Deep learning-assisted proteomic dissection reveals sex-biased and shared proteomic patterns in Populus deltoides...

Sexual dimorphism in dioecious tree species and proteomic responses under stress represents an underappreciated axis of stress resilience. Here, we investigated sex-biased proteomic patterns in Populus deltoides exposed to long-term waterlogging and subsequent recovery. By integrating isobaric tags for relative and absolute quantification (isobaric tags for relative and absolute quantification, iTRAQ)-based quantitative proteomics with machine learning approaches, including autoencoders, graph neural networks (GNNs) and Shapley Additive exPlanations (SHAP)-based model explainability, we explored latent structure within high-dimensional protein abundance data. Unsupervised autoencoders captured dominant stress-related variation but showed limited resolution of sex-associated differences. In contrast, GNN and GraphSAGE embeddings disentangled sex-biased proteins embedded within broader waterlogging and recovery responses, highlighting a structured and topologically coherent proteomic differentiation. SHAP analyses applied to Random Forest, and XGBoost models trained on latent features confirmed that waterlogging is the most decisive contributor to sex bias compared with recovery. Further, in males, selective retention of photosynthetic capacity was evident through the increased protein abundance of Photosystem II components (PsbH, PsbR) and light-harvesting proteins (Lhcb1, Lhcb2), coupled with targeted decreased protein abundance of cytochrome b6f and Photosystem I subunits during stress and a partial rebound post-waterlogging. In contrast, females displayed widespread suppression across the photosynthetic apparatus, suggesting a reduced capacity for functional recovery. Stress-specific modulation of phenylpropanoid and proline pathways further revealed divergent metabolic strategies. Together, our findings revealed sex-biased proteomic plasticity as a critical determinant of recovery potential in P. deltoides and established a graph-based machine-learning framework for decoding the hidden layers of functional dimorphism under abiotic stress.