Combining environmental and genomic information in artificial neural networks for phenotypic prediction

Abstract

Artificial neural networks (ANNs) have received increasing attention for phenotypic and genomic prediction because they can accommodate non-linear relationships and high-dimensional inputs. Their performance has nevertheless been inconsistent when using genomic data, and many reported studies have relied on phenotypes pre-corrected for fixed and non-genetic random effects. The objective of this study was to evaluate the feasibility and predictive ability of ANNs that use environmental and genomic information jointly for phenotypic prediction, thereby avoiding phenotype pre-correction while allowing non-linear relationships among effects to be captured. Three levels of complexity were investigated. The first baseline model (ANN 1) was an ANN including a single input layer combining environmental and genomic data followed by a dropout layer with a 30 % rate, and then a fully connected (FC) layer with a single node and a linear activation function. The second model (ANN 2) consisted in a modular three-network architecture that separated genomic and environmental inputs, each with one FC layer and one node, before combining their predictions to predict phenotypes with a third ANN, also with one FC layer. The third model (ANN 3) was built on the structure of ANN 2 and included additional locally connected layers and non-linear activation functions. The predictive abilities of the ANNs were evaluated using data from the Dutch routine genomic evaluations for dry matter intake (DMI) and enteric methane (CH4) emissions, partitioned into training, validation and test subsets. As benchmark, phenotypes were also predicted with Single Nucleotide Polymorphism Best Linear Unbiased Prediction (SNPBLUP) applied on the training sets only. Modelling environmental and genomic data separately (ANNs 2 and 3) improved predictive ability relative to the single combined-input ANN 1. For DMI, ANN 3 achieved a Pearson correlation of 0.85 with the observed phenotypes for the test subset but did not exceed SNPBLUP (0.87). For CH4 and the test subset, ANN 3 attained the highest Pearson correlation (0.30), exceeding SNPBLUP (0.22). These results indicate that ANN architectures with networks dedicated to a source of information may improve phenotypic prediction, particularly for CH4, although the computational cost of our implementation with the most complex model was substantial compared to SNPBLUP.