Comparison of machine learning methods for genomic prediction of selected Arabidopsis thaliana traits

We present a comparison of machine learning methods for the prediction of four quantitative traits in Arabidopsis thaliana. High prediction accuracies were achieved on individuals grown under standardized laboratory conditions from the 1001 Arabidopsis Genomes Project. An existing body of evidence suggests that linear models may be impeded by their inability to make use of non-additive effects to explain phenotypic variation at the population level. The results presented here use a nested cross-validation approach to confirm that some machine learning methods have the ability to statistically outperform linear prediction models, with the optimal model dependent on availability of training data and genetic architecture of the trait in question. Linear models were competitive in their performance as per previous work, though the neural network class of predictors was observed to be the most accurate and robust for traits with high heritability. The extent to which non-linear models exploit interaction effects will require further investigation of the causal pathways that lay behind their predictions. Future work utilizing more traits and larger sample sizes, combined with an improved understanding of their respective genetic architectures, may lead to improvements in prediction accuracy.