Adapting to Complexity: Deep Learnable Architecture for Protein-protein Interaction Predictions

Protein-protein interactions play an important role in the development of new therapeutic treatments and prophylactic vaccines. For instance, the efficacy of a vaccine strongly depends to what extent an antibody may form a stable bond with an antigen. In-laboratory experiments are both time-consuming and expensive, which limits their scope to only the most relevant interactions. Computational experiments, on the other hand, have the potential to explore and screen a vast number of possibilities, thus providing experimentalists with the most promising cases. Protein-protein interactions may be learned by deep learning networks. Nonetheless, the training of these networks requires a large number of instances which may not be readily available, as in the case, of rare or new diseases. Furthermore, the learning process is made more complex by the scarcity of data about non-interacting proteins, making the network prone to overfitting. These two shortcomings are addressed in this paper. A new learnable pyramid network architecture is proposed in which the depth and the complexity of the network are directly learned from the data, which makes the network suitable for both small and large datasets. Our network outperforms state-of-the-art neural networks for protein-protein interaction predictions and is capable of adapting its architecture to datasets whose size varies from a few thousand to seventeen million. A classification accuracy of over 96% is achieved for all datasets.