Background

Bacteria become antibiotic resistant due to the excessive use of drugs in healthcare and agriculture. In the United States, around 3-million people get infected and approximately 35,000 individuals die because of antibiotic-resistant organisms [1]. Therefore, the resistance nature of bacteria drives the need for inventing novel antimicrobial compounds to treat antibiotic-resistant patients. Researchers developed several approaches to extract natural products as antimicrobial compounds by mining the bacterial genomes [2]. Bacteriocin is one type of natural antimicrobial compound which is a bacterial ribosomal product. As bacteriocins have both broad and narrow killing spectra depending on their specific structure and mode of action, they became attractive choices in the discovery of novel drugs that can produce less resistance in bacteria [3-5]. Current whole genome sequencing technology provides many genes that encode bacteriocins and these sequences are publicly available for future research. Researchers introduced several methods to identify bacteriocins from bacterial genomes based on bacteriocin precursor genes or context genes. For example, BAGEL [6] and BACTIBASE [7] are two publicly available online tools that curate experimentally validated and annotated bacteriocins. Like the widely used protein searching tool BLASTP [8, 9], these methods also allow users to identify putative bacteriocin sequences based on the homogeneity of known bacteriocins. However, these similarity-based approaches often fail to detect useful sequences that have high dissimilarity with known bacteriocin sequences; thereby, generating an undesired number of false negatives. To resolve this problem, some prediction tools, such as BOA (Bacteriocin Operon Associator) [10], were developed based on locating conserved context genes of the bacteriocin operon, but they still rely on homology-based genome searches.

Machine learning technique can be applied as a substitute for sequence similarity and context-based methods that can utilize potential peptide (protein) features of bacteriocin and non-bacteriocin to make strong prediction in identifying novel bacteriocin sequences. Recently some machine learning-based bacteriocin prediction techniques were proposed that utilized the presence or absence of k-mer (i.e., subsequences of length k) as potential features and represented peptide sequences using word embedding [11, 12]. There are also deep learning-based methods for bacteriocin prediction, for example RMSCNN [13] used a convolutional neural network [14, 15] for identifying marine microbial bacteriocins. However, these existing approaches did not consider the primary and secondary structure information of peptides that are crucial to find highly dissimilar bacteriocins. Also, those strategies did not apply any feature evaluation algorithm to eliminate the unnecessary features that may reduce the achievement of a machine learning classifier.

You are viewing: Which Of The Following Statements About Bacteriocins Is False

Read more : Which Adventure Time Character

In this work we present a predictive pipeline for identifying bacteriocins by generating features from the physicochemical and structural characteristics of peptide sequences. We evaluated and selected subsets of the candidate features based on Pearson correlation coefficient, t − test, mean decrease Gini (MDG), and recursive feature elimination (RFE) analyses. The reduced feature sets called optimal feature sets are then used to predict bacteriocins using support vector machine (SVM) [16] and random forest (RF) [17] machine learning models. The main objective was to develop a software package called Bacteriocin Prediction Software (BaPreS) using the best machine learning model with a simple and intuitive graphical user interface (GUI) that can generate all required optimal features to get prediction results for testing protein sequences. The software provides options to users to test multiple sequences and add new training bacteriocin or non-bacteriocin sequences to the machine learning model for improving the prediction capability. BLASTP, a sequence matching tool and RMSCNN, a deep learning model were used to compare the performance of our software tool.

Source: https://t-tees.com
Category: WHICH