Explainable AI for Bioinformatics: Methods, Tools, and Applications

Machine learning models in bioinformatics often achieve high predictive accuracy while offering little insight into why a prediction was made. Explainable AI (XAI) addresses this gap by providing methods to interpret model decisions, identify important features, and expose potential biases. For researchers working with genomic, transcriptomic, proteomic, and clinical data, XAI methods such as SHAP, LIME, and integrated gradients offer practical routes to understand model behavior, validate biological relevance, and build trust in computational findings. This article explains the main XAI techniques, their strengths and limitations in bioinformatics contexts, and provides a decision framework for selecting appropriate methods based on model type and research question.

The Interpretability Problem in Bioinformatics Machine Learning

Bioinformatics models operate on high-dimensional data where the number of features often far exceeds the number of samples. A typical gene expression dataset may contain measurements for 20,000 genes across only a few hundred patients. Deep learning models can identify complex patterns in such data, but the internal representations they learn are difficult to inspect directly. This opacity, often called the black-box problem, creates practical challenges for researchers who need to understand which biological features drive predictions, whether the model has learned meaningful biology or dataset artifacts, and whether findings can be generalized to new cohorts.

The need for explainability in biomedical AI has been recognized across multiple specialties. A qualitative study involving expert interviews defined six competency domains for health care professionals working with AI-based tools, including basic knowledge of AI, evidence-based evaluation of AI tools, and understanding of social and ethical implications [5]. These competencies reflect a broader expectation that AI outputs in biomedical contexts must be interpretable and evaluable by domain experts, beyond technically validated by computer scientists.

In bioinformatics research, interpretability serves several distinct purposes. First, it supports hypothesis generation by revealing which genes, variants, or pathways contribute most to a prediction. Second, it enables model debugging by identifying cases where the model relies on confounding variables or batch effects. Third, it facilitates communication with collaborators, reviewers, and regulatory bodies who require mechanistic explanations for computational findings. Fourth, it supports the integration of computational predictions with experimental validation by prioritizing features for follow-up study.

The consequences of ignoring explainability extend beyond scientific inconvenience. Models that achieve high accuracy on training data but rely on non-biological features can produce misleading results when applied to new datasets. A review of deep learning in cancer research noted that the lack of phenotypically rich data and the need for more explainable deep learning models remain current limitations for clinical application [7]. Similarly, work on sequence-to-function models in regulatory genomics has shown that strong predictive accuracy can fail to translate into robust regulatory understanding, with interpretability methods revealing systematic failure modes [17].

Core XAI Methods Used in Bioinformatics

SHAP (SHapley Additive exPlanations)

SHAP assigns each feature an importance value for a particular prediction based on cooperative game theory. The method computes how much each feature contributes to the difference between the model prediction and a baseline prediction, averaging over all possible feature orderings. This produces additive feature attributions that sum to the model output.

SHAP has become the most frequently used XAI technique in quantitative prediction tasks. A systematic review of 44 Q1 journal articles applying XAI to prediction problems found SHAP used in 35 of the 44 studies, reflecting its widespread adoption for feature-importance ranking and model interpretation [11]. The same review identified LIME, partial dependence plots, and permutation feature importance as the next most popular methods.

In bioinformatics, SHAP has been applied to variant pathogenicity prediction. The MetaXVP framework, built on the XGBoost algorithm, used SHAP and LIME to make variant classification interpretable, achieving high classification performance on ClinVar test sets while providing feature-level explanations for pathogenicity predictions [13]. SHAP analysis has also been used to identify microbial genera predictive of feeding stage and diet in hybrid sturgeon, where the method revealed the contribution and directionality of each genus to group classification [20].

For gene expression models, SHAP can identify genes that drive predictions for individual samples. A study applying SHAP to deep learning models for influenza onset prediction identified five genes shared across influenza subtypes, including immune response and sex-linked genes, demonstrating how SHAP can generate actionable biological insights from complex models [21]. In diabetes prediction, SHAP analysis revealed the significant contributions of glucose, BMI, and age to model decisions, with pregnancies showing a particularly positive impact on diabetes risk [22].

