Transcriptome-based prediction of nitrogen use efficiency in Arabidopsis and maize - I

Inferring phenotypic outcome from genomics features is the promise and challenge for basic and applied research. Functionally testing whether the feature(s) with predictive power can shed light on the underlying mechanism remains largely unexplored. We address these gaps by applying a machine learning approach to predict phenotypes based on transcriptome data. We validate the predictive models in silico using out-of-sample varieties and in planta by reverse genetics. Our study primarily focuses on a agronomic trait nitrogen use efficiency (NUE), but the described approach can be applied to any phenotype across biology, agriculture and medicine. We measure NUE and gene expression from natural and cultivated varieties of Arabidopsis and maize, respectively, under low and high N environments. We integrate transcriptomics across a model and a crop and demonstrate that identifying the evolutionarily conserved N-responsive differentially expressed genes is an efficient technique to reduce the feature dimensionality, which ultimately improves the predictive power of the gene-to-trait models learned. Furthermore, including a model organism with a comprehensive mutant collection has enabled the efficient validation of the role candidate genes with predictive power in NUE outcomes. Using the matched transcriptomic and phenotypic data from a model and crop species, we present an analysis pipeline to infer a complex trait from transcript abundance. This analysis pipeline can be applied to any other organisms to uncover novel genes underpinning physiological traits of interest. Arabidopsis: 108 samples were analyzed, including 18 accessions in two nitrogen conditions with 3 replicates each. Maize: 91 samples were analyzed, including 16 genotypes in two nitrogen conditions.