Machine Learning for Prediction of Resistance Scores in Wheat (Triticum aestivum L.)

Abstract

Machine learning methods were shown to improve the prediction accuracies of genomic prediction of resistance scores compared to methods like RR-BLUP, which were originally designed for metric rather than ordinal response values. We conducted a cross-validation study with 361 wheat genotypes evaluated for five fungal diseases. Our objective was to compare the prediction accuracy and the ability to identify the most resistant genotypes of 19 genomic prediction approaches. Each approach consisted of a different combination of prediction method (RR-BLUP, an alternative method with heterogeneous marker variances, Bayesian generalized linear regression with an ordinal response, support vector machine, gradient boosting machine and random forest), predictor (single SNP markers, LD-based haplotype blocks, 250 variables generated with an autoencoder and SNPs identified with incremental feature selection) and response value (untransformed and logit-transformed resistance scores). In our dataset, RR-BLUP was consistently among the methods with the largest prediction accuracies and the best abilities to identify resistant genotypes in four of five investigated traits. However, in P. triticina, using gradient boosting machine and random forest instead of RR-BLUP increased the prediction accuracy from 0.64 to 0.71, indicating that machine learning methods may have an advantage over linear models in genomic prediction. We also found that even though there was a positive correlation between the prediction accuracy and Cohen's κ, a measure to judge how well the most resistant genotypes can be identified, the correlation is not perfect and a large value for the prediction accuracy does not necessarily translate into an equally large κ value.