SHAP also has documented limitations. The systematic review noted theoretical limitations of SHAP and related model-agnostic methods, particularly their additive nature [11]. A robustness study applying SHAP to breast cancer subtype classification found that complex models like XGBoost and MLP produced different sets of patient-specific features across runs, even when overall accuracy remained constant [23]. This instability raises concerns about using SHAP local interpretations for precision medicine applications where individual patient features guide treatment decisions.

LIME (Local Interpretable Model-Agnostic Explanations)

LIME explains individual predictions by fitting a simple interpretable model, such as a linear model or decision tree, to approximate the complex model's behavior in the neighborhood of a specific instance. The method perturbs the input features, observes how predictions change, and builds a local surrogate model that reveals which features matter most for that particular prediction.

LIME has been applied alongside SHAP in bioinformatics contexts. The MetaXVP variant predictor used both methods to make the XGBoost model interpretable, providing complementary perspectives on feature importance [13]. However, LIME has shown notable variability in some applications. A benchmarking study comparing Grad-CAM, SHAP, and LIME for brain MRI classification found that LIME exhibited pronounced variability due to its super-voxel segmentation approach, leading the authors to caution that LIME requires careful parameter tuning before clinical deployment [19].

The local nature of LIME explanations means they can differ substantially from global feature importance measures. Researchers using LIME should verify that local explanations are stable across perturbations and consistent with domain knowledge.

Integrated Gradients and Gradient-Based Methods

Integrated gradients is a model-aware attribution method that computes feature importance by integrating gradient values along a path from a baseline input to the actual input. Unlike SHAP and LIME, which treat the model as a black box, integrated gradients uses information about the model's internal structure to compute attributions efficiently.

Gradient-based methods are particularly relevant for deep learning models in bioinformatics. The OmiXAI pipeline integrated ensemble model-aware XAI methods for deep learning models trained on omics feature matrices, incorporating integrated gradients, InputXGradients, guided backpropagation, and deconvolution for convolutional and graph neural networks [16]. The pipeline demonstrated that these methods could reduce a critical feature set from almost 2,000 features to just 50, enabling feature engineering and model simplification.

Gradient-based methods have computational advantages over perturbation-based approaches. The OmiXAI developers noted that model-agnostic approaches relying on combinatorial feature perturbations are often computationally prohibitive for deep learning models, making model-aware gradient methods a practical alternative [16].

Grad-CAM and Saliency Methods

Gradient-weighted Class Activation Mapping (Grad-CAM) generates visual explanations for convolutional neural networks by highlighting the regions of an input image that most influenced a prediction. In bioinformatics, Grad-CAM has been applied to medical imaging tasks including brain MRI classification.

A controlled comparison of Grad-CAM, Deep-SHAP, and LIME using a standardized pipeline found that Grad-CAM exhibited the highest consistency and stability for both Alzheimer's disease and brain tumor classification tasks, accurately highlighting hippocampal atrophy and tumor cores [19]. The study recommended gradient-based explanations for rapid focal-lesion screening while reserving SHAP for diffuse pathology requiring voxel-level justification.

Model-Specific Interpretability Approaches

Some bioinformatics models are designed with interpretability as a primary objective. Biologically constrained neural networks incorporate prior knowledge about gene regulatory relationships into the model architecture, making the learned representations more interpretable by construction.

The TF-GateNet framework exemplifies this approach, combining transcription factor-aware feature integration with sample-specific dynamic gating on a hierarchical pathway backbone [18]. The model achieved strong performance for primary-metastatic state prediction in prostate and breast cancer cohorts while enabling interpretability analysis that traced a route from transcription factor-related genomic perturbation cues to genes, pathways, and phenotype-associated predictions.

Similarly, the senescence-related gene signature for gastric cancer prognosis used a dual-model interpretable feature selection strategy that integrated a biologically informed Kolmogorov-Arnold Network with a tabular foundation model [15]. This approach identified cancer-associated senescence genes that were then distilled into a core four-gene signature using an ensemble of machine learning algorithms.

At a Glance: XAI Method Selection

The following table summarizes the main XAI methods, their model requirements, computational costs, and typical bioinformatics applications.

