Enhancing Type 1 Diabetes Polygenic Risk Prediction Through Neural Networks and Entropy-Derived Insights.

Area of Science:

• Genetics and Genomics
• Immunology
• Computational Biology

Background:

• Type 1 diabetes (T1D) is an autoimmune disease with significant genetic heritability (~70%).
• Early risk identification is critical for timely intervention and management.
• Existing polygenic risk scores (PRS) often rely on linear assumptions and have limited generalizability.

Purpose of the Study:

• To develop a novel neural network classifier for predicting T1D genetic risk.
• To evaluate the utility of including an entropy-derived feature representing genetic variability.
• To assess the model's performance and generalizability in internal and external cohorts.

Main Methods:

• A feed-forward neural network was trained using genotype data from 67 T1D-associated single nucleotide polymorphisms (SNPs).
• Data from 11,909 UK Biobank participants (546 T1D cases, 11,363 controls) were analyzed.
• Model performance was validated using five-fold cross-validation, a held-out UK Biobank test set, and an external German cohort.

Main Results:

• The neural network achieved a median Area Under the Receiver Operating Characteristic Curve (AUC) of 0.903 during cross-validation.
• The model demonstrated strong performance on the UK Biobank test set (AUC = 0.8889) and external validation.
• A probability-based risk framework provided high negative predictive value (98.9%) for the 'very low' risk group and high specificity (97.3%) for the 'very high' risk group.

Conclusions:

• The proposed neural network reliably predicts T1D genetic risk using a compact set of 67 SNPs.
• The model maintains accuracy across internal and external European cohorts, enabling clinically interpretable risk thresholds.
• This approach offers comparable discriminative performance to established genetic risk models, with potential for future feature integration.