Multi-Omics Integration in Disease Research: A Workflow for Biomarker Discovery

Multi-omics integration combines genomic, transcriptomic, proteomic, metabolomic, and other molecular data layers to identify biomarkers and understand disease mechanisms. This article provides a practical workflow for researchers applying multi-omics approaches to disease studies, covering study design, data integration strategies, machine learning methods, validation requirements, and common pitfalls. The workflow emphasizes reproducible practices, quality control at each stage, and clear criteria for deciding when findings are robust enough to warrant further investigation.

At a Glance

Scope and Context for Multi-Omics Studies

Multi-omics integration addresses a fundamental limitation of single-omics analyses. A single molecular layer often captures only part of the biological variation driving disease. Genomic data reveal risk variants and mutations, transcriptomics show gene expression changes, proteomics identify protein abundance alterations, and metabolomics capture downstream metabolic consequences. Each layer provides complementary information about disease state.

The value of integration has been demonstrated across multiple disease areas. In inflammatory bowel disease, researchers analyzing nine metagenomic and four metabolomics cohorts identified consistent gut microbiota characteristics and constructed multi-omics biological correlation maps that highlighted microbial biotransformation deficiencies and alterations in aminoacyl-tRNA synthetases. The resulting multi-omics biomarkers achieved diagnostic performance with AUROC values ranging from 0.92 to 0.98 across multiple global cohorts [5]. This example illustrates how integration across data types can produce biomarkers with cross-population reliability that single-omics approaches may miss.

In chronic kidney disease research, a comprehensive analysis pipeline integrating plasma proteome and transcriptome data identified 32 proteins associated with the disease, with 12 proteins having prior Mendelian randomization support and 20 novel causal proteins not previously reported [6]. The integration of protein quantitative trait loci and expression quantitative trait loci enabled causal inference that neither data layer alone could provide.

For thyroid eye disease, a systematic review of 69 studies involving 1,363 patients and 1,504 controls found that multi-omics integration was pivotal in understanding molecular mechanisms and identifying diagnostic and prognostic markers with clinical potential [7]. The review highlighted how genomics, transcriptomics, proteomics, metabolomics, and microbiomics each contributed distinct biomarker candidates across various biological matrices.

Researchers should consider multi-omics integration when the research question requires understanding relationships between molecular layers, when single-omics findings are inconsistent across studies, or when the goal is patient stratification into molecular subtypes. The approach is particularly valuable for complex diseases where multiple biological processes contribute to pathogenesis.

Core Principles of Multi-Omics Integration

Biological Layer Complementarity

Each omics layer captures a different aspect of molecular biology. Genomics provides relatively stable information about genetic variants and mutations. Transcriptomics reflects dynamic gene expression in response to disease and environment. Proteomics measures the functional molecules that execute cellular processes. Metabolomics captures the downstream metabolic state influenced by both host and microbial activity.

In neurodegenerative disease research, integration of these layers has revealed how genetic risk variants such as APOE ε4 in Alzheimer's disease relate to downstream molecular changes including amyloid beta accumulation, hyperphosphorylated tau, and metabolic alterations like mitochondrial dysfunction [12]. The connections between layers provide mechanistic insight that isolated measurements cannot deliver.

For atherosclerosis, integrated analysis of microbiome and host transcriptomic data identified five microbe-metabolite-host gene tripartite associations involving five microbial genera, two metabolites, and two host genes. These associations were validated for diagnostic potential as noninvasive biomarkers using multiple validation strategies including 5-fold cross-validation and leave-one-study-out validation [11].

Data Heterogeneity and Harmonization

Multi-omics data arrive in fundamentally different formats. Genomic variant calls are categorical, gene expression measurements are continuous and often count-based, proteomic data may be intensity-based, and metabolomic data can be relative abundances. These differences require careful preprocessing before integration.

The FAIR Guiding Principles provide a framework for managing research data to ensure it is Findable, Accessible, Interoperable, and Reusable [4]. Applying these principles to multi-omics projects means using standard file formats, comprehensive metadata, persistent identifiers, and clear data use agreements. The EMBL-EBI Training resources offer guidance on data standards and submission practices for various omics data types.