Decision Tree for Selecting an XAI Method

Selecting an appropriate XAI method depends on three primary factors: the type of model being interpreted, the research question being addressed, and the computational resources available.

Step 1: Determine the model type. If the model is a neural network, gradient-based methods such as integrated gradients or Grad-CAM are computationally efficient options. If the model is tree-based, such as XGBoost or random forest, SHAP provides native support and efficient computation. If the model is a support vector machine or other non-differentiable model, model-agnostic methods like SHAP or LIME are required.

Step 2: Define the interpretation scope. Global interpretation seeks to understand which features matter most across the entire dataset. SHAP summary plots and permutation importance serve this purpose. Local interpretation seeks to explain individual predictions, which is relevant for patient-level analyses or variant classification. LIME and SHAP force plots provide local explanations.

Step 3: Consider the data modality. For imaging data, Grad-CAM provides intuitive visual explanations. For tabular omics data, SHAP and integrated gradients are more appropriate. For sequence data, attribution methods that operate on nucleotide-level features may be needed.

Step 4: Assess computational constraints. SHAP can be computationally expensive for models with thousands of features, particularly when computing exact Shapley values. Gradient-based methods are generally faster. The OmiXAI pipeline was developed specifically because model-agnostic perturbation approaches were computationally prohibitive for deep learning on omics data [16].

Step 5: Validate explanation stability. Run the explanation method multiple times and assess whether the same features are consistently identified. The robustness study of SHAP in precision medicine found that complex models produced inconsistent patient-specific features across runs, highlighting the need for stability checks [23].

Practical Workflow for Applying XAI in Bioinformatics

Data Preparation and Feature Definition

The quality of explanations depends on the quality of input features. Before applying any XAI method, researchers should ensure that features are well-defined, properly normalized, and free from obvious artifacts. The FAIR Guiding Principles provide a framework for making data findable, accessible, interoperable, and reusable, which supports reproducibility in XAI applications [4].

For genomic data, feature definitions should align with biological meaning. Variant-level features might include conservation scores, functional predictions, and allele frequencies. Gene-level features might include expression values, methylation levels, or copy number alterations. Pathway-level features aggregate gene-level information into biological processes.

Normalization choices can affect explanation outcomes. The SHAP robustness study found that z-score and min-max normalization had minimal effect on XGBoost model performance but influenced feature contribution scores [23]. Researchers should document normalization procedures and test whether explanations remain stable across reasonable preprocessing choices.

Model Training and Validation

XAI methods interpret a trained model, so model quality directly affects explanation quality. Models should be trained and validated using appropriate procedures, including cross-validation and independent test sets. The TF-GateNet study used a fixed 80/10/10 split with repeated runs and external validation cohorts to ensure robust performance assessment before interpretability analysis [18].

For clinical applications, the competencies defined for health care professionals emphasize evaluating the quality, accuracy, safety, contextual appropriateness, and biases of AI-based tools and their underlying datasets [5]. This evaluation should occur before explanations are generated, because explanations of a biased model will reflect and potentially amplify those biases.

Generating and Interpreting Explanations

When generating explanations, researchers should use multiple methods when feasible. Complementary methods can reveal different aspects of model behavior. The MetaXVP framework used both SHAP and LIME to make variant predictions interpretable, providing cross-validation of feature importance findings [13].

For global interpretation, SHAP summary plots show the distribution of feature importance across all samples. These plots reveal whether a feature consistently contributes in one direction or has mixed effects depending on sample context. For local interpretation, SHAP force plots and LIME explanations show which features drove a specific prediction.

Interpretation should connect computational findings to biological knowledge. When SHAP identifies a gene as important for a cancer prognosis model, researchers should check whether that gene has known roles in relevant biological pathways. The gastric cancer study validated the oncogenic role of SERPINE1, a top-ranked core gene, through in vitro functional assays [15]. This experimental validation strengthens confidence in computational findings.

Documentation and Reporting

Reproducibility requires detailed documentation of the XAI workflow. Reports should include the model architecture, training procedure, hyperparameters, explanation methods, and parameter settings for those methods. The OmiXAI pipeline provides an open-source implementation that supports reproducible XAI benchmarking [16].

