From molecules to minds: Integrative multi-omics in psychiatry Authors: Hossein Abbasi, Sage E Hawn, Arash Javanbakht, Soraya Seedat, Kyle Bourassa, Sinead M Sinnott, Antonia V Seligowski, Sian Hemmings, Nathan A Kimbrel, Erika Wolf, Alicia K Smith, Leslie Brick, Divya Mehta Published: 2026 ABSTRACT AbstractPsychiatric disorders are biologically complex conditions arising from interactions across genomic, epigenomic, transcriptomic, proteomic, metabolomic, and metagenomic layers. Single-omics approaches rarely capture more than a fraction of the variance in complex conditions, underscoring the importance of integrative multi-omics frameworks. This mini-review summarizes key methodologies and their application in psychiatric research, with a focus on systems-level integration of genomic risk scores, transcriptomic networks, and neuroimaging data to advance biological understanding of disorders such as depression, schizophrenia, and Alzheimer’s disease. We also outline the infrastructural requirements for effective multi-omics research, including standardized biobanking, Laboratory Information Management Systems, adherence to FAIR data principles, and federated learning approaches for privacy-preserving analysis. Importantly, we highlight the need for greater global inclusivity in psychiatric genomics. Current datasets are heavily biased toward relatively high-resourced and predominantly White, non-Hispanic populations, limiting generalizability. Initiatives such as the Psychiatric Genomics Consortium-Africa and H3ABioNet demonstrate how locally led efforts can strengthen capacity, promote data sovereignty, and support equitable research practices. Advancing multi-omics psychiatry will require coordinated investment in infrastructure, training, and inclusive international collaboration. This mini-review serves primarily as a conceptual roadmap, highlighting what integrative approaches have demonstrated so far and future directions for the field.Keywords: Multi-omic, Dimensional, Systems-Level, Integrated Biological Models Overview of multi-omics approaches Multi-omics approaches integrate genomics, epigenomics, transcriptomics, proteomics, metabolomics, and/or metagenomics to deliver a comprehensive understanding of disease mechanisms [1]. In this review, we define integrative multi-omics as the joint analysis of two or more biological omics layers (e.g., genomics, epigenomics, transcriptomics, proteomics, metabolomics, or metagenomics) within a unified analytical framework to investigate disease biology. We distinguish this from multimodal integration, which combines omics data with non-molecular modalities such as neuroimaging or clinical phenotypes, and from single-omics studies that interrogate only one molecular layer. That said, selected multimodal and single-domain examples are included throughout this review where they illustrate important conceptual advances relevant to systems-level psychiatry. Genomics (e.g., genome-wide association studies, polygenic risk scores, copy number variants, and structural variants) identifies genetic variants associated with psychiatric risk; pharmacogenetic analyses further inform how genetic variation shapes drug metabolism and treatment response [2]. Epigenomic analyses, such as DNA methylation and chromatin accessibility, help show how environmental exposures and development shape gene regulation and expression; importantly, methylation risk scores (MRSs) can also serve as proxies for unmeasured environmental exposures, capturing cumulative epigenetic burden linked to early-life adversity and other environmental risks [3]. Transcriptomics includes both bulk RNA sequencing (RNA-seq) and single-cell/spatial transcriptomic approaches. Bulk RNA-seq characterizes differential gene expression and reveals how genetic and epigenetic variation influence transcription, whereas single-cell and spatial methods resolve cell-type-specific expression patterns linked to psychiatric phenotypes [4]. Proteomics captures post-translational modifications and cellular states that often align more closely with disease states than gene expression alone. Metabolomics highlights biochemical pathways (e.g., inflammation) involved in psychiatric disorders; untargeted approaches can additionally detect exogenous metabolites reflecting environmental exposures such as toxicants, while targeted metabolomics offers greater quantitative precision for specific pathways [5]. Metagenomic studies of the gut microbiome have linked microbial communities to psychiatric outcomes via the gut–brain axis. Large collaborative projects, such as the PsychENCODE consortium, illustrate the power of multi-omics by mapping regulatory networks and showing how diverse molecular layers contribute to psychiatric pathology [6]. Studies cited in this review were identified through searches of PubMed and Google Scholar using terms including 'multi-omics psychiatry,' 'integrative genomics,' and 'psychiatric machine learning,' prioritizing peer-reviewed work published between 2015 and 2025 that illustrates key methodological approaches or integration strategies relevant to psychiatric research. Table 1 summarizes the key multi-omics methods and their applications in psychiatric research. Table 1Major multi-omics modalities and their contributions to psychiatric research.Multi-Omics MethodUnique ContributionGenomics (GWAS, polygenic risk scores, CNVs/structural variants, pharmacogenetics)Captures inherited liability, polygenic burden, and rare structural variation. Pharmacogenetics links genetic variation to drug metabolism and treatment response [2].Epigenomics (DNA methylation, chromatin accessibility)Reflects environmentally responsive regulation of gene expression and cumulative epigenetic burden. MRSs serve as proxies for unmeasured environmental exposures such as early-life adversity [3].Transcriptomics (Bulk RNA-seq, Single-cell and spatial)Measures dynamic gene expression across