Predicting growth, water-use efficiency and drought response through machine learning, GWAS and differential expression in Ponderosa pine

Understanding the molecular basis of phenotypic trait variation is key in improving field performance in plants. Many plants have high within seed source phenotypic variation, making trait-based inferences for performance difficult and inaccurate. Our study combined machine learning methods along with genomics and transcriptomics to understand the molecular drivers of important seedling traits in ponderosa pine. We measured height, specific leaf area, biomass related traits, d13C, d15N, percent carbon, percent nitrogen in well-watered and drought conditions using species' range-wide seed sources. Seedlings from California's seed sources were the fastest growing, while the ones from Montana and Wyoming were the slowest. Despite differences in growth, common responses to drought were seen across all regions. Needles per bundle was shown to be an extremely useful trait to screen for growth strategies of a seed source. We identified one to 36 unique genes (2-209 SNPs) per trait that provided accurate predictions for most traits (2-37% mean absolute percent error). We show that prediction accuracy is trait dependent, mostly higher for traits with high heritability and lower in traits sensitive to environmental change. Drought-stressed seed sources from contrasting elevations showed differential expression of phenylpropanoids, terpenoids and carotenoids genes. Our predictive models show promise for future studies to predict phenotypes upon germination instead of waiting several years to measure specific traits. This will allow for a faster, more accurate selection of best suited individuals and seed sources for any site, resulting in more efficient and successful outplanting.