CNN-MGP: Convolutional Neural Networks for Metagenomics Gene Prediction

Abstract

Accurate gene prediction in metagenomics fragments is a computationally challenging task due to the short-read length, incomplete, and fragmented nature of the data. Most gene-prediction programs are based on extracting a large number of features and then applying statistical approaches or supervised classification approaches to predict genes. In our study, we introduce a convolutional neural network for metagenomics gene prediction (CNN-MGP) program that predicts genes in metagenomics fragments directly from raw DNA sequences, without the need for manual feature extraction and feature selection stages. CNN-MGP is able to learn the characteristics of coding and non-coding regions and distinguish coding and non-coding open reading frames (ORFs). We train 10 CNN models on 10 mutually exclusive datasets based on pre-defined GC content ranges. We extract ORFs from each fragment; then, the ORFs are encoded numerically and inputted into an appropriate CNN model based on the fragment-GC content. The output from the CNN is the probability that an ORF will encode a gene. Finally, a greedy algorithm is used to select the final gene list. Overall, CNN-MGP is effective and achieves a 91% accuracy on testing dataset. CNN-MGP shows the ability of deep learning to predict genes in metagenomics fragments, and it achieves an accuracy higher than or comparable to state-of-the-art gene-prediction programs that use pre-defined features.

Keywords: Convolutional neural network; Deep learning; Gene prediction; Metagenomics; ORF.

Fig. 1

One-hot Encoding for DNA sequence. Each nucleotide is represented as a one-hot vector: A = 1000, T = 0001, C = 0100, and G = 0010

Fig. 2

CNN-MGP Architecture. First, an ORF is encoded numerically using one-hot encoding; then, a matrix of numbers is inputted into an appropriate CNN-MGP model based on its fragment GC content. The CNN-MGP model consists of six layers. The first layer is a convolutional layer with 64 filters and a filter window size of 21. The second layer is a max-pooling layer with a pool size of 2. The third layer is a convolutional layer with 200 filters and a filter window size of 21, and the fourth layer is a max-pooling layer with a pool size of 2. Then, the output is flattened to a 1D vector before being inputted into a fully connected layer with 128 neurons. Then, the output layer produces a final gene probability