A comprehensive narrative review evaluating the integration of artificial intelligence into genome-wide association studies indicates that machine learning and deep learning architectures offer powerful capabilities for capturing nonlinear genetic interactions, yet they encounter significant challenges regarding interpretability, data dependency, and clinical translation. High-throughput sequencing technologies generate vast, high-dimensional datasets wherein complex traits are rarely governed by single mutations but rather by polygenic architectures. Conventional statistical approaches in genome-wide association studies, which historically tested single nucleotide polymorphisms one at a time via univariate regression or linear mixed models, often struggle with non-additive interactions and linkage disequilibrium. To evaluate how modern computational intelligence addresses these limitations, the authors systematically examined thirty methodological approaches from nineteen publications spanning sixty-nine individual studies published between January 2019 and December 2024, grouping them into four core paradigms: directly interpretable models, post-hoc interpretability methods, representation learning frameworks, and performance-optimized architectures.
Standard genome-wide association pipelines follow a structured sequence extending from cohort data collection through quality control, imputation, association testing, meta-analysis, and post-GWAS functional annotation. In these classic frameworks, researchers commonly apply a rigorous genome-wide significance threshold of p < 5 × 10⁻⁸ to control false-positive rates arising from multiple hypothesis testing across millions of markers. As whole-genome sequencing expands the catalog to rare variants, thresholds can become even more stringent, reaching p < 2.5 × 10⁻⁸ or lower. Despite advancements, linear mixed models struggle to differentiate true causal variants from neighboring variants in strong linkage disequilibrium. Furthermore, classical additive models primarily measure narrow-sense heritability, leaving a substantial proportion of broad-sense heritability unexplained. The review emphasizes that machine learning provides computationally tractable heuristic approaches to detect epistasis and higher-order gene-gene interactions, which remain practically infeasible to search exhaustively across entire genomes using traditional brute-force statistical testing.
Within the first reviewed category, directly interpretable models maintain explicit mathematical relationships between input variants and predicted phenotypes. Statistical learning frameworks such as shrinkage regression, Bayesian ridge regression, and Elastic Net regularizations combine Least Absolute Shrinkage and Selection Operator penalties with ridge regression to resolve multicollinearity and limit overfitting. Ensemble structures, including bagging-based decision trees and gradient boosting classifiers, provide inherent feature importance metrics that permit the direct ranking of single nucleotide polymorphisms. Specialized hybrid frameworks, such as GASVeM pairing genetic algorithms with support vector machines, or JADBio-Gen employing Epilogi selection, output statistically equivalent biosignatures. A critical feature across these methods is the incorporation of biological knowledge into model architecture, exemplified by GenNet utilizing hierarchical variant-to-gene pathways, SLEM incorporating cellular constraints, and GNN-AD using linkage disequilibrium graphs to visualize epistatic relationships.
In contrast to intrinsically transparent models, post-hoc interpretability methods decouple high-capacity predictive modeling from subsequent biological explanation. Researchers frequently train complex architectures—such as multilayer perceptrons, deep neural networks, convolutional networks, and gradient boosting frameworks—using modern optimization tools like dropout, adaptive learning rates, and regularizations. Because these predictive models operate as non-transparent systems, post-hoc feature attribution algorithms are deployed to calculate variant importance scores. Techniques include Layer-wise Relevance Propagation, which propagates classification relevance backward through network layers, connection weight analysis, and permutation feature importance, which quantifies accuracy degradation when individual variants are permuted. To prevent false discoveries, investigations integrate these machine learning relevance scores with established statistical validation pipelines, applying false discovery rate corrections, p-value thresholds, or Mahalanobis distance outlier detection.
Representation learning frameworks represent a distinct conceptual shift by aiming to compress high-dimensional genotype matrices into informative, lower-dimensional latent embeddings rather than deriving direct single-marker associations. Fully connected networks extract successive hierarchical abstractions, while convolutional neural networks apply filters across variant sequences to exploit spatial correlations and capture local haplotype blocks. Autoencoder architectures reconstruct genomic inputs through compressed latent bottlenecks, facilitating tasks like phenotype imputation across large cohorts or generating pre-trained variant weights for downstream prediction. Sequence- and graph-based models, including Long Short-Term Memory units and Gated Recurrent Units, model long-range dependencies across distant loci. However, because the resulting latent embeddings rarely carry direct biological identities, representation learning models exhibit low intrinsic interpretability and serve primarily as feature extraction steps preceding formal association testing.
