Genetic risk predictions using deep learning models with summary data.

Area of Science:

• Genomics
• Bioinformatics
• Artificial Intelligence

Background:

• Deep learning drives the Fourth Industrial Revolution, showing promise in genetic and genomic studies.
• Privacy concerns and data-sharing limitations restrict access to individual-level genetic data for research.
• Alternative data sources are needed for deep learning applications in genomics.

Purpose of the Study:

• To investigate deep learning model performance using genetic summary data (e.g., linkage disequilibrium matrices).
• To compare deep learning predictive accuracy with individual-level versus summary genetic data.
• To explore deep learning as an alternative for genetic risk prediction with limited data.

Main Methods:

• Applied various deep learning models: deep neural networks, convolutional neural networks, recurrent neural networks, and transformers.
• Utilized the bootstrap method to approximate test error for model evaluation.
• Conducted simulation studies and real data analyses to compare performance metrics.

Main Results:

• Most deep learning models demonstrated comparable test mean squared errors (MSEs) between individual-level and summary genetic data.
• Deep learning approaches showed robust performance even with aggregated genetic information.
• The findings validate the utility of summary data in deep learning for genetic prediction.

Conclusions:

• Deep learning models can effectively predict disease-related traits using only linkage disequilibrium matrices.
• Genetic summary data presents a viable alternative when individual-level data is inaccessible.
• This research expands the applicability of deep learning in genomic studies under data-sharing constraints.