Predicting Protein-Protein Interaction via Convolutional Adaptive Dot Product

Abstract

We propose a deep-learning method for predicting protein-protein interaction (PPI) via an adaptive dot product in combination with dilated convolutions. Protein-protein interac- tion is a crucial biological process. Being able to predict when two proteins will interact with each other can be crucial information for developing drug targets. Adaptive convo- lutions have proven effective at determining some relationship between two inputs in the context of video interpolation, and dilated convolutions in WaveNet have proven effective at understanding sequence information in the context of audio generation. Our proposed method combines the dilations in one dimension to interpret sequence information from a WaveNet approach (though without the conditional nature) and uses it to produce an adap- tive dot product kernel while simultaneously learning an embedding of the input via an autoencoder. The dot product of the generated kernel and embedding can then be used to predict PPI. We show that this method is able to extract useful features and obtain perfor- mance very near the state-of-the-art that uses hand-engineered features that take advantage of human knowledge of proteins. We anticipate that this method will more generally be an efficient approach to any problem requiring sequence comparison.