For studies with clinical implications, reporting should follow established guidelines for AI research. The competencies framework emphasizes that health care professionals need to understand how social, economic, and political systems influence AI-based tools and how these relationships impact justice, equity, and ethics [5]. This context should inform how XAI findings are presented and interpreted.

Applications in Genomics

Variant Pathogenicity Prediction

Variant interpretation remains a significant challenge in clinical genomics, particularly for variants of uncertain significance. Computational predictors can help prioritize variants for experimental validation, but their utility depends on interpretability. The MetaXVP framework addressed this need by combining XGBoost with SHAP and LIME explanations, achieving high classification performance while providing feature-level insights into pathogenicity predictions [13].

The framework used 32 features spanning conservation metrics, functional scores, ensemble predictors, and allele frequencies, selected through correlation-based feature selection. SHAP analysis revealed which features contributed most to pathogenicity classifications, helping researchers understand the biological basis of predictions.

Gene Expression and Biomarker Discovery

Deep learning models applied to gene expression data can predict disease outcomes, drug responses, and other phenotypes. SHAP analysis can identify the genes that drive these predictions, generating candidate biomarkers for experimental validation.

The influenza study applied SHAP to deep learning models trained on pre-exposure gene expression data, identifying five genes shared across influenza subtypes [21]. The finding that elevated XIST and reduced RPS4Y1 and KDM5D expression suggested higher susceptibility in females demonstrates how XAI can generate testable hypotheses about biological mechanisms.

For cancer research, SHAP analysis of gene expression models can reveal subtype-specific biomarkers. The breast cancer subtype classification study used SHAP to identify patient-specific feature contributions, though the authors noted concerns about the stability of these local interpretations for complex models [23].

Regulatory Genomics

Sequence-to-function models predict molecular regulatory readouts directly from DNA sequence. These models support variant effect prediction, mechanistic interpretation, and regulatory sequence design. However, generalization across genetic variation and cellular contexts remains inconsistent [17].

Interpretability methods have helped elucidate cis-regulatory organization and highlighted systematic failure modes in these models. The review of seq2func models suggested that progress requires reframing these models as continually refined systems, with targeted perturbation experiments, systematic evaluation, and iterative model updates coupled through AI-experiment feedback loops [17].

Applications in Proteomics and Multi-Omics

Protein Structure and Function Prediction

Machine learning models for protein structure prediction and function annotation benefit from interpretability when researchers need to understand which sequence features drive predictions. Gradient-based attribution methods can identify critical residues or sequence motifs that contribute to predicted structures or functions.

Multi-Omics Integration

Integrating genomics, transcriptomics, epigenomics, proteomics, and metabolomics data presents unique challenges due to high dimensionality and heterogeneous data types. AI methods can bridge these gaps through modeling, feature extraction, and multimodal learning [14]. Explainable AI frameworks are emerging as essential tools for interpreting these integrated models.

The OmiXAI pipeline was designed specifically for deep learning models trained on explicit omics feature matrices, integrating multiple gradient-based attribution methods [16]. The pipeline demonstrated that ensemble approaches can reduce feature sets dramatically while maintaining model performance, supporting both interpretation and model simplification.

Disease Subtyping and Prognosis

Multi-omics integration with machine learning has been applied to disease subtyping and prognosis prediction. The gastric cancer study integrated multi-omics data with an interpretable framework to identify a senescence-related gene signature for prognosis [15]. The framework used a dual-model interpretable feature selection strategy, followed by an ensemble of ten machine learning algorithms to distill a core four-gene signature.

The resulting Senescence Risk Score stratified patients into high- and low-risk groups with distinct overall survival across multiple cohorts. SHAP-style interpretability revealed the biological context of the signature, including immune cell infiltration patterns, checkpoint expression, and metabolic reprogramming [15].

Applications in Medical Imaging

Pathology and Histopathology

Deep learning has demonstrated potential for analyzing histological images and molecular profiling data for disease classification, tissue biomarker quantification, and clinical outcome prediction [10]. Explainable AI methods help pathologists understand which image regions and features drive model decisions.

The review of explainable AI for precision pathology emphasized the need to foster mutual understanding between the biomedical and AI communities, presenting worked-through examples for practical understanding of what AI can achieve and how it should be done [10].

