Multi-protein complexes carry out a large amount of essential cellular functions. Dynamic protein-protein interactions (PPIs) play a crucial role in regulating the assembly/disassembly of protein complexes. The fast accumulation of proteomics datasets has enabled the computational integration of large-scale protein abundance and known interaction features to construct/refine PPI networks. However, the dynamic PPI changes that underlie transitions of cellular states are often obscured in noise and susceptible to heterogeneity across samples. To address this challenge, we developed DsINE (Dynamic sample Interaction Network Examination), a deep learning framework that utilizes contrastive graph representation to calculate fine-grained molecular networks of individual samples in large biological datasets. DsINE integrates the co-expression degrees of interacting partners with various features of cellular states to determine the dynamic landscape of PPI complexes. It maximized mutual information across specific sub-graphs and homogenized background to capture graphic structures and define state-specific PPIs. In practice, DsINE could illustrate the refined landscape of PPIs in breast cancer cell lines and accurately determine the featuring molecular complexes of an individual sample. Moreover, DsINE dissected neuroblastoma subtypes' PPIs at the single-cell level, revealing dynamic alterations of molecular networks during neuroblastoma differentiation and development. Finally, we exploited DsINE to analyze extracellular vesicles (EV) and discovered distinctive PPI landscapes in EVs derived from cancer and normal tissues, which demonstrated robust diagnostic performance. In conclusion, DsINE is a powerful tool that exploits fast-growing omics datasets to delineate the PPI networks in various biological contexts.