Integrating Artificial Intelligence with Global Genomic Resources: A Narrative Review of Implications for Precision Medicine

The rapid evolution of genomic science has fundamentally reshaped our understanding of human biology, revealing an extraordinary level of genetic diversity across populations worldwide. 1 Over the past two decades, large-scale initiatives have generated diverse global genomic resources, including population biobanks, reference datasets, disease-specific repositories, and genomic databases, rather than a single unified atlas. 2 In this review, the term global genomic resources are used to refer collectively to heterogeneous large-scale population genomic databases and biobank-based resources rather than one specific named atlas. These resources provide a comprehensive view of how genetic differences are distributed across populations and how they contribute to health and disease. 3 , 4 However, the sheer scale and complexity of such data present substantial analytical challenges. 5 Traditional statistical approaches, while valuable, are often insufficient to capture the intricate, multidimensional relationships embedded within genomic datasets. 6 As a result, there is a growing need for advanced computational frameworks capable of interpreting these complex patterns in a meaningful and clinically relevant manner. 7 Multi-omics integration can improve AI-based genomic interpretation, but it also raises challenges related to harmonization, interoperability, and bias in underrepresented populations. 8 Artificial intelligence has emerged as a powerful enabler in this context, offering innovative tools to process, integrate, and analyze large-scale genomic data. 9 Techniques such as machine learning and deep learning are particularly well suited for identifying hidden structures and associations within high-dimensional datasets. 10 By leveraging these approaches, researchers can move beyond linear models and uncover nonlinear interactions between genetic variants, environmental factors, and disease phenotypes. 11 This capability is especially important in the era of precision medicine, where the goal is to tailor healthcare interventions based on individual genetic profiles. 12 , 13 The integration of artificial intelligence with genomic resources therefore represents more than a technological advancement; it reflects a paradigm shift toward data-driven and individualized healthcare systems that are responsive to biological complexity. 14 At the same time, the concept of global genomic resources underscores the importance of population diversity in genomic research. 15 Human genetic variation is not evenly distributed, and differences across populations can significantly influence disease susceptibility, drug response, and clinical outcomes. 16 , 17 Historically, many genomic studies have been disproportionately focused on specific populations, leading to gaps in knowledge and potential biases in clinical applications. 18 , 19 The integration of diverse datasets into a unified analytical framework is essential for ensuring that genomic insights are broadly applicable and equitable. 20 Artificial intelligence provides a means to harmonize and analyze such heterogeneous data, enabling the identification of population-specific patterns while maintaining a global perspective. 21 This approach not only enhances scientific understanding but also supports the development of more inclusive and representative precision medicine strategies. 22 , 23 Although several reviews have examined artificial intelligence in precision medicine, cancer genomics, multi-omics integration, pharmacogenomics, and genomic data interpretation, 2 , 4 , 7 , 9 , 10 the existing review literature remains relatively fragmented across specific diseases, computational techniques, or clinical applications. In contrast, this review focuses on the broader integration of artificial intelligence with global genomic resources, with particular attention to how population diversity, large-scale genomic variation, predictive modeling, and clinical translation collectively shape precision medicine. This focus provides a timely synthesis that extends beyond technique-centered or disease-specific discussions by emphasizing the role of diverse genomic resources in developing more generalizable and equitable AI-supported precision medicine frameworks. Despite these promising developments, the integration of artificial intelligence with global genomic resources is not without challenges. Issues related to data standardization, interoperability, and quality remain significant barriers to effective data integration. 24 Furthermore, the use of sensitive genetic information raises important ethical and legal considerations, including concerns about privacy, consent, and data security. 25 Addressing these challenges requires a multidisciplinary effort that brings together expertise from genomics, computer science, clinical medicine, and ethics. 26 , 27 Accordingly, this narrative review was guided by the following question: how can artificial intelligence be integrated with global genomic resources to support precision medicine, and what implications does this integration have for population diversity, predictive modeling, clinical translation, and ethical governance? Specifically, the review aims to synthesize current evidence on AI-enabled genomic data integration, examine the contribution of diverse population data to model generalizability and equity, identify major applications in disease risk prediction and treatment response, and discuss key barriers affecting responsible clinical implementation.

This narrative review was conducted to explore the integration of artificial intelligence with global genomic resources and human genetic variation and its implications for precision medicine. An interpretive and integrative approach was adopted to address the interdisciplinary nature of the topic, allowing the synthesis of evidence from genomics, computational science, and clinical research. This design provided a flexible yet structured way to examine how artificial intelligence supports the analysis and application of complex genomic data. The conceptual framework guiding this review is illustrated in Figure 1 , which presents the integration process as a continuum from diverse data inputs such as genomic, multi omics, and electronic health record data through an artificial intelligence layer incorporating machine learning, deep learning, graph-based models, and explainable AI. These approaches support key processes including data integration, pattern recognition, and predictive modeling, ultimately leading to clinical outputs such as disease prediction, risk stratification, personalized treatment, and decision support. This framework provides a structured basis for synthesizing the literature and interpreting how artificial intelligence enables the translation of complex genomic data into clinically meaningful insights, while also emphasizing the importance of scalability, diversity, and real-world applicability in precision medicine. A narrative and integrative review design was selected instead of a systematic or scoping review because the objective was to synthesize concepts, applications, methodological patterns, and translational implications across heterogeneous evidence rather than to answer a narrowly defined intervention question or map a single bounded evidence field. No protocol was registered because the review was not designed as a formal systematic or scoping review. Figure 1 Integration of AI with global genomic resources for precision medicine. The figure illustrates how genomic data, multi-omics data, electronic health record data, and biobank data can be integrated through AI-based approaches to support precision medicine. The green panel represents input data sources, the blue panel represents the AI layer, the purple panel represents analytical processes, and the red panel represents clinical outputs. Arrows indicate the direction of information flow from inputs to AI analysis, analytical processing, and clinical outputs. Numbered circles indicate the sequential analytical processes of data integration, pattern recognition, and predictive modeling. Bold text indicates the main categories and subcategories within the framework. A diagram of AI integration with genomic data for precision medicine. The diagram shows AI′s role in precision medicine using genomic data. It has four sections: Inputs, AI Layer, Processes and Outputs. Inputs cover Genomic Data (whole-genome sequencing, GWAS, transcriptomics), Multi-Omics Data (proteomics, epigenomics, metabolomics) and EHR & Biobank Data (health records, clinical phenotypes, multi-ethnic populations). The AI Layer includes Machine Learning (supervised, unsupervised, ensemble methods), Deep Learning (neural networks, transformers, multimodal architectures), Graph-Based Models (gene networks, knowledge graphs) and Explainable AI (feature importance, model interpretability). Processes involve Data Integration (harmonization, cross-omics), Pattern Recognition (variant association, subtype discovery) and Predictive Modeling (risk stratification, treatment prediction). Outputs are Disease Prediction, Risk Stratification, Personalized Treatment and Clinical Decision Support. Integration of AI with global genomic resources for precision medicine. The figure illustrates how genomic data, multi-omics data, electronic health record data, and biobank data can be integrated through AI-based approaches to support precision medicine. The green panel represents input data sources, the blue panel represents the AI layer, the purple panel represents analytical processes, and the red panel represents clinical outputs. Arrows indicate the direction of information flow from inputs to AI analysis, analytical processing, and clinical outputs. Numbered circles indicate the sequential analytical processes of data integration, pattern recognition, and predictive modeling. Bold text indicates the main categories and subcategories within the framework. A comprehensive and carefully structured search strategy was employed to identify relevant literature. Major scientific databases and supplementary search sources were used, including PubMed, Scopus, Web of Science, IEEE Xplore, Embase, and Google Scholar. The literature search was conducted from 5 March to 8 April 2026. Full database-specific search strings are provided in Table 1 to improve reproducibility. The search was designed to capture studies at the intersection of artificial intelligence and genomic variation. Keywords were selected based on their relevance to the research topic and combined using Boolean operators to ensure a broad yet focused retrieval of studies. Terms such as artificial intelligence, machine learning, genomic diversity, population genomics, global genomic resources, and precision medicine were used in various combinations. The search process was iterative, meaning that initial findings informed subsequent refinements of keywords and search strings. Additionally, reference lists of selected articles were manually reviewed to identify further relevant studies that might have been missed during database searches. Table 1 Search Strategy Database Search String Keywords Used Boolean Operators Filters Applied Search Fields Search date Notes PubMed (“Artificial Intelligence”[MeSH] OR “Machine Learning” OR “Deep Learning” OR “Neural Networks”) AND (“Genomic Variation”[MeSH] OR “Population Genomics” OR “Genetic Diversity” OR “Global Genetic Atlas”) AND (“Precision Medicine”[MeSH] OR “Personalized Medicine”) Artificial intelligence; Machine learning; Deep learning; Neural networks; Genomic variation; Population genomics; Genetic diversity; Precision medicine AND; OR English; Humans; Journal Articles Title/Abstract; MeSH Terms 5 March to 14 March 2026 Controlled vocabulary (MeSH) used for high sensitivity and specificity Scopus (TITLE-ABS-KEY(“artificial intelligence” OR “machine learning” OR “deep learning” OR “neural networks”) AND TITLE-ABS-KEY(“genomic variation” OR “population genomics” OR “genetic diversity” OR “global genetic atlas”) AND TITLE-ABS-KEY(“precision medicine” OR “personalized medicine”)) Artificial intelligence; Machine learning; Deep learning; Neural networks; Genomic variation; Genetic diversity; Precision medicine AND; OR English; Articles; Reviews Title/Abstract/Keywords 15 March to 22 March 2026 Broad interdisciplinary coverage Web of Science (TS=(“artificial intelligence” OR “machine learning” OR “deep learning” OR “neural networks”) AND TS=(“genomic variation” OR “population genomics” OR “genetic diversity” OR “global genetic atlas”) AND TS=(“precision medicine” OR “personalized medicine”)) Artificial intelligence; Machine learning; Deep learning; Genomic variation; Population genomics; Precision medicine AND; OR English; Articles; Reviews Topic (Title/Abstract/Keywords) 23 March to 30 March 2026 High-quality indexed journals IEEE Xplore (“All Metadata”:“artificial intelligence” OR “machine learning” OR “deep learning”) AND (“All Metadata”:“genomic data” OR “genomic variation” OR “population genomics”) AND (“All Metadata”:“precision medicine” OR “personalized medicine”) Artificial intelligence; Machine learning; Deep learning; Genomic data; Precision medicine AND; OR English; Conferences; Journals All Metadata 31 March to 3 April 2026 Focus on computational and engineering studies Embase (“artificial intelligence”/exp OR “machine learning” OR “deep learning”) AND (“genomic variation”/exp OR “population genomics” OR “genetic diversity”) AND (“precision medicine”/exp OR “personalized medicine”) Artificial intelligence; Machine learning; Genomic variation; Genetic diversity; Precision medicine AND; OR English; Humans; Articles Title/Abstract; Emtree Terms 4 April to 5 April 2026 Extensive biomedical and pharmacological database Google Scholar “artificial intelligence” AND “genomic variation” AND “precision medicine” Artificial intelligence; Genomic variation; Precision medicine AND English Full text 6 April to 8 April 2026 Used for grey literature and additional records Search Strategy To ensure the relevance and quality of the included literature, clear inclusion and exclusion criteria were established. Studies were included if they addressed at least one of the following areas: application of artificial intelligence in genomic data analysis, development or use of large-scale genetic variation datasets, or implications of such integration for personalized or precision medicine. Both original research studies and high-quality review articles were considered, as they provided complementary perspectives on the topic. Studies were excluded if they focused solely on genetic variation without any computational or analytical integration, or if they discussed artificial intelligence in contexts unrelated to genomics or healthcare. Non-scholarly sources, such as blog posts or unverified reports, were also excluded to maintain academic rigor. The criteria were applied consistently throughout the selection process. The process of selecting studies involved multiple stages to ensure a systematic yet flexible approach. Initially, titles and abstracts of retrieved articles were screened to identify those that appeared relevant to the research objectives. This step helped to narrow down the large pool of studies to a more manageable number. Subsequently, full-text articles were reviewed in detail to confirm their eligibility. During this stage, particular attention was paid to how well each study aligned with the central theme of integrating artificial intelligence with genomic variation for clinical applications. When uncertainties arose, studies were re-evaluated to determine their contribution to the overall narrative. This iterative process allowed for thoughtful inclusion of diverse perspectives while maintaining coherence and focus. Two reviewers independently screened titles and abstracts, followed by independent full-text assessment. Disagreements were resolved through discussion, and unresolved cases were adjudicated by a third reviewer. Because this was a narrative review, no formal inter-rater agreement statistics were calculated. A structured data extraction approach was adopted to systematically capture key information from the selected studies. For each article, details such as study objectives, type of genomic data used, artificial intelligence techniques applied, and reported outcomes were recorded. Additional information regarding study design, population characteristics, and clinical implications was also documented. The extraction process emphasized clarity and consistency, ensuring that comparable data could be analyzed across studies. Rather than simply listing findings, the extracted data were organized into meaningful categories, enabling deeper interpretation and synthesis. This approach helped to identify recurring patterns and relationships within literature. Because the included studies used heterogeneous designs, a structured narrative quality and risk-of-bias consideration was conducted to guide interpretation of the evidence. This appraisal was not used to exclude studies but to distinguish more robust empirical evidence from preliminary or exploratory findings. For quantitative and model-development studies, attention was given to study design, dataset size and transparency, population representation, validation strategy, external or prospective testing, reporting of model performance, and potential risks of overfitting or bias. For qualitative stakeholder studies, appraisal considered the relevance of participants, clarity of data collection, transparency of analysis, and credibility of interpretation. For proof-of-concept studies, bibliometric analyses, educational evaluations, and conceptual papers, emphasis was placed on methodological transparency, relevance to AI-genomics integration, and whether conclusions were proportionate to the evidence presented. This approach allowed the synthesis to retain the breadth of a narrative review while interpreting findings according to the strength and maturity of the supporting evidence. The analysis of the extracted data was conducted using a thematic synthesis approach, informed by the quality and risk-of-bias consideration described above. Studies were grouped according to key themes that emerged during the review process. Themes were derived through repeated reading of extracted study characteristics and findings, followed by grouping of recurring concepts related to genomic data integration, population diversity, predictive modeling, clinical translation, and ethical governance. The synthesis compared consistencies and limitations across studies, with attention to validation quality, population representation, and translational maturity. These themes included the role of artificial intelligence in genomic data integration, the importance of population diversity in global genomic resources, advancements in predictive modeling, and challenges related to clinical translation. Each theme was examined in detail to understand how different studies contributed to the broader narrative. The synthesis process involved comparing findings across studies, identifying consistencies and discrepancies, and interpreting their implications for precision medicine. This approach allowed for a nuanced and comprehensive understanding of the topic, moving beyond simple description to critical interpretation.