Brain Imaging

Grad-CAM and other visualization methods have been applied to brain MRI classification. The benchmarking study comparing Grad-CAM, SHAP, and LIME found that Grad-CAM exhibited the highest consistency and stability, accurately highlighting hippocampal atrophy in Alzheimer's disease and tumor cores in brain tumor classification [19].

The study recommended gradient-based explanations for rapid focal-lesion screening and SHAP for diffuse pathology requiring voxel-level justification, while cautioning that LIME requires careful parameter tuning before clinical deployment [19].

Cardiology and Other Clinical Imaging

AI and machine learning are poised to influence nearly every aspect of medicine, including cardiology. A guide for clinicians on AI in cardiology reviewed predictive modeling concepts, common supervised learning algorithms, and deep learning methods, noting the importance of feature selection and common pitfalls such as improper dichotomization [6].

Tools and Implementation Considerations

Available Software Frameworks

Several open-source libraries implement XAI methods. The SHAP library provides implementations of Shapley value-based explanations for tree-based models, neural networks, and other model types. The LIME library implements local surrogate models. Integrated gradients and related methods are available in deep learning frameworks such as TensorFlow and PyTorch.

The OmiXAI pipeline provides an integrated framework for ensemble model-aware XAI methods specifically designed for omics data [16]. The pipeline is freely available and supports integration of additional attribution methods, ensuring adaptability beyond omics to diverse problem domains.

Computational Considerations

The computational cost of XAI methods varies substantially. SHAP can be expensive for models with many features, particularly when computing exact Shapley values. Approximate methods and tree-specific implementations reduce this cost. Gradient-based methods are generally more efficient for neural networks.

The OmiXAI developers noted that model-agnostic approaches relying on combinatorial feature perturbations are often computationally prohibitive for deep learning models, motivating their development of an ensemble model-aware approach [16].

Reproducibility and Stability

Explanation stability is a critical concern for bioinformatics applications. The SHAP robustness study found that complex models like XGBoost and MLP produced different sets of patient-specific features across runs, even when accuracy remained constant [23]. This instability challenges the reliability of local interpretations for precision medicine applications.

Researchers should run explanation methods multiple times and assess stability before drawing conclusions. The benchmarking study for brain MRI classification used standardized preprocessing and patient-stratified splits to minimize scanner bias and prevent information leakage, providing a foundation for reliable explanation assessment [19].

Common Failure Patterns and How to Avoid Them

Overinterpreting Local Explanations

Local explanations describe why a model made a specific prediction, not necessarily why the prediction is biologically correct. A SHAP force plot showing that a particular gene drove a cancer prediction does not prove that the gene is causally involved in cancer. Researchers should treat local explanations as hypotheses for experimental validation instead of definitive biological findings.

Ignoring Explanation Instability

If an explanation method produces different results across runs, the explanations are not reliable. The SHAP robustness study demonstrated this problem for complex models [23]. Researchers should assess stability before using explanations to guide decisions.

Applying Model-Agnostic Methods to Large Feature Sets

Model-agnostic methods that rely on feature perturbations become computationally prohibitive as feature counts increase. The OmiXAI developers highlighted this limitation for deep learning on omics data [16]. Gradient-based methods provide a practical alternative for these cases.

Neglecting Biological Validation

Computational explanations should be connected to biological knowledge. When SHAP identifies important genes, researchers should check whether those genes have known roles in relevant pathways. The gastric cancer study validated the oncogenic role of SERPINE1 through in vitro functional assays [15], providing a model for integrating computational and experimental approaches.

Confusing Correlation with Causation

XAI methods reveal which features the model uses for predictions, not causal relationships. A model might rely on a correlated proxy feature instead of the causal feature. Researchers should be cautious about interpreting feature importance as biological importance.

Limitations and Safety Considerations

Theoretical Limitations of Additive Methods

SHAP and related model-agnostic methods assume that feature contributions are additive. The systematic review of XAI techniques noted theoretical limitations of SHAP and related methods, particularly their additive nature [11]. For models with strong feature interactions, additive explanations may not capture the full complexity of model behavior.

