Precision medicine shifts healthcare from a one-size-fits-all model to interventions tailored to individual genetic, environmental, and lifestyle factors [1, 2]. This approach promises improved diagnostic accuracy, prognostic stratification, and therapeutic efficacy, particularly in oncology [3, 4]. Its realization depends on comprehensive molecular characterization, enabled by advances in high-throughput biotechnologies.

The past two decades have seen an explosion in biomedical data from omics technologies (genomics, transcriptomics, proteomics, etc.) [5]. However, this data deluge presents significant analytical challenges: high dimensionality (often tens of thousands of features) relative to sample size, complex inter-feature correlations, and high noise levels. These factors contribute to the “curse of dimensionality,” complicating statistical inference and machine learning (ML) model development [6–8].

Single omics layers offer only a partial view of biological systems. A holistic understanding requires integrating multiple modalities to capture the interplay from DNA to metabolism [9–11]. Multi-omics integration improves disease subtyping, prognostic accuracy, and therapeutic target identification compared to single-omics analyses [12, 13]. However, integration is challenging due to heterogeneous data types, scale disparities, missing data, batch effects, and a lack of standardized preprocessing pipelines [14–17].

Conventional ML models, often used for disease classification and biomarker discovery, frequently prioritize accuracy over interpretability, functioning as “black boxes” [18, 19]. This opacity is problematic in clinical settings where understanding model predictions is crucial for trust and safety [20]. Moreover, conventional feature selection often fails to capture complex, non-linear interactions across omics layers. Recent work also highlights that no single feature attribution method can satisfy all desirable properties for interpretability, leading to potential inconsistencies [21, 22].

Explainable artificial intelligence (XAI) aims to make AI models transparent, interpretable, and trustworthy [23, 24]. In precision medicine, XAI bridges predictive performance and clinical utility by elucidating the molecular drivers of model predictions [20, 25, 26]. XAI includes inherently interpretable models and post-hoc explanation methods [27, 28]. Applied to multi-omics, XAI can reveal cross-omics interactions and generate testable hypotheses [29]. For instance, attention mechanisms can highlight important features across omics layers, while graph neural networks (GNNs) can leverage biological pathways for interpretability [30]. In clinical settings, XAI fosters trust, facilitates regulatory compliance, and supports decision-making [31, 32]. However, rigorous evaluation is essential to avoid false assurances of understanding [19, 33].

While recent reviews have begun to explore connections between XAI and multi-omics integration, few provide a critically oriented examination that systematically addresses the translational gap from methodology to clinical deployment. Existing reviews have primarily focused either on technical integration methods, explainability techniques in isolation, or disease-specific applications, often without systematically addressing translational barriers such as clinical workflow integration, regulatory requirements, explanation faithfulness, and deployment readiness. In contrast, the present review integrates computational, biological, and translational perspectives within a unified framework. This review provides a holistic, critically oriented synthesis of XAI strategies for multi-omics integration, with a dedicated focus on clinical translation. Our objectives are threefold: (1) to systematically categorize computational strategies for explainable multi-omics integration; (2) to critically assess challenges hindering clinical adoption; and (3) to identify future research directions for accelerating clinical translation. The novelty lies in its integrated focus on both computational methodologies and practical clinical deployment, providing a unique framework for developing trustworthy AI systems [28–30]. This review critically examines the translational gap, emphasizing the need for rigorous evaluation, prospective validation, and clinical workflow integration [31, 34–36], serving as a roadmap for advancing clinically impactful, interpretable multi-omics solutions.

While spatial omics technologies are rapidly advancing, this review focuses on molecular multi-omics data (genomics, transcriptomics, proteomics, metabolomics, epigenomics, microbiomics) and does not extensively cover spatial resolution or spatial omics integration. The principles discussed are extendable, but a dedicated treatment of spatial omics is beyond our scope. Figure 1 illustrates the overall workflow from multi-omics data acquisition to clinical decision support, highlighting where XAI methods intervene to enable interpretable predictions.