Public data repositories such as NCBI Data Resources provide access to large-scale datasets including The Cancer Genome Atlas and Gene Expression Omnibus. These resources enable researchers to validate findings across independent cohorts and increase statistical power through meta-analysis.

Statistical Considerations

Multi-omics data present a high-dimensional problem. A typical study may measure tens of thousands of transcripts, thousands of proteins, and hundreds of metabolites from a relatively small number of patients. This creates the multiple testing problem and risk of overfitting.

The Study of a Prospective Adult Research Cohort in inflammatory bowel disease analyzed genomics, transcriptomics from gut biopsies, and proteomics from blood plasma across hundreds of patients. The machine learning model trained on this data successfully discriminated between ulcerative colitis and Crohn's disease, with the most predictive features representing both known and novel omics signatures [8]. The study design demonstrates that adequate sample sizes and careful feature selection are essential for reliable results.

Researchers should plan for sample sizes that support the intended analysis. Cross-validation and external validation are mandatory for any biomarker discovery claim. The atherosclerosis study used multiple validation approaches including study-to-study transfer validation and leave-one-study-out validation to confirm diagnostic performance [11].

Study Design and Cohort Selection

Defining the Research Question

The research question determines which omics layers to collect and how to integrate them. Biomarker discovery for diagnosis requires different data than biomarker discovery for treatment response prediction or patient stratification.

For esophageal cancer, researchers studying neoadjuvant chemoradiotherapy response noted that achieving pathologic complete response is significantly associated with improved survival outcomes. The need for biomarkers to predict treatment response drove the integration of genomic and transcriptomic sequencing with proteomic and metabolomic mass spectra [9]. The research question focused on prediction of treatment response, which shaped the study design and analytical approach.

In ovarian cancer, the OMICS-FUSE model integrated proteomic, transcriptomic, and methylomic data to predict treatment response. The early-fusion Random Forest model achieved an AUC of 0.939 with accuracy of 0.896 and F1 of 0.939, with performance comparable to or surpassing the best-performing single-omics models [14]. The multi-omics framework yielded superior balance across accuracy and F1 score, demonstrating the value of integration for treatment response prediction.

Cohort Selection and Sample Collection

Cohort selection should consider disease heterogeneity, population diversity, and availability of appropriate control groups. The inflammatory bowel disease cross-cohort analysis included nine metagenomic and four metabolomics cohorts from different populations, which enabled identification of consistent characteristics despite inter-study variation [5].

Sample types must be matched to the omics layers being measured. Blood plasma is commonly used for proteomics and metabolomics. Tissue biopsies provide transcriptomic and genomic data but are invasive. Fecal samples enable microbiome analysis. The choice of sample matrix affects which biomarkers can be discovered and their clinical utility.

For thyroid eye disease, the systematic review found biomarker candidates across various biological matrices, reflecting the systemic nature of the disease [7]. Researchers should consider which sample types are feasible for eventual clinical use when designing discovery studies.

Data Generation Planning

Data generation should follow established protocols with appropriate technical replicates and quality controls. Batch effects are a major concern in multi-omics studies, particularly when samples are processed over extended periods or across multiple centers.

The Genomic Data Sharing Policy from the National Institutes of Health provides requirements for data sharing and management that apply to NIH-funded research. Researchers should review these requirements during study design to ensure compliance with data deposition and access expectations.

Data Preprocessing and Quality Control

Per-Omics Quality Assessment

Each omics data type requires specific quality control procedures before integration. Genomic data need variant calling quality filters. Transcriptomic data require assessment of mapping rates, gene coverage, and library complexity. Proteomic data need evaluation of protein identification confidence and quantification accuracy. Metabolomic data require assessment of peak quality and batch effects.

The EMBL-EBI Training resources provide detailed guidance on quality assessment for various omics data types. Researchers should document all quality control steps and decisions to ensure reproducibility.

