Abstract
Recent advances in AI for science have highlighted the power of contrastivelearning in bridging heterogeneous biological data modalities. Building on thisparadigm, we propose HIPPO (HIerarchical Protein-Protein interaction predictionacross Organisms), a hierarchical contrastive framework for protein-proteininteraction(PPI) prediction, where protein sequences and their hierarchicalattributes are aligned through multi-tiered biological representation matching.The proposed approach incorporates hierarchical contrastive loss functions thatemulate the structured relationship among functional classes of proteins. Theframework adaptively incorporates domain and family knowledge through adata-driven penalty mechanism, enforcing consistency between the learnedembedding space and the intrinsic hierarchy of protein functions. Experimentson benchmark datasets demonstrate that HIPPO achieves state-of-the-artperformance, outperforming existing methods and showing robustness in low-dataregimes. Notably, the model demonstrates strong zero-shot transferability toother species without retraining, enabling reliable PPI prediction andfunctional inference even in less characterized or rare organisms whereexperimental data are limited. Further analysis reveals that hierarchicalfeature fusion is critical for capturing conserved interaction determinants,such as binding motifs and functional annotations. This work advancescross-species PPI prediction and provides a unified framework for interactionprediction in scenarios with sparse or imbalanced multi-species data.