Multi-modal features-based human-herpesvirus protein-protein interaction prediction by using LightGBM

Area of Science:

• Virology
• Bioinformatics
• Computational Biology

Background:

• Identifying human-herpesvirus protein-protein interactions (PPIs) is crucial for understanding viral infection mechanisms, especially in cancer patients.
• Existing natural language processing (NLP) methods for predicting human-herpesvirus PPIs are limited, particularly in multi-modal feature fusion.
• Herpesvirus infections are common in malignant tumor patients, necessitating advanced methods for studying viral-host interactions.

Purpose of the Study:

• To develop and evaluate a multi-modal embedding feature fusion method for predicting human-herpesvirus PPIs.
• To improve the accuracy and comprehensiveness of PPI prediction by integrating sequence, network, and functional data.
• To assess the model's generalizability across different herpesvirus subtypes using transfer learning.

Main Methods:

• A LightGBM model was developed, integrating multi-modal features from human and herpesviral proteins.
• Document and graph embedding approaches were used to represent sequence, network, and functional characteristics.
• Models were trained and validated on rigorous and non-rigorous benchmarking datasets, comparing performance against individual modal features and other machine learning methods.

Main Results:

• The multi-modal embedding fusion method significantly outperformed models using individual modal features.
• The developed LightGBM model demonstrated superior performance compared to traditional feature encoding methods and state-of-the-art deep learning approaches.
• Transfer learning showed the model could reliably predict human-cytomegalovirus PPIs even without cytomegalovirus-specific data, indicating broad applicability.

Conclusions:

• Multi-modal feature fusion using LightGBM is a powerful strategy for predicting human-herpesvirus PPIs.
• The method effectively captures complex protein interaction features across diverse herpesvirus subtypes.
• This approach offers a valuable tool for studying viral infections and host-pathogen interactions in clinical settings.