Normalization and Batch Correction

Normalization methods differ by data type. Transcriptomic data commonly use library size normalization or more sophisticated methods that account for composition effects. Proteomic data may use intensity-based normalization or spike-in controls. Metabolomic data often require log transformation and scaling.

Batch correction is essential when samples are processed in multiple batches or centers. The choice of batch correction method depends on the experimental design and whether batch information is balanced across disease groups. Researchers should be cautious about overcorrecting, which can remove biological signal along with technical variation.

Feature Filtering

Feature filtering reduces the dimensionality of each omics layer before integration. Common approaches include removing features with low detection rates, low variance, or high missingness. For genomic data, filtering may focus on variants with sufficient minor allele frequency or genes with known functional relevance.

The lung adenocarcinoma study used bulk, single-cell, and spatial transcriptomics analyses to identify consensus differentially expressed genes before applying machine learning algorithms. The integration of multiple datasets and filtering steps identified 159 consensus differentially expressed genes, from which four core genes were selected by all four machine learning algorithms [10].

Data Integration Strategies

Early Integration

Early integration concatenates all omics features into a single matrix before analysis. This approach preserves all information but creates a very high-dimensional feature space. Early integration works well when the number of samples is large relative to the number of features or when feature selection methods can effectively reduce dimensionality.

The OMICS-FUSE model in ovarian cancer used early fusion of proteomic, transcriptomic, and methylomic data. The early-fusion Random Forest model achieved excellent predictive accuracy, demonstrating that this approach can work effectively when implemented carefully [14].

Intermediate Integration

Intermediate integration transforms each omics layer into a common representation before combining them. This may involve dimensionality reduction of each layer separately, followed by concatenation of the reduced representations. Intermediate integration can reduce noise and computational burden while preserving layer-specific structure.

Late Integration

Late integration builds separate models for each omics layer and combines their predictions. This approach is useful when omics layers have very different data structures or when interpretability of individual layer contributions is important. Late integration can be more robust to missing data in individual layers.

Graph-Based and Network Approaches

Graph-based methods represent biological entities as nodes and their relationships as edges. These approaches can incorporate prior knowledge about molecular interactions and pathways. An adaptive graph learning approach with attention mechanisms has been developed for patient classification and biomarker identification [25]. Explainable graph neural networks can integrate multi-omics data with prior knowledge to identify biomarkers while providing interpretable results [27].

For Alzheimer's disease research, the MOGAD framework uses integrated multi-omics and graph attention for biomarker discovery [26]. Graph-based approaches are particularly valuable when the research question involves understanding relationships between molecular entities across omics layers.

Deep Learning Approaches

Deep learning methods can capture complex nonlinear relationships across omics layers. An autoencoder deep neural network combined with model-based clustering was used to integrate gene expression and somatic mutation data in lung cancer, identifying molecularly defined subtypes that enhanced understanding of tumor heterogeneity [21].

A graph-based deep learning framework with gating and omics-linked attention has been developed for multi-omics integration and biomarker discovery [24]. These methods can identify patterns that linear methods miss but require careful validation to avoid overfitting.

Machine Learning for Biomarker Discovery

Algorithm Selection

The choice of machine learning algorithm depends on the data characteristics and research question. The head and neck squamous cell carcinoma study employed four machine learning algorithms including LASSO, SVM-RFE, XGBoost, and Boruta to screen for core candidate genes. The four core genes selected by all four algorithms were considered robust candidates [10].

For the inflammatory bowel disease study, a machine learning model was trained to classify ulcerative colitis versus Crohn's disease samples. The high performance of the model showed that multi-omics signatures can discriminate between the two indications [8].

Feature Selection and Interpretability

Feature selection identifies the most informative features for the prediction task. SHAP analysis provides interpretability by quantifying each feature's contribution to predictions. In the ovarian cancer study, SHAP analysis identified key determinants of treatment response including CLEC2A, MYH4, and methylation of SYT12_1, with functional enrichment implicating immune regulation, metabolic pathways, and drug resistance signaling [14].

For the lung adenocarcinoma study, the Lasso plus StepCox algorithm identified nine super-enhancer-related prognostic genes and built a prognostic model. The model was linked to genomic mutations, oncogenic pathway enrichment, tumor microenvironment regulation, and immune response patterns [13].

