Predicting individual performance from genomic information using Neural Networks

Published on
June 5, 2020

Neural networks are interesting models for prediction of individual phenotypes in livestock using genomic data, because these models can fit complex relationships between a large number of variables. In collaboration with researchers from Wageningen University, Machine Learning experts from Radboud University Nijmegen developed a Neural network that had robust performance across many simulated scenarios. However, the computation time of the Neural network was much longer than of traditional models, hindering its use in practice.

Predicting individual phenotypes

Decisions in daily farm management practices may be optimized by predicting the performance (i.e. phenotype) of young individuals. For example, pig farmers can optimize farm management by housing piglets together that have a similar predicted growth rate. The additive genetic component of a phenotype can be predicted with genotype data, using traditional genomic prediction models. Such models ignore the non-additive component that arise due to complex interactions between genes. Thus, the prediction of phenotypes from genotype data may be improved by modelling interactions between genotypes.

Neural networks

Neural networks are a class of prediction models that can fit complex relationships between a large number of variables, and may therefore be a promising tool to predict phenotypes from genomic marker data. These models are commonly applied in image classification problems, for example in identifying objects in pictures. Machine Learning experts from Radboud University worked in close collaboration with researchers from Wageningen University to develop a Neural network that can predict individual phenotypes from genotype data. They evaluated the performance of the neural network in simulated livestock datasets.

Performance of the Neural network

The results showed that the Neural network was slightly more accurate than traditional genomic prediction models in scenarios where the phenotype was influenced by only a small number of genes, or when interactions between genes were abundant and strong. In other scenarios, the accuracies were comparable. The downside of the Neural network was its computation time, which was about 200 times as long as that of traditional methods. In conclusion, although the Neural network had robust performance across scenarios, computation time may hinder use in practice.

This research was supported by the Netherlands Organisation of Scientific Research (NWO) and the Breed4Food consortium partners Cobb Europe, CRV, Hendrix Genetics, and Topigs Norsvin.