This systematic review was conducted and reported in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA 2020) guidelines. The review methodology follows established frameworks for systematic mapping studies in computational biology and biomedical AI. This review was not registered in a prospective database such as PROSPERO.

A comprehensive literature search was performed across three major scientific databases: PubMed/MEDLINE, Scopus, and Web of Science. The search covered publications from 2020 to 2025. The search strategy was designed to capture the intersection of three conceptual domains: (1) multi-omics integration, (2) XAI, and (3) precision medicine applications. A combination of controlled vocabulary (e.g., MeSH terms) and free-text keywords was used. The search terms included:

1.

Multi-omics domain: “multi-omics,” “multiomics,” “integrative omics,” “genomics AND transcriptomics AND proteomics,” “multi-modal omics,” “multi-platform omics”

These terms were combined using Boolean operators (AND, OR), with the core query structure:

(“multi-omics” OR “integrative omics”) AND (“explainable AI” OR “interpretable machine learning”) AND (“precision medicine” OR “clinical application”).

In addition, manual screening of reference lists from relevant review articles and citation tracking using Google Scholar were performed to identify additional studies. Records identified through these methods were included within the overall dataset but were not separately quantified.

Studies were included if they met the following criteria:

(1) addressed multi-omics data integration involving at least two molecular layers;

(2) employed XAI methods or explicitly discussed model interpretability;

(3) were applied within precision medicine contexts (e.g., oncology, cardiovascular, neurological, or complex diseases);

(4) reported quantitative performance metrics or provided sufficient methodological detail;

(5) were published in peer-reviewed journals or conference proceedings;

(6) were written in English.

Studies were excluded if they:

(1) focused solely on single-omics analysis;

(2) used AI models without any interpretability component;

(3) lacked biomedical or clinical relevance;

(4) were editorials, opinion pieces, or abstracts without full methodological detail;

(5) relied exclusively on simulated data without validation on real datasets;

(6) focused primarily on spatial omics without molecular multi-omics integration.

The study selection process was conducted in accordance with PRISMA 2020 guidelines. Screening was performed in three stages.

First, titles and abstracts were screened independently by two reviewers (S.H. and M.H.) to exclude irrelevant studies. Disagreements were resolved through discussion, and when necessary, a third reviewer (M.W.) adjudicated. Second, full-text articles were assessed against the eligibility criteria. Third, a methodological quality assessment was performed to ensure inclusion of robust and clinically relevant studies. A total of 847 records were identified through database searches. After title and abstract screening, 412 records were excluded, leaving 435 articles for full-text assessment. Following full-text review and quality assessment, 116 studies were included in the final synthesis. The study selection process is illustrated in the PRISMA 2020 flow diagram (Figure 3).

Data extraction was performed using a standardized template. The following information was collected from each included study:

(1) publication details (authors, year, journal);

(2) omics layers integrated;

(3) integration strategy (early, intermediate, late, hierarchical fusion);

(4) AI/ML methodology;

(5) XAI methods employed;

(6) application domain;

(7) datasets used;

(8) evaluation metrics;

(9) validation approach;

(10) reported limitations and future directions.

Due to heterogeneity in datasets, study designs, and evaluation metrics, a quantitative meta-analysis was not feasible. Therefore, a narrative synthesis approach was adopted, organizing findings into thematic categories including computational strategies, XAI methodologies, clinical applications, and translational challenges.

Risk of bias was assessed using an adapted version of the PROBAST (Prediction model Risk of Bias Assessment Tool) framework. The following domains were evaluated:

(1) participant selection and data representativeness;

(2) predictor assessment and preprocessing;

(3) outcome definition;

(4) model development and overfitting control;

(5) appropriateness and faithfulness of XAI methods;

(6) performance evaluation;

(7) validation strategy.

Each study was categorized as having low, moderate, or high risk of bias. A subset of studies (25%) was independently evaluated by two reviewers to ensure consistency, yielding substantial inter-rater agreement (κ = 0.78). Discrepancies were resolved through consensus.

