Prediction of gene ontology by deep learning

The prediction of gene functions is a major challenge in biology and bioinformatics. Gene function prediction aims to predict a set of its associated GO terms along with the confidence of the association for a given gene. Existing databases of known gene functions are incomplete and prone to errors and experiments needed to improve these databases are costly. Conventional methods of gene function prediction are linear methods that are not effective in handling data of nonlinear structures. Recently a few studies attempted to incorporate nonlinear techniques into gene function prediction, but there still exists limitations. For example, (Khatri et al. ۲۰۰۵) used truncated singular value decomposition method (tSVD) for GO annotation prediction. In (Žitnik and Blaž. ۲۰۱۵), a method based on matrix factorization presented to predict gene functions with data fusion. In this study, to predict associated GO terms of genes, we proposed a novel method based on using deep neural networks (DNN). First, we find latent representations of genes and GO terms. Then, the similarity between latent representation vectors is measured by cosine function. This similarity is the probability of estimating the relation for a previously-unseen pair (Gene, GO Term).

Keywords:

Gene Function Prediction , Gene Ontology , Latent representation vector , Deep neural networks , Deep learning