Friday, November 20, 2026
Science
No Result
View All Result
  • Login
  • HOME
  • SCIENCE NEWS
  • CONTACT US
  • HOME
  • SCIENCE NEWS
  • CONTACT US
No Result
View All Result
Scienmag
No Result
View All Result
Home Science News

Knowledge Graphs Improve Diabetes Prediction by Integrating Incompatible Gene Expression Datasets

November 20, 2026
in Science News
Juliet Wilcox
By Juliet Wilcox Scienmag Editorial Profile - Human Genetics
Reading Time: 6 mins read
0
Knowledge Graphs Improve Diabetes Prediction by Integrating Incompatible Gene Expression Datasets

Knowledge Graphs Improve Diabetes Prediction by Integrating Incompatible Gene Expression Datasets

65
SHARES
587
VIEWS
Share on FacebookShare on Twitter
ADVERTISEMENT

Integrating diverse gene expression datasets with domain-specific knowledge through biomedical knowledge graphs can substantially boost the performance of machine learning models in predicting diabetes, according to a study published in the Journal of Biomedical Semantics by researchers Rita T. Sousa and Heiko Paulheim. By mapping disparate clinical data collections alongside structured biological knowledge, the researchers achieved gains in weighted F1-scores of up to 13 percent over traditional methods. The framework overcomes a persistent bottleneck in computational biology: the inability to directly combine gene expression datasets from distinct clinical studies because different platforms measure non-identical sets of genes across relatively small patient cohorts.

Diabetes represents a major global health challenge that directly accounted for 1.5 million deaths worldwide in 2019, with nearly half of those fatalities occurring in individuals under the age of 70, according to data from the World Health Organization. The chronic condition arises when the pancreas produces insufficient insulin or when the body cannot effectively utilize the insulin it generates, often leading to severe secondary complications such as blindness, kidney failure, cardiovascular disease, stroke, and lower limb amputation. While machine learning holds promise for detecting complex regulatory mechanisms and identifying individuals at risk, typical gene expression studies are severely constrained by limited participant numbers, creating an obstacle for data-driven supervised algorithms that demand large volumes of labeled training examples.

Combining independent datasets has traditionally presented substantial technical hurdles because differing experimental designs, microarrays, and platforms produce incompatible feature spaces. A naive concatenation of such matrices introduces widespread missing or null values for genes measured in one study but omitted in another. Previous studies typically either restricted their scope to a single cohort or trimmed their data down to a small intersection of shared genes, discarding valuable biological information from unshared transcripts and ignoring functional relationships between gene products. To overcome these limitations, the authors constructed a unified knowledge graph incorporating microarray data from the public Gene Expression Omnibus alongside biological hierarchies from the Gene Ontology and molecular networks from the STRING database.

The data processing pipeline began by collecting raw microarray values from three separate datasets: GSE123658, GSE140627, and GSE143143. Each dataset cataloged subjects diagnosed with type 1 diabetes alongside non-diabetic control individuals. In the GSE123658 cohort, which provided detailed demographic data, participants included 40 female and 42 male patients with ages ranging between 19 and 73 years, showing a median age of 41 years among diabetic patients and 30 years in the control cohort. Microarray probes lacking associated gene annotations were systematically filtered out, and the remaining measurements were evaluated across three normalization strategies: raw unnormalized values, gene-level min-max normalization scaling values between 0 and 1 across patients, and patient-level min-max normalization scaling across genes.

To convert continuous numerical expression values into machine-readable graph structures, the team implemented two distinct knowledge graph design strategies. The first employed an expression binning strategy using blank nodes in the Resource Description Framework schema to link individual patient entities to discrete value intervals for specific genes based on the distribution of unique expression values. The second strategy established direct relational links between patients and genes whenever an observed expression level exceeded a calculated benchmark. This thresholding compared gene expression either against the patient’s own mean expression across all genes or against the mean expression of that specific gene across the entire patient population.

To enrich these patient measurements with biological context, the graph incorporated functional annotations from the Gene Ontology, which delineates biological processes, molecular functions, and cellular components as three independent root hierarchies. The researchers mapped gene identifiers to corresponding protein entities using the UniProt mapping service and incorporated high-confidence protein-protein interactions from the STRING database, filtering for interactions with combined scores exceeding 0.7. The resulting semantic graph was mapped into directed, labeled Resource Description Framework triples, establishing a unified relational space where patient phenotypes, transcript levels, gene functions, and physical protein interactions were formally interconnected.

The researchers transformed these rich graph structures into dense vector spaces using six established knowledge graph embedding algorithms: the walk-based RDF2vec method, translational distance models TransE and TransR, and semantic matching models ComplEx, DistMult, and HolE. For each embedding model, eight different patient representation schemes were tested. Patients were represented either directly via node embeddings generated during graph processing or indirectly by taking the weighted average of gene embeddings weighted by their measured expression levels. The resulting 100-dimensional patient vectors were subsequently supplied as inputs to a multi-layer perceptron classifier evaluated using a five-fold stratified cross-validation setup on the GSE123658 dataset, augmented during training by data from the other two cohorts.

