Machine Learning Identifies Shared Targets Between Diabetes and Aggressive Breast Cancer

A computational investigation combining microarray data mining, topological network analysis, and explainable machine learning has identified shared molecular signatures bridging type 2 diabetes and triple-negative breast cancer, proposing two existing pharmaceutical compounds as candidates for dual-disease drug repurposing. The research, published in BioData Mining by a multidisciplinary team led by Md. Feroj Ahmed and Md. Nurul Haque Mollah, addresses an urgent medical challenge arising from the complex clinical intersection of metabolic dysregulation and oncology. Triple-negative breast cancer, which represents approximately 15 to 20 percent of newly diagnosed breast cancers, is characterized by an absence of estrogen receptors, progesterone receptors, and human epidermal growth factor receptor 2 expression, leaving patients without standard targeted endocrine or HER2-directed interventions. Concurrently, type 2 diabetes affects hundreds of millions of adults worldwide and constitutes a documented risk factor for aggressive malignancies. By integrating multiple independent transcriptomic datasets, the study identified thirty-six shared differentially expressed genes and prioritized four central key genes that govern intersecting pathogenetic axes, offering a bioinformatic foundation for evaluating cooperative therapeutic strategies.

The clinical motivation for resolving common molecular pathways stems from the severe prognostic penalties observed when type 2 diabetes and triple-negative breast cancer occur concurrently. Epidemiological studies documented in the source indicate that individuals with type 2 diabetes experience a 20 to 30 percent elevated risk of developing triple-negative breast cancer and represent roughly 50 percent of breast cancer-related mortality. In postmenopausal women, breast cancer arising in the presence of diabetes is frequently diagnosed at advanced stages, exhibiting heightened invasiveness, distant disease relapse, and elevated mortality compared to disease presentations in patients without diabetes. Clinical management is further exacerbated by conflicting therapeutic regimens, where medications prescribed for glycemic control may inadvertently accelerate oncogenic behaviors or vice versa. Additionally, polypharmacy regimens that rely on independent single-drug protocols for each illness elevate the hazard of adverse drug-drug interactions and systemic toxicities. Discovering unified therapeutic agents capable of mitigating the pathophysiological drivers of both disorders concurrently could reduce pharmaceutical burdens and circumvent opposing pharmacological effects in this vulnerable patient demographic.

To explore these common biological foundations, the investigators retrieved ten distinct microarray-based gene expression datasets from the National Center for Biotechnology Information Gene Expression Omnibus database. For triple-negative breast cancer, the team gathered profiles GSE27447, GSE38959, GSE37751, GSE43358, and GSE65194, while datasets GSE25724, GSE9006, GSE29221, GSE29231, and GSE64998 were assembled for type 2 diabetes. These datasets represented diverse human populations, tissue origins, and experimental microarray platforms, providing broad biological and technological heterogeneity. To maximize statistical power, the researchers merged the five independent cohorts for each disease into unified case and control pools. The consolidated triple-negative breast cancer dataset encompassed 107 disease samples and 89 control samples, whereas the aggregated diabetes dataset comprised 61 disease samples and 85 control samples. To harmonize cross-platform variations and remove batch effects across disparate cohorts, the investigators applied quantile normalization and log-base-two transformations where appropriate, followed by empirical Bayes batch correction using the PyCombat implementation of the ComBat algorithm, with correction efficiency verified via boxplots and principal component analyses.

Following comprehensive data preprocessing, differential gene expression analyses were conducted independently for each disease cohort using the Linear Models for Microarray Data empirical Bayes framework implemented in the limma package. Strict statistical thresholds were enforced to ensure biological consistency: genes were classified as significantly dysregulated only if they demonstrated a Benjamini-Hochberg adjusted p-value of 0.05 or lower alongside an absolute log-base-two fold change of at least 2. In the triple-negative breast cancer dataset, this filtering yielded 2,117 significantly upregulated genes and 914 downregulated genes, whereas the type 2 diabetes dataset revealed 217 upregulated and 259 downregulated genes. Intersecting the disease-specific expression profiles isolated 36 shared differentially expressed genes displaying concordant dysregulation across both pathological states. Specifically, the analysis identified 24 genes that were consistently upregulated and 12 genes that were consistently downregulated in both patient populations. These 36 overlapping transcripts were hypothesized to represent core molecular disruptions jointly implicated in the metabolic and oncological manifestations of the dual disease state.

