• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Friday, November 6, 2026
BIOENGINEER.ORG
No Result
View All Result
  • Login
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
No Result
View All Result
Bioengineer.org
No Result
View All Result
Home NEWS Science News Technology

Imaging, Genes and Proteins Combine to Predict Liver and Pancreas Disease Risk

by
October 30, 2026
in Technology
Reading Time: 6 mins read
0
Imaging, Genes and Proteins Combine to Predict Liver and Pancreas Disease Risk

Imaging, Genes and Proteins Combine to Predict Liver and Pancreas Disease Risk

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

Combining three different kinds of biological data—brain-scanning precision applied to the abdomen, genomes and blood proteins—can predict disorders of the liver, gallbladder, biliary tract and pancreas far better than any single data type alone, according to a new study published in BMC Artificial Intelligence. Researchers led by Mohamad Koohi-Moghadam and Kyongtae Tyler Bae analyzed records from 38,844 UK Biobank participants who had undergone both abdominal MRI scans and whole-genome sequencing, and found that a machine learning model integrating all available data types reached an area under the ROC curve of 0.903 and a test accuracy of 0.86 for classifying these hepato-pancreato-biliary disorders.

The work sits at the intersection of radiogenomics, a field that pairs quantitative measurements extracted from medical images with genetic information to probe the biology of disease. The imaging features in this study were not the high-dimensional texture descriptors often associated with radiomics, but six focused quantitative phenotypes measured from MRI: liver iron content, liver proton density fat fraction (PDFF), liver volume, pancreas iron content, pancreas PDFF and pancreas volume. These measurements, generated through standardized image analysis pipelines on the UK Biobank Research Analysis Platform, serve as non-invasive biomarkers of organ fat deposition, iron overload and structural change—processes implicated in conditions ranging from non-alcoholic fatty liver disease to pancreatitis, diabetes and cancers.

The first analytical stage was a genome-wide association study (GWAS) linking genetic variants to each MRI-derived trait. Participants were classified as cases or controls for each trait depending on whether their measurement exceeded the mean plus one standard deviation for that feature, and the team used PLINK and REGENIE to test millions of variants after extensive quality control, including filters on missingness, minor allele frequency, Hardy-Weinberg equilibrium, relatedness and population stratification. The analysis was restricted to participants of White British ancestry, with the top ten principal components included as covariates.

Many of the strongest signals replicated established biology. The variant rs1800562 within the HFE gene on chromosome 6, famous for its role in hereditary hemochromatosis, showed the most significant association with liver iron levels, with a -log10 p-value of 60.437. For liver fat, the strongest hit was rs58542926 near TM6SF2 (-log10 p = 45.923), followed by rs738409 within PNPLA3 (-log10 p = 37.377), both genes previously implicated in fatty liver disease. Other significant liver iron loci included rs855791 within TMPRSS6 and several nearby variants on chromosome 6. Liver volume was most strongly associated with rs4665972 near SNX17, and pancreas PDFF with rs1341982 near CDKN2C.

The study also reported a potentially novel finding: rs3734626, located near the RSPO3 gene on chromosome 6, was significantly associated with pancreas volume (-log10 p = 6.322), an association the authors state has not previously been reported in this context. RSPO3 is known to influence body fat distribution and adipose cell biology, suggesting the link to pancreas morphology may reflect broader metabolic roles. Pancreas volume was also associated with rs2287990 near CTRB1 and rs77581903 near AGAP12P. Enrichment analyses reinforced the iron theme: the largest fraction of implicated genes connected to hemoglobin measurement, and the most significantly enriched biological pathway was

the “Hfe effect on hepcidin production,” a pathway central to iron regulation and homeostasis, with “iron metabolism disorders” also showing moderate enrichment. The gene-trait association analysis found that nine of the implicated genes connected to hemoglobin measurement, seven to type II diabetes mellitus, and five each to hematocrit and total cholesterol measurement, while four genes each linked to aspartate aminotransferase measurement, liver fat measurement, liver disease biomarker, mean corpuscular hemoglobin concentration, reticulocyte count and serum alanine aminotransferase measurement. The authors interpret this clustering around iron-related traits as supporting their hypothesis that the identified variants play critical roles in the iron-related measurements visible on MRI.

Beyond the genetic associations, the study examined whether these multimodal features carried prognostic information. The primary endpoint for the survival analysis was all-cause mortality, with follow-up measured from the date of MRI acquisition to death or censoring at the end of available follow-up data. The researchers applied Kaplan-Meier survival analysis stratified by demographic factors, radiomics traits and significant SNPs, with differences assessed by log-rank tests, and used both univariate and multivariate Cox proportional hazards regression to quantify associations through hazard ratios and 95% confidence intervals.