Experimental evaluations revealed that models trained on representations generated through the knowledge graph consistently outperformed conventional baselines. While direct gene expression baselines yielded a top weighted F1-score of 0.771, the knowledge graph methodology achieved weighted F1-scores as high as 0.870. The walk-based RDF2vec algorithm paired with patient-gene links based on individual mean expression thresholds delivered the most robust predictive performance. The authors noted that RDF2vec effectively traverses multi-hop paths within ontologies, capturing indirect semantic connections between distinct genes across the graph that traditional translational and matrix factorization methods struggled to leverage.

Comparative analyses against established multi-omics integration frameworks, including SIMBA and MultiMAP, as well as an untrained Graph Neural Network architecture, further highlighted the advantages of the knowledge graph approach. Both SIMBA and MultiMAP underperformed relative to the primary gene expression baselines, indicating that aggressive dimensionality reduction without domain-specific biological context harms downstream diagnostic accuracy. Similarly, the Graph Neural Network encountered difficulties processing the heterogeneous schema of the knowledge graph, in part because the input nodes lacked the uniform intrinsic feature vectors typically required by graph convolutional layers.

Clustering analyses based on Calinski-Harabasz, Davies-Bouldin, and silhouette scores, alongside two-dimensional t-SNE visualizations, demonstrated that the learned patient vectors formed distinct groupings reflecting diabetic and non-diabetic phenotypes. Visualizations also indicated that patient-level normalization successfully harmonized patients across differing source datasets, whereas certain gene-level strategies retained noticeable cohort-specific clustering. While the present methodology focused on binary classification within type 1 diabetes gene expression profiles, the authors concluded that the knowledge graph architecture provides a versatile blueprint that can readily accommodate multi-class disease subtyping, multi-omics measurements, and diverse clinical applications across medicine.

A detailed examination of the representation techniques highlighted why specific graph designs succeeded while others struggled. While patient-gene linking proved most effective for predictive tasks, the binning approach consistently underperformed, often failing to match baseline metrics. In that design, blank nodes linked patients to specific expression bins and genes, creating separate triples that disrupted path exploration. This structure prevented walk-based algorithms from tracing unbroken routes between genes and their measured expression levels, hindering downstream classification. In contrast, representing patients through a weighted average of gene embeddings showed moderate success, though prior studies had reported greater predictive power with that strategy. The researchers noted that this weighted averaging method loses efficacy when applied to datasets containing higher counts of genes, suggesting that the total number of evaluated genes directly influences representation quality.

The evaluation framework also incorporated rigorous baseline comparisons that clarified the impact of naive data merging. When constructing direct gene expression baselines without graph structures, the researchers observed that simply merging datasets and assigning zero values to unmeasured genes yielded modest performance gains over training on an isolated cohort. This unexpected improvement occurred because the independent datasets shared a subset of common genes, allowing models to benefit slightly from an expanded training volume despite the absence of systematic data integration. However, semantic integration within the knowledge graph yielded vastly superior predictive capacity, verifying that structuring domain-specific relationships across disparate cohorts provides advantages beyond basic sample pooling.

Beyond predictive accuracy, clustering evaluations across the three cohorts provided deeper insights into how different algorithms organized patient representations. Computing Calinski-Harabasz, Davies-Bouldin, and silhouette scores over the raw patient embeddings revealed consistent patterns across models. Higher Calinski-Harabasz and silhouette scores indicate better cluster separation, whereas lower Davies-Bouldin scores denote tighter clustering. Across these metrics, RDF2vec, HolE, and TransR consistently exhibited superior clustering behavior, whereas ComplEx, DistMult, and TransE delivered comparatively weaker results. RDF2vec achieved its strongest separation scores when paired with patient-gene links based on individual patient average expression thresholds, demonstrating an ability to capture entity relationships without suffering from difficulties that translational and semantic matching methods encounter when learning representations of concrete entities rather than abstract ontology classes.

Despite these encouraging outcomes, the authors outlined several key methodological constraints that frame future investigations. The current architecture operates exclusively on gene expression data and does not yet incorporate additional omics layers, such as proteomics or metabolomics, or comprehensive clinical records that could offer a more holistic biological perspective. Furthermore, structuring disease detection as a binary classification problem simplifies the clinical reality of complex conditions. The investigators emphasized that expanding the methodology to support multi-class or multi-label classification will be essential for distinguishing between specific disease stages or distinct clinical subtypes, such as varying forms of cancer, across broader patient populations.