tissues, cell types, and spatial contexts. Bulk RNA-seq identifies system-level dysregulation; single-cell/spatial approaches resolve cell-type specificity [4].ProteomicsCaptures functional protein states and post-translational modifications that may track disease biology more closely than transcription alone.Metabolomics (Targeted and Untargeted)Highlights biochemical pathways (e.g., inflammation) involved in psychiatric disorders. Untargeted metabolomics can capture exogenous metabolites (e.g., toxicant exposures), while targeted metabolomics offers quantitative precision for specific pathways [5].Metagenomics (Gut Microbiome)Links gut microbial communities to psychiatric outcomes via the gut–brain axis. Major multi-omics modalities and their contributions to psychiatric research. Systems-level thinking: toward integrated biological models A systems biology perspective views psychiatric disorders as the result of dynamic, multi-layered interactions across environmental, molecular, and neural systems, shifting classification from symptom-based categories to biologically informed dimensions. Because psychiatric diagnosis often relies entirely on subjective conscious symptoms and observable clinical symptoms due to a lack of objective markers, employing a systems-level approach is crucial [7]. Initiatives like PsychENCODE and BrainSeq have generated integrated regulatory maps that link noncoding genomic variation to transcriptional effects in neuropsychiatric disorders through chromatin interactions and epigenetic modifications [6], [8]. By combining genomic risk scores with brain transcriptomic networks and neuroimaging data, integrative models have helped link genetic risk to effects on specific neural cell types and developmental pathways, marking an important shift toward biologically informed frameworks. This represents an important shift from symptom-based diagnosis toward frameworks informed by underlying biology. A clear example of this approach integrates GWAS summary statistics for schizophrenia, bipolar disorder, and major depressive disorder with a large-scale single-cell transcriptomic atlas of the adult human brain [9]. This approach highlighted neuronal subtypes showing enriched associations with disorder-associated genetic risk, including deep-layer excitatory neurons, hippocampal CA1–3 neurons, and amygdala excitatory neurons. These cell types were then mapped onto brain regions, and the implicated regions were examined using resting-state fMRI, where hippocampal and amygdala connectivity consistently differentiated schizophrenia cases from controls across two independent datasets [9]. Together, these findings illustrate how genomic, transcriptomic, and neuroimaging data can be combined to explore patterns relevant to psychiatric risk. However, the findings remain associative, and longitudinal or experimental studies will be required to determine causal relationships. Building on this, schizophrenia research has also explored how molecular subtyping can inform treatment stratification. Because schizophrenia is highly heterogeneous and many studies report immune system abnormalities that may relate to treatment response, one study used machine learning on blood-based inflammatory markers to classify patients into treatment-related subtypes: antipsychotic responsive, clozapine responsive, and clozapine resistant [10]. A support vector machine classifier distinguished the subgroups with AUC values up to 0.88. While this study relied on a single biological domain and was cross-sectional in design, it suggests that molecule profiles may offer a way to explore patterns related to treatment heterogeneity. Integrating such inflammatory markers with genomic, transcriptomic, and epigenomic data in future longitudinal studies could help map correlational patterns related to treatment resistance in schizophrenia, a question that single-domain approaches alone cannot fully address. Importantly, the predictive model in this study was evaluated using leave-one-out cross-validation within a single cohort, and independent external validation in separate samples will be essential before such classifiers can be considered clinically meaningful. Beyond psychiatry, multi-omics integration has also shown promise in closely related neurological conditions such as Alzheimer's disease, which shares important features with major psychiatric disorders, including polygenic risk architecture and the need for early biological stratification, making it a useful proof of concept [11]. One study combined polygenic risk scores and DNA methylation profiles in a Graph Convolutional Network model, using cognitive test scores to define relationships between samples, with the goal of classifying disease stages across cognitively normal, mild cognitive impairment, and Alzheimer's disease groups [11]. Adding genomic and epigenomic data on top of clinical information improved classification at each step, and the graph-based model performed better than traditional machine learning methods across all comparisons. This shows that combining multiple biological layers can capture patterns that single-domain models tend to miss. Whether similar approaches can work for psychiatric conditions like depression or schizophrenia, where biological complexity is equally high, is an important question for future research. Table 2 summarizes representative examples spanning true multi-omics, multimodal, and single-omics approaches to illustrate the current spectrum of biological data integration in psychiatry and related brain disorders. Table 2Representative examples illustrating multi-omics, multimodal, and single-domain biological integration in psychiatric and related brain disorders.StudyBiological Layers / ModalitiesType of IntegrationPrimary ObjectiveMain Finding (Evidence Type)Yao et al.Genomics (GWAS) + single-cell transcriptomics + resting-state fMRIMulti-modal (omics + neuroimaging)Biologically