Performance-optimized models prioritize maximal predictive accuracy over biological interpretability, treating genetic risk assessment as an advanced pattern recognition challenge. These approaches utilize multilayer perceptrons, neural architecture searches guided by evolutionary algorithms, and stacked ensemble configurations combining feedforward and recurrent networks. They frequently integrate multi-layered data streams, combining genotype matrices with summary statistics, linkage disequilibrium scores, and functional annotations to predict variant pathogenicity or refine polygenic risk scores. Automated hyperparameter optimization, transfer learning, and large-scale pretraining enhance generalizability across independent cohorts. While effective for predicting clinical outcomes or adverse drug reactions from targeted variant panels, their complex multi-layered mechanics prevent direct attribution of phenotypic traits to specific single nucleotide polymorphisms, necessitating that they operate alongside traditional association pipelines.
The review identifies eight major methodological vulnerabilities that affect these artificial intelligence models to varying degrees. Confounding variables, notably population stratification arising from divergent ancestral backgrounds between cases and controls, can artificially inflate apparent predictive accuracy. If confounders like age, sex, or principal components are processed through standard hidden layers rather than dedicated adjustment pathways or adversarial deconfounding, neural networks may predict outcomes solely from demographic signals. The curse of dimensionality represents an equally severe barrier, as genomic datasets feature hundreds of thousands to millions of variant features against comparatively small sample sizes. Without aggressive preliminary feature filtering, standard dropout and regularization techniques often fail to prevent severe overfitting in architectures such as Deep-COMBI, DeepGenomeScan, and hybrid stacking networks.
Additional structural challenges include distributional differences across populations, hyperparameter tuning complexities, information leakage, sample dependency, and class imbalance. Variations in allele frequencies and linkage disequilibrium structures across cohorts impair models like DEEP*HLA and spatial recognition networks when tested outside their training populations. Information leakage arises when data transformations, feature selection, or imputation are performed across entire datasets rather than strictly within isolated cross-validation folds. Violations of independent and identically distributed assumptions frequently occur when case-control cohorts from disparate collection centers are pooled without accounting for platform or ascertainment discrepancies. Furthermore, severe class imbalance in rare disease studies can bias models toward majority classes, necessitating precision-recall evaluations over standard accuracy or receiver operating characteristic curves.
Data infrastructure plays an essential supporting role in training and validating these algorithms, with biobanks providing individual-level genotypic and phenotypic records while repositories distribute aggregated summary statistics. Major biobanks include the UK Biobank, which houses records on 488,377 individuals across 4,024 phenotypes, BioBank Japan with 199,982 participants spanning 85 disease categories, the H3Africa initiative covering 118,065 participants across diverse noncommunicable and infectious traits, and the United States All of Us initiative, which has enrolled 245,388 individuals to date. Repositories such as the GWAS Catalog host standardized summary statistics, compiling approximately 400,000 associations across 45,000 studies. The review notes a persistent geographic imbalance, with European ancestries representing 86.3% of historic studies, while East Asian and African cohorts comprise only 5.9% and 1.1% respectively, emphasizing the necessity of domain adaptation and equitable data collection.
Ultimately, the review concludes that while artificial intelligence offers substantial potential to decipher non-additive architectures and epistasis, current implementations have yet to achieve demonstrated clinical utility. The field suffers from a dual translational bottleneck characterized by an absence of rigorous external validation across ethnically diverse cohorts and a lack of prospective clinical utility trials. A high predictive accuracy or area under the curve does not inherently translate into actionable medical interventions or improved patient outcomes. Bridging this translational gap will require decentralized privacy-preserving architectures such as federated learning, comprehensive multi-omics data integration combining transcriptomic and proteomic layers, and rigorous real-world evaluation within electronic health record systems.
Subject of Research: Evaluation of machine learning and deep learning methodologies applied to genome-wide association studies
Article Title: Exploring the use of machine and deep learning in genome-wide association studies: a comprehensive review
Article References: D’Antona, S., Abdalla Abakar, M. E., Ramazzotti, D., Antoniotti, M., & Graudenzi, A. (2026). Exploring the use of machine and deep learning in genome-wide association studies: a comprehensive review. BioData Mining, 19(1), Article 72. https://doi.org/10.1186/s13040-026-00584-8
Image Credits: AI Generated
DOI: 10.1186/s13040-026-00584-8
Keywords: Genome-Wide Association Studies, Artificial Intelligence, Machine Learning, Deep Learning, Statistical Learning, Epistasis, Linkage Disequilibrium, Curse of Dimensionality, Model Interpretability, Feature Selection, Polygenic Risk Scores, Population Stratification, Biobanks, Genomic Data, Representation Learning, Post-Hoc Interpretability
News Source: Blake Davidson. (November 16, 2026). Machine Learning Enhances Genomic Association Studies but Faces Key Interpretability Hurdles. Scienmag.