Data Quality and Bias

Explanations of biased models will reflect and potentially amplify those biases. The competencies framework for health care professionals emphasizes evaluating the biases of AI-based tools and their underlying datasets [5]. Researchers should assess whether training data adequately represents the populations to which the model will be applied.

Regulatory and Ethical Considerations

For clinical applications, regulatory requirements may mandate specific approaches to explainability. The Genomic Data Sharing Policy from the National Institutes of Health provides guidance on responsible sharing of genomic data [3]. Researchers should be aware of applicable regulations and ensure that XAI workflows comply with data sharing and privacy requirements.

Professional Escalation Criteria

When XAI findings have clinical implications, researchers should involve domain experts in interpretation. The competencies framework emphasizes that health care professionals need basic knowledge of AI and the ability to evaluate AI-based tools [5]. For variant classification, findings should be reviewed by clinical geneticists or molecular pathologists before influencing patient care decisions.

Records and Measurements for XAI Studies

Documentation Requirements

Reproducible XAI studies require detailed documentation of the entire workflow. Records should include data sources and preprocessing steps, model architecture and hyperparameters, training and validation procedures, explanation methods and parameter settings, and stability assessments.

Quality Metrics

Several metrics support quality assessment of XAI outputs. Explanation stability measures whether repeated runs produce consistent attributions. Explanation fidelity measures how well the explanation approximates the model's behavior. Biological plausibility assesses whether explanations align with domain knowledge.

Benchmarking

Benchmarking XAI methods requires standardized pipelines. The brain MRI study used uniform preprocessing, patient-stratified splits, and multiple consistency and stability measures to compare explanation methods [19]. Similar approaches can be applied to omics data.

Frequently Asked Questions

What is the difference between model-aware and model-agnostic XAI methods?

Model-aware methods, such as integrated gradients and Grad-CAM, use information about the model's internal structure to compute attributions. They require access to model gradients or internal representations and are typically more computationally efficient. Model-agnostic methods, such as SHAP and LIME, treat the model as a black box and rely on input perturbations to assess feature importance. They work with any model type but can be computationally expensive for large feature sets [16].

When should I use SHAP versus integrated gradients for genomic data?

SHAP is well-suited for tree-based models and provides both global and local explanations with a solid theoretical foundation. Integrated gradients is more appropriate for neural networks, where gradient information is readily available and perturbation-based approaches become computationally prohibitive [16]. For deep learning on omics data, integrated gradients and related gradient-based methods offer practical advantages.

How can I assess whether my XAI explanations are stable and reliable?

Run the explanation method multiple times and compare the resulting feature attributions. The SHAP robustness study found that complex models produced different patient-specific features across runs even when accuracy remained constant [23]. If explanations vary substantially across runs, consider using simpler models or gradient-based methods that may provide more stable attributions.

Can XAI methods identify causal biological relationships?

No. XAI methods reveal which features the model uses for predictions, not causal relationships. A model might rely on correlated proxy features instead of causal features. Computational explanations should be treated as hypotheses for experimental validation. The gastric cancer study validated the oncogenic role of SERPINE1 through in vitro functional assays [15], demonstrating the importance of experimental confirmation.

What are the main limitations of SHAP for bioinformatics applications?

SHAP has theoretical limitations related to its additive nature [11]. It can be computationally expensive for models with many features. Local SHAP interpretations may be unstable for complex models, producing different patient-specific features across runs [23]. Researchers should assess stability and consider complementary methods when appropriate.

How do I choose between global and local interpretation methods?

Global interpretation methods, such as SHAP summary plots and permutation importance, describe which features matter most across the entire dataset. Local interpretation methods, such as SHAP force plots and LIME, explain individual predictions. Use global methods for biomarker discovery and model validation. Use local methods for patient-level analyses or variant classification where understanding individual predictions is important.

What role does biological knowledge play in interpreting XAI outputs?

Biological knowledge is essential for validating and contextualizing XAI outputs. When SHAP identifies a gene as important, researchers should check whether that gene has known roles in relevant pathways. Biologically constrained models, such as TF-GateNet, incorporate prior knowledge about gene regulatory relationships into the model architecture, making learned representations more interpretable by construction [18].