A total of 54 studies were included in this review ( Figure 2 ), representing a diverse and multidisciplinary body of evidence on the integration of artificial intelligence with global genomic resources and human genetic variation. The included studies comprised a range of research designs, including retrospective cohort studies, machine learning model development and validation studies, multi-omics integration analyses, and database or knowledge-graph construction, alongside a smaller number of conceptual investigations ( Table 2 ). Across these studies, a wide spectrum of genomic data types was utilized, including whole-genome sequencing, whole-exome sequencing, genome-wide association study data, and transcriptomics, often integrated with proteomics, epigenomics, and metabolomics ( Figure 3 ). Notably, many studies linked genomic datasets with electronic health records and large-scale biobank data, particularly from multi-ethnic populations, underscoring the growing emphasis on incorporating global genetic diversity to improve the representativeness and clinical relevance of findings. Table 2 Overview of Included Studies Authors Study Type Population Characteristics Study Objectives Type of Genomic Data Used Artificial Intelligence Techniques Applied Reported Outcomes Clinical Implications Johnson et al, 28 Observational genomic cohort study N=36,736 individuals from UCLA ATLAS biobank; diverse ancestries including European American, African American, Hispanic/Latino American, East Asian American, South Asian American; linked EHR data To explore genetic ancestry diversity and its relationship with disease phenotypes using EHR-linked genomic data; to assess interplay between genetic ancestry and self-identified race/ethnicity; to conduct GWAS and PheWAS analyses Genome-wide SNP data (≈667,191 genotyped SNPs; imputed to >8 million variants using the TOPMed panel) linked with electronic health records. Principal Component Analysis (PCA); K-nearest neighbors (KNN) clustering; Identity-by-Descent (IBD) analysis; ADMIXTURE (unsupervised clustering); GWAS; PheWAS Identification of 5 major continental genetic ancestry clusters and subcontinental groups; 259 significant associations between genetic ancestry and disease phenotypes; GWAS identified loci associated with liver disease; PheWAS revealed associations with neurologic and neoplastic conditions; discrepancies between genetic ancestry and self-identified race/ethnicity Demonstrates value of integrating genomic data with EHRs for precision medicine; highlights importance of diverse populations; shows genetic ancestry provides additional clinical insight beyond self-identified race/ethnicity; supports improved disease risk prediction and reduction of healthcare disparities Bakhshayeshi et al, 29 Empirical computational study (model development and validation) 47,318 samples including neurocognitive disorder (ND) patients and healthy controls; additional schizophrenia cohort (n=886); prostate cancer dataset for external validation To develop and validate an interpretable, uncertainty-aware multi-modal deep learning framework (CLinNET) for gene prioritization and variant interpretation in neurocognitive disorders Copy number variations (CNVs), single nucleotide variants (SNVs), gene expression data, biological pathway data (Reactome), gene ontology (GO) annotations Multi-modal deep neural network (GPINN architecture), dual-branch pathway-informed networks, SHAP (Shapley Additive Explanations), uncertainty quantification, confidence-based classification, cross-validation (20-fold) F1-score 76.4%, accuracy 77.2%, AUC-PR 84%, AUC-ROC 84%; precision improved to 87% with uncertainty filtering; identification of 78 ND-associated genes (p=1.2e-11); improved diagnostic efficiency (45% reduction in testing burden) Enhances gene prioritization and variant interpretation in neurocognitive disorders; supports precision medicine; improves diagnostic accuracy and confidence; enables biologically interpretable AI for clinical genomics Ahmed et al, 30 Original research (database and application development) Not reported (no human subjects; aggregated genomic databases) To develop a comprehensive gene-SNP-disease-drug database with mobile access for clinical genomics and precision medicine Genes (>59,000), germline SNPs (>67,000), somatic mutations (>1 million), gene-disease associations, ICD and NDC codes from multiple databases Not explicitly applied; bioinformatics/database integration (AI mentioned conceptually) Centralized searchable database and iOS application enabling gene–disease–drug associations Supports precision medicine, improves disease prediction, enhances clinical decision-making and drug targeting Kearns et al, 31 Technical advance / experimental genomic study 12 patients with endometriosis and matched controls; women aged 28–49 years; all of White ethnicity To develop and apply a biomimetic digital twin ecosystem combined with phenotype-driven variant analysis and exome sequencing to identify molecular mechanisms and biomarkers of endometriosis Whole-exome sequencing (NGS), variant call data (VCF), phenotype-ranked genomic variants, knowledge graph-integrated genomic data Phenotype-driven variant ranking (QIAGEN Clinical Insight), biomimetic digital twin ecosystem, knowledge engineering, AI-assisted variant interpretation Identified 8 pathogenic mutations and 4 variants of unknown significance (MUC20, USP17L1, FAM66B, DEFB109B); potential biomarker detected; chromosomal hotspot identified on chromosome 8 Potential for improved diagnosis of endometriosis, biomarker discovery, understanding disease mechanisms, and enabling personalized medicine and virtual clinical trials Zhang et al, 32 Original research article; dual-database cross-validated bibliometric analysis No human participants or patient samples; bibliographic dataset of 1,809 publications on AI-driven precision medicine, including 810 Web of Science Core Collection records and 999 Scopus records published from 2015–2025. To map the intellectual structure, thematic evolution, collaboration patterns, knowledge base, and emerging frontiers of AI-enabled precision care/personalized medicine using independent WoSCC and Scopus analyses with cross-database consistency checks. No primary genomic dataset was analyzed; the bibliometric corpus included literature themes related to genomics, multi-omics, proteomics, radiomics, biomarkers, imaging, and clinical data integration. Bibliometric science-mapping and visualization approaches using bibliometrix, VOSviewer, CiteSpace, BiblioVista, Scimago Graphica, keyword co-occurrence networks, co-citation analysis, burst detection, thematic evolution, Spearman rank correlations, and Jaccard similarity. The field shifted from algorithm-centric studies toward multimodal/multi-omics integration, mechanism-informed interpretability, patient stratification, therapy optimization, imaging AI, hybrid clinical decision support, and emerging LLM-agent paradigms. The United States and China were central research hubs; Harvard Medical School, Stanford University, and major Chinese institutions were key institutional contributors. Findings support prioritizing interpretable multimodal AI, mechanism-constrained integration, patient stratification, therapy optimization, workflow-integrated clinical decision support, and deployable digital-health systems for precision medicine. Murcia Pienkowski et al, 33 Original research 71 patients with advanced colorectal adenocarcinoma; mean age 68; 65% male; stages II–IV; Poland Identify novel therapeutic targets for precision medicine in colorectal cancer using AI-based approaches Whole exome sequencing (tumor and blood), RNA sequencing (tumor), copy number variations, single nucleotide variants, methylation data AI-based neoantigen prediction (ARDentify, ARDisplay); machine learning models; multimodal variational autoencoder (MVAE); deep learning integration of multi-omics data ~8,700 neoepitopes identified; synthetic lethality pairs (APC–CTNNB1, BRAF–DUSP4, APC–TCF7L2, VPS4A–VPS4B); improved tumor–cell line alignment using MVAE Supports personalized cancer therapy; enables patient-specific targets; highlights need for individualized immunotherapy and synthetic lethality approaches Gao and Cui. 34 Empirical machine learning/model-development study Multi-ancestral dbGaP datasets for lung cancer, prostate cancer, and Alzheimer’s disease, including European, East Asian, African American, and Latin American populations To improve clinico-genomic disease prediction across ancestry groups and reduce performance disparities for data-disadvantaged populations Genome-wide SNP/genotype data combined with clinical variables; synthetic multi-ancestry genotype datasets also used Deep neural networks, deep transfer learning, logistic regression, LR-based transfer learning, PRS-CSx comparison Deep transfer learning improved prediction accuracy for data-disadvantaged ancestry groups, while linear transfer learning showed weaker improvement Supports equitable AI-based precision medicine by improving disease prediction across underrepresented ancestry groups without reducing performance in other groups Foksinska et al, 35 Original research Rare disease patients; includes pediatric and adult cases; >500 families supported; specific cases include TMLHE variant (pediatric male), RHOBTB2 variants (two pediatric patients), and cyclic vomiting case To describe and evaluate the mediKanren AI-driven precision medicine process for identifying therapeutic options in rare and complex diseases Genetic reports, whole-genome sequencing, whole-exome sequencing, genetic variants, phenotype data (HPO terms) Knowledge graph†based reasoning (mediKanren), logic programming (miniKanren), biomedical knowledge integration, query-based inference (one-hop and two-hop queries) Identification of therapeutic candidates (eg, levocarnitine, celecoxib, isopropyl alcohol); improved patient symptoms in some cases; >1,300 drug candidates identified for SARS-CoV-2; successful hypothesis generation for clinical trials Supports AI-assisted drug repurposing in precision medicine; enhances rare disease treatment discovery; facilitates clinician decision-making; demonstrates scalable AI-driven translational research approach Smith et al, 36 Empirical computational method-development and validation study using retrospective multi-dataset cancer genomics data. TCGA breast, thyroid, and uterine cancer cohorts were used: BRCA 842 individuals split into 21 ancestry-specific datasets; THCA 436 individuals split into 28 datasets; UCEC 491 individuals split into 15 datasets. External BRCA validation included 1039 individuals from 14 breast cancer studies across African, American, Asian, European, and unknown ancestry groups. To develop and validate PhyloFrame, an ancestry-aware equitable machine-learning framework designed to reduce ancestral bias and improve precision-medicine predictions across diverse populations. Bulk RNA-seq transcriptomic training data, TCGA cancer data, population genomics data, ancestry information, enhanced allele frequency (EAF) variants, SNP-level ancestry-enrichment information, and functional interaction network data. PhyloFrame framework integrating logistic regression with LASSO for initial disease signatures, functional interaction network projection/propagation, EAF-based ancestry-aware gene filtering, ridge regression to force inclusion of equitable genes, and comparison against a benchmark logistic-regression model without population-genomics amplification. PhyloFrame improved predictive performance across most metrics, including MCC, AUC, F-measure, Cohen’s kappa, recall, and accuracy; reduced overfitting; produced more stable disease signatures; identified known cancer-related genes more consistently; improved prediction among underrepresented and admixed ancestry groups; and outperformed the benchmark in external BRCA validation. Equitable AI methods can mitigate ancestral bias in omics-based precision medicine, improve prediction for underrepresented populations, support more generalizable disease signatures, and contribute to fairer clinical decision-making in cancer genomics. Saghapour et al, 37 Original research; proof-of-concept bioinformatics/transcriptomic visualization study LNCaP prostate cancer cell line treated with DMSO control, mefloquine, tamoxifen, withaferin A, and pairwise/triple combinations; triplicate samples across 0, 3, 6, 9, 12, and 24 h To introduce and evaluate Temporal GeneTerrain for visualizing dynamic temporal gene-expression changes and drug-induced transcriptomic perturbations in precision medicine Time-course transcriptomic gene-expression data from GEO dataset GSE149428 ; 1,000 most variably expressed genes selected, with 999 strongly co-expressed genes retained after Pearson correlation filtering (r ≥ 0.5) Temporal GeneTerrain visualization framework using Z-score normalization, Pearson correlation filtering, protein-protein interaction network integration, Kamada-Kawai force-directed embedding, Gaussian density field mapping, adaptive smoothing, temporal integration, and pathway enrichment analysis using PAGER Temporal GeneTerrain revealed dynamic transcriptional shifts, delayed NGF-stimulated transcription, unfolded protein response activation, zinc homeostasis changes, and stronger perturbations under combination treatments, especially TMW, compared with minimal changes for mefloquine and tamoxifen alone Supports precision medicine by improving interpretation of time-dependent drug responses, identifying phase-specific pathway activation, and informing time-sensitive therapeutic targeting and personalized treatment strategies in cancer research Chandak et al, 38 Data descriptor / resource development study Not reported (no human participants; integrated biomedical datasets) To develop a multimodal precision medicine knowledge graph (PrimeKG) integrating diverse biomedical data sources for disease understanding and drug-disease relationships Multi-omics and biomedical data including genomics, transcriptomics, proteomics, phenotypic data, drug