Metagenomic Gene Prediction Using Bidirectional LSTM

Genomics and microbial research have been revolutionized by the beginning of low-cost, high-throughput sequencing technologies. Which, in turn, brings a large amount of genomes to public archives today. Annotation tools are essential to understanding these microorganisms. The metagenomic sequences are fragmented, which makes accurate gene prediction challenging. Most computational gene predictor models use machine learning (ML) and deep learning (DL) to predict genes in metagenomic sequences. However, to capture the sequential dependencies and contextual information within the sequences, recurrent neural networks are more popular. This study uses a bi-directional long short-term memory (LSTM) model to classify input ORF sequences into coding or non-coding classes. The proposed model is compared with other DL methods, such as convolutional neural networks (CNN) and LSTM models. It achieved an area under the curve (AUC) value of 99%, Accuracy of 95.3%, Precision of 96.53%, Recall of 94.57% and F1-score of 95.22%.