In the multivariate Cox models, liver iron and pancreas PDFF emerged as strong risk factors, with elevated levels associated with significantly increased hazard of disease progression. The SNP rs3734626, the variant near RSPO3 identified in the GWAS as associated with pancreas volume, also showed a significant prognostic effect, with a hazard ratio of 1.355 and a p-value of 1.109e-02. Other significant predictors included liver volume, BMI, smoking status, sex and age. Higher liver volume and BMI were associated with increased risk, while smoking, male sex and older age correlated with worse outcomes. Kaplan-Meier curves for the top nine features showed clear separation between high and low groups; individuals with high pancreas PDFF or high liver iron exhibited significantly lower survival probabilities over time. Log-rank tests confirmed the statistical significance of these stratifications, with p-values ranging from below 1e-12 for age, sex and pancreas PDFF to below 0.05 for rs3734626 and liver volume.

The predictive modeling component focused on a narrower clinical question. The researchers built a multimodal dataset targeting disorders of the gallbladder, biliary tract and pancreas, defined by ICD-10 codes K80 through K87. Among 3,582 participants with available data, 189 had a diagnosis within this category. To create a balanced training set, they randomly selected 200 control participants without these diagnoses, producing a final dataset of 389 individuals characterized by 27 features: the six MRI-derived radiomics measures, 16 SNPs selected from the GWAS hits, and five proteomics features drawn from the Olink panel corresponding to proteins encoded by genes linked to significant SNPs, including interacting proteins identified through the IntAct database.

Eight classification algorithms were evaluated: a deep neural network, a support vector machine with a radial basis function kernel, logistic regression, k-nearest neighbors, gradient boosting, random forest, Naive Bayes and XGBoost. Data were split into 80% training and 20% testing sets using stratified sampling, with five-fold cross-validation on the training set for tuning and performance reported on the unseen test set. The SVM with RBF kernel achieved the highest performance, with a test AUC of 0.903 and accuracy of 0.86, followed by the deep neural network and logistic regression, while Naive Bayes consistently underperformed.

The feature-combination comparisons illustrated the additive value of each data layer. A model trained solely on demographics achieved an AUC of only 0.287, which the authors cite as evidence that basic patient characteristics alone have minimal predictive value for these disorders. Adding radiomics features raised the AUC to 0.744, incorporating SNP features further improved it to 0.878, and combining radiomics, SNPs and proteomics produced the best performance at 0.903. DeLong tests confirmed the statistical significance of these differences: the full multimodal model significantly outperformed the demographics-only, radiomics-only and radiomics-plus-SNPs models, with p-values below 0.001 for the first two comparisons and below 0.01 for the third. Radiomics plus SNPs significantly outperformed both demographics-only and radiomics-only models, and radiomics alone was significantly superior to demographics alone.

Detailed metrics for the SVM-RBF models traced the same progression. The radiomics-only model achieved precision of 0.591, recall of 0.684, an F1-score of 0.634 and accuracy of 0.615, with 68% sensitivity and 55% specificity in its confusion matrix. Adding SNP features raised precision to 0.794, recall to 0.711, the F1-score to 0.749 and accuracy to 0.769, with sensitivity of 71% and specificity of 83%. The full three-modality model reached precision of 0.861, recall of 0.816, an F1-score of 0.838 and accuracy of 0.846, corresponding to 82% sensitivity and 88% specificity. Cross-validation and test performance were closely aligned across models, which the authors present as indicating minimal overfitting and reasonable generalization.

The authors justify combining gallbladder, biliary and pancreatic disorders under a single outcome on mechanistic grounds: both the liver and pancreas are central to iron metabolism and lipid processing, the enriched pathways in their analysis concerned hepcidin production and iron metabolism disorders, and K80–K87 conditions frequently co-occur and share metabolic and inflammatory risk factors. On this view, the imaging measures of iron deposition and fat accumulation capture processes that unite these diagnoses rather than treating them as separate entities.

The study carries several limitations that the authors acknowledge. The GWAS and much of the analysis were restricted to participants of White British ancestry, a constraint of the UK Biobank population that may limit generalizability to more diverse groups; the authors call for replication in multi-ethnic cohorts. The analysis was also cross-sectional in design, so longitudinal studies with serial imaging and biomarker measurements would be needed to understand temporal changes in organ traits and any causal relationship with disease onset and progression. The survival endpoint was all-cause mortality rather than disease-specific death, and the machine learning dataset, while balanced by design, was small relative to the full cohort, with only 189 diagnosed cases available among participants with complete data.