Cross-Validation and Avoiding Data Leakage

Proper cross-validation is essential for reliable performance estimates. Data leakage occurs when information from the test set influences model training, leading to inflated performance estimates. Common sources of leakage include normalization performed on the full dataset before splitting, feature selection using all samples, and duplicate samples across train and test sets.

The atherosclerosis study used 5-fold cross-validation, study-to-study transfer validation, and leave-one-study-out validation to confirm diagnostic performance of the five microbial genera identified as noninvasive biomarkers [11]. This multi-layered validation approach provides stronger evidence than a single cross-validation scheme.

Case Study: Multi-Omics Integration in Inflammatory Bowel Disease

The inflammatory bowel disease research provides a practical example of multi-omics integration for biomarker discovery. The disease comprises Crohn's disease and ulcerative colitis, which are complex conditions with diverse manifestations. Recent advances in multi-omics technologies have helped researchers unravel molecular characteristics to develop targeted treatments [8].

Study Design

The Study of a Prospective Adult Research Cohort analyzed genomics, transcriptomics from gut biopsy samples, and proteomics from blood plasma from hundreds of patients. This multi-omics cohort provided the data foundation for both classification and patient subgroup analysis [8].

Analytical Approach

The researchers trained a machine learning model to classify ulcerative colitis versus Crohn's disease samples. In parallel, they integrated multi-omics data to unveil patient subgroups in each indication independently and analyzed the molecular phenotypes of these patient subpopulations [8].

Findings and Validation

The high performance of the model demonstrated that multi-omics signatures can discriminate between the two indications. The most predictive features included both known and novel omics signatures for inflammatory bowel disease, with potential use as diagnostic biomarkers. Patient subgroup analysis uncovered omics features associated with disease severity in ulcerative colitis patients and with tissue inflammation in Crohn's disease patients. The analysis culminated in the observation of two Crohn's disease subpopulations characterized by distinct inflammation profiles [8].

Cross-Cohort Validation

A separate cross-cohort integrative analysis of nine metagenomic and four metabolomics cohorts from different populations identified consistent characteristics of commensal gut microbiota. Three bacteria rarely reported in inflammatory bowel disease were identified, including Asaccharobacter celatus, Gemmiger formicilis, and Erysipelatoclostridium ramosum. Metagenomic functional analysis revealed that an essential gene of the two-component system pathway, linked to fecal calprotectin, is implicated in the disease. Metabolomics analysis showed 36 identified metabolites with significant differences. Multi-omics biomarkers for diagnosis were validated across multiple global cohorts with AUROC values ranging from 0.92 to 0.98 [5].

Validation and Reproducibility

External Validation Requirements

External validation in independent cohorts is the gold standard for biomarker discovery. The chronic kidney disease study validated 32 proteins associated with the disease across diverse datasets, kidney function indicators, and clinical types. Twelve proteins with prior Mendelian randomization support were confirmed, including fibroblast growth factor 5, uromodulin, and complement components [6].

For the Alzheimer's disease study, the diagnostic potential of individual feature genes was evaluated in an external validation dataset. Spatial transcriptomic analyses revealed cell type-specific expression patterns in brain tissue, with IGF1R, ASCC3, and TWSG1 showing potential co-localization in oligodendrocytes [16].

Experimental Validation

Computational findings require experimental validation to confirm biological relevance. The head and neck squamous cell carcinoma study validated the expression of the key gene SASH1 at the protein level via Western blot in cell lines. The clinical and therapeutic value was assessed through survival, clinical correlation, and drug sensitivity analyses [10].

The ovarian cancer study confirmed six hub genes including CASP8, AQP8, CAV1, FN1, CREB1, and KDR, exhibiting expression patterns associated with drug resistance, immune regulation, and prognosis [14].

Reproducibility Practices

Reproducibility requires documentation of all analytical steps, version control for code, and deposition of data and code. The lung cancer deep learning study made all source code, figures, latent features, cluster labels, and instructions for data download available on GitHub [21]. This level of transparency enables other researchers to reproduce and build upon the work.

