Computational Molecular Biology 2025, Vol.15, No.5, 218-226 http://bioscipublisher.com/index.php/cmb 22 2 superior to that of single antigens. In terms of drugs, interaction prediction can lock onto interfacial targets. For example, blocking the binding of Escherichia coli Tir to host actin can prevent infection. Meanwhile, network analysis is also used in drug combination design to guide combination medication by identifying the synergistic interaction module. Screening projects for broad-spectrum vaccines and multi-drug combinations have entered the validation stage, demonstrating the potential of machine learning prediction to move from theory to application. Figure 1 AlphaFold produces highly accurate structures (Adopted from Jumper et al., 2021) 6.3 Cross-species interaction prediction and integration with systems biology Cross-species interaction prediction enables us to systematically understand the infection process. The model has been able to predict the binding of bacterial effector proteins to host targets, explaining how pathogens evade immunity or manipulate host signals. Furthermore, machine learning has also been used to infer the interaction relationship between pathogens and symbiotic bacteria. For example, the inhibition of pathogen interaction modules by short-chain fatty acids suggests probiotic potential. After integrating multi-omics information, interaction prediction becomes more biologically significant and can reveal the dynamic changes of interaction networks under infection. Currently, graph neural networks and attention mechanisms are used to integrate multi-source data, bringing us closer to the overall map of the infection system. In the future, regulating the microbiota or multi-target intervention may become a new strategy to weaken the pathogenicity of pathogens. 7 Case Study 7.1 Dataset construction and model design Salmonella is a typical intestinal pathogen. Studying its protein interaction network helps understand the complex regulation of virulence. Here, Salmonella Typhimurium is taken as an example, using deep learning to predict its whole-genome interaction network. The data were obtained from the SalmoNet database and literature-based experimental records, with approximately 1,000 verified interactions as positive samples. For negative samples, a combination of localization differences and random pairing was adopted to select an equal number of non-interacting protein pairs from about 4,000 proteins. In terms of model design, we combined convolutional and
RkJQdWJsaXNoZXIy MjQ4ODYzNA==