informed interpretation: linking genetic risk to brain cell types and regionsAssociative: Risk-enriched neuronal subtypes mapped to specific brain regions; hippocampal and amygdala connectivity differentiated schizophrenia cases across two independent datasetsYee et al.Proteomics (blood-based inflammatory markers only)Single-omics (proteomics)Patient stratification by treatment responsePredictive: SVM classifier distinguished antipsychotic-responsive, clozapine-responsive, and clozapine-resistant subtypes (AUC up to 0.88); included to illustrate stratification potential, not as a multi-omics exampleLi et al.Genomics (PRS) + epigenomics (DNA methylation) + clinical/cognitive dataMulti-omics and multimodal (genomic + epigenomic layers integrated with clinical cognitive data)Staging and early detection: classifying individuals across cognitively normal, mild cognitive impairment, and Alzheimer's disease groupsPredictive: graph convolutional network integrating genomic and epigenomic layers outperformed single-omics and traditional ML methods across all disease-stage comparisonsNote. AUC = area under the receiver operating characteristic curve; fMRI = functional magnetic resonance imaging; GWAS = genome-wide association study; ML = machine learning; PRS = polygenic risk score; SVM = support vector machine. Representative examples illustrating multi-omics, multimodal, and single-domain biological integration in psychiatric and related brain disorders. Note. AUC = area under the receiver operating characteristic curve; fMRI = functional magnetic resonance imaging; GWAS = genome-wide association study; ML = machine learning; PRS = polygenic risk score; SVM = support vector machine. Methodological challenges and considerations Advances in next-generation multi-omics technologies and computational methods, such as network-based models (e.g., WGCNA, Bayesian networks), multi layered networks, graph-based AI, and meta-analytical strategies, have dramatically accelerated integrative research. Nevertheless, several methodological challenges continue to limit the robustness, reproducibility, and clinical translation of these approaches. First, challenges surrounding biological variability are substantial. Molecular signatures derived from peripheral tissues (e.g., blood or saliva) may not accurately reflect processes occurring in the brain. Beyond general tissue-type differences, there is also potential for sample-level mismatch, where peripheral and brain tissue from the same individual are not truly equivalent biological replicates, further complicating cross-modal inference. Moreover, molecular profiles are dynamic across development, disease progression, and treatment, complicating biological interpretation [1], [12]. Second, most published multi-omics studies rely on associative designs, and distinguishing biological associations from causal mechanisms remains difficult even in longitudinal studies, since temporal precedence alone does not establish causality. Experimental and intervention-based designs will ultimately be needed to support causal conclusions [13]. Third, computational integration presents significant challenges because multi-omics datasets are high dimensional, frequently contain missing data, and are susceptible to batch effects introduced during sample collection, laboratory processing, sequencing, and data preprocessing. Without rigorous harmonization, appropriate feature selection, independent external validation, and careful prevention of overfitting and data leakage, machine learning models may produce overly optimistic estimates of predictive performance. Finally, important translational challenges remain [1], [12]. Most psychiatric genomic studies have been conducted in populations of predominantly European ancestry, limiting the portability of polygenic risk scores and other predictive models across diverse populations. Addressing these biological, computational, and translational challenges through standardized analytical workflows, larger longitudinal cohorts, more diverse study populations, and independent replication will be essential before integrative multi-omics approaches can be routinely translated into clinical psychiatry [14] Infrastructure needs: biobanking, sample tracking, and data management The success of multi-omics approaches fundamentally relies on robust infrastructure encompassing biospecimen management, data tracking, and computational integration. This critically includes the implementation of standardized, large-scale biobanking initiatives to ensure quality control across diverse biospecimens [15], coupled with ethical frameworks that comprehensively address informed consent, data privacy, and cultural sensitivity for truly inclusive research [16]. Furthermore, Laboratory Information Management Systems (LIMS) are indispensable for precisely linking biospecimens to their detailed metadata and processing parameters, thereby significantly enhancing both data integrity and traceability [17]. Cloud-based infrastructure plays a pivotal role by enabling harmonization and accelerating data processing across multiple sites [18]. Moreover, the adoption of FAIR (Findable, Accessible, Interoperable, and Reusable) data principles and standardized Application Programming Interfaces is crucial, as they facilitate seamless integration and interoperability of information [19]. In terms of data privacy, Federated learning enables multi-institutional analysis without sharing patient data, preserving privacy while maintaining performance [20], [21]. Looking to the future, this field is likely to move towards building real-time data integration pipelines that link biobanks directly with electronic health records and genomic databases and also include advanced AI-based quality control systems. In addition, ongoing training in both biobanking, data science, and bioethics will be essential to support the long-term development and success of