The FAIR Guiding Principles provide a framework for making data Findable, Accessible, Interoperable, and Reusable [4]. Applying these principles to multi-omics projects ensures that data can be shared and integrated across studies.

Common Failure Patterns

Overfitting to Discovery Cohorts

The most common failure in multi-omics biomarker discovery is overfitting to the discovery cohort. Models that perform excellently in cross-validation often fail in external validation. This occurs when models capture cohort-specific noise instead of disease-relevant signal.

Prevention strategies include using multiple cross-validation schemes, external validation in independent cohorts, and regularization methods that penalize complex models. The atherosclerosis study's use of study-to-study transfer validation and leave-one-study-out validation provides a model for rigorous performance assessment [11].

Batch Effects Confounded with Disease Status

When batch information is confounded with disease status, batch correction methods cannot distinguish technical from biological variation. This situation arises when all cases are processed in one batch and all controls in another. Study designs should balance cases and controls across batches whenever possible.

Inconsistent Findings Across Studies

The inflammatory bowel disease cross-cohort analysis noted that inconsistent findings across studies impede comprehensive understanding of disease mechanisms [5]. Methodological heterogeneity across studies and the absence of consensus regarding disease-specific microbial signatures were identified as research limitations in the atherosclerosis study [11].

Standardization of analytical pipelines and reporting guidelines can reduce inter-study variability. Researchers should document all preprocessing and analysis decisions to enable comparison across studies.

Ignoring Missing Data Structure

Multi-omics studies frequently have missing data, particularly when some samples fail quality control for specific omics layers. The pattern of missingness matters. Missing completely at random can be handled with simple imputation. Missing not at random, such as when degraded samples fail for specific assays, requires more sophisticated approaches.

Limitations and Interpretation

Correlation Versus Causation

Multi-omics integration identifies associations between molecular features and disease. These associations may or may not be causal. The chronic kidney disease study used Mendelian randomization and summary-based Mendelian randomization to strengthen causal inference for identified proteins [6]. The Alzheimer's disease study integrated Mendelian randomization, summary-based Mendelian randomization, and differential expression analysis to identify core genetically associated genes [16].

Researchers should be cautious about claiming causal relationships without appropriate study designs or analytical methods that support causal inference.

Generalizability Across Populations

Biomarkers identified in one population may not perform equally in other populations. The inflammatory bowel disease cross-cohort analysis included cohorts from different populations to address this concern [5]. Researchers should validate biomarkers across diverse populations before clinical use.

Technical Limitations

Each omics technology has inherent limitations. Mass spectrometry-based proteomics may miss low-abundance proteins. RNA sequencing captures transcript abundance but not post-transcriptional regulation. Metabolomics detects only a fraction of the metabolome. These limitations affect which biomarkers can be discovered with each approach.

The esophageal cancer review noted current limitations, challenges, and future perspectives of multi-omics platforms to assist in experimental designs and clinical use [9]. Researchers should understand the technical capabilities and limitations of each platform when interpreting results.

Records and Documentation

Data Management Plans

Data management plans should be established at study initiation. The Genomic Data Sharing Policy from the National Institutes of Health specifies expectations for data sharing, including timelines for data deposition and access levels. Researchers should review applicable policies and funder requirements.

Analysis Documentation

Documentation should include software versions, parameter settings, and analysis scripts. The EMBL-EBI Training resources provide guidance on bioinformatics best practices including documentation standards.

Metadata Standards

Comprehensive metadata is essential for data interpretation and reuse. Metadata should include sample collection protocols, processing dates, batch information, and clinical annotations. The FAIR Guiding Principles emphasize the importance of rich metadata for data discoverability and reuse [4].

Professional Escalation Criteria

Researchers should escalate findings for further investigation when specific criteria are met. A candidate biomarker panel warrants additional validation when it demonstrates consistent performance across multiple cross-validation schemes and external cohorts. The inflammatory bowel disease biomarkers achieved AUROC values from 0.92 to 0.98 across multiple global cohorts, providing strong evidence for further development [5].

