DePolymerase Predictor (DePP): a machine learning tool for the targeted identification of phage depolymerases

Abstract

Biofilm production plays a clinically significant role in the pathogenicity of many bacteria, limiting our ability to apply antimicrobial agents and contributing in particular to the pathogenesis of chronic infections. Bacteriophage depolymerases, leveraged by these viruses to circumvent biofilm mediated resistance, represent a potentially powerful weapon in the fight against antibiotic resistant bacteria. Such enzymes are able to degrade the extracellular matrix that is integral to the formation of all biofilms and as such would allow complementary therapies or disinfection procedures to be successfully applied. In this manuscript, we describe the development and application of a machine learning based approach towards the identification of phage depolymerases. We demonstrate that on the basis of a relatively limited number of experimentally proven enzymes and using an amino acid derived feature vector that the development of a powerful model with an accuracy on the order of 90% is possible, showing the value of such approaches in protein functional annotation and the discovery of novel therapeutic agents.

Keywords: Bacteriophage; Depolymerase; Machine-learning.

Fig. 1

Heatmap of pairwise similarity scores calculated for the training dataset. Grayscale colours correspond to percentage identity as provided in the associated legend. The negative and positive components of the dataset are highlighted with braces and associated labels. As highlighted by the scale of the legend, the global identities of the matrix are rather low, showing a high level of dissimilarity between the sequences

Fig. 2

Normalised confusion matrices summarising model performance on test data. Matrices give the proportion of depolymerase (DP) and non-depolymerase (Not DP) that are correctly identified by the model, corresponding thus to the true/false positive and true/false negative proportions. Matrices are shown for non-optimised (before hyperparameter tuning) SVM (a), optimised SVM (after hyperparameter tuning) (b), and optimised RF (c) models

Fig. 3

Top Predictions of Pseudomonas phage pf16 depolymerases. The graph highlights that probability reported by the model of the gene product being a depolymerase. Gene products are labelled accordingly. The putative depolymerase previously reported is highlighted on the graph and the modelling of this protein shown with respect to a known EPS depolymerase and endopolygalacturonase as reported in Magill et al. [29]

Fig. 4

Graphs showing ranking of depolymerases predicted by the model. Rankings performed on depolymerase predictions from genomes described by Pires et al. [11]. Rankings are coloured by depolymerase domains (a), family of the phage described (b), whether the host is Gram-positive or negative (c), and by the host genus (d)