Several limitations should be acknowledged. Despite a comprehensive search strategy, relevant studies may have been missed due to database coverage or non-standard terminology. Publication bias may have favored studies reporting positive results. Rapid methodological advancements may not be fully captured despite inclusion up to 2025. Heterogeneity across studies precluded quantitative meta-analysis. Additionally, assessment of XAI methods involves subjective interpretation due to the absence of standardized evaluation benchmarks.

Despite these limitations, this methodology provides a robust and reproducible synthesis of current research on XAI for multi-omics integration in precision medicine.

The increasing complexity of AI models in biomedicine has created a fundamental tension between predictive accuracy and interpretability. This tension is a critical barrier to clinical adoption, regulatory approval, and scientific discovery [23, 24]. XAI addresses this by making model decisions transparent and trustworthy. Importantly, the required depth and form of explainability vary according to clinical application. Screening and risk stratification tasks may only require global feature importance or simplified explanations, whereas high-risk decisions such as treatment selection or intensive care interventions demand more rigorous, causally informed, and patient-specific explanations. Therefore, explainability should be considered context-dependent rather than universally standardized across all clinical tasks. The importance of explainability in biomedicine stems from several imperatives:

Clinical reasoning: Physicians must understand why a model recommends a particular diagnosis or treatment to integrate it into their reasoning and communicate it to patients [32, 33].

In multi-omics integration, explainability is particularly critical. Models must not only identify important features but also contextualize them within biological pathways, distinguish signal from noise, and communicate uncertainty—demanding XAI approaches that are both technically sophisticated and biologically meaningful [28].

Explainability encompasses approaches that differ in scope, timing, and relationship to the underlying model.

Model-agnostic methods interpret any model by analyzing input-output relationships without accessing internal parameters. They are widely used in multi-omics for their flexibility.

While these methods enable the interpretation of any black-box model, they are not without drawbacks. SHAP, though theoretically grounded, can be computationally expensive, and its approximations may be unstable. LIME offers fast, local explanations but can be sensitive to perturbations. Permutation importance is efficient for global feature ranking but ignores interactions, and partial dependence plots assume feature independence—an assumption rarely met in multi-omics data. Given these limitations, practitioners are advised to use multiple methods and assess convergence of explanations [21, 45].

Model-specific methods leverage internal architecture components to generate explanations that are inherently faithful to the model’s computation. The following content summarizes key approaches, including attention weights, saliency maps, gradient-based methods, and rule-based/sparse models.

Attention weights, prevalent in transformers and GATs, offer intrinsic interpretability and can highlight cross-modal interactions. However, their faithfulness is debated, as attention may not always reflect true feature importance [21]. Saliency and gradient-based methods provide fine-grained, computationally efficient attributions but can be noisy and unstable near sharp decision boundaries. Rule-based and sparse models (e.g., LASSO, decision trees) are intrinsically interpretable and produce transparent decision logic, though they may underperform on highly complex tasks. Hybrid approaches that combine deep feature extraction with sparse or rule-based final layers are increasingly used to bridge this performance-interpretability gap [62, 63]. Table 3 provides a comparative overview of the most common model-agnostic techniques—SHAP, LIME, permutation importance, and partial dependence analysis—detailing their core ideas, advantages, and limitations.

Translating XAI outputs into biological knowledge requires bridging computational attributions with established frameworks—from individual genes to pathways and disease mechanisms.

Although these challenges are discussed separately, they are highly interconnected. Data heterogeneity and batch effects constitute foundational bottlenecks that propagate downstream problems including overfitting, poor generalization, unstable explanations, and reproducibility failures. Similarly, limited sample sizes amplify both model instability and interpretability challenges. Consequently, improving data quality, harmonization, and preprocessing standardization should be considered a primary priority before optimization of advanced model architectures. Despite significant progress in XAI for multi-omics integration, the field faces formidable challenges that impede clinical translation and scientific reproducibility. These challenges span data characteristics, methodological limitations, evaluation practices, and ethical considerations.

