Using the Multi-instance Learning Method to Predict Protein-Protein Interactions with Domain Information

Identifying protein-protein interactions (PPIs) can help us to know the protein function and is critical for understanding the mechanisms of proteome. Recently, lots of computational methods such as the domain-based approach have been developed for predicting the protein-protein interactions. The conventional domain-based methods usually need to infer the interacting domain pairs from already known interacting sets of proteins, and then to predict the PPIs. However, it is difficult to provide the detailed information that which of the domain pairs will actually interact for the PPIs prediction. Therefore, it is of great importance to develop a new computational model which can ignore the information whether a domain pair is interacting or not. In this paper, we propose a novel method using multi-instance learning (MIL) for predicting protein-protein interactions based on the domain information. Firstly, the domain pairs of two proteins were composed. Then, we use the amino acid composition feature encoding method to encode the domain pairs. Finally, two multi-instance learning methods were used for training the data. The experiment results demonstrate that the proposed method is effective.