Findings warrant experimental validation when computational analyses identify consistent signals across multiple omics layers and analytical approaches. The head and neck squamous cell carcinoma study progressed from computational identification to Western blot validation when four machine learning algorithms consistently selected the same core genes [10].

Clinical translation should be considered when biomarkers demonstrate robust performance in external validation and have plausible biological mechanisms. The chronic kidney disease study identified proteins with prior Mendelian randomization support and novel causal proteins, providing a basis for therapeutic target development [6].

Frequently Asked Questions

What is the difference between early, intermediate, and late integration in multi-omics analysis?

Early integration concatenates all omics features into a single matrix before analysis, preserving all information but creating a high-dimensional feature space. Intermediate integration transforms each omics layer into a common representation before combining them, often through dimensionality reduction. Late integration builds separate models for each omics layer and combines their predictions. The choice depends on data structure, sample size, and whether interpretability of individual layer contributions is important. The OMICS-FUSE model in ovarian cancer used early fusion and achieved strong predictive accuracy [14].

How many samples are needed for a multi-omics biomarker discovery study?

Sample size requirements depend on the number of omics features measured, the expected effect size, and the analysis approach. Multi-omics data are high-dimensional, creating risk of overfitting with small sample sizes. The inflammatory bowel disease study analyzed hundreds of patients from the SPARC IBD cohort [8]. The thyroid eye disease systematic review encompassed 69 studies involving 1,363 patients and 1,504 controls [7]. Researchers should perform power calculations based on the specific analysis planned and consider external validation requirements.

How do I handle batch effects in multi-omics data?

Batch effects should be assessed before integration. Methods include principal component analysis to visualize batch structure and statistical tests for batch association. Batch correction methods should be applied carefully to avoid removing biological signal. Study designs should balance cases and controls across batches whenever possible. The cross-cohort integrative analysis in inflammatory bowel disease addressed inter-study variation by analyzing multiple cohorts from different populations [5].

What machine learning algorithms work best for multi-omics integration?

No single algorithm works best for all situations. The head and neck squamous cell carcinoma study used four algorithms including LASSO, SVM-RFE, XGBoost, and Boruta, selecting genes identified by all four as robust candidates [10]. The ovarian cancer study evaluated five machine learning algorithms with the early-fusion Random Forest model achieving the best performance [14]. Deep learning approaches such as autoencoders have been used for feature extraction and clustering in lung cancer [21]. Algorithm selection should be based on data characteristics and validation performance.

How do I validate multi-omics biomarkers?

Validation requires multiple approaches. Internal validation uses cross-validation schemes such as k-fold cross-validation. External validation tests the model in independent cohorts. The atherosclerosis study used 5-fold cross-validation, study-to-study transfer validation, and leave-one-study-out validation [11]. Experimental validation confirms biological relevance through methods such as Western blot or quantitative PCR. The head and neck squamous cell carcinoma study validated SASH1 expression at the protein level via Western blot [10].

What is data leakage and how do I avoid it?

Data leakage occurs when information from the test set influences model training, leading to inflated performance estimates. Common sources include normalization performed on the full dataset before splitting, feature selection using all samples, and duplicate samples across train and test sets. Avoid leakage by performing all preprocessing steps within cross-validation folds and ensuring no sample overlap between training and test sets.

How do I interpret multi-omics integration results biologically?

Biological interpretation requires mapping identified features to known pathways and processes. The chronic kidney disease study combined protein-protein interaction, Gene Ontology, and single-cell annotation to explore potential biological roles of identified proteins [6]. The inflammatory bowel disease study constructed multi-omics biological correlation maps to elucidate relationships between gut microbiota and metabolites [5]. Spatial transcriptomics can reveal cell type-specific expression patterns, as demonstrated in the Alzheimer's disease study where IGF1R, ASCC3, and TWSG1 showed potential co-localization in oligodendrocytes [16].

What are the main limitations of multi-omics integration?