data, and clinical text Knowledge graph construction; multimodal data integration; ontology mapping; natural language processing; network-based AI analytics PrimeKG includes 129,375 nodes and 4,050,249 relationships across 10 biological scales; improved disease coverage; integrated drug-disease relationships Supports precision medicine, drug repurposing, clinical decision support, and personalized treatment strategies Fu et al, 39 Observational genetic risk prediction study using real-world EHR-linked biobank data with external validation Hispanic Latino Americans (N=610; 126 dementia cases), African Americans (N=440; 84 dementia cases), and East Asian Americans (N=673; 75 dementia cases) from UCLA ATLAS; validation in All of Us cohort among Hispanic Latino American and African American groups To improve dementia genetic risk prediction in underrepresented populations by integrating functional genomic prioritization, GWAS-derived SNPs, and machine learning models Genome-wide SNP/genotype data prioritized from Alzheimer’s disease and neurodegenerative disease GWASs; APOE-ε4 status; polygenic risk score SNPs; EHR-linked genotype data; overlapping UCLA ATLAS and All of Us SNPs Elastic Net SNP model; APOE-ε4 count model; Alzheimer’s disease PRS models; multi-PRS models; sensitivity testing with Gradient Boosting Machine and XGBoost; FUMA-based SNP prioritization and gene mapping Elastic Net gene-annotated SNP model outperformed APOE and PRS models across ancestry groups, improving AUPRC by 31–84% and AUROC by 11–17%; identified shared and ancestry-specific dementia risk SNPs, genes, and pathways Supports more accurate ancestry-aware dementia risk prediction, improves precision medicine strategies in underrepresented populations, and may aid earlier dementia diagnosis and targeted intervention planning Marth et al, 40 Empirical genomic sequencing study; 1000 Genomes Exon Pilot Project using targeted exon-capture sequencing and bioinformatics variant analysis 697 samples from seven populations representing Africa (LWK, YRI), Asia (CHB, CHD, JPT), and Europe (CEU, TSI); approximately 1.43 Mb of exonic sequence covering 8,279 exons and 942 genes To develop effective informatics pipelines for targeted capture sequencing data, characterize coding variation across populations, and assess functional properties of common, intermediate, and low-frequency coding variants Deep-coverage exon-capture sequencing data from approximately 1,000 human genes; exonic SNPs and short insertion/deletion variants from targeted coding regions Bioinformatics variant-calling and functional prediction pipelines, including MOSAIK, MAQ/SSAHA2, GigaBayes, GATK Unified Genotyper, SIFT, and PolyPhen-2 Identified 12,758 exonic SNPs, 70% novel and 74% below 1% allele frequency; validation rate 96.6%; rare coding variants showed increased population specificity and enrichment for functional/damaging variants; identified 21 insertions and 75 deletions Supports the clinical and research value of targeted exome sequencing for detecting rare functional coding variants and informs disease-gene discovery, variant interpretation, and population-specific genetic risk assessment Thorn et al, 41 Observational cohort study 7,253 female breast cancer patients from Genomics England, TCGA, BCN Biobank-Barts, and Genes & Health; ancestry groups included European, African, and South Asian populations. To evaluate molecular and clinical effects of genetic ancestry in African and South Asian women with breast cancer compared with European women using multimodal cohort data. Whole-genome sequencing; germline sequencing; somatic mutation data; ancestry inference; tumour mutational burden; mutational signatures; homologous recombination deficiency indicators; EHR-linked clinical data. Machine learning-based variant calling was noted for Strelka2; HRDetect supervised lasso logistic regression model was used to classify HR-deficient tumours. African and South Asian patients presented earlier and died at younger ages than European patients; African patients had higher grade and hormone receptor-negative disease; South Asian patients showed lower tumour mutational burden; ancestry-associated somatic and germline mutation differences were identified; BRCA1/BRCA2/PALB2 and HR-deficiency-related therapeutic opportunities were highlighted. Findings support ancestry-informed precision oncology, more equitable screening windows, improved genetic testing strategies, and identification of pharmacogenomic targets such as PARP inhibitor-related opportunities. Rhead et al, 42 Observational molecular epidemiology study 8454 colorectal cancer patients; primarily late-stage; median age 60.7 years; diverse ancestry groups (AFR, AMR, EAS, EUR, SAS) To identify associations between genetic ancestry, race/ethnicity, and molecular tumor profiles in colorectal cancer Targeted NGS (Tempus xT 648-gene panel); RNA-Seq; ancestry-informative markers Statistical modeling (multivariable logistic regression, likelihood ratio tests); ancestry estimation algorithms; gene set analysis (GSVA, mROAST) Higher AFR ancestry associated with increased APC, KRAS, and PIK3CA mutations and decreased BRAF mutations; EAS ancestry associated with lower KRAS mutations; differences in gene expression and CMS subtypes Genetic ancestry influences tumor biology and mutation patterns; supports personalized oncology approaches and highlights need for diversity in genomic datasets Boquet-Pujadas et al, 43 Original research (multi-cohort observational study) 129,340 participants from MULTI consortium including UK Biobank and independent cohorts; multi-organ disease populations (brain, eye, heart) To develop AI-derived multi-organ endophenotypes to characterize disease heterogeneity and shared etiology across brain, eye, and heart diseases Genome-wide SNP data (GWAS), proteomics (Olink), and multi-omics integration Weakly supervised deep learning (Surreal-GAN), multi-modal integration, GWAS, PWAS, ProWAS Identification of 11 AI-derived endophenotypes; associations with Alzheimer’s disease progression, mortality, and disease risk; gene–drug–disease links Improved patient stratification, precision medicine applications, disease risk prediction, and drug discovery insights Taiwan Precision Medicine Initiative. 44 Empirical cohort study; large-scale population-based precision medicine cohort 565,390 consented participants recruited through 16 partner medical centres and affiliated hospitals in Taiwan; 486,956 participants had matched genotype and EMR data; predominantly Han Chinese ancestry; male and female participants; most aged 20–90 years To establish the Taiwan Precision Medicine Initiative cohort by integrating population-optimized genetic profiling with electronic medical records for large-scale precision medicine research, disease-risk prediction, pharmacogenetic response assessment, and longitudinal health studies Population-optimized SNP genotyping arrays (TPMv1 and TPMv2); imputed genome-wide variants; genotype data linked with longitudinal electronic medical records; GWAS, PheWAS, PRS, and pharmacogenetic data resources Polygenic risk score modeling; genome-wide association and phenome-wide association analysis pipelines; imputation and population-structure analysis; NLP models and regular expressions for extracting lifestyle and cognitive-test information from free-text EMRs 565,390 participants enrolled; 486,956 participants had matched genotype and EMR data; 388,545 novel variants identified; GWAS results replicated and identified novel T2D-associated SNPs; PRS for T2D achieved AUC 0.65 alone and 0.86 after adding age, sex, and BMI; 265 dichotomized phecode PRSs and 24 quantitative-trait PRSs developed Provides a major non-European precision medicine resource; supports validation of genetic risk prediction models; enables risk-based health management trials, pharmacogenetic implementation, targeted prevention, and population-specific precision health policy development Xu et al, 45 Observational cohort study with machine learning modeling INTERVAL study; ~50,000 participants; multi-ancestry validation cohorts including European, Asian, and African American populations; subset sizes varied by omics platform To develop and validate genetic scores for predicting multi-omic traits and assess their utility in disease association and biological pathway analysis Genome-wide genotyping (~10.5 million variants); RNA sequencing; proteomics data from SomaScan and Olink; metabolomics data from Metabolon and Nightingale Bayesian ridge regression; machine learning-based genetic score modeling Development of 17,227 genetic scores; 10,522 significant traits; prediction performance up to R 2 > 0.5; cross-ancestry validation; disease associations identified through PheWAS; pathway insights including JAK–STAT signaling Enables cost-effective prediction of multi-omic traits; supports disease risk stratification; facilitates biomarker discovery; enhances precision medicine and large-scale biobank analyses Abdullah et al, 46 Experimental proof-of-concept AI framework study 126 subjects; 1305 respiratory recordings; included pediatric and adult patients across asthma, bronchiectasis, bronchiolitis, COPD, healthy respiratory state, LRTI, pneumonia, and URTI To develop and evaluate a multimodal AI framework for automated multiclass lung disease diagnosis from respiratory sounds, simulated biomarker fusion, and personalized medication recommendation Simulated molecular biomarker profiles, including CRP, IL-6, MMP-9, TGF-beta, IL-5, and YKL-40, fused with audio-derived respiratory features Hybrid deep learning model combining CNN, LSTM, hierarchical attention mechanisms, multimodal molecular-audio fusion, synthetic prescription modeling, and SHAP explainability LOSO accuracy 99.99%; external held-out accuracy 98.99%; class AUCs above 0.98; biomarker-audio fusion improved macro accuracy by approximately 0.5% and macro AUC by 1.1%; prescription model achieved high agreement with clinicians, including kappa=0.91 for drug selection, kappa=0.78 for dosage, and kappa=0.96 for frequency Demonstrates potential for AI-assisted respiratory diagnosis, biomarker-informed decision support, and personalized prescription recommendation; requires validation with real patient records and regulatory review before clinical use Fritzsche et al, 47 Qualitative study 28 stakeholders including researchers, pharmaceutical representatives, and patient board members; AD/Pso context Explore ethical challenges in biomarker research and precision medicine for AD/Pso Not reported (clinical, molecular, and genetic data referenced broadly) AI and data-driven biomarker approaches (not specified) Ethical challenges identified: harm, injustice, uncertainty; issues in stratification and data use Supports ethical policy development and patient-centered biomarker implementation Breeze et al, 48 Empirical observational study; epigenome-wide association study (EWAS) Discovery sample included up to 5,428 multi-ethnic adults: 2,879 African American, 1,737 European/European American, and 812 Hispanic/Latino participants; replication sample included up to 8,109 participants from independent cohorts. To identify trans-ethnic and ethnic-specific differentially methylated positions associated with estimated glomerular filtration rate (eGFR) using a genome-wide approach. Whole-blood DNA methylation data measured using Illumina 450K and EPIC methylation arrays; kidney tissue methylation and meQTL data were also used for integrative analyses. No explicit AI or machine-learning model reported; analyses used statistical EWAS models, fixed-effect inverse-variance meta-analysis, integrative epigenomics, eFORGE/eFORGE-TF, and pathway analysis. Identified 93 eGFR-associated DMPs at FDR 0.05; 13 trans-ethnic DMPs and 1 African American-specific DMP replicated; DMPs were enriched in kidney regulatory elements and kidney-development pathways. Findings suggest epigenetic mechanisms linked to kidney function and chronic kidney disease risk, supporting future biomarker development and precision-medicine research across diverse populations. Ge et al, 49 Observational study with methodological development and validation Multi-ethnic populations: eMERGE (European n≈54,793; African n≈12,472; Hispanic/Latino n≈2,374), UAB Black cohorts (REGARDS, GenHAT, HyperGEN, WPC), Taiwan Biobank (~110,926 participants); includes T2D cases and controls across diverse ancestries. To develop and validate a trans-ancestry polygenic risk score (PRS) for type 2 diabetes using multi-ethnic GWAS data and assess its predictive performance across diverse populations Genome-wide SNP data from multi-ethnic GWAS; imputed genotype data (HRC, TOPMed, 1000 Genomes); HapMap3 variants (~1.26 million variants) Bayesian polygenic modeling (PRS-CSx), PRS-CS-auto, LDpred2-auto, machine learning-based phenotyping (PheCAP), statistical modeling Top 2% PRS associated with ~2.5–4.5-fold increased T2D risk; AUC up to ~0.70; variance explained up to ~15.3%; improved predictive performance over covariates. Enables early risk stratification of T2D across diverse populations; supports clinical integration of PRS; highlights need for equitable genomic representation Rahman et al, 50 Mixed-methods educational evaluation study of an AI-enabled precision medicine training platform 1,495 graduated learners evaluated for learning confidence; 72 certified respondents surveyed for satisfaction; learners came from broad educational and socioeconomic backgrounds across all 50 US states and 92 countries; 90.2% were low- and middle-income learners with free access. To evaluate Stanford Data Ocean precision medicine training outcomes, AI Tutor performance, learner satisfaction, post-program academic and career impacts, and AI Data Visualization capabilities. Multi-omics and biomedical datasets including genomics, transcriptomics/RNA-seq, epigenomics, proteomics, metabolomics, microbiome/16S, clinical test data, wearable data, VCF files, and 107 genome sequences from iPOP and public datasets. Large language models including GPT, Claude, Gemini, and LLaMA; LLM-powered AI