbiome in irritable bowel syndrome and its bowel habit subtypes.. Microbiome, 2023.
• Oxidative stress gene expression, DNA methylation, and gut microbiota interaction trigger Crohn's disease: a multi-omics Mendelian randomization study.. BMC medicine, 2023.
• Multi-omics integration and machine learning reveal gut-immune signatures in idiopathic pulmonary fibrosis: insights from bulk RNA-seq, single-cell profiles, spatial transcriptomics, and experimental validation.. Frontiers in immunology, 2026.
• Machine Learning and Artificial Intelligence in the Multi-Omics Approach to Gut Microbiota.. Gastroenterology, 2025.
• Multi-omic integration of microbiome data for identifying disease-associated modules.. Nature communications, 2024.
• Leveraging complementary multi-omics data integration methods for mechanistic insights in kidney diseases.. 2025.
• Multi-omics data integration methods and their applications in psychiatric disorders.. 2023.
• Evaluation and Comparison of Multi-Omics Data Integration Methods for Subtyping of Cutaneous Melanoma.. 2022.
• Unsupervised Multi-Omics Data Integration Methods: A Comprehensive Review.. 2022.
• Evaluation and comparison of multi-omics data integration methods for cancer subtyping.. 2021.
• More Is Better: Recent Progress in Multi-Omics Data Integration Methods.. 2017.
• Metagenomics-metabolomics integration reveals the impacts and action mechanisms of sulfamethoxazole in a continuous-flow aerobic granular sludge membrane bioreactor: System performance and microbial functional shifts. Chemical Engineering Journal, 2025.
• Enzyme-assisted ultrasonication pretreated co-fermentation (Lactobacillus plantarum and Acetobacter pasteurianus) of jujube-based vinegar production: Optimization, volatile metabolomics, sensory, and integration of microbial pathways using metagenomics. Food Bioscience, 2025.
• Integration of metagenomics and targeted metabolomics reveals the flavor metabolism network of the microbial community in traditional watermelon soybean paste.. Food Research International, 2025.
• Advances in the integration of metabolomics and metagenomics for human gut microbiome and their clinical applications. Trends in Analytical Chemistry (TrAC), 2023.
• Integration of volatile flavor metabolomics and metagenomics reveals microbial-enzymatic pathways governing key aromatic volatile compound biosynthesis in Hongqujiu fermentation. Food chemistry: X, 2026.
• EasyMultiProfiler: an efficient multi-omics data integration and analysis workflow for microbiome research. Science China Life Sciences, 2026.
• Multi-omics decodes host-specific and environmental microbiome interactions in sepsis. Frontiers in Microbiology, 2025.
• AI/ML-enabled multi-omics integration of host genetics, immunity, and the gut microbiome in Crohn's disease: From diagnosis to theranostics. Slas Technology, 2026.

This article is educational and does not replace validated analysis plans, institutional policy, clinical interpretation, or specialist review.