Key limitations include data heterogeneity across studies, lack of standardization, high dimensionality relative to sample size, and correlation versus causation ambiguity. The atherosclerosis study noted methodological heterogeneity across studies and absence of consensus regarding disease-specific microbial signatures [11]. The neurodegenerative disease chapter highlighted challenges such as data heterogeneity and limited standardization [12]. Researchers should design studies with these limitations in mind and use appropriate validation strategies.

Related Bioinformatics Guides

• Multi-Omics Integration Strategies
• Network Pharmacology Approaches to Multi-Target Drug Discovery
• Machine Learning Approaches in Structural Virology: Predicting Viral Protein Dynamics
• Machine Learning-Guided Design of Pan-Coronavirus Spike Protein Inhibitors: From Sequence to Structure
• The Protein Data Bank (PDB): Archival Standards, Structural Validation Metrics, and Bioinformatics Integration Protocols

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.
• Identification of novel therapeutic targets for chronic kidney disease and kidney function by integrating multi-omics proteome with transcriptome.. Genome medicine, 2024.
• Multi-Omics Approaches to Discover Biomarkers of Thyroid Eye Disease: A Systematic Review.. International journal of biological sciences, 2024.
• Multi-omics data integration identifies novel biomarkers and patient subgroups in inflammatory bowel disease.. Journal of Crohn's & colitis, 2025.
• Multi-omics approaches for biomarker discovery in predicting the response of esophageal cancer to neoadjuvant therapy: A multidimensional perspective.. Pharmacology & therapeutics, 2024.
• A multi-omics pipeline integrating machine learning and spatial-cellular analysis identifies SASH1 as a prognostic biomarker and therapeutic target in head and neck squamous cell carcinoma.. International journal of surgery (London, England), 2025.
• Multi-omics integration reveals functional signatures of gut microbiome in atherosclerosis.. Gut microbes, 2025.
• Multi-omics integration in disease research.. Progress in brain research, 2025.
• Multi-omics machine learning-driven investigation of super-enhancers signatures and prognostic biomarkers in lung adenocarcinoma.. 2026.
• Multi-omics fusion with machine learning enables robust prediction of treatment response in ovarian cancer for precision population health.. 2026.
• Editorial: Refining precision medicine through AI and multi-omics integration.. 2026.
• A multi-omics framework integrating gut microbiota, blood metabolites, and immune cells to elucidate the pathogenesis of Alzheimer's disease.. 2026.
• The Nutri-Exposome Intelligence Framework: Integrating Multi-Omics, Machine Learning, and Digital Nutrition for Precision Chronic Disease Prevention.. 2026.
• Multi-Modal, Machine Learning-Driven Framework Integrating Multi-Omics for Personalized Chronic Kidney Disease Management.. 2026.
• AI-Driven Multi-Omics Integration for Enhanced Drug Discovery Pipelines. 2025 International Conference on Multi-Agent Systems for Collaborative Intelligence (ICMSCI), 2025.
• Multi-Omics Pipeline and Omics-Integration Approach to Decipher Plant’s Abiotic Stress Tolerance Responses. Genes, 2023.
• Deep learning multi-omics integration identifies new molecular subtypes of lung cancer. BioData Mining, 2026.
• Precision Medicine and Multi-Omics Integration: Transforming Drug Discovery Through FAIR-Enabled Systems. The EuroBiotech Journal, 2026.
• Integrated multi-omics analysis reveals distinct microbiota-metabolite signatures and a novel HCN2-2-hydroxybutyric acid interaction in inflammatory bowel disease. Frontiers in Nutrition, 2026.
• A Graph-Based Deep Learning Framework with Gating and Omics-Linked Attention for Multi-Omics Integration and Biomarker Discovery. Biology, 2025.
• Integration of multi-omics data using adaptive graph learning and attention mechanism for patient classification and biomarker identification. Computers in Biology and Medicine, 2023.
• MOGAD: Integrated Multi-Omics and Graph Attention for the Discovery of Alzheimer’s Disease’s Biomarkers. Informatics, 2025.
• Effective integration of multi-omics with prior knowledge to identify biomarkers via explainable graph neural networks. Npj Systems Biology and Applications, 2025.

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