Relationships among phenotypes, gene expression, and genomic similarity across environments.

(a) Proportion of phenotypic variance explained by genotype (G), environment (E), and genotype-by-environment interaction (G×E) for six traits: green-up (GR), emergence (EM), flowering time (FLT), tiller count (TC), panicle height (HT), and log-transformed biomass. Pairwise expression correlation (eCor) among all genotypes based on transcriptomic data from (b) TX and MI (yellow rectangles highlight subpopulation-specific genotype clusters that shift across environments) and (c) gene expression differences between TX and MI (Diff), with heatmap colors ranging from blue (low) to white (intermediate) to brown (high) indicating the degree of correlation. (d) Pearson correlation coefficients (PCCs) between genomic relationship (K) or eCor based on TX, MI, or Diff transcript levels and phenotypic values in TX or MI or the difference in values between TX and MI (Diff). Heatmap colors ranging from white (zero) to red (positive) indicate PCC. Note that K calculated was using all genotypes, and the same value was used for all PCC calculations.

Model performance across locations and traits.

(a-f) XGBoost model performance (measured as R² using a testing set that has not been subjected to the training process) for six traits using genetic variants (G), transcriptomic variants (T), and their combination (G+T). Model performance based solely on population structure (first five principal components based on SNPs) is indicated by red dotted lines. Heritability estimates (black dotted lines, 0–1 scale) are shown for reference and are not equivalent to R².

Correlation of transcript importance across MI, TX, and plasticity models for flowering time and biomass

(a-b) Feature importance scores for benchmark genes for flowering time (FLT) (a) and benchmark genes plus QTL candidate genes for biomass (b). Feature importance was assessed using SHAP values and Bayesian ridge regression (BRR) for two individual environments (TX and MI) and their difference (Diff). Heatmap intensity represents the percentile rank of feature importance within each model, with yellow indicating highest importance. (c-h) Comparison of SHAP values between the TX, MI, and Diff models for FL (c-e) and biomass (f-h). SHAP values < 10⁻⁶ were set to 10⁻⁶ for visualization. Red dashed lines indicate the 99th percentile of SHAP values. Red features represent benchmark genes (c-e) and QTL candidate genes (f-h). Correlations are shown as Spearman rank coefficients (rho).

Genotype-dependent effects of top transcriptomic predictors of trait plasticity

(a,d) SHAP values for the top 20 genes (y-axis) across genotypes (x-axis) for flowering time (a) and biomass (d) plasticity models. For visualization, SHAP values were clipped to ±0.6 for flowering time and ±0.2 for biomass (b–c, e–f) Relationship between gene expression plasticity (difference in expression level [TPM] between TX and MI; x-axis) and trait plasticity (difference in trait value between TX and MI; y-axis) for the top two predictors of FLT (b–c) and biomass (e–f). Colors indicate the SHAP values.

Gene–gene interaction profiles associated with flowering time and biomass.

(a) Distribution of SHAP interaction values for the flowering time (FLT; left), and biomass (right) models. (b) Number of gene–gene interactions per gene for FLT (left) and biomass (right). (c-d) examples of top interactions for FLT (c), and biomass (d).