A Multivariate Poisson Deep Learning Model for Genomic Prediction of Count Data.

Area of Science:

• Genetics and Animal Breeding
• Statistical Genomics
• Machine Learning Applications

Background:

• Genomic selection (GS) is a predictive methodology revolutionizing plant and animal development.
• Existing GS models lack universal applicability, necessitating specific methods for diverse output types.
• Efficient methodologies for multivariate count data outcomes in GS are currently lacking.

Purpose of the Study:

• To propose a novel multivariate Poisson deep neural network (MPDN) model.
• To enable simultaneous genomic prediction of multiple count outcomes.
• To address the gap in efficient methodologies for multivariate count data in GS.

Main Methods:

• Developed an MPDN model utilizing Poisson distribution's minus log-likelihood as a loss function.
• Employed rectified linear unit (RELU) activation in hidden layers for nonlinear pattern capture.
• Used exponential activation in the output layer to ensure count-scale predictions.
• Implemented models in Tensorflow (back-end) and Keras (front-end).

Main Results:

• The proposed MPDN model demonstrated superior performance compared to univariate Poisson deep neural network models.
• MPDN prediction accuracy was comparable to univariate generalized Poisson regression models.
• Deep learning models were implementable on moderate to large datasets, a key advantage.

Conclusions:

• The MPDN model offers a viable approach for multivariate count data in genomic prediction.
• While competitive, MPDN did not surpass established generalized Poisson regression models in predictive power.
• The Tensorflow/Keras implementation provides scalability for large-scale genomic datasets.