Fig. 4.

Phenotype assignment and random forest analysis of spongy mesophyll organization. Image processing and lattice measurements for (a–e) honeycomblike (Illicium anisatum shown) and (f–j) nonhoneycomb (no lattice; Aesculus californica shown) species. (a) Paradermal grayscale image taken with micro‐computed tomography. Bar, 50 µm. Inset: three‐dimensional (3D) rendering of tissue showing vertical stacking of cells. (b) Binary thresholded image. (c) Tissue skeleton image. (d) Intercellular airspace (IAS) pore nearest‐neighbor image; gray cells not included to minimize edge effects. (e) IAS pore outlines. (f–j) Analogous images for a representative nonhoneycomb species showing the absence of lattice properties. Phenotype assignments are approximate and do not preclude intermediate forms. (k) Scaling relationship (solid line) between cell arm length A _L and flow rate directionality (axes are log transformed). Species with honeycomb and nonhoneycomb spongy mesophyll are shown with blue and red circles, respectively. Spinacia oleracea, which was treated as nonhoneycomb in the analysis yet had a unique phenotype, is indicated with a brown circle. The 3D renderings above the horizontal axis show spongy mesophyll (white) and IAS (gray) for (from left to right) Vitis vinifera, Ficus microcarpa, Nepenthes ventricosa, and Platycerium andinum. (l) Power law relationship (solid line) between A _L and cell packing density. Images above the horizontal axis show spongy mesophyll for (from left to right) Quercus kelloggi, Castanea dentata, F. microcarpa, Illicium anisatum, and P. andinum. Figure inset shows the log–log transformed data and linear fit (solid line). Color scheme is same as in (k). (m) Variable importance plot for random forest classification between honeycomb and nonhoneycomb phenotypes. Predictors at the top of the ranking have a higher relative importance in classification, as determined by mean decrease in accuracy when these variables are removed from the model. (n–q) Partial dependence plots showing the natural logarithm of the odds of classification over the value range for the four traits with the highest relative importance. Blue shading shows approximate values over which probability favors the honeycomb phenotype; red shading shows approximate values over which probability favors the nonhoneycomb phenotype.