Tutor; embedding-based retrieval; prompt engineering; AI-driven data visualization; automated code generation; multi-modal data analysis; guardrail mechanisms. Overall Bioinformatics program completion rate was 50.5%, and structured cohort completion rate was 85.7%; learner confidence increased substantially, including Cloud Computing from 33.9% to 97.6% and Bioinformatics from 47.8% to 96.1%; 97.5% reported positive program impact; AI Tutor guardrail precision and specificity were 100%, recall was 93.4%, and F1 score was 96.6%; 76.56% of AI Tutor responses were positively rated. Supports scalable and inclusive precision medicine education; improves access to multi-omics data interpretation and AI-assisted biomedical analytics; helps clinicians, researchers, and learners from under-resourced settings develop bioinformatics and genomic data skills for personalized healthcare and research. Wen and Wang, 51 Multi-center retrospective cohort study with machine learning model development and validation 1110 hepatocellular carcinoma (HCC) patients across six cohorts (TCGA-LIHC, CHCC, GSE14520 , GSE116174 , GSE144269 , LIRI-JP) To develop and validate a consensus AI-derived prognostic signature (CAIPS) for predicting prognosis and guiding therapy in HCC Multi-omics data including mRNA expression, SNVs, CNVs, and pathway-level transcriptomic data Combination of 10 machine learning algorithms (101 models) including Stepwise Cox regression and Gradient Boosting Machine (GBM), Random Survival Forest (RSF), LASSO, CoxBoost High predictive accuracy (C-index highest among models), improved survival prediction (OS, DSS, PFI, DFI), identification of genomic instability, immune differences, and therapy response patterns; drug candidates identified (Irinotecan, BI-2536) Provides a robust biomarker for risk stratification, supports personalized therapy selection, predicts response to immunotherapy, TACE, and targeted therapy in HCC Yang et al, 52 Retrospective bioinformatics and machine learning study Glioma patients from CGGA 325 and 693 datasets, TCGA, and GEO GSE18494 ; includes low-grade gliomas (LGGs) and high-grade gliomas (HGGs). To identify clinically significant differentially expressed genes in gliomas; establish molecular subtype classification; develop and validate a prognostic scoring system; and screen key diagnostic/prognostic genes for precise glioma classification and individualized treatment. RNA-seq transcriptomic gene expression data, including log2(TPM+1)-normalized expression profiles from CGGA, TCGA, and GEO cohorts. Differential expression analysis using limma; non-negative matrix factorization (NMF); weighted gene coexpression network analysis (WGCNA); protein-protein interaction (PPI) analysis; single-sample GSEA; and machine learning models including GLM, random forest (RF), SVM-RFE, and XGBoost (XGB). Identified 117 DEGs; classified gliomas into two molecular subgroups (C1 and C2); developed the MGP_Score prognostic model; validated the model in TCGA and GSE18494 ; identified B2M, SRPX2, and SERPINH1 as diagnostic/prognostic biomarkers; high MGP_Score and elevated biomarker expression were associated with poorer survival. Supports precision glioma classification, prognosis evaluation, biomarker-driven risk stratification, early diagnosis, individualized treatment planning, and potential therapeutic target discovery. Khan et al, 53 Methodological computational study using multiple genomic datasets Multiple public datasets of mental disorders; includes SNPs, eQTLs, CNVs, and patient genomic data (eg, schizophrenia, ASD datasets) To develop iMEGES, a deep learning-based tool for prioritizing susceptibility genes and variants for mental disorders in personal genomes Whole genome sequencing (WGS), whole exome sequencing (WES), SNPs, CNVs, eQTLs, epigenomic data (PsychENCODE, Roadmap Epigenomics), GTEx data Deep neural networks (TensorFlow/Keras), multilayer neural network with dropout, sigmoid output; integration of multiple genomic scoring systems (CADD, DANN, GWAVA, FATHMM, EIGEN) Improved variant prioritization performance (AUC up to ~91%); ncDeepBrain AUC ~80%, outperforming individual predictors (~52–59%); gene prioritization AUC ~57–58% Supports identification of disease-relevant variants and genes in mental disorders; enables personalized genomic analysis and may guide targeted therapeutic strategies Brunk et al, 54 Computational systems biology resource study Human metabolic genes (3,288 genes), ~10,000 SNPs, 889 cancer mutations from 178 tumor samples To develop Recon3D integrating genomic, proteomic, metabolomic, and structural data for disease understanding SNP data (dbSNP), cancer genomic data (TCGA), UniProt variation data, PharmGKB Machine learning (genetic algorithm, ROC-AUC optimization), 3D hotspot analysis, network modeling Mutation hotspot identification; AUC up to ~0.80 for drug prediction; clustering of deleterious mutations Supports precision medicine, biomarker discovery, drug repurposing, and improved disease mechanism understanding Singha et al, 55 Experimental study with computational AI modeling and in vitro validation Cancer cell lines from LINCS/CCLE/CMap resources, including 359 cell lines in the LINCS-3549 training dataset; live-cell validation was performed on Panc 04.03 pancreatic cancer, DU145 prostate cancer, and HCC70 breast cancer cell lines. To develop and validate CancerOmicsNet, an artificial intelligence-based system for predicting the therapeutic effects of multitargeted kinase inhibitors across cancer types and guiding precision anticancer therapy. Differential gene expression profiles from CCLE, drug-perturbed gene expression profiles from CMap, gene-disease associations, protein-protein interaction networks, pharmacogenomics data, and kinase inhibitor profiling data. CancerOmicsNet using graph neural networks (GNNs), deep graph learning, attention propagation mechanisms, cosine similarity-based node embeddings, graph reduction, Set2Set pooling, and multimodal biomedical data integration. CancerOmicsNet outperformed the gene signature-based method on the LINCS-87 dataset (AUC-ROC 0.761 vs 0.475; accuracy 0.714 vs 0.437). Selected predictions were supported by literature and live-cell time-course assays; several kinase inhibitors, including JNJ-7706621 and PP1, showed dose-dependent antiproliferative effects. Supports AI-guided precision oncology by improving prediction of anticancer drug response, identifying effective kinase inhibitor therapies, enabling drug repurposing, and integrating multi-omics and network-based data for personalized cancer treatment. Barki et al, 56 Methodological study with retrospective dataset evaluation More than 150 cancer cases; ~82 detailed cases; diverse cancer types; includes demographics (age, gender, weight) Develop an ontology-based AI model to predict radiation therapy adverse effects and support decision-making Not reported Ontology-based modeling; semantic modeling; rule-based reasoning; ML (DT, NB, KNN, SVM, ANN, DL); OWL-DL; SWRL/SQWRL 80.1% prediction compatibility; 92% accuracy (potential 97%); 20% reduction in adverse events; 25% reduction in radiation exposure Improves radiation safety; enables personalized treatment; enhances clinical decision-making; reduces risks Yang et al, 57 Development and implementation study (applied experimental study) Colorectal cancer datasets (TCGA COAD); 366 patients in case studies; includes clinical, genomic, and simulated SDOH variables To develop and evaluate a conversational AI system (AI-HOPE-PM) for integrating genomic, clinical, and SDOH data for multidimensional cancer analysis Genomic mutation data (eg, TP53, APC, KRAS), multiomics cancer datasets from TCGA, AACR GENIE, and cBioPortal Large Language Models (LLaMA 3), Natural Language Processing (NLP), Retrieval-Augmented Generation (RAG), Python-based statistical modeling workflows 92.5% accuracy in natural language query interpretation; identification of survival disparities (P =0.03–0.05); odds ratio analyses; faster execution (28.02 seconds vs 46–58 seconds); detection of associations between genomic mutations, SDOH factors, and treatment outcomes. Enables real-time integrative precision oncology analysis; reduces technical barriers; supports equity-focused research by incorporating SDOH; improves hypothesis generation and clinical decision support in cancer research Abualigah et al, 58 Experimental machine learning study UK Biobank: ~500,000 participants with genetic, clinical, and lifestyle data; MIMIC-IV: critically ill patients (demographics, vitals, labs, diagnoses) Develop an AI-driven framework to predict disease outcomes and optimize patient-centric care Genetic data (UK Biobank), clinical data, lifestyle data; multi-modal biomedical data Gradient Boosting Machines (GBM); Deep Neural Networks (DNN); hybrid AI framework with feature selection AUROC up to 0.96; Accuracy up to 92.4%; Precision up to 91.8%; Recall up to 92.9%; improved predictive performance over traditional ML models Supports real-time clinical decision support; enables personalized medicine; improves disease prediction, risk stratification, and treatment optimization Mak et al, 59 Observational cohort study with multi-omics analysis 56 patients with MN1 truncating variants (43 CTT, 13 NTT); mixed sex; neurodevelopmental disorders; international multi-center cohort To differentiate phenotypic subgroups (CTT vs NTT) within MN1-related syndromes using AI-driven genotype–epigenotype–phenotype integration DNA methylation (DNAm) data; RNA sequencing (RNA-seq) from fibroblasts; facial image data GestaltMatcher; Support Vector Machine (SVM); statistical clustering methods Distinct DNAm episignature for CTT; clear clustering between CTT and NTT; high classification performance; no significant RNA expression difference Improves syndrome delineation, supports early diagnosis, enables objective classification, advances precision medicine Silva et al, 60 Proof-of-concept study with virtual patient case Virtual patient case; includes 60-year-old female with chronic diseases and polypharmacy in primary care Develop and validate an integrated informatic and bioanalytic platform combining pharmacogenomics and AI for medication management ~90 SNPs; CYP2D6 copy number variation; pharmacogenomic panel data Clinical Semantic Network; predictive analytics on EHR; integration of pharmacokinetic and pharmacogenomic data Identification of drug–gene and drug–drug interactions; detection of CYP pathway overload; improved symptom interpretation Supports clinical decision-making; enables personalized medication management; reduces adverse drug reactions De La Vega et al, 61 Retrospective benchmarking and validation study of an AI-based electronic clinical decision support tool 119 pediatric benchmark cases, mostly critically ill NICU infants with diagnosed rare genetic diseases; 60 validation cases from five academic medical centers, mostly non-NICU rare disease patients; included singleton, duo, and trio cases. To assess the diagnostic performance of Fabric GEM for AI-enabled genome interpretation and candidate diagnosis nomination in rare genetic diseases, and to compare it with existing variant prioritization tools. Whole-genome sequencing (WGS) and whole-exome sequencing (WES) data, including SNVs, small insertions/deletions, copy number variants, and structural variants. Fabric GEM AI-based clinical decision support using Bayesian scoring/Bayes factors, variant prioritization, phenotype-driven gene and condition matching, clinical natural language processing-derived HPO terms, and ab initio structural variant inference. GEM ranked more than 90% of causal genes as the top or second candidate and prioritized a median of 3 candidate genes per case. In structural variant cases, it identified the causal SV as the top candidate in 17/20 cases and within the top five in 19/20 cases. AI-enabled genome interpretation can automate and shorten candidate review for rare genetic disease diagnosis, support WGS/WES interpretation across variant types, reduce manual workload, and potentially decrease cost and speed clinical reporting. Morin et al, 62 Observational cohort study with AI framework development ~175,000 cancer patients (UCSF, 2010–2019); subsets: 3,782 breast, 2,054 lung; multimodal EHR data Develop and validate MEDomics AI framework integrating longitudinal EHR for cancer prognostication Not reported Random forest, SVM, LASSO, GBM, CART, penalized Cox regression, NLP (TF-IDF) AUC 0.75–0.85; improved F1 (eg, 0.80 vs 0.67); identification of prognostic factors (Framingham risk score) Supports AI-driven personalized oncology care and continuous prognosis updating Sanjaya et al, 63 Empirical deep learning/model-development and validation study PCAWG whole-genome sequencing dataset: 2,587 tumours across 24 tumour types; TCGA whole-exome sequencing dataset: 7,352 tumours across 20 tumour types; additional independent whole-genome validation cohorts To develop and validate Mutation-Attention (MuAt), a deep learning model for tumour typing and molecular subtyping from somatic mutation profiles Whole-genome and whole-exome somatic mutation data, including SNVs, MNVs, indels, structural variants, and mobile element insertions Mutation-Attention (MuAt) deep neural network with attention mechanism; comparison with DNN and random forest models MuAt achieved 89% accuracy for whole-genome tumour typing and 64% accuracy for whole-exome tumour typing, with strong top-5 accuracy and validation in independent cohorts Supports AI-assisted tumour classification and molecular subtyping, with potential relevance for precision cancer medicine Gimeno et al, 64 Original research; computational/explainable AI precision medicine study with external in-silico validation BeatAML cohort: 319 acute myeloid leukemia (AML) ex-vivo tumor samples/patients with drug screening and whole-exome sequencing; after filtering, 111 drugs and 64 candidate biomarkers were analyzed; validation used AML cell-line datasets from CERES (n=14), DEMETER2 (n=18), GDSC (n=23), and additional DEMETER RNAi (n=15). To develop and validate an