Within those boundaries, the study offers a template for how population-scale imaging, genomic and proteomic resources can be layered: quantitative MRI phenotypes anchor the analysis, GWAS identifies variants associated with those phenotypes, and the resulting features feed both prognostic survival models and diagnostic classifiers. The convergence of iron-related genetic signals, imaging-measured iron and fat deposition, and improved prediction when all modalities are combined suggests, in the authors’ framing, that integrating diverse biological data can enhance disease risk stratification and support personalized healthcare strategies for liver and pancreatic disorders.

Subject of Research: Multimodal integration of MRI-derived radiomics traits, genome-wide association study variants and proteomic markers for prognostic and predictive modeling of liver and pancreas disorders in the UK Biobank.

Article Title: Multimodal radiogenomic integration enhances prognostic and predictive modeling of liver and pancreas disorders

Article References: Koohi-Moghadam, M., Vardhanabhuti, V., Wang, J., Chan, A. C.-Y., & Bae, K. T. (2026). Multimodal radiogenomic integration enhances prognostic and predictive modeling of liver and pancreas disorders. BMC Artificial Intelligence, 2(1), Article 4. https://doi.org/10.1186/s44398-026-00022-z

Image Credits: AI Generated

DOI: 10.1186/s44398-026-00022-z

Keywords: radiogenomics, UK Biobank, GWAS, machine learning, liver iron, pancreas PDFF, MRI biomarkers, proteomics, survival analysis, SNP rs3734626, RSPO3, hepato-pancreato-biliary disorders

News Source: Juliet Wilcox. (October 30, 2026). Imaging, Genes and Proteins Combine to Predict Liver and Pancreas Disease Risk. Scienmag.

Tags: GWAShepato-pancreato-biliary disordersliver ironMachine LearningMRI biomarkerspancreas PDFFProteomicsRadiogenomicsRSPO3SNP rs3734626survival analysisUK Biobank
Share12Tweet7Share2ShareShareShare1

Related Posts

Research report: The Diffusion of Digital Empowerment: The Relationship Between Digital Strategy, Digital Routinization, and Digital Capability

Research report: The Diffusion of Digital Empowerment: The Relationship Between Digital Strategy, Digital Routinization, and Digital Capability

November 6, 2026
Fourier-Enhanced TecoGAN Improves Video Super-Resolution Quality and Temporal Stability

Fourier-Enhanced TecoGAN Improves Video Super-Resolution Quality and Temporal Stability

November 6, 2026

Optimal Transport Augmentation Preserves Node Degree Information in Graph Contrastive Learning

November 6, 2026

GF-1 Satellite Leads in Modeling Salt Lake Potassium and Magnesium Concentrations

November 6, 2026

POPULAR NEWS

  • KRAS Phosphorylation Drives Colorectal Cancer Invasion by Boosting Cathepsin B Secretion

    KRAS Phosphorylation Drives Colorectal Cancer Invasion by Boosting Cathepsin B Secretion

    29 shares
    Share 12 Tweet 7
  • Platelets Undergo Lifespan Remodeling That Promotes Thromboinflammation and Fuels Tumor Progression

    29 shares
    Share 12 Tweet 7
  • Social Sciences Expose Hidden Inequities Shaping Neglected Tropical Diseases

    29 shares
    Share 12 Tweet 7
  • Fourier-Enhanced TecoGAN Improves Video Super-Resolution Quality and Temporal Stability

    29 shares
    Share 12 Tweet 7

About

We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.

Follow us

Recent News

Research report: The Diffusion of Digital Empowerment: The Relationship Between Digital Strategy, Digital Routinization, and Digital Capability

Beetle Diversity Surveyed Across Elevational Gradients in Colombia’s Coastal Sierra Nevada

KRAS Phosphorylation Drives Colorectal Cancer Invasion by Boosting Cathepsin B Secretion

Subscribe to Blog via Email

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

Join 86 other subscribers
  • Contact Us

Bioengineer.org © Copyright 2023 All Rights Reserved.

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
  • Homepages
    • Home Page 1
    • Home Page 2
  • News
  • National
  • Business
  • Health
  • Lifestyle
  • Science

Bioengineer.org © Copyright 2023 All Rights Reserved.