Improved protein interaction models

Protein complexes, formed through networks of protein-protein interactions (PPIs), are vital for maintaining cellular function and adaptability. These complexes drive processes such as signal transduction, genome maintenance, and metabolic regulation. Despite significant advancements from large-scale protein interaction studies, a comprehensive mapping of the human interactome remains out of reach, limited by gaps in proteome coverage and systemic error resulting from modality-specific biases inherent to all experimental techniques. Such limitations impede a holistic understanding of complex protein organization, dynamics, and context-specific roles within different cellular environments. To address these challenges, we developed a novel computational framework to improve the prediction of PPIs and complex assembly. Our approach leverages heterogeneous data sources while addressing the critical issue of incomplete datasets, namely the absence of available features across a massive feature matrix. By partitioning data into subsets where complete features are available, individual models were trained and then integrated to generate robust interaction likelihoods. Our method significantly enhanced the accuracy and coverage of predicted protein interactions and complexes, surpassing the performance of previous computational techniques. Moreover, our approach extended its utility further by incorporating cell line-specific data, exposing dynamic differences in the protein complexes assembled across distinct cellular contexts. These findings offer critical insights into the adaptability of cellular processes and the specific configurations of protein complexes within different environments. By improving our ability to predict and model PPIs, this work contributes to a more complete understanding of the human interactome, providing a foundation for studying protein complexes' roles in normal physiology and disease. The advancements demonstrated here mark a step forward in addressing the challenges of incomplete data and fragmented interaction networks, enabling deeper insights into cellular machinery.