Without traditional cultures, metagenomics studies the microorganisms sampled from the environment. In those studies, the binning step results serve as an input for the next step of metagenomic projects such as assembly and annotation. The main challenging issue of this process is due to the lack of explicit features of metagenomic reads, especially in the case of short-read datasets. There are two approaches, namely, supervised and unsupervised learning. Unfortunately, only about 1% of microorganisms in nature is annotated. That can cause problems for supervised approaches when an under-study dataset contains unknown species. It is well-known that the main challenging issue of this process is due to the lack of explicit features of metagenomic reads, especially in the case of short-read datasets. Previous studies usually assumed that reads in a taxonomic label have similar k-mer distributions. Our new method is to use Natural Language Processing (NLP) techniques in generating feature vectors. Additionally, the paper presents a comprehensive unsupervised framework in order to apply different embeddings categorized as notable NLP techniques in topic modeling and sentence embedding. The experimental results present our proposed approach's comparative performance with other previous studies on simulated datasets, showing the feasibility of applying NLP for metagenomic binning. The program can be found at https://github.com/vandinhvyphuong/NLPBimeta.
PR domain-containing 9 (PRDM9) is a zinc-finger protein that binds to specific DNA motifs and induces the crossing-over between chromosomes, resulting in a high recombination rate around binding sites. Currently, the binding sites of PRDM9 are predicted with methods based on motif matching and Position-specific Weight Matrix (PWM). Meanwhile, the Convolutional Neural Network (CNN) has shown superior performance in recent studies to identify protein-binding regions in general, and it is expected to perform well in PRDM9 binding site prediction. In this study, we compared the performance of PWM and CNN for predicting PRDM9 binding sites with not only test data but also the correlation between the prediction score for a fragment and the local recombination rate to evaluate the performance without overfitting effects. Approximately 170,000 genomic DNA fragments of the human genome containing the Chromatin Immuno-Precipitation with high-throughput sequencing (ChIP-seq) peak of PRDM9 were used for constructing PWM and CNN. We found that CNN outperformed PWM in terms of area under the ROC curve and other metrics. Furthermore, the prediction scores of CNN correlated more strongly with the local recombination rate than PWM. We discuss that the superior performance of CNN would be in part due to the ability of CNN to capture the feature of surrounding sequences of actual PRDM9-binding sites.