The convergence of multi-omics integration and XAI has catalyzed significant advances across the precision medicine continuum. From cancer subtyping to drug response prediction, XAI-enhanced multi-omics models are enabling more accurate, interpretable, and clinically actionable insights. This section reviews key application domains, highlighting how XAI methods contribute to both predictive performance and biological understanding, while critically examining the evidence base and remaining challenges.

Translating a research model into a clinically deployable tool requires a paradigm shift from optimizing predictive accuracy to ensuring safety, usability, and sustainability. The process typically begins with retrospective validation on curated cohorts, but clinical implementation demands prospective evaluation in real-world settings [34, 60]. This includes assessing model performance across diverse patient populations, handling missing data encountered in routine care, and integrating with existing information technology infrastructure [29, 35].

A critical early step is defining the intended use case: whether the model will serve as a screening tool, a diagnostic aid, a prognostic marker, or a treatment selection guide. Each use case carries different performance requirements and regulatory pathways [18]. For example, a model used to identify patients for clinical trial enrollment may tolerate lower specificity than one used to recommend high-risk treatments. Establishing clear clinical utility—evidence that the model improves patient outcomes or reduces costs—is essential but rarely achieved in academic studies [34]. Clinical deployment should also include mechanisms for continuous performance monitoring, recalibration, and periodic updating to address data drift, evolving patient populations, and changing clinical practices throughout the AI lifecycle.

Clinician trust is a prerequisite for adoption, and explainability is a key driver of trust [32, 33]. When physicians understand why a model makes a recommendation, they are more likely to integrate it into their clinical reasoning and to accept responsibility for the decision. XAI provides a mechanism for clinicians to verify that the model’s reasoning aligns with their own knowledge and with the patient’s context [23, 31].

However, trust is not automatically conferred by explanations. If explanations are inconsistent, overly complex, or biologically implausible, they may erode rather than build trust [19]. Moreover, explanations must be tailored to the clinical audience. Visualizations of pathway-level contributions, attention maps, or simplified feature summaries may be more useful than raw Shapley values [28, 64]. In antimicrobial stewardship, for instance, explainable models that highlight resistance genes and pharmacokinetic parameters have been shown to support clinician acceptance [105]. Systematic reviews indicate that explainability can increase trust when explanations are accurate, timely, and presented in a user-friendly manner, but poorly designed explanations can have the opposite effect [33].

Excessively detailed or computationally intensive explanations may inadvertently increase clinician cognitive burden, particularly in time-sensitive clinical environments. Therefore, explanation interfaces should prioritize concise, clinically relevant summaries that support rapid interpretation without overwhelming end users.

For a multi-omics AI model to be used at the point of care, it must be integrated with EHRs and clinical workflows. This requires interoperability with hospital information systems, secure data transfer, and minimal disruption to existing practices [29, 35]. Many institutions are exploring real-time integration of omics data (e.g., genomic sequencing results) into EHRs, but challenges remain in standardizing data formats, ensuring data quality, and enabling automated model invocation [106].

Workflow integration also involves determining when and how model outputs are presented. A model that provides a prognostic score upon diagnosis may be deployed as a discrete alert, a clinical decision support panel, or a report integrated into the patient chart. Each approach has implications for user acceptance and cognitive load [32]. Successful implementation requires co-design with clinicians, iterative testing, and attention to how the model fits into existing decision-making processes [31].

Some clinical applications demand real-time or near-real-time predictions. For example, predicting acute deterioration, surgical complications, or immediate drug–drug interactions requires models that can process incoming data within seconds to minutes. For multi-omics models, real-time inference is complicated by the time required to generate omics measurements (e.g., sequencing or mass spectrometry), which can take hours to days [36]. Therefore, many multi-omics XAI systems are currently designed for pre-treatment planning or retrospective risk stratification rather than acute decision support.

Where real-time support is needed, strategies include using proxies for omics data (e.g., rapid gene expression assays), pre-computing model predictions for common scenarios, or deploying models that can operate on incomplete data while providing uncertainty estimates [106, 107]. Advances in portable sequencing and mass spectrometry may eventually enable same-day omics profiling, making real-time multi-omics AI more feasible [36].

