Protein-Protein Interaction (PPI) prediction plays a critical role in biological processes, driving a rising interest in identifying both the existence and types of PPIs. With advancements in computational techniques and the growing availability of protein data, computational PPI prediction methods have emerged as a new paradigm. However, most existing methods fail to fully exploit the multi-modal protein information, such as functional descriptive texts, and overlook the mutual enhancement among different modalities. To address these challenges, we propose a Cross-modal HIerarchical contrastive learning fraMEwork for Protein-Protein Interaction prediction (CHIME-PPI). Our approach employs an internal-to-external perspective to extract protein features from both intra-protein and inter-protein views. Protein functional descriptions are integrated as a novel modality to enrich the semantic representation space. To strengthen cross-modal alignment, we introduce a hierarchical contrastive learning framework across three semantic levels. Extensive experiments on performance comparisons and domain generalization tests provide empirical evidence of CHIME-PPI’s effectiveness, which opens new avenues for advancing PPI prediction research.