explainable AI method, multi-dimensional module optimization (MOM), that links drug sensitivity screening with genetic biomarkers to generate interpretable AML treatment stratification rules. Whole-exome sequencing (WES) genetic variants from AML samples, including SNVs, indels, fusion genes/genomic aberrations; RNA-seq data were also used for functional subgroup analysis. Multi-dimensional Module Optimization (MOM); IC50* drug-sensitivity normalization; Independent Hypothesis Weighting (IHW) for drug-biomarker associations; Mixed-Integer Linear Programming (MILP) to generate an interpretable decision-tree treatment guideline; k-nearest-neighbor imputation; Wilcoxon testing and functional enrichment analysis. MOM identified a four-subgroup AML treatment strategy using FLT3 mutation, inv(16)/CBFb-MYH11, and NRAS mutation status: Quizartinib for FLT3Mut patients (n=103), Trametinib for FLT3WT and inv(16) patients (n=15), Selumetinib for FLT3WT/no inv(16)/NRASMut patients (n=42), and Crizotinib for remaining FLT3WT/no inv(16)/NRASWT patients (n=159). The strategy was robust in 10-fold cross-validation and externally supported in CERES, DEMETER2, and GDSC datasets, with significant sensitivity differences reported for CERES, DEMETER2, and GDSC. The study suggests that interpretable XAI can support precision oncology by assigning AML treatments through a simple biomarker-based decision tree, improving transparency for clinicians and regulators, and identifying potential drug repositioning opportunities beyond currently approved AML therapies. Ahmadi et al, 65 Original research; methodological development and validation study 61,697 patients from two German university hospitals across endocrinology, gastroenterology, pneumonology, and hematology domains, including an acute myeloid leukemia validation dataset. To develop and customize an OMOP-based rare diseases common data model (RD-CDM) integrating clinical, phenotypic, and genotypic data for multi-center rare disease research and AI-enabled applications. Genotype and phenotype data, including gene mutation features, HPO-coded symptoms, genomic variant-related elements, and linked clinical-genomic rare disease datasets. Data harmonization using OMOP CDM, FHIR/MII-CDS mapping, FHIR-to-OMOP ETL, Genotype/Phenotype-to-OMOP ETL, OMOP Genomic terminology, and preparation for downstream AI/machine-learning tools such as PLP and HADES. Developed an OMOP-based RD-CDM; integrated 61,697 patient records; tested hematology validation data; achieved 100% mapping success for gene mutation elements to OMOP concepts, although some mappings required many-to-one concept mapping. Supports large-scale multi-center rare disease studies, genotype-phenotype correlation analysis, AI-based decision support, improved diagnostic workflows, and personalized medicine research. Hindhede and Andersen. 66 Qualitative empirical study using document analysis, semi-structured qualitative interviews, in situ interviews, classroom observations, and thematic analysis; grounded in the sociology of expertise. Educators, course organizer, and master’s students/course participants in a Danish personalized medicine program across five universities; participants were experienced healthcare, science, public-sector, and medical-industry professionals, including physicians, pharmacists, chemists, molecular biologists, data analysts, bioanalysts, senior consultants, and specialist consultants; many course participants were oncology physicians. The study included 24 qualitative interviews, including nine longer interviews and 15 in situ interviews, plus 28 hours of observation across two modules. To examine how educators and master’s students in personalized medicine perceive and engage with the knowledge needed to integrate artificial intelligence into patient treatment and care, especially regarding algorithmic decision support, clinical judgment, expertise, opportunities, challenges, and accountability in clinical decision-making. No primary genomic dataset was analyzed. The study context was personalized/precision medicine in Denmark, including discussion of complex health data, big data, patient data, and whole-genome sequencing as part of the Danish National Genome Center and personalized medicine education. No AI model was developed or validated in the study. AI was examined sociologically as clinical decision support and algorithmic guidance, including predictive algorithms, automated screening of blood-test results, machine learning/deep-learning or black-box algorithms, data-driven algorithms for diagnosis, prognosis, treatment selection, triage, and risk prediction. Participants and educators generally viewed AI and big data as potentially transformative for personalized treatment and diagnostics. Key themes were: understanding expertise in the context of AI; the art of medicine and trust in AI; ethical implications and accountability; interprofessional collaboration; and balancing data use with human care. Participants emphasized AI as a supportive tool rather than a replacement for clinical judgment, while expressing concerns about black-box opacity, bias, representativeness, accountability, over-reliance, erosion of clinical skills, and loss of patient-centered care. AI integration in personalized medicine should preserve clinician autonomy and clinical judgment while building interactional expertise so professionals can interpret, challenge, and critically evaluate algorithmic outputs. Implementation should include transparency, explainability, bias assessment, accountability frameworks, patient-data governance, interdisciplinary collaboration, and continued emphasis on empathy and practical clinical experience. Yagin et al, 67 Observational cohort study with machine learning model development 234 patients (93 COVID-19 positive, 141 negative); adults ≥18 years; UCSF dataset Develop an explainable AI model for identifying COVID-19 gene biomarkers and predicting disease status Metagenomic next-generation sequencing (mNGS) gene expression data (15,979 genes) LASSO; SVM-SMOTE; Logistic Regression; SVM; Random Forest; XGBoost; LIME; SHAP XGBoost accuracy 0.93; F1-score 0.933; key biomarkers IFI27, LGR6, FAM83A Improves interpretable diagnosis; supports personalized risk assessment and clinical decision-making Alrefaei et al, 68 Cross-sectional survey study 804 Saudi adults (≥18 years); 38.4% female, 61.6% male; diverse age and education levels To assess awareness, knowledge, and attitudes toward the Saudi Human Genome Program (SHGP), genetic data sharing/privacy, and the role of AI in genomic data analysis Population-level genomic data from the Saudi Human Genome Program (NGS-based genomic database; not directly analyzed, but referenced conceptually) General AI approaches including supervised, unsupervised, and deep learning algorithms for genomic analysis (conceptual discussion, not empirically applied) Moderate awareness of SHGP; low knowledge of genetic data privacy; high positive attitudes toward AI use (92.8%) and data sharing with privacy safeguards Need for public education, improved data governance policies, and integration of AI for efficient genomic analysis and personalized medicine advancement Briguglio et al, 69 Empirical computational method-development and validation study using genomic cohort data 10,336 patients with tumor tissue samples from the MSK-IMPACT clinical sequencing dataset; >100,000 somatic mutations; includes diverse cancer types To develop a privacy-preserving machine learning with encryption (MLE) framework for cancer prediction using genomic data while maintaining data confidentiality Somatic mutation data including copy number alterations, gene fusions, mutation signatures, and clinical genomic sequencing data from MSK-IMPACT Machine learning models (logistic regression, SVM, random forest) combined with fully homomorphic encryption (FHE) using Microsoft SEAL library Prediction accuracy up to 77.47% (logistic regression); improved performance over prior studies (~75%); secure computation on encrypted genomic data; feasible privacy-preserving analytics with acceptable computational cost Enables secure and privacy-preserving precision oncology analytics; supports clinical decision-making without exposing sensitive genomic data; facilitates multi-institutional data sharing and AI deployment in healthcare Qu et al, 70 Empirical original research; multi-center genomic database development and validation study 6,031 samples from Chinese populations across six laboratories; 66,485 BRCA variant submissions; 156 registered users; 15 expert curators involved in variant interpretation To develop and validate BRCA-CN, a consortium blockchain-based framework for standardized BRCA1/2 variant interpretation, secure data sharing, and regulatory-grade genomic database development for Chinese populations NGS-derived BRCA1/BRCA2 variant data, including SNVs and indels, with clinical and population annotations AI-based pathogenicity prediction using PrimateAI, REVEL, and EVE; combined ensemble validation requiring agreement across at least two algorithms 66,485 submissions reduced to 950 unique variants; 462 variants achieved definitive consensus; 83.6% AI concordance with expert interpretation; combined AI AUC=0.912; VUS rate reduced from 52.3% to 41.5%; 16 novel pathogenic variants identified Improves BRCA variant interpretation consistency, supports Chinese population-specific precision medicine, reduces uncertain classifications, strengthens genetic counseling and clinical decision-making, and enables secure multi-institutional genomic data sharing Elvas et al, 71 Original research (framework development and real-world validation using clinical data) 512,764 patients; multi-source hospital data including cardiology, ICU, and respiratory units; real-time physiological signals (eg, heart rate, SpO2, temperature) To develop a secure Information Sharing Infrastructure (ISI) integrating AI and homomorphic encryption for privacy-preserving health data sharing and predictive analytics Electronic Health Records (EHR); real-time physiological data; structured clinical datasets (non-genomic) Logistic Regression; Random Forest; Machine Learning; Statistical correlation analysis; AI-based predictive modeling under homomorphic encryption High predictive accuracy: Logistic Regression (96.88%), Random Forest (97.62%); successful encrypted data processing; real-time prediction of abnormal clinical events Enables privacy-preserving AI-driven clinical decision support; improves early detection of abnormal patient conditions; supports secure multi-institutional data sharing; enhances patient safety and healthcare outcomes Harrison et al, 72 Qualitative study using semi-structured focus groups and inductive content analysis 34 Australian adults aged 22–74 years; recruited via social media; participants included men and women from metropolitan, regional, and rural areas; seven focus groups with 4–6 participants each To explore public perceptions of the use of artificial intelligence in genomic medicine, with specific attention to automated reanalysis of genomic data from undiagnosed patients Next-generation sequencing-derived genomic data and clinical genomic datasets used for genomic analysis and reanalysis Artificial intelligence and machine learning-based genomic data analysis/reanalysis; no specific AI model or algorithm was empirically tested in participants Participants were generally comfortable with AI analysing genomic data; perceived benefits included faster analysis, greater efficiency, reduced workload, improved diagnostic potential, and possible reduction of human bias; concerns included data security, misdiagnosis, algorithmic bias, discrimination, and unclear liability Findings support the need for transparent governance, human oversight, audit/checking mechanisms, data security safeguards, bias mitigation, and public education before broader clinical implementation of AI in genomic medicine Nwebonyi and McKay. 73 Exploratory qualitative study (semi-structured interviews) 11 expert stakeholders (bioethics, AI, targeted medicine); UK-based researchers and industry specialists Explore bias risks in AI-driven targeted medicines manufacturing Not directly genomic dataset; discusses targeted medicine and gene/cell therapy-related data General AI algorithms for data processing, predictive modeling, and manufacturing optimization (no specific model) Bias identified in upstream and downstream processes; demographic, geographic, and economic biases; concept of ‘corrective bias’ Need for diverse datasets, transparency, patient involvement, and ethical AI design to reduce health inequalities Hampel et al, 74 Phase 2a clinical trial with AI-based secondary genomic analysis 32 patients with mild-to-moderate Alzheimer’s disease aged 50–85 years; genomic data available for 21 patients; multicenter study in Australia To identify and confirm genomic biomarkers associated with therapeutic response to blarcamesine (ANAVEX2-73) using an AI-driven precision medicine framework Whole-exome sequencing (WES); RNA sequencing (RNA-seq); genomic variants; transcriptomic expression profiles Unsupervised Formal Concept Analysis (FCA); Knowledge Extraction and Management (KEM) platform; association rule mining; linear mixed-effect models SIGMAR1 p.Gln2Pro and COMT p.Leu146fs variants were associated with treatment response; high blarcamesine concentration was linked to improved MMSE and ADCS-ADL outcomes; effect sizes increased up to Cohen’s d = 1.05 after biomarker-based subgrouping Supports AI-enabled biomarker discovery and patient stratification for Alzheimer’s disease therapy; informs precision medicine approaches and future confirmatory Phase 2b/3 trials Bahado-Singh et al, 75 Original research (observational epigenomic + AI prediction study) 24 neonates with coarctation of the aorta (CoA) and 16 controls; newborn blood samples To identify epigenetic changes and evaluate AI-based prediction of CoA using DNA methylation Genome-wide DNA methylation (Illumina HumanMethylation450 BeadChip; CpG loci) Deep learning, Random Forest, Support