PhageScanner, a flexible machine learning pipeline for automated bacteriophage genomic and metagenomic feature annotation

Abstract

Even though bacteriophages are the most plentiful organisms on Earth, many of their genomes and assemblies from metagenomic sources lack protein sequences with identified functions. Most proteins in bacteriophages are structural, known as Phage Virion Proteins (PVPs), but a considerable number remain unclassified. Complicating matters further, conventional lab-based methods for PVP identification are time-consuming and tedious. To expedite the process of identifying PVPs, machine-learning models are increasingly being employed. While existing tools have developed models for predicting PVPs from protein sequences as input, none of these efforts have built software allowing for genomic and metagenomic as input. In addition, there isnt a framework available for easily curating data and creating new types of models. In response, we introduce PhageScanner, an open-source platform that streamlines data collection, model training and testing, and includes a prediction pipeline for annotating genomic and metagenomic data. PhageScanner also features a graphical user interface (GUI) for visualizing annotations on genomic and metagenomic data. We also introduce a BLAST-based classifier that outperforms ML-based models (achieving an F1 score of 94% for multiclass PVP detection and 97% for binary PVP detection) and an efficient Long Short-Term Memory (LSTM) classifier. We showcase the capabilities of PhageScanner by predicting PVPs in six previously uncharacterized bacteriophage genomes. In addition, showing the utility of the framework, we create a new model that predicts phage-encoded toxins within bacteriophage genomes.