Stacked neural network for predicting polygenic risk score.

Area of Science:

• Genetics
• Computational Biology
• Bioinformatics

Background:

• Polygenic risk scores (PRS) are valuable for predicting disease susceptibility using genome-wide association studies (GWAS).
• Traditional PRS methods can suffer from overfitting and effect size overestimation due to correlated genetic variants.
• Existing models may not fully capture the complex interplay of genetic factors influencing disease risk.

Purpose of the Study:

• To introduce a novel Stacked Neural Network Polygenic Risk Score (SNPRS) to overcome limitations of traditional PRS.
• To enhance the accuracy and nuance of genetic risk prediction by integrating diverse genetic variant data.
• To evaluate the performance of SNPRS against existing methods in large-scale genetic datasets.

Main Methods:

• Developed SNPRS, a method synthesizing outputs from multiple neural networks trained on variants selected via different p-value thresholds.
• Utilized diverse p-value thresholds to capture a broader spectrum of genetic variants.
• Applied SNPRS to UK Biobank and Korea Genome and Epidemiology Study (KoGES) datasets for validation.

Main Results:

• SNPRS demonstrated superior predictive accuracy compared to traditional PRS models.
• The novel approach outperformed an isolated deep neural network in predicting disease susceptibility and quantitative traits.
• SNPRS effectively integrated information from a wider array of genetic variants for improved risk assessment.

Conclusions:

• SNPRS represents a significant advancement in polygenic risk prediction methodology.
• The stacked neural network approach offers a more robust and accurate way to assess genetic susceptibility.
• SNPRS holds promise for refining the efficacy and clinical relevance of PRS in genetic research and personalized medicine.