Vector Machine, Linear Discriminant Analysis, Prediction Analysis for Microarrays, Generalized Linear Model Deep learning AUC 0.97, sensitivity 95%, specificity 98%; identification of 65 differentially methylated CpG sites across 75 genes Potential for early non-invasive newborn screening and improved prediction of congenital heart defects using AI and epigenomics DeGroat et al, 76 Empirical observational transcriptomic study with machine learning model development and validation 61 CVD patients (40 male, 21 female; age 45–92 years) and 10 healthy controls (5 male, 5 female; age 28–78 years); participants included White, Black/African American, Asian, and unknown race groups To identify significant transcriptomic biomarkers and develop a predictive AI/ML model for accurate cardiovascular disease (CVD) classification using integrated clinical, demographic, and transcriptomic data Whole-transcriptome RNA-seq gene expression data integrated into a Clinically Integrated Genomics and Transcriptomics (CIGT) dataset Recursive Feature Elimination (RFE), Pearson correlation, Chi-square test, ANOVA, Random Forest, Support Vector Machine, XGBoost, k-Nearest Neighbors, and Soft Voting Classifier Identified 18 significant transcriptomic biomarkers; CVD prediction accuracy up to 96%; RF accuracy 95%, SVM 91%, XGBoost 91%, k-NN 91%; ROC-AUC up to 0.99; weighted F1-scores approximately 0.91–0.96 Supports early CVD detection using biomarker profiles; provides a precision-medicine framework for personalized risk prediction and clinical decision support Costa et al, 77 Comparative data analysis study using genomics databases Genomic data from six databases (ClinVar, Ensembl, GWAS Catalog, LOVD, CIViC, and CardioDB); includes oncology- and cardiology-disorder-associated DNA variations. To evaluate the consequences of data dispersion across genomic data sources in precision medicine by assessing completeness and concordance. DNA variation data/genomic variant information from multiple public repositories, including clinical interpretation and GWAS/statistical association data. ETL workflow, data integration, conceptual schema mapping, and statistical/data-quality metric analysis; no predictive machine-learning model was applied. ClinVar showed the highest completeness; no single data source contained all variants; high overall concordance was reported, but discordant interpretations existed; exclusive clinically relevant variants were found across databases. Reliance on a single genomic data source may lead to incomplete or inaccurate genomics-based diagnosis; integration of multiple sources is important for precision medicine. Natarajan et al, 78 Original research; machine learning-based bioinformatics framework study Secondary gene expression datasets related to Parkinson’s disease patients and healthy/control samples; no newly recruited participant cohort explicitly reported To propose and evaluate a Random Forest-based bioinformatics analysis in gene expression profiling (BA-GEP) framework for identifying Parkinson’s disease-related gene expression patterns and supporting personalized healthcare recommendations Gene expression profiling data from microarray/RNA-seq sources; differentially expressed mRNA, lncRNA, and miRNA; whole-genome/protein expression sequence information Random Forest classifier; machine learning-based gene expression classification; Discrete Cluster Analysis; k-means-based feature selection; PCA and bioinformatics network/pathway analysis Identification of differentially expressed genes and gene clusters; PD vs control classification; 524 DE genes reported at FDR 0.2; feature selection reduced genes to approximately 85 and then approximately 21 genes; identification of candidate PD-related genes/biomarkers for personalized healthcare May support biomarker discovery, earlier Parkinson’s disease detection, individualized treatment planning, personalized medication recommendations, and improved understanding of genetic variation in treatment response Isong et al, 79 Empirical computational method-development and validation study using federated genomic cohorts Three cohorts: CINECA synthetic cohort (n=483), 1000 Genomes Phase 3 (n=2,504; five continental populations), and msprime simulated cohort (n=50,000; Out-of-Africa demography) To develop and evaluate XS-FedPRS, an explainable and secure federated learning framework for privacy-preserving polygenic risk prediction across heterogeneous healthcare networks SNP-based genomic data; polygenic risk scores for common variants; allele dosage vectors for rare variants; allele frequency and Hardy-Weinberg equilibrium summary statistics; GWAS-derived effect sizes Federated learning; hierarchical two-pathway neural network; attention-based rare variant modeling; Isolation Forest for gradient plausibility analysis; genetic-aware byzantine defense; blockchain-based Proof-of-Authority-and-Association trust management; explainable AI modules; differential privacy AUC up to 0.844 on the msprime cohort; 0.731 accuracy on 1000 Genomes; mean attack detection F1 around 87.0% with 1.9% false positive rate; near-perfect detection for several byzantine attacks; membership inference attack advantage about 0.026; communication overhead about 3.1–3.3% Supports trustworthy, privacy-preserving, interpretable federated genomic analytics for precision medicine; enables multi-institutional polygenic risk prediction without centralizing patient genomic data; improves robustness against malicious updates and supports clinical decision interpretation DeGroat et al, 80 Empirical original research; multimodal multi-omics machine learning study 71 participants: 61 cardiovascular disease (CVD) patients (40 male, 21 female; aged 45–92; diverse ethnicity/race) and 10 healthy controls (5 male, 5 female; aged 28–78) with no clinical manifestation of CVD. To discover novel biomarkers and predict cardiovascular disease by integrating transcriptomic expression, whole-genome SNP data, and clinical demographic information into patient-specific multi-omics profiles. RNA-seq transcriptomic expression data, whole-genome sequencing (WGS)-derived SNPs, CIGT-formatted integrated genomic/transcriptomic profiles, and clinical demographic data. k-nearest neighbors (k-NN) imputation; DESeq2 differential expression analysis; minimum redundancy–maximum relevance (MRMR) feature selection; CADD and gnomAD-based pathogenic variant filtering; Random Forest, Logistic Regression, and XGBoost classifiers; Bayesian hyperparameter optimization; SHAP explainability. Identified 28 differentially expressed genes, 10 MRMR-selected biomarkers, and 17 rare deleterious SNPs; RF and XGBoost achieved 100% accuracy on the small test set, while Logistic Regression achieved 93% accuracy; XGBoost had the best overall performance considering Brier score; RPL36AP37 and HBA1 were the most important predictors by SHAP. The integrated multi-omics AI/ML framework may support early CVD risk prediction, biomarker discovery, patient-specific risk profiling, interpretable clinical decision support, and personalized cardiovascular medicine. Nelson et al, 81 Empirical computational method-development and validation study using de-identified EHR data and biomedical knowledge-network integration De-identified structured EHR data from 816,504 UCSF patients; 292,753 patients with at least one of 137 complex diseases represented in SPOKE; BMI cohort example included 100,187 obese patients To connect EHR concepts to the SPOKE biomedical knowledge network and create biologically meaningful machine-readable embeddings to support precision medicine Biomedical genomic and knowledge-network data integrated from SPOKE, including GWAS Catalog, STARGEO gene expression data, ChEMBL, LINCS, Gene Ontology, Disease-Gene associations, gene expression relationships, and other biomedical databases; EHR diagnosis, medication, and laboratory concepts mapped to SPOKE nodes Unsupervised machine learning using a modified PageRank/random-walk algorithm to generate Propagated SPOKE Entry Vectors (PSEVs); network embedding and benchmarking through deleted-edge recovery and random/shuffled network comparisons Generated 3,233 PSEVs from EHR-linked SPOKE entry points; obese BMI PSEV enriched for known BMI-associated genes (p = 2.19E−10); preserved known SPOKE relationships; recovered deleted Disease–Gene, Disease–Disease, Compound–Gene, Compound–Compound, and SideEffect–Anatomy relationships; inferred biologically meaningful and traceable disease–gene relationships Provides a clear-box, biologically informed embedding approach for precision medicine; may support biomarker discovery, patient characterization, explainable clinical AI, and drug repurposing by linking clinical EHR features with genomic and biomedical mechanisms Figure 2 PRISMA flow diagram. The diagram summarizes the process used to identify, screen, assess, and include studies in this narrative review. A PRISMA flow diagram showing study identification, screening and inclusion process. The PRISMA flow diagram outlines the process of identifying, screening and including studies via databases. In the identification phase, 824 records were identified from databases, with none from manual citation searching. Before screening, 257 duplicate records, 24 marked as ineligible by automation tools and 11 removed for other reasons were excluded. This left 532 records for screening. During screening, 229 did not meet inclusion criteria, 149 were non-research articles, 31 had irrelevant interventions or populations and 7 were in non-English languages, leaving 116 records sought for retrieval. Four records were not retrieved due to incorrect links. In the eligibility assessment, 112 records were evaluated, with 27 excluded for methodological flaws, 22 for insufficient or unclear data, 6 for inadequate intervention description and 3 for ethical concerns. Finally, 54 studies were included in the review. Figure 3 Distribution of genomic data types used in included studies (n = 54). Bars show the number and percentage of studies utilizing each type of genomic data, based on the “Type of genomic data used” column in Table 2 . Categories are not mutually exclusive, as many studies incorporated more than one type of genomic data; therefore, percentages sum to more than 100%. Clinical/phenotypic/EHR data (22/54, 40.7%) and transcriptomic/RNA-seq data (19/54, 35.2%) were the most commonly used data types, followed closely by other genetic/molecular data such as eQTL and pathway-level information (19/54, 35.2%) and genome-wide SNP/genotype data from GWAS or SNP arrays (18/54, 33.3%). Less commonly used data types included copy number variation data (5/54, 9.3%) and metagenomic/microbiome data (2/54, 3.7%). A horizontal bar graph showing distribution of genomic data types used in included studies. A horizontal bar graph shows the percentage of studies using different types of genomic data. The X-axis represents the percentage (0-50%) and the Y-axis lists 12 data categories. The data categories and their respective percentages are: Clinical/Phenotypic/EHR Data (40.7%), Transcriptomic/RNA-seq Data (35.2%), Other Genetic/Molecular Data (eQTL, Pathways, etc.) (35.2%), Genome-wide SNP/Genotype Data (GWAS, SNP arrays) (33.3%), Somatic Mutation/Variant Data (SNVs, Indels) (22.2%), Whole Exome Sequencing (WES) Data (16.7%), Whole Genome Sequencing (WGS) Data (13.0%), Multi-omics Data (Integrative genomics) (13.0%), Epigenomic Data (Methylation, meQTL) (11.1%), Protein/Proteomics Data (11.1%), Copy Number Variation (CNV) Data (9.3%) and Metagenomic/Microbiome Data (3.7%). Overview of Included Studies PRISMA flow diagram. The diagram summarizes the process used to identify, screen, assess, and include studies in this narrative review. Distribution of genomic data types used in included studies (n = 54). Bars show the number and percentage of studies utilizing each type of genomic data, based on the “Type of genomic data used” column in Table 2 . Categories are not mutually exclusive, as many studies incorporated more than one type of genomic data; therefore, percentages sum to more than 100%. Clinical/phenotypic/EHR data (22/54, 40.7%) and transcriptomic/RNA-seq data (19/54, 35.2%) were the most commonly used data types, followed closely by other genetic/molecular data such as eQTL and pathway-level information (19/54, 35.2%) and genome-wide SNP/genotype data from GWAS or SNP arrays (18/54, 33.3%). Less commonly used data types included copy number variation data (5/54, 9.3%) and metagenomic/microbiome data (2/54, 3.7%). Artificial intelligence techniques applied across the studies were highly varied ( Figure 4 ), with machine learning and deep learning approaches being the most dominant. Commonly used methods included neural networks, random forests, support vector machines, and gradient boosting algorithms, while more advanced techniques such as multimodal deep learning architectures, graph neural networks, Bayesian modeling, and knowledge graph–based reasoning systems were increasingly reported. In addition, emerging approaches, including explainable artificial intelligence and large language models, were used to enhance interpretability and support clinical translation. Reported outcomes across studies demonstrated promising but uneven performance, with many models achieving high predictive accuracy and robust discrimination metrics. These approaches were effective in identifying disease-associated variants, predicting clinical outcomes, stratifying patient risk, and discovering novel biomarkers, while also improving diagnostic efficiency and reducing analytical complexity. However, several studies relied on relatively small, retrospective, or demographically imbalanced datasets, increasing the risk of overfitting and limiting generalizability. Figure 4 Distribution of artificial intelligence techniques in included studies (n = 54). Bars show the number and percentage of studies applying each AI technique, based on the “Artificial intelligence techniques applied” column in Table 2 . Categories are not mutually exclusive, as many studies applied