To identify topologically influential nodes within the identified expression signature, the 36 shared genes were mapped into a functional protein-protein interaction network using the STRING database with an interaction confidence threshold set at 0.5. The resulting interactome was imported into the Cytoscape platform and examined using the CytoHubba analytical plugin. To prevent analytical bias toward single topological definitions of network importance, the authors evaluated interactome nodes across six distinct topological metrics based on shortest paths: Degree, Betweenness centrality, Closeness centrality, Edge Percolated Component, Maximal Clique Centrality, and Stress. Genes that repeatedly achieved high rankings across these six algorithms were considered central structural hubs possessing elevated functional connectivity. Through this multi-metric evaluation, the authors designated the fifteen most frequently appearing hub genes as topologically significant shared genes, isolating candidates whose central positions suggested essential roles in mediating the structural integrity and signaling flow of the comorbid protein interaction network.

Parallel to the topological screening, the researchers applied an explainable machine learning approach to discover shared feature genes based on predictive importance. Gene expression matrices of the 36 shared differentially expressed genes were extracted from both disease cohorts and modeled using extreme gradient boosting algorithms. Models were trained separately on each disease dataset using five-fold cross-validation to tune hyperparameters, limit overfitting, and isolate non-linear transcriptional relationships capable of separating diseased samples from healthy controls. To interpret feature importance, the authors computed SHapley Additive exPlanations values, ranking the top 20 contributing genes for each condition. Intersecting these ranked outputs revealed nine shared feature genes: S100A9, CIRBP, ARL15, USP10, GCOM1, AASS, PLAUR, DHRS12, and PSMD1. The trained models demonstrated robust classification performance, which was further confirmed on independent validation cohorts (GSE113865 for breast cancer and GSE38642 for diabetes), achieving receiver operating characteristic areas under the curve of 0.89 and 0.96, respectively. Intersecting the nine machine-learning feature genes with the fifteen topological network hubs yielded four consensus shared key genes: S100A9, CIRBP, USP10, and PSMD1.

To delineate upstream molecular switches governing the expression of these four shared key genes, the authors constructed multi-layered regulatory interaction networks using NetworkAnalyst. Transcriptional control mechanisms were evaluated through a transcription factor network derived from experimentally validated binding profiles in the JASPAR database. This analysis identified three prominent transcriptional regulators: TP53, USF1, and EGR1, selected based on Degree and Betweenness topological centrality. The study noted that TP53 regulates apoptosis and cellular metabolism, with its function frequently compromised under chronic diabetic hyperglycemia and oxidative stress, thereby accelerating genomic instability. USF1 serves as a critical regulator of lipid and glucose homeostasis whose disruption promotes insulin resistance and generates nutrient-rich environments that support cancer proliferation, while EGR1 acts as a stress-responsive factor driving inflammatory and angiogenic pathways. Concurrently, post-transcriptional regulatory networks built using TarBase identified three key microRNAs: hsa-miR-20a-5p, hsa-miR-18a-5p, and hsa-let-7b-5p. The former two microRNAs are known oncogenic members of the miR-17-92 cluster that impair insulin signaling while driving cancer cell proliferation, whereas let-7b-5p represents a metabolic and tumor-suppressing regulator whose downregulation under metabolic stress relieves oncogene repression.

Functional enrichment analyses were subsequently performed using the Enrichr and GeneCodis4 computational platforms to identify biological themes and signaling cascades linked to the four shared key genes. Statistical significance was defined by a false discovery rate adjusted p-value below 0.05 across both tools. Gene Ontology analysis revealed that the shared genes significantly contribute to biological processes including autocrine signaling, chronic inflammatory response, and the regulation of autophagy. Key cellular component localizations included the cell nucleus, secretory granule lumen, and the proteasome complex, while molecular function terms highlighted RAGE receptor binding, Toll-like receptor binding, ubiquitin protein ligase binding, and calcium-dependent protein binding. At the pathway level, Kyoto Encyclopedia of Genes and Genomes enrichment highlighted IL-17 signaling, the proteasome, and pathways of neurodegeneration as central mechanistic links. Among the key genes, S100A9 was characterized as an amplifier of chronic inflammation via TLR4 and RAGE interactions, PSMD1 was linked to proteasome-dependent turnover of metabolic and tumor-suppressive proteins, CIRBP was tied to stress-responsive mRNA stabilization, and USP10 was implicated in deubiquitination processes governing autophagy and protein homeostasis.

