Graph convolutional network based virus-human protein-protein interaction prediction for novel viruses.

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning

Background:

• Accurate identification of human-virus protein-protein interactions (PHIs) is essential for understanding viral pathogenesis.
• Experimental methods for PHI detection are costly and time-consuming, necessitating computational approaches.
• Leveraging protein network topology can enhance the performance of predictive models.

Purpose of the Study:

• To develop and evaluate a novel machine learning pipeline for predicting human-virus protein-protein interactions.
• To generate hybrid protein embeddings that incorporate both sequence and network topological features.
• To improve the accuracy and efficiency of computational PHI prediction.

Main Methods:

• A three-stage machine learning pipeline was developed.
• Numerical features were extracted from amino acid sequences using Doc2Vec and Byte Pair Encoding.
• A modified GraphSAGE model was trained using amino acid embeddings, followed by a binary classifier using hybrid embeddings.

Main Results:

• The proposed method demonstrated superior performance compared to state-of-the-art techniques.
• Experimental results showed a 3-23% improvement in Area Under Curve (AUC) scores on a benchmark dataset.
• The hybrid embeddings effectively captured relevant protein features for interaction prediction.

Conclusions:

• The developed machine learning pipeline offers an efficient and accurate approach for PHI prediction.
• The integration of sequence-derived and topological features in hybrid embeddings is key to the model's success.
• This method advances computational strategies for studying viral infection mechanisms.