Implementation of a TensorFlow Convolutional Neural Network to Discriminate Similar Gene Functions

Convolutional neural networks are used widely in natural language processing [2, 3], image recognition [4], and recommender systems [9]. And, while deep NNs have been used for protein function prediction [1, 5], we present a novel method of using a convolutional neural network to discriminate among a set of similar protein-encoding genes from amino acid sequences. Our CNN uses a method often seen in those used for image recognition, i.e. it applies a sliding window-like technique to obtain convolutions that can be pooled before fully-connected layers. However, there are no images here, only one-hot representations of amino acids that form genes with interesting functions. This method had a 97% accuracy rate on our primary validation set and compares favorably to recurrent NNs and ordinary (i.e., not convolutional or recurrent) deep NNs which we developed for comparison.