more than one technique; therefore, percentages sum to more than 100%. Machine learning approaches, encompassing traditional statistical and algorithmic methods such as logistic regression, gradient boosting, and clustering, were the most frequently applied (32/54, 59.3%), followed by deep learning (13/54, 24.1%) and random forest (10/54, 18.5%). Less commonly applied techniques included convolutional neural networks and transfer learning (1/54, 1.9% each). A horizontal bar graph showing distribution of artificial intelligence techniques across studies. A horizontal bar graph illustrates AI techniques by study percentage. X-axis: Percentage of Studies (%), ranging from 0 to 60. Y-axis: AI Techniques. Bars and labels: Machine Learning (ML) leads with 59.3%, followed by Deep Learning (DL) at 24.1%. Random Forest (RF) accounts for 18.5%, Support Vector Machine (SVM) at 14.8%, Natural Language Processing (NLP) and Explainable AI (XAI) both at 11.1%. Graph Neural Networks (GNN) represent 7.4%, Ensemble Learning at 5.6%, Generative Models (GAN, VAE) at 3.7% and both Convolutional Neural Networks (CNN) and Transfer Learning at 1.9%. Distribution of artificial intelligence techniques in included studies (n = 54). Bars show the number and percentage of studies applying each AI technique, based on the “Artificial intelligence techniques applied” column in Table 2 . Categories are not mutually exclusive, as many studies applied more than one technique; therefore, percentages sum to more than 100%. Machine learning approaches, encompassing traditional statistical and algorithmic methods such as logistic regression, gradient boosting, and clustering, were the most frequently applied (32/54, 59.3%), followed by deep learning (13/54, 24.1%) and random forest (10/54, 18.5%). Less commonly applied techniques included convolutional neural networks and transfer learning (1/54, 1.9% each). From a clinical and translational perspective, the findings highlighted the significant potential of artificial intelligence to advance precision medicine. AI-driven models supported personalized treatment selection, pharmacogenomic decision-making, and real-time clinical decision support across various disease domains, including cancer, neurological disorders, and rare genetic conditions. Several studies also demonstrated that incorporating diverse population data improved model generalizability and reduced bias, contributing to more equitable healthcare outcomes. However, challenges related to data heterogeneity, lack of standardization, interoperability issues, and ethical concerns regarding data privacy and governance were frequently reported. Overall, the included studies indicate that while artificial intelligence is transforming genomic research and enabling more precise and individualized care, the current evidence remains constrained by limited prospective clinical validation, inconsistent external testing, and reproducibility concerns across published machine learning genomic models. Artificial intelligence has emerged as a potentially important force in the integration of large-scale genomic datasets, particularly those derived from global genomic resources. 28 Across the reviewed literature, AI-driven approaches such as machine learning, deep learning, and probabilistic modeling were consistently applied to manage and interpret highly complex genomic information. These techniques demonstrated the ability to process multi-layered datasets, including whole genome sequencing, transcriptomics, and epigenomic profiles, thereby enabling the identification of subtle patterns that would be difficult to detect through conventional statistical methods. 29 Studies highlighted that AI algorithms could efficiently harmonize heterogeneous datasets collected from diverse populations, overcoming challenges related to data fragmentation and variability. 30 , 31 This integration has facilitated a more unified understanding of genetic variation at a global scale, supporting the development of comprehensive genomic resources that are both scalable and adaptable. 32 In addition to technical integration, artificial intelligence contributed significantly to enhancing the interpretability of genomic data. 33 Several studies reported the use of explainable AI frameworks to uncover biologically meaningful associations between genetic variants and disease phenotypes. 34 , 35 These models were particularly effective in identifying rare variants and population-specific mutations that are often underrepresented in traditional datasets. 36 By enabling dynamic data integration and continuous learning, AI systems were shown to improve the accuracy of genomic interpretation over time. 37 The findings suggest that the integration of AI with global genomic resources not only improves data processing efficiency but also deepens biological insight, thereby laying the foundation for more informed clinical decision-making in precision medicine. 80 Nevertheless, these benefits depend strongly on data quality, cohort diversity, and transparent model development; otherwise, AI systems may amplify existing biases or produce findings that are difficult to reproduce. A prominent theme across the reviewed studies was the critical importance of population diversity in the construction and utilization of global genomic resources. 38 , 39 The results revealed that many existing genomic databases have historically been skewed toward populations of European ancestry, leading to gaps in knowledge regarding genetic variation in underrepresented groups. 40 , 41 Artificial intelligence played a pivotal role in addressing this imbalance by enabling the inclusion and analysis of diverse datasets from multiple geographic and ethnic populations. 42 AI-driven clustering and classification techniques were used to identify population-specific genetic signatures, thereby improving the representativeness and inclusivity of genomic resources. 43 These advancements have contributed to a more equitable understanding of human genetic diversity and its implications for health and disease. 44 Furthermore, the integration of diverse population data into AI models significantly enhanced the generalizability and robustness of predictive algorithms. Studies demonstrated that models trained on multi-ethnic datasets performed better in identifying disease risk and treatment responses across different populations compared to those trained on homogeneous data. 45–47 This finding underscores the importance of diversity not only as an ethical consideration but also as a scientific necessity. 48 The results indicate that incorporating global genetic variation into AI frameworks may reduce bias and improve the clinical applicability of precision medicine. 34 , 49 At the same time, models trained on underrepresented or unevenly sampled populations may still produce biased predictions, and diversity alone does not eliminate the need for rigorous validation across ancestry groups. 29 However, challenges related to data accessibility, ethical governance, and cross-border data sharing were also highlighted, suggesting the need for collaborative efforts to ensure equitable representation in future genomic research. 50 Together, these findings show that population diversity is central to both the scientific validity and equity of AI-supported precision medicine. The reviewed literature also showed geographic and institutional clustering, with many studies relying on large biobanks, institutional datasets, or high-resource research settings. This may limit the applicability of AI-genomics models in underserved populations and resource-limited health systems. The application of artificial intelligence in predictive modeling was one of the most extensively reported outcomes in the reviewed literature ( Figure 5 ). AI-based models demonstrated promising accuracy in predicting disease susceptibility, progression, and treatment outcomes by leveraging integrated genomic data. 51 , 52 However, high accuracy reported in retrospective studies may not always translate into reliable clinical performance, particularly when models are trained on small or biased cohorts. Techniques such as neural networks, support vector machines, and ensemble learning methods were frequently employed to analyze complex interactions between genetic variants and environmental factors. 53 These models were particularly effective in identifying polygenic risk scores, which quantify an individual’s genetic predisposition to specific diseases. 54 The integration of large-scale population genomic data further enhanced the predictive power of these models by incorporating a broader range of genetic influences. 36 Figure 5 Predictive modeling framework. The figure shows how genomic data, clinical and environmental factors, and pharmacogenomic information can be processed through AI techniques and modeling approaches to support disease susceptibility prediction, drug response prediction, and treatment outcome prediction. The green panel represents input data, the blue panel represents the AI model, and the red panel represents output domains. Arrows indicate the direction of analytical flow from inputs to AI modeling and outputs. Bold text indicates the main categories and subcategories within the framework. The bottom blue bar summarizes the overall goal of the framework. A diagram of AI predictive modeling using genomic, clinical and environmental data. Predictive modeling framework. The figure shows how genomic data, clinical and environmental factors, and pharmacogenomic information can be processed through AI techniques and modeling approaches to support disease susceptibility prediction, drug response prediction, and treatment outcome prediction. The green panel represents input data, the blue panel represents the AI model, and the red panel represents output domains. Arrows indicate the direction of analytical flow from inputs to AI modeling and outputs. Bold text indicates the main categories and subcategories within the framework. The bottom blue bar summarizes the overall goal of the framework. In clinical contexts, AI-driven predictive models were shown to support personalized treatment strategies by identifying patients who are more likely to benefit from specific therapies. 38 For example, studies reported improved prediction of drug response and adverse effects through the analysis of pharmacogenomics data. 55 , 56 This has significant implications for reducing trial-and-error approaches in treatment selection and improving patient outcomes. 57 Moreover, the use of longitudinal data and real-time analytics allowed AI systems to adapt predictions based on evolving patient information. 58 These findings highlight the potential of AI-enhanced predictive modeling to support precision medicine by enabling proactive and individualized healthcare interventions. 59 At present, this potential remains limited by overfitting risks, insufficient prospective validation, and variability in model reproducibility. The translation of AI-integrated genomic insights into clinical practice was a key focus of the reviewed studies. Results indicated that decision support systems powered by artificial intelligence are increasingly being used to assist clinicians in interpreting complex genomic data. 60 , 61 These systems integrate patient-specific genetic information with clinical guidelines and evidence-based knowledge to provide actionable recommendations. 62 In many cases, AI tools were embedded within electronic health record systems, allowing seamless access to genomic insights during clinical workflows. 63 This integration has the potential to improve the efficiency and accuracy of clinical decision-making, particularly in areas such as oncology, rare diseases, and pharmacogenomics. 64 However, many systems remain at the development or retrospective validation stage rather than being prospectively tested in routine clinical care. Despite these advancements, the studies also highlighted several challenges in the clinical implementation of AI-driven systems. Issues related to data standardization, interoperability, and user trust were frequently reported. 65 , 66 Clinicians expressed concerns about the transparency of AI algorithms and the potential for over-reliance on automated recommendations. 63 , 64 To address these challenges, some studies emphasized the importance of developing explainable AI models and providing adequate training for healthcare professionals. 67 , 68 Overall, the results suggest that while AI has significant potential to enhance clinical decision support, its successful integration into practice requires careful consideration of technical, ethical, and human factors. Particular attention is needed to ensure that model outputs are reproducible, clinically validated, and not used as substitutes for professional judgment. The integration of artificial intelligence with global genomic resources raises important ethical, legal, and social considerations, which were extensively discussed in the reviewed literature. One of the primary concerns was the protection of individual privacy and the secure handling of sensitive genomic data. 69 AI systems require access to large datasets to function effectively, but this raises the risk of data breaches and unauthorized use. 70 Studies emphasized the need for robust data governance frameworks, including encryption, anonymization, and controlled access mechanisms, to ensure the security of genomic information. 71 , 72 Additionally, the potential for misuse of genetic data in areas such as insurance and employment discrimination was highlighted as a significant ethical challenge. 73 Another key issue identified was the need for equitable access to AI-driven precision medicine. 74 While technological advancements have the potential to improve healthcare outcomes, there is a risk that these benefits may not be distributed equally across populations. 75 Studies pointed out that resource-limited settings may face barriers in adopting AI technologies due to infrastructure and cost constraints. Furthermore, the lack of representation of certain populations in genomic datasets could exacerbate existing health disparities. 