Machine learning-based leaf trait selection and PTN architecture for characterizing DFE-tolerant phenotypes in semi-commercial clones of Eucalyptus spp.

Brazilian Eucalyptus plantations are affected by a physiological disorder (EPD, Eucalypt Physiological Disorder) whose tolerance has so far been linked mainly to wood and bark traits. Whether the leaf-blade and the petiole anatomy carry tolerance signatures, and how trait coordination is organized across tolerance phenotypes, remains untested. Therefore, the hypothesis of foliar and petiole traits being able to discriminate DFE tolerance phenotypes was tested. Further, traits aiming the strongest predictors, and whether trait-coordination networks are able to reorganize along the tolerance gradient was evaluated. Multi-source measurements were integrated into a single curated matrix of 43 leaf-blade and petiole anatomical and morphological traits scored on 27 trees from nine semi-commercial clones spanning three divergent eucalyptus genotypes (EUR, EGR, GRUR), and three EPD phenotypes (tolerant, mid-tolerant, susceptible; n = 9 each). We applied PCA, MANOVA, leave-one-out linear discriminant analysis (LDA), Random Forest, and phenotype-specific Plant Trait Networks (PTNs). Multivariate trait composition differed significantly among phenotypes (P = 0.004). Random Forest classification reached 74% out-of-bag accuracy, and leaf length and leaf area (upper and mid-crown), petiole epidermis area and abaxial leaf-face thickness were the strongest discriminants, indicating that gross leaf architecture, not internal mesophyll anatomy, carries the dominant tolerance signal (LDA cross-validated accuracy, 59%). The tolerant phenotype formed a compact, tightly connected network (clustering coefficient = 0.92, diameter = 2, edge density = 0.14), whereas the susceptible phenotype presented a scattered network (edge density = 0.08) with longer path lengths and higher modularity (0.80), suggesting fragmentation into weakly connected trait modules. Machine learning identified gross leaf shape and size as the strongest and lowest-cost markers of DFE tolerance, complementing the known wood-based markers. Because each network was built from only nine plants, and a different set of connected traits, these links are best read as descriptive coordination patterns specific to this dataset, rather than causal or regulatory relationships.