A split-and-merge deep learning approach for phenotype prediction

Abstract

Background: Phenotype prediction with genome-wide markers is a critical but difficult problem in biomedical research due to many issues such as nonlinearity of the underlying genetic mapping and high-dimensionality of marker data. When using the deep learning method in the small-n-large-p data, some serious issues occur such as over-fitting, over-parameterization, and biased prediction. Methods: In this study, we propose a split-and-merge deep learning method, named SM-DL method, to learn a neural network on the dimension reduce data by using the split-and-merge technique. Conclusions: Numerically, the proposed method has significant performance in phenotype prediction for a simulated example. A real example is used to demonstrate how the proposed method can be applied in practice.

Keywords

• deep learning
• genomic prediction
• high-dimensionality data
• machine learning
• neural networks