Multi-phase phenotyping and machine learning predict grain quality profiles in a population of temperate

Full description

This dataset supports the manuscript “Multi-phase phenotyping and machine learning predict grain quality profiles in a population of temperate japonica rice”. The study investigated the phenotypic architecture of rice grain quality in the Australian Rice Diversity Panel, a pre-breeding panel of rice (Oryza sativa L.) accessions representing grain-quality diversity relevant to Australian temperate rice production. The dataset includes figures, supplementary tables, processed phenotypic outputs and analysis materials used to characterise physical, milling, starch-pasting, cooked-texture, biochemical and derived grain-quality traits.

The field experiment was conducted during the 2017-2018 growing season at Leeton Field Station, New South Wales, Australia, using a randomised complete block design with five replicates. A total of 300 accessions were field-evaluated, of which 278 accessions had sufficient complete laboratory data for the final multi-phase analysis. Field measurements included days to flowering and harvest weight. Laboratory phenotyping included grain morphology and chalkiness using digital image analysis, colour measurement using the CIE Lab* colour space, milling-quality traits, starch pasting properties measured using a Rapid Visco Analyser, cooked rice texture measured using a texture analyser, amylose content measured using iodine colorimetric and INQR protocols, gelatinisation temperature measured using differential scanning calorimetry, and nitrogen, protein and sulphur content measured using FT-NIR calibrated against Dumas combustion reference data.

Primary measurements were used to derive additional functional traits, including length-to-width ratio, whole grain millout, degree of milling, RVA breakdown viscosity, RVA setback viscosity, yellowness index and texture density. In total, the final analytical dataset comprised 49 grain quality traits, consisting of 44 primary traits and five derived functional traits. Phenotypic data were adjusted using linear mixed models to account for spatial field effects, laboratory day, instrument effects and processing order where relevant. Best linear unbiased predictors were extracted for downstream multivariate and machine-learning analyses.

The associated analyses included distribution assessment, correlation-network analysis, principal component analysis, hierarchical clustering, UMAP visualisation, quality ideotype classification and cross-phase machine-learning prediction. Machine-learning models included Ridge regression, LASSO, ElasticNet, Random Forest and XGBoost. Predictive modelling was conducted using raw phenotypes, phenotypic BLUPs and z-normalised phenotypic BLUPs to compare preprocessing strategies. To reduce algebraic information leakage, cross-phase prediction models restricted predictors to upstream analytical groups within the broader four-phase experimental framework. Feature selection was performed using LASSO to identify reduced predictor sets for streamlined grain-quality evaluation.

The deposited files are intended to support interpretation and reproducibility of the manuscript figures, supplementary analyses and predictive-modelling workflow. They include figure files, supplementary tables and/or analysis outputs describing trait distributions, correlation structures, principal components, quality ideotype clusters, machine-learning performance and feature importance. Where code files are included or linked, they provide the computational workflow used for data processing, statistical analysis, visualisation and machine-learning model development.

This research used plant germplasm and grain-quality phenotyping data and did not involve human participants, human biological materials, clinical data or animal experimentation. Therefore, human or animal research ethics approval was not required. Field trials and sample collection were conducted through established agricultural research activities involving the Australian Rice Diversity Panel, with support from the NSW Department of Primary Industries and Regional Development, Yanco Agricultural Institute and AgriFutures Australia. Access to restricted DArTseq genotype data is controlled by AgriFutures Australia and Rice Breeding Australia and is not publicly released in this Figshare record. Requests for restricted genotype data should be directed to the relevant data owners and are subject to their approval.