The computational overhead associated with generating explanations, particularly for SHAP-based or attention-intensive models, may limit deployment in real-time settings. Efficient approximation strategies and lightweight explanation frameworks will therefore be essential for time-critical clinical applications.

Regulatory approval is a major gatekeeper for clinical AI deployment. In the United States, the FDA has issued guidance on software as a medical device (SaMD), emphasizing the need for transparency, validation, and post-market surveillance. The European Union’s Medical Device Regulation and the proposed Artificial Intelligence Act similarly require that high-risk AI systems be interpretable and provide explanations for their outputs [18, 34].

For multi-omics XAI models, regulatory bodies are likely to require evidence of generalizability across populations, robustness to input variability, and validation of explanation faithfulness. The lack of standardized XAI evaluation frameworks complicates this process [28]. Moreover, the dynamic nature of AI models—where continuous learning may change behavior—poses challenges for traditional approval pathways [34].

Ethical considerations extend beyond regulation to include informed consent, data privacy, and equitable access. Multi-omics data are uniquely sensitive; patients must be informed about how their data will be used and protected [60]. Models must be audited for bias across demographic groups, and deployment plans must address disparities in healthcare access that could exacerbate existing inequities [32].

Emerging regulatory frameworks increasingly emphasize ethical AI principles including transparency, traceability, non-discrimination, human oversight, and explainability. In clinical environments, explainability is not only a technical requirement but also an ethical necessity that enables informed clinical judgment and patient-centered care. Future multi-omics AI systems should therefore incorporate ethics-by-design principles throughout model development and deployment.

Even when models are technically sound and regulatory approvals are obtained, adoption in hospitals and laboratories faces numerous barriers. Infrastructure limitations—including lack of computational resources, data storage, and trained personnel—can preclude deployment in smaller or resource-limited institutions [34]. Hospital IT departments may be reluctant to integrate external AI tools due to security concerns, vendor compatibility issues, and maintenance burdens [35].

Laboratory adoption of multi-omics workflows is similarly constrained by the need for standardized protocols, quality control, and turnaround times. Clinical laboratories may be hesitant to implement new assays that feed into AI models without clear evidence of cost-effectiveness and reimbursement pathways [94]. Moreover, the rapid evolution of AI technologies can conflict with the slower pace of clinical validation and adoption cycles, leading to “pilot purgatory” where promising tools never scale [36].

To bridge the gap between research and clinical practice, multi-omics XAI systems must meet a set of translational readiness criteria [34]. These include:

Prospective validation on multi-center, real-world cohorts that reflect the target population, with predefined performance thresholds and prespecified subgroup analyses [60].

Interoperability with existing EHR and laboratory information systems through standard data formats [e.g., Health Level Seven Fast Healthcare Interoperability Resources (HL7 FHIR)] and secure application programming interfaces [29].

Usability and human factors engineering, ensuring that explanations are presented in a format that clinicians can quickly understand and act upon without adding excessive cognitive burden [32].

Scalability to handle increasing volumes of data and users, with cloud or on-premises architectures that maintain performance under peak loads [35].

Sustainability including clear funding models for maintenance, updates, and technical support, as well as pathways for reimbursement of AI-guided services [34].

Continuous monitoring for performance drift, bias, and safety issues, with mechanisms for updating models, when necessary, without violating regulatory requirements [57].

Ethical and legal frameworks that address data ownership, consent, liability, and equity.

Achieving these requirements demands close collaboration among computational scientists, clinicians, hospital administrators, regulators, and patients. Multi-stakeholder consortia and public–private partnerships have begun to address these challenges, but widespread clinical deployment of explainable multi-omics AI remains in its infancy. Future progress will depend on sustained investment, pragmatic evaluation frameworks, and a shared commitment to patient-centered innovation [12, 36].

The key stages required to move a multi-omics XAI model from research to clinical practice are summarized in Figure 5, which outlines the pipeline from retrospective validation through regulatory approval and post-market surveillance.