Subject of Research: Biomedical data integration and diabetes prediction using knowledge graphs and machine learning

Article Title: Gene expression knowledge graph for patient representation and diabetes prediction

Article References: Sousa, R. T., & Paulheim, H. (2025). Gene expression knowledge graph for patient representation and diabetes prediction. Journal of Biomedical Semantics, 16(1), Article 2. https://doi.org/10.1186/s13326-025-00325-6

Image Credits: AI Generated

DOI: 10.1186/s13326-025-00325-6

Keywords: Knowledge Graphs, Diabetes Prediction, Gene Expression, Machine Learning, RDF2vec, Graph Embeddings, Biomedical Semantics, Bioinformatics, Type 1 Diabetes, Gene Ontology, Protein Protein Interactions, STRING Database, Multi Layer Perceptron, Cross Validation, Data Integration, Translational Embeddings

Cite Scienmag News

Juliet Wilcox. (November 20, 2026). Knowledge Graphs Improve Diabetes Prediction by Integrating Incompatible Gene Expression Datasets. Scienmag. https://scienmag.com/knowledge-graphs-improve-diabetes-prediction-by-integrating-incompatible-gene-expression-datasets/

Juliet Wilcox. "Knowledge Graphs Improve Diabetes Prediction by Integrating Incompatible Gene Expression Datasets." Scienmag, 20 November 2026, https://scienmag.com/knowledge-graphs-improve-diabetes-prediction-by-integrating-incompatible-gene-expression-datasets/. Accessed 20 November 2026.

Juliet Wilcox. "Knowledge Graphs Improve Diabetes Prediction by Integrating Incompatible Gene Expression Datasets." Scienmag. November 20, 2026. https://scienmag.com/knowledge-graphs-improve-diabetes-prediction-by-integrating-incompatible-gene-expression-datasets/

Tags: bioinformaticsbiomedical semanticscross-validationdata integrationdiabetes predictiongene expressionGene Ontologygraph embeddingsknowledge graphsMachine learningmulti-layer perceptronprotein-protein interactionsRDF2vecSTRING databaseTranslational Embeddingstype 1 diabetes
Share26Tweet16
Previous Post

Cofactor Engineering Increases Succinic Acid Productivity in Engineered Yarrowia lipolytica

Next Post

Nationwide Study Tests Implementation of Acupuncture and Massage for Cancer Pain Management

Related Posts

Graph Reinforcement Learning Curbs Power Grid Cyber Attacks in Distribution Systems
Technology and Engineering

Graph Reinforcement Learning Curbs Power Grid Cyber Attacks in Distribution Systems

November 20, 2026
Male Bumblebees Help Incubate Developing Brood and Accelerate Nest Growth
Biology

Male Bumblebees Help Incubate Developing Brood and Accelerate Nest Growth

November 20, 2026
Cucurbitacin B and Resveratrol Combination Demonstrates Synergistic Antitumor Activity in Breast Cancers
Medicine

Cucurbitacin B and Resveratrol Combination Demonstrates Synergistic Antitumor Activity in Breast Cancers

November 20, 2026
India Launches National HPV Vaccination Programme
Cancer

India Launches National HPV Vaccination Programme

November 20, 2026
Comparative Study Maps Narrative Functions Across Premodern Chinese and European Drama
Psychology & Psychiatry

Comparative Study Maps Narrative Functions Across Premodern Chinese and European Drama

November 20, 2026
Inpatient Mental Health Units Face Severe Staff Shortages and Safety Crises
Medicine

Inpatient Mental Health Units Face Severe Staff Shortages and Safety Crises

November 20, 2026
Next Post
Nationwide Study Tests Implementation of Acupuncture and Massage for Cancer Pain Management

Nationwide Study Tests Implementation of Acupuncture and Massage for Cancer Pain Management

  • Mothers who receive childcare support from maternal grandparents show more optimized

    Mothers who receive childcare support from maternal grandparents show more parental warmth, finds NTU Singapore study

    27656 shares
    Share 11059 Tweet 6912
  • University of Seville Breaks 120-Year-Old Mystery, Revises a Key Einstein Concept

    1061 shares
    Share 424 Tweet 265
  • Bee body mass, pathogens and local climate influence heat tolerance

    682 shares
    Share 273 Tweet 171
  • Researchers record first-ever images and data of a shark experiencing a boat strike

    546 shares
    Share 218 Tweet 137
  • Groundbreaking Clinical Trial Reveals Lubiprostone Enhances Kidney Function

    531 shares
    Share 212 Tweet 133
Science

Embark on a thrilling journey of discovery with Scienmag.com—your ultimate source for cutting-edge breakthroughs. Immerse yourself in a world where curiosity knows no limits and tomorrow’s possibilities become today’s reality!

RECENT NEWS

  • Graph Reinforcement Learning Curbs Power Grid Cyber Attacks in Distribution Systems
  • Male Bumblebees Help Incubate Developing Brood and Accelerate Nest Growth
  • Cucurbitacin B and Resveratrol Combination Demonstrates Synergistic Antitumor Activity in Breast Cancers
  • India Launches National HPV Vaccination Programme

Categories

  • Agriculture
  • Anthropology
  • Archaeology
  • Athmospheric
  • Biology
  • Biotechnology
  • Blog
  • Bussines
  • Cancer
  • Chemistry
  • Climate
  • Earth Science
  • Editorial Policy
  • Marine
  • Mathematics
  • Medicine
  • Pediatry
  • Policy
  • Psychology & Psychiatry
  • Science Education
  • Science News
  • Social Science
  • Space
  • Technology and Engineering

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 5,151 other subscribers

© 2026 Scienmag - Science Magazine

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • HOME
  • SCIENCE NEWS
  • CONTACT US

© 2026 Scienmag - Science Magazine

Discover more from Science

Subscribe now to keep reading and get access to the full archive.

Continue reading