Understanding the protein-protein interaction (PPI) network in Arabidopsis thaliana is critical for elucidating the molecular basis of plant growth, development, and stress response. Protein sequence-based PPI study is advantageous over experimental techniques like yeast2hybrid, mass spectrometry and protein structure-based docking study. In this study, a novel architecture of sequence-based PPI prediction, ARACoFusion has been proposed using a fusion-based deep learning technique tailored for Arabidopsis thaliana. The proposed framework integrates contextual embeddings from a protein language model with a reciprocal cross-attention encoder, interaction projection, and multi-source feature fusion for robust classification. The model incorporated focal loss with label smoothing, variance-based uncertainty regularization, and probability calibration via temperature scaling to address extreme class imbalance. The model was trained on gold standard Arabidopsis PPIs and validated using stratified splits and a true network dataset. The performance of the model outperforms existing plant-specific models across multiple metrics, achieving high precision-recall area (AUPRC), balanced accuracy, and Matthew’s correlation coefficient (MCC). Cross-species generalizability was demonstrated using an independent PPI dataset. The prediction separability and confidence estimation have been evaluated using t-SNE visualization and calibration plots. ARACoFusion supports network-level inference and is available as an accessible web tool (https://ARAcofusion.compbiosysnbu.in/), facilitating a large-scale plant PPI study.
