Deep Learning the Protein Function in Protein Interaction Networks

One of the essential challenges in proteomics is the computational function prediction. In Protein Interaction Networks (PINs) this problem is one of proper labeling of corresponding nodes. In this paper a novel three-step approach for supervised protein function learning in PINs is proposed. The first step derives continuous vector representation for the PIN nodes using semi-supervised learning. The vectors are constructed so that they maximize the likelihood of preservation of the graph topology locally and globally. The next step is to binarize the PIN graph nodes (proteins) i.e. for each protein function derived from Gene Ontology (GO) determine the positive and negative set of nodes. The challenge of determining the negative node sets is solved by random walking the GO acyclic graph weighted by a semantic similarity metric. A simple deep learning six-layer model is built for the protein function learning as the final step. Experiments are performed using a highly reliable human protein interaction network. Results indicate that the proposed approach can be very successful in determining protein function since the Area Under the Curve values are high (>0.79) even though the experimental setup is very simple, and its performance is comparable with state-of-the-art competing methods.