Multi-environment Genomic Prediction of Plant Traits Using Deep Learners With Dense Architecture

Abstract

Genomic selection is revolutionizing plant breeding and therefore methods that improve prediction accuracy are useful. For this reason, active research is being conducted to build and test methods from other areas and adapt them to the context of genomic selection. In this paper we explore the novel deep learning (DL) methodology in the context of genomic selection. We compared DL methods with densely connected network architecture to one of the most often used genome-enabled prediction models: Genomic Best Linear Unbiased Prediction (GBLUP). We used nine published real genomic data sets to compare a fraction of all possible deep learning models to obtain a “meta picture” of the performance of DL methods with densely connected network architecture. In general, the best predictions were obtained with the GBLUP model when genotype×environment interaction (G×E) was taken into account (8 out of 9 data sets); when the interactions were ignored, the DL method was better than the GBLUP in terms of prediction accuracy in 6 out of the 9 data sets. For this reason, we believe that DL should be added to the data science toolkit of scientists working on animal and plant breeding. This study corroborates the view that there are no universally best prediction machines.

Exploring Deep Learning for Complex Trait Genomic Prediction in Polyploid Outcrossing Species

“…pattern recognition (Drayer and Brox, 2014;Liang and Hu, 2015;Işin et al, 2016;Badrinarayanan et al, 2017) and natural language processing (NLP) (Deng and Liu, 2018). The DL implementation in regression tasks is less abundant and the benefit of using these methods remains uncertain (Bellot et al, 2018;Montesinos-López et al, 2018a;Azodi et al, 2019). Most GP problems fall into the regression task due to the complex nature of quantitative traits (MacKay, 2009).…”

“…DL is relatively straightforward to implement (https://keras.io/whyuse-keras/), but optimum performance depends on an adequate hyperparameter choice, which is not trivial and requires considerable computational resources (Young et al, 2015;Chan et al, 2018). Although previous, limited evidence does not show a consistent advantage of DL over penalized linear methods for genomic prediction (GP) purposes (González-Recio et al, 2014;Ma et al, 2017;Bellot et al, 2018;Montesinos-López et al, 2018a;Montesinos-López et al, 2018b;Montesinos-López et al, 2019a), more efforts are needed to fully understand the behavior and potential constraints and capabilities of DL in GP scenarios.…”

A multiple-trait Bayesian Lasso for genome-enabled analysis and prediction of complex traits

“…It is risky to make sweeping statements arguing in favor of a specific treatment of data as outcomes are heavily dependent on the biological architecture of the traits considered, and on the data structure as well. The picture emerging from two decades of experience in genome-enabled prediction in the fields of animal and plant breeding is that is largely futile to categorize methods in terms of expected predictive performance using broad criteria, in view of the large variability of performance with respect to data structure for any given prediction machine (Morota and Gianola 2014; Gianola and Rosa 2015; Momen et al 2018; Montesinos-López et al 2019 a,b,c,d; Azodi et al 2019).…”

“…However, it was found that MBL was better than MT Bayesian BLUP for the two pine tree traits. After almost two decades of genome-enabled prediction it is now clear that no universally best prediction machine exists (Gianola et al 2011; Heslot 2012; de los Campos et al 2013; Momen et al 2018; Bellot et al 2018; Montesinos-López et al 2018a, b, c, d) even when non-parametric or deep learning techniques are brought into the comparisons.…”

CGIAR modeling approaches for resource‐constrained scenarios: I. Accelerating crop breeding for a changing climate

“…In addition, it increases the precision of genetic correlation parameter estimates between traits, which helps crop improvement teams perform indirect selection. Multivariate models have been implemented using Bayesian analysis (Montesinos-López et al, 2016b) as well as deep machine learning regression ,2018c. Notably, report that the performance of multi-trait and multi-environment deep learning (MTDL) is commensurate with that of the Bayesian multi-trait and multi-environment approach.…”