Transfer learning via multi-scale convolutional neural layers for human-virus protein-protein interaction prediction

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning

Background:

• Machine learning methods are crucial for predicting human-virus protein-protein interactions (PPIs).
• Transfer learning leverages existing data to improve predictions on smaller datasets.

Purpose of the Study:

• To develop an accurate computational method for predicting human-virus PPIs.
• To introduce and evaluate transfer learning strategies for enhancing PPI prediction.

Main Methods:

• Combined evolutionary sequence profile features with a Siamese convolutional neural network (CNN) and multi-layer perceptron.
• Developed and applied two transfer learning methods ('frozen' and 'fine-tuning') for domain adaptation.
• Utilized the 'frozen' transfer learning approach for human-SARS-CoV-2 PPI prediction.

Main Results:

• The proposed Siamese CNN architecture outperformed existing machine learning and state-of-the-art methods.
• Transfer learning methods reliably predicted interactions in target human-virus domains.
• Predicted human-SARS-CoV-2 PPIs showed topological and functional similarity to known interactions.

Conclusions:

• The developed method and transfer learning strategies offer a robust approach for predicting human-virus PPIs.
• This work provides valuable insights for understanding virus-host interactions and developing therapeutic strategies.