How should I report XAI findings in publications?

Report the model architecture, training procedure, hyperparameters, explanation methods, and parameter settings. Document stability assessments and any preprocessing choices that might affect explanations. Connect computational findings to biological knowledge and clearly distinguish between model attributions and validated biological relationships.

Related Bioinformatics Guides

• The Rise of Omics: Genomics, Proteomics, and Metabolomics
• Epigenetics and Computational DNA Methylation Analysis: Mechanisms, Methods, and Veterinary Applications
• The Bioinformatics Revolution in Structural Proteomics and Computational Drug Discovery: A Unified Paradigm
• Structure-Based Drug Design in Bioinformatics: Computational Pipelines, Active Site Grid Mapping, and Virtual Screening Workflows
• The History and Evolution of Bioinformatics

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.
• Competencies for the Use of Artificial Intelligence-Based Tools by Health Care Professionals.. Academic medicine : journal of the Association of American Medical Colleges, 2023.
• Artificial Intelligence in Cardiology.. Journal of the American College of Cardiology, 2018.
• Deep learning in cancer diagnosis, prognosis and treatment selection.. Genome medicine, 2021.
• Accelerating scientific discovery with Co-Scientist.. Nature, 2026.
• Artificial intelligence for detecting keratoconus.. The Cochrane database of systematic reviews, 2023.
• Toward Explainable Artificial Intelligence for Precision Pathology.. Annual review of pathology, 2024.
• eXplainable Artificial Intelligence (XAI): A Systematic Review for Unveiling the Black Box Models and Their Relevance to Biomedical Imaging and Sensing.. Sensors (Basel, Switzerland), 2025.
• Emerging Trends in Artificial Intelligence-Integrated Biochip Technologies for Biomedical Applications.. Micromachines, 2026.
• MetaXVP: an interpretable machine learning framework for deep insight into variant pathogenicity and VUS classification.. 2026.
• Artificial intelligence-driven multi-omics integration for plant enhancement: advances, challenges, and future perspectives.. 2026.
• Exploring the prognostic role of senescence-related genes in gastric cancer through multi-omics integration and machine learning.. 2026.
• OmiXAI: An ensemble XAI pipeline for interpretable deep learning in omics data.. 2026.
• Toward generalizable and interpretable AI in regulatory genomics.. 2026.
• TF-GateNet: An Interpretable and Biologically Guided Framework for Primary-Metastatic State Prediction from Somatic Genomic Alterations.. 2026.
• Consistency and stability benchmarking of Grad-CAM, SHAP, and LIME for diffuse and focal brain MRI classification. International Conference on Biomedicine and Bioinformatics Engineering, 2025.
• Integrating random forest and SHAP analysis to uncover gut microbial dynamics across growth stages in hybrid sturgeon. International Conference on Biomedicine and Bioinformatics Engineering, 2025.
• SHAP-Driven Deep Learning Identifies Pre-Symptomatic Influenza A Biomarkers: An Explainable AI Framework for Clinical Genomics. IEEE Symposium on Computational Intelligence in Bioinformatics and Computational Biology, 2025.
• Research on Diabetes Prediction Based on Multi-Model Comparison and SHAP Interpretability Analysis. 2025 17th International Conference on Bioinformatics and Biomedical Technology (ICBBT), 2025.
• Evaluating SHAP’s Robustness in Precision Medicine: Effect of Filtering and Normalization. IEEE International Conference on Bioinformatics and Biomedicine, 2023.
• Implementation-oriented machine learning for three-level maternal risk triage in primary antenatal care: stability, error profiles, and SHAP-based explainability. Network Modeling Analysis in Health Informatics and Bioinformatics, 2026.
• Explainable artificial intelligence for biomedical applications. Explainable Artificial Intelligence for Biomedical Applications, 2023.
• Explainable Artificial Intelligence in Medical Imaging: Fundamentals and Applications. Explainable Artificial Intelligence in Medical Imaging Fundamentals and Applications, 2025.
• Explainable artificial intelligence (XAI) in the veterinary and animal sciences field. Explainable Artificial Intelligence for Biomedical Applications, 2023.

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