multi-omics research. Need for more work among under-resourced populations In future studies, it will be important to focus more on low-income and displaced populations. At the same time, researchers need additional training to work with underserved groups in a culturally and contextually sensitive way. This includes understanding language-related nuances and avoiding simple generalizations based on race, nationality, ethnicity, socioeconomic status, or lived experience. Meaningful involvement of members of the target communities and those with lived experience is highly recommended in understanding the specific needs and challenges of the communities, as well as legal, political, cultural, literacy and language barriers. We highly recommend inclusive participation of members of the researched communities in research teams. Within consortium based genomic initiatives, locally led initiatives (e.g., Psychiatric Genomics Consortium-Africa, H3ABioNet) can serve to ensure relevance, capacity, data sovereignty, equitable data distribution, capacity strengthening, and advance LMIC leadership [22], [23]. For forcibly displaced populations and immigrant communities, it is important to involve bilingual and bicultural community members in both data collection and interpretation. Building mutual trust requires genuine collaboration with trusted local partners, such as resettlement agencies that work directly with refugees and other community service organizations. Sharing research results in ways that support community organizations for instance, by helping with educational materials or funding applications can create a mutually beneficial relationship. Partnerships between institutions in high-income countries and local clinics or universities in host countries can also encourage resource sharing, mutual learning, co-authorship, and shared leadership. Research infrastructure must be adapted to low-resource settings, including using mobile data collection tools, translated and culturally adapted assessments, and accessible biospecimens such as saliva, dried blood spots, or mobile psychophysiology measures. To scale up the biological data neuroscience ecosystem in LMICs, there must be investment in laboratories, biobanks, core facilities, and software, as well as the use of metadata-rich protocols that follow FAIR principles and respect local ethical and data governance requirements. Long-term funding and mentoring programmes that connect less resourced local researchers with teams in well-established labs in the global north and facilitate retention of scientific talent in LMICs are also important. Finally, workshops and short courses on advanced tools, data management, secure data sharing, and FAIR-compliant workflows are key for building local capacity. In sum, integrated capacity-building approaches are foundational to replicable, culturally rooted, and internationally relevant biological data science [24]. Conclusions Psychiatric disorders are highly polygenic and heterogeneous, and although single-omics studies have provided important insights, they cannot fully capture the biological complexity of these conditions. Team science efforts have highlighted that focusing on just one biological factor within a complex system is like carefully studying one star in the sky: it can reveal important mechanisms, but it does not show the whole picture. In the same way, analyzing a single biomarker can deepen our understanding of specific pathways, but many other interacting processes also shape psychiatric phenotypes. Therefore, a systems-level approach is needed to capture the dynamic and interconnected mechanisms that influence mental health and disease risk. To achieve this, standardized protocols, rigorous measurement frameworks, continuous monitoring, and interoperable data formats are essential. Finally, equitable collaborations that support long-term training, mentorship, infrastructure, and cultural competency are crucial to ensure that researchers in LMICs can not only participate in, but also lead, global mental health research. Declaration of Competing Interest The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: This manuscript represents the collaborative efforts of members of the Physical Health Working Group within the Psychiatric Genomics Consortium Posttraumatic Stress Disorder Working Group (PGC-PTSD). The PGC-PTSD is supported by R01MH106595 (PI: Nievergelt). Drs. Antonia Seligowski and Sinead Sinnot report financial support was provided by National Institutes of Health. Dr. Antonia Seligowski’s time is supported by NIH K23MH125920. Dr. Sinead M. Sinnott (NCCIH 2T32AT000051) received support from the NIH National Center for Complementary Integrative Health. Drs. Nathan Kimbrel and Kyle Bourassa report financial support was provided by US Department of Veterans Affairs. Dr. Nathan Kimbrel was supported by a VA Research Career Scientist (RCS) Award (#IK6BX006523) from the Biomedical Laboratory Research and Development Service of the Department of Veterans Affairs (VA) Office of Research and Development. Dr. Kyle Bourassa was supported by Award #IK2CX002694 from the Clinical Science Research and Development (CSR&D) Service of the Department of Veterans Affairs (VA) Office of Research and Development. The views expressed in this article are those of the authors and do not necessarily reflect the position or policy of the VA, the U.S. government, or any other affiliated institution. Drs. Erika J. Wolf and Sage E. Hawn serve on the editorial board of the Journal of Mood and Anxiety Disorders. Given their roles on the editorial board, Dr. Wolf and Dr. Hawn had no involvement in the peer review of this article and had no access to information regarding its peer review. Full responsibility for the editorial process for this article was delegated to another journal editor. If there are other authors, they declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. References