Can Deep Learning Improve Genomic Prediction of Complex Human Traits?

Area of Science:

• Genetics
• Bioinformatics
• Artificial Intelligence

Background:

• Artificial intelligence (AI), particularly deep learning (DL), is gaining traction in complex trait genetic analysis.
• The performance of DL models, such as Multilayer Perceptrons (MLPs) and Convolutional Neural Networks (CNNs), for genomic prediction of human traits remains underexplored.

Purpose of the Study:

• To comprehensively evaluate the predictive performance of MLPs and CNNs for complex human traits using UK Biobank data.
• To compare DL methods against established Bayesian linear regression models in genomic prediction.

Main Methods:

• Utilized UK Biobank data from approximately 100,000 individuals with ~500,000 SNPs.
• Analyzed five complex human phenotypes (height, bone mineral density, BMI, systolic blood pressure, waist-hip ratio).
• Compared MLPs and CNNs against Bayesian linear regressions after hyperparameter optimization and SNP preselection.

Main Results:

• For highly heritable traits like height, DL methods showed similar performance to linear models, with CNNs slightly underperforming.
• For other traits, some CNNs performed comparably or slightly better than linear models.
• MLP performance varied significantly based on SNP set and phenotype; no DL method substantially outperformed linear models.

Conclusions:

• Deep learning models, including CNNs, demonstrate competitive performance against linear models for genomic prediction of complex human traits.
• Current DL approaches do not offer a significant advantage over linear models for the evaluated traits.
• Adaptation of CNNs for genetic data analysis is necessary for them to surpass linear model performance.