As the fields of multi-omics integration and XAI continue to evolve, emerging paradigms promise to address current limitations while opening new frontiers for precision medicine. Table 5 summarizes these future directions alongside the challenges they aim to resolve.

Centralized data aggregation raises significant privacy concerns. Federated learning enables model training across distributed institutions without exchanging raw data, addressing privacy regulations while enabling larger-scale analyses [108, 114]. For multi-omics integration, federated learning must handle heterogeneous feature spaces and missing modalities across sites. Future work should focus on developing federated XAI frameworks that provide faithful explanations without compromising privacy.

Current multi-omics AI models predominantly capture correlations, whereas clinical decision-making increasingly requires causal evidence. Causal AI aims to infer cause–effect relationships from observational data and support reasoning about interventions [110, 115]. Mendelian randomization uses genetic variants as instrumental variables to investigate causal relationships [73, 74]. Recent disease-focused studies have combined Mendelian randomization with multi-omics, ML, and SHAP to identify causally associated metabolites and interpretable molecular signatures [116].

Foundation models—large-scale pre-trained models adaptable to diverse downstream tasks—are transforming AI across domains [108, 111]. For multi-omics integration, foundation models offer the potential to overcome small sample size problems by leveraging pre-trained representations that encode rich biological knowledge. Future directions include developing truly multimodal foundation models that jointly process genomics, transcriptomics, proteomics, and clinical data, with built-in interpretability mechanisms.

Digital twins—virtual representations of individual patients integrating multi-scale data—represent a frontier in precision medicine [113]. A patient’s digital twin would incorporate multi-omics, clinical, and longitudinal data, enabling personalized simulations of therapeutic interventions. XAI is essential for digital twins to provide interpretable predictions and allow clinicians to interrogate model behavior. Future work must address computational demands, standardized interfaces, and validation against real-world outcomes.

The reliance on large, well-annotated datasets is a major bottleneck. Self-supervised learning enables models to learn representations from unlabeled data by solving pretext tasks such as masked reconstruction [79]. Weakly supervised learning uses noisy, incomplete labels. For multi-omics, these approaches can leverage vast amounts of unlabeled data and incorporate diverse evidence sources. Future directions include developing self-supervised objectives tailored to multi-omics data and ensuring learned features are biologically interpretable.

The proliferation of XAI methods has outpaced the development of rigorous evaluation frameworks, making it difficult to compare methods or trust their outputs [28]. Standardized benchmarks are urgently needed to assess explanation faithfulness, stability, and biological validity. Future initiatives should develop common metrics—faithfulness, stability, biological alignment, computational cost—and establish community challenges (e.g., DREAM-style) specifically for XAI in multi-omics.

The integration of AI into clinical workflows should be conceived as a partnership between human experts and machines. Human-in-the-loop AI systems augment clinician capabilities, with explanations serving as a medium for collaboration [32]. Clinicians can validate explanations against patient context, provide feedback, and retain final decision authority. Future directions include developing intuitive user interfaces that present explanations in clinically meaningful ways and designing workflows for efficient clinician–AI interaction.

The convergence of multi-omics technologies and XAI marks a decisive shift toward truly interpretable and data-driven precision medicine. By integrating heterogeneous molecular layers with transparent computational frameworks, these approaches enable the discovery of complex biological relationships while providing clinically meaningful explanations for model predictions. Such capabilities not only improve disease stratification, biomarker identification, and therapeutic response prediction, but also strengthen trust, accountability, and regulatory acceptance of AI-driven medical tools. Nevertheless, the path to clinical implementation remains constrained by persistent challenges, including data heterogeneity, limited cohort sizes, reproducibility concerns, and insufficient real-world validation. Overcoming these barriers will require standardized benchmarks, robust validation across diverse populations, and seamless integration with clinical workflows. Ultimately, the next generation of precision medicine will depend on the development of scalable, interpretable, and clinically reliable AI systems that transform complex multi-omics data into actionable knowledge, thereby enabling more accurate diagnoses, personalized treatments, and improved patient outcomes.