Crafting negative training datasets for a high-performance classification of bacterial proteins secreted via non-classical pathways

Abstract

Understanding protein secretion pathways is paramount in studying diseases caused by bacteria and their respective treatments. Most such paths must signal ways to identify secretion. However, some proteins, known as non-classical secreted proteins, do not have signaling ways. This study aims to classify such proteins from predictive machine-learning techniques. We collected a set of physical-chemical characteristics of amino acids from the AA index site, bolding known protein motifs, like hydrophobicity. We developed a six-step method (Alignment, Preliminary classification, mean outliers, two Clustering algorithms, and Random choice) to filter data from raw genomes and compose a negative dataset in contrast to a positive dataset of 141 proteins from the literature. Using a conventional Random Forest machine-learning algorithm, we obtained an accuracy of 91% on classifying non-classical secreted proteins in a validation dataset with 14 positive and 92 negative proteins - sensitivity and specificity of 91 and 86%, respectively, performance compared to state of the art for non-classical secretion classification. However, this works novelty resides in the fastness of executing non-CSP classification: instead of dozens of seconds to just one second considering a few dozen protein samples or only ten seconds to classify one hundred thousand proteins. Such fastness is more suitable for pan-genomic analyses than current methods without losing accuracy. Therefore, this research has shown that selecting an appropriate descriptors set and an expressive training dataset compensates for not using an advanced machine learning algorithm for the secretion by non-classical pathways purpose. Available at https://github.com/santosardr/non-CSPs.