76 , 77 The findings underscore the importance of developing inclusive policies and international collaborations to ensure that the benefits of AI-integrated genomic research are accessible to all. Ethical considerations should be integrated into every stage of research and implementation to promote fairness and social responsibility. Therefore, responsible governance should not be treated as a separate concern but as a core requirement for equitable AI-genomics integration. The reviewed studies identified several technical and methodological challenges that continue to hinder the full realization of AI integration with global genomic resources. One of the most important challenges is the complexity and scale of genomic data, which requires advanced computational resources and sophisticated algorithms. 38 , 45 Issues related to data quality, missing information, and heterogeneity were frequently reported, affecting the reliability of AI models. 41 These limitations can contribute to overfitting, poor transportability across populations, and inconsistent performance when models are applied outside the original training setting. Additionally, the lack of standardized data formats and interoperability between different systems poses a barrier to effective data integration. 78 Researchers emphasized the need for the development of unified frameworks and standards to facilitate seamless data exchange and analysis. 79 Looking forward, the studies highlighted several promising directions for future research and development. Advances in federated learning and decentralized AI were identified as potential solutions for addressing data privacy concerns while enabling collaborative analysis across institutions. 69 , 80 The integration of multi-omics data, including proteomics and metabolomics, was also suggested to enhance the depth and accuracy of genomic insights. 73 Furthermore, the incorporation of real-world data, such as electronic health records and wearable device data, could provide a more comprehensive understanding of patient health. 81 However, future research should prioritize prospective validation, external replication, standardized reporting, and transparent evaluation before AI genomic models are implemented clinically. The findings suggest that continued innovation and interdisciplinary collaboration will be essential for overcoming current challenges and unlocking the full potential of AI in precision medicine.

The integration of artificial intelligence with global genomic resources represents a promising but still evolving advancement in precision medicine, offering new opportunities to understand complex biological systems and improve individualized healthcare. By enabling the analysis of large-scale, diverse genomic datasets, artificial intelligence can enhance disease prediction, support targeted therapeutic interventions, and facilitate more inclusive research practices. However, current AI applications in genomics should not be interpreted as fully established clinical solutions, as many models remain limited by overfitting risks, small or biased cohorts, insufficient external validation, limited prospective clinical testing, and variable reproducibility. These conclusions should also be interpreted in light of the methodological heterogeneity of the included studies, which varied in design, data sources, validation strategies, and translational maturity. The successful translation of these innovations into clinical and public health contexts therefore depends on addressing persistent challenges related to data quality, interoperability, ethical governance, equitable access, model transparency, validation rigor, and reproducibility. A collaborative and multidisciplinary approach that aligns technological development with policy and clinical readiness is essential to maximize the benefits of this integration. Ultimately, the future of precision medicine will depend on the ability to harness artificial intelligence in a manner that is scientifically robust, ethically sound, clinically validated, reproducible, and globally inclusive.

The integration of artificial intelligence with global genomic resources represents an important but still developing shift in how biomedical knowledge is generated and translated into clinical practice. 82 Rather than functioning solely as a computational tool, artificial intelligence is increasingly shaping the way researchers conceptualize genetic complexity and disease mechanisms. 83 , 84 Existing literature suggests that conventional analytical approaches are often insufficient to capture the multidimensional nature of genomic data, particularly when dealing with large-scale and heterogeneous datasets. 85 , 86 In contrast, artificial intelligence can introduce adaptive and iterative learning processes that allow for continuous refinement of biological insights. 87 However, these advantages should be interpreted cautiously, as model performance may be overestimated when training data are small, biased, or insufficiently separated from validation data. This aligns with broader developments in biomedical informatics, where learning health systems are designed to evolve through ongoing data integration and feedback. 88 , 89 The findings of this review therefore reflect not only technical progress but also the need for rigorous validation, reproducibility, and careful clinical interpretation within data-driven precision medicine frameworks. A key aspect of this transformation is the increasing emphasis on population diversity in genomic research. 90 Numerous studies have demonstrated that limited representation of diverse populations can lead to biased findings and reduced generalizability of predictive models. 91 , 92 Artificial intelligence offers a practical solution by enabling the integration and analysis of data from multiple populations, thereby improving the inclusiveness and robustness of genomic research. 93 Nevertheless, the inclusion of diverse datasets does not automatically eliminate bias, particularly when sample sizes remain uneven across ancestry groups or when models are not externally validated in underrepresented populations. However, the importance of diversity extends beyond methodological considerations. It is closely linked to global disparities in research participation and access to healthcare innovation. 94 Prior studies have highlighted that underrepresentation of certain populations is often driven by structural inequalities, including limited research infrastructure and funding in low- and middle-income regions. 95–97 Consequently, while artificial intelligence can facilitate the technical inclusion of diverse datasets, addressing the root causes of inequity requires coordinated global efforts, policy interventions, and investment in capacity building. 98 The advancement of predictive modeling through artificial intelligence has significant implications for the evolution of precision medicine. 99 Research across various clinical domains indicates that machine learning models can enhance the prediction of disease risk, progression, and treatment response by integrating genomic and clinical data. 100 , 101 These developments are contributing to a shift from reactive to proactive healthcare, where interventions are tailored based on individual risk profiles. 102 The findings of this review support this transition, particularly in the context of incorporating global genetic variation into predictive frameworks. 103 However, this transition remains constrained by important current limitations, including overfitting, inconsistent external testing, limited prospective clinical validation, and variable reproducibility across published machine learning genomic models. The discussion should also consider the challenges associated with model interpretability and clinical acceptance. 104 Studies have consistently reported that healthcare professionals are more likely to adopt predictive tools when the underlying mechanisms are transparent and align with clinical reasoning. 105 , 106 This underscores the importance of explainable artificial intelligence, which aims to provide insights into how predictions are generated and to enhance trust in algorithmic outputs. 107 Even explainable models require cautious interpretation, as transparency alone does not guarantee clinical validity or reproducible performance across different settings. Translating artificial intelligence-driven genomic insights into routine clinical practice remains a complex and multifaceted process. 108 While decision support systems have shown promise in improving diagnostic accuracy and treatment planning, their implementation is often hindered by practical challenges. 109 These include issues related to data interoperability, integration with existing electronic health record systems, and variability in digital literacy among healthcare providers. 110 , 111 Furthermore, the introduction of artificial intelligence into clinical workflows raises important questions about responsibility and decision-making authority. 112 When algorithms provide recommendations based on complex genomic analyses, clinicians should balance these insights with their professional judgment. 113 This is particularly important because many AI genomic tools remain at the retrospective development or internal validation stage rather than being tested prospectively in real-world clinical workflows. Previous research suggests that successful integration requires not only technological innovation but also organizational readiness, training, and clear governance structures. 114 , 115 Without these elements, the potential benefits of artificial intelligence may not be fully realized in clinical settings. Ethical, legal, and social considerations are central to the responsible integration of artificial intelligence with global genomic data. 1 , 116 The use of large-scale genetic information raises concerns related to privacy, data security, and informed consent. 2 , 13 , 117 Studies have emphasized the need for robust data governance frameworks that ensure the protection of sensitive information while enabling meaningful research. 7 , 8 , 118 At the same time, there is a growing recognition that the benefits of artificial intelligence-driven precision medicine should be distributed equitably. 119 There is a risk that technological advancements may disproportionately benefit populations with greater access to healthcare resources, thereby widening existing disparities. 120 , 121 These concerns are closely linked to methodological issues, such as biased cohorts, poorly documented training datasets, and non-reproducible models that can reinforce existing inequities in genomic medicine. Addressing these challenges requires commitment to inclusive research practices, transparent policies, and international collaboration. Ultimately, the integration of artificial intelligence with genomic data should be guided not only by scientific and clinical objectives but also by principles of fairness, accountability, reproducibility, validation rigor, and social responsibility.

The integration of artificial intelligence with global genomic resources has significant implications for clinical practice, particularly in advancing personalized and data-driven care. Healthcare providers can leverage AI-enhanced genomic insights to improve diagnostic precision, identify disease risks earlier, and tailor treatment strategies based on individual genetic profiles. However, clinical use should be approached cautiously and limited to models that have undergone adequate external and prospective validation, with clear evidence of reproducibility and fairness across populations. Practice and policy recommendations should therefore be weighted toward studies with stronger empirical designs, transparent model reporting, external validation, and clinically relevant outcomes. Findings from qualitative stakeholder studies, proof-of-concept demonstrations, bibliometric analyses, educational evaluations, and conceptual papers should be interpreted as preliminary or supportive rather than as direct evidence for implementation. This shift toward proactive and individualized care requires clinicians to develop competencies in interpreting AI-supported outputs and integrating them into routine decision-making. Clinicians should also remain aware of the risks of overfitting, biased training cohorts, and limited generalizability when interpreting AI-generated genomic predictions. In addition, healthcare systems should invest in interoperable digital infrastructures that enable seamless integration of genomic data with electronic health records. Training programs for healthcare professionals are essential to ensure that AI tools are used effectively and responsibly, fostering trust and improving patient outcomes across diverse clinical settings. From a policy perspective, the findings highlight the need for comprehensive frameworks that address data governance, equity, and cross-border collaboration. Policymakers should establish clear regulations for data privacy, consent, and ethical use of genomic information while promoting standardized data formats to enhance interoperability. Equally important is the need to ensure equitable access to AI-driven precision medicine, particularly in low-resource settings where disparities may be exacerbated. Investment in infrastructure, capacity building, and international data-sharing agreements will be critical to support inclusive genomic research. Policies should also encourage the development of transparent and explainable AI systems to enhance accountability and public trust. In addition, regulatory and institutional frameworks require documentation of training data composition, validation strategy, overfitting controls, external testing, and reproducibility before clinical deployment. These coordinated efforts will be essential to translate technological advancements into sustainable and equitable healthcare improvements.