Unsupervised representation learning on high-dimensional clinical data improves genomic discovery and prediction.

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• High-dimensional clinical data (HDCD) in large biobanks offer potential for genetic discovery but present analytical challenges.
• Existing methods struggle to fully leverage the complexity and richness of HDCD for identifying genetic associations.
• Unsupervised learning approaches are needed to extract meaningful features from HDCD.

Purpose of the Study:

• To introduce Representation Learning for Genetic Discovery on Low-Dimensional Embeddings (REGLE), an unsupervised deep learning model.
• To enable the discovery of genetic variant associations with HDCD.
• To improve disease prediction using polygenic risk scores (PRSs) derived from novel genetic discoveries.

Main Methods:

• Developed REGLE, an unsupervised deep learning model utilizing variational autoencoders to create nonlinear, disentangled embeddings of HDCD.
• Integrated these embeddings as inputs for genome-wide association studies (GWAS).
• Applied REGLE to analyze spirograms (respiratory) and photoplethysmograms (circulatory) from biobank data.

Main Results:

• REGLE successfully identified known and novel genetic loci associated with respiratory and circulatory HDCD.
• The model uncovered clinically relevant features not captured by traditional expert-defined variables.
• PRSs constructed using REGLE-identified loci demonstrated improved disease prediction accuracy across multiple biobanks.
• REGLE-derived genetic associations were predictive of overall survival.

Conclusions:

• REGLE effectively analyzes HDCD for genetic discovery, overcoming limitations of existing methods.
• The model extracts clinically relevant information, enhancing the accuracy of genetic association studies.
• REGLE facilitates the development of more powerful PRSs, improving disease risk prediction and patient stratification.