Automatic gene annotation with artificial intelligence: binary classification between enzymes and non-enzymes

Abstract

Genomic annotation, a pivotal step in genomics, entails uncovering functional elements such as genes and regulatory components within DNA sequences. This process is crucial for comprehending biological processes and pinpointing disease-related mutations. Integrating highthroughput DNA sequencing and computational tools has revolutionized genetic annotation, ensuring heightened accuracy through data integration. Manual genetic annotation, involving the identification and annotation of diverse genomic elements, is labor-intensive, making traditional methods challenging due to the intricate nature of genomic data, species diversity, and the continual influx of new genomic information with different annotation practices among research groups also makes the problem harder. In this work, we created a dataset of amino acid sequences with the binary class of enzymes and non-enzymes and a model for classifying enzymes and non-enzymes sequences, eliminating some of these current problems of the genomic annotation pipeline. The dataset was compiled from The UniProt Consortium, encompassing both enzyme and non-enzyme amino acid sequences represented in standard FASTA format. The core of the model architecture adapted the Transformer’s encoder segment, renowned for its ability to capture intricate dependencies within sequential data. We treated each amino acid in the sequence as an analogous token, and the adapted architecture excludes the decoder component as it is unnecessary for the problem formulation. Model size considerations are based on both computing budget and token quantity