Artificial Intelligence Methods for Protein Complex Identification

Protein complexes are fundamental structural elements in cells, which perform vital biological roles and are central to mechanisms such as signal transduction, metabolic control and gene expression. As high-throughput experimental technologies have developed, extensive data on protein-protein interactions have been collected and often in the form of complex networks. These datasets are, however, usually marred by problems of noise and incompleteness and these present serious problems in the identification of protein complexes. In recent years, AI techniques have provided new ways to solve this issue. With the identification of protein complexes in the spotlight, the paper will provide a systematic review of two main dimensions: First, it discusses graph embedding-based approaches with the focus on their strategies of feature representation and integration with multi-source data and clustering algorithms. Second, it is a summary of graph neural network-based approaches, and their benefits in nonlinear feature learning, as well as multi-modal information fusion. Moreover, the paper types and classifies popular datasets, and provides detailed analysis of the existing studies under the prism of data quality, multi-source fusion approaches and model complexity. Studies show that graph neural networks are better at expressing relationships of complexity; they are however, not without problems of being costly to compute, and highly dependent on data.