Predicting Enzyme Commission Numbers Using Recurrent Neural Networks with Amino Acid Sequence Shift and Consistency Loss

Abstract

Countless species of plants have been employed for the development of medications for diseases in all fields, ranging from viruses, cancer, diarrhea, and various skin diseases. However, the percentage of plant-based new molecular entity products compared to all types of medications has dropped by more than 50% since 1950, a major reason being the lack of information scientists and industries have for countless plant species. The majority of botanical medicines have been created off of plants that have already been known for their medicinal properties for centuries and have been utilized by specific cultures to treat illnesses. Therefore, the process of developing a plant-based medicine without much information about the plant can be incredibly expensive and time-consuming, leading to many industries avoiding doing so. Ergo, I proposed a method to predict the Enzyme Commission (EC) number of plant genomes, allowing scientists to gain an understanding of the functionality of the genes in various plant species. The amino acid sequences will be represented on a 2D matrix, and I proposed a method to shift the matrix multiple times followed by stacking these respective matrices so the convolutional neural network can analyze a larger part of the amino acid sequence. As the model will be predicting a sequence of numbers, I used a recurrent neural network to preserve the context in the sequence. Moreover, I proposed a method of adding a classifier for the prediction of the first digit of the EC number to decrease the sequential error.