Leveraging these molecular targets, the researchers implemented a computational drug repurposing pipeline to screen approved and investigational compounds. Candidate molecules associated with the diseases and identified genes were compiled from DrugBank and DGIdb, with druggability assessments conducted using DoGSiteScorer, confirming that all four key proteins and candidate transcription factors possessed druggability scores exceeding 0.5. Blind molecular docking performed in AutoDock Vina evaluated binding affinity scores across three-dimensional protein structures retrieved from the Protein Data Bank and SWISS-MODEL. The five compounds demonstrating the strongest average binding affinities (all exceeding negative 7.0 kilocalories per mole) were tucatinib, pimozide, tepotinib, ursodiol, and zanubrutinib. Subsequent pharmacokinetic and safety assessments via Deep-PK, pkCSM, and ADMETlab 3.0 evaluated Caco-2 permeability, intestinal absorption, blood-brain barrier penetration, and cytochrome P450 inhibition. Both tepotinib and ursodiol demonstrated zero violations of Lipinski’s Rule of Five and showed no predicted inhibition of CYP3A4 or CYP2D6 enzymes, indicating lower risks of metabolic drug interactions. Cross-validation docking against the top ten disease-specific key genes for diabetes and breast cancer confirmed consistent binding affinities below negative 7.0 kilocalories per mole.

To evaluate dynamic receptor-ligand interactions beyond static docking calculations, the authors performed 100-nanosecond molecular dynamics simulations in YASARA utilizing the AMBER14 force field under simulated physiological conditions, focusing on complexes formed with the hub protein CIRBP. Trajectory analyses measured backbone root mean square deviation, radius of gyration, solvent-accessible surface area, and residue root mean square fluctuation, complemented by MM-PBSA binding free energy estimations. In structural stability assessments, the CIRBP-ursodiol complex exhibited consistently lower backbone root mean square deviation and maintained a lower, more stable radius of gyration, reflecting superior structural compactness and fold stability compared to CIRBP-tepotinib. Conversely, MM-PBSA calculations indicated comparable average binding free energies, with tepotinib showing slightly stronger energetic interactions at specific intervals. While both compounds showed promising in silico stabilization and occupy known binding pockets involving residues like phenylalanine-16 and lysine-91, the authors emphasized significant study limitations: the analyses relied entirely on independent datasets because combined comorbidity transcriptomes were unavailable, and in silico docking inherently carries risks of false-positive predictions, meaning these findings remain preliminary until verified through rigorous in vitro, in vivo, and clinical investigations.

Subject of Research: Bioinformatic and machine learning analysis of shared molecular mechanisms between type 2 diabetes and triple-negative breast cancer

Article Title: Large-scale transcriptomic data mining using explainable XGBoost and SHAP reveals shared biomarkers and molecular mechanisms between type-2 diabetes and triple-negative breast cancer for drug repurposing

Article References: Ahmed, M. F., Sharma, O., Noor, T., Latif, M. A., Ahmed, M. F., Bashak, M., Ahmed, D. M., Hossain, M. S., & Mollah, M. N. H. (2026). Large-scale transcriptomic data mining using explainable XGBoost and SHAP reveals shared biomarkers and molecular mechanisms between type-2 diabetes and triple-negative breast cancer for drug repurposing. BioData Mining, 19(1), Article 70. https://doi.org/10.1186/s13040-026-00575-9

Image Credits: AI Generated

DOI: 10.1186/s13040-026-00575-9

Keywords: Triple-Negative Breast Cancer, Type 2 Diabetes, Transcriptomics, Machine Learning, XGBoost, Drug Repurposing, Molecular Dynamics, Tepotinib, Ursodiol, Biomarkers, Large-scale, transcriptomic