ProphET, prophage estimation tool: A stand-alone prophage sequence prediction tool with self-updating reference database

Background Prophages play a significant role in prokaryotic evolution, often altering the function of the cell that they infect via transfer of new genes e.g., virulence or antibiotic resistance factors, inactivation of existing genes or by modifying gene expression. Recently, phage therapy has gathered renewed interest as a promising alternative to control bacterial infections. Cataloging the repertoire of prophages in large collections of species’ genomes is an important initial step in understanding their evolution and potential therapeutic utility. However, current widely-used tools for identifying prophages within bacterial genome sequences are mainly web-based, can have long response times, and do not scale to keep pace with the many thousands of genomes currently being sequenced routinely. Methodology In this work, we present ProphET, an easy to install prophage predictor to be used in Linux operation system, without the constraints associated with a web-based tool. ProphET predictions rely on similarity searches against a database of prophage genes, taking as input a bacterial genome sequence in FASTA format and its corresponding gene annotation in GFF. ProphET identifies prophages in three steps: similarity search, calculation of the density of prophage genes, and edge refinement. ProphET performance was evaluated and compared with other phage predictors based on a set of 54 bacterial genomes containing 267 manually annotated prophages. Findings and conclusions ProphET identifies prophages in bacterial genomes with high precision and offers a fast, highly scalable alternative to widely-used web-based applications for prophage detection.


Methodology
In this work, we present ProphET, an easy to install prophage predictor to be used in Linux operation system, without the constraints associated with a web-based tool. ProphET predictions rely on similarity searches against a database of prophage genes, taking as input a bacterial genome sequence in FASTA format and its corresponding gene annotation in GFF. ProphET identifies prophages in three steps: similarity search, calculation of the density of prophage genes, and edge refinement. ProphET performance was evaluated and compared with other phage predictors based on a set of 54 bacterial genomes containing 267 manually annotated prophages.

Findings and conclusions
ProphET identifies prophages in bacterial genomes with high precision and offers a fast, highly scalable alternative to widely-used web-based applications for prophage detection. PLOS

Introduction
Prophages play an important role in the evolution of bacterial genomes and their pathogenicity [1][2][3]. Sequence alterations resulting from a prophage integration event can change or knock out gene functions, as well as alter gene expression [1][2][3]. In addition, they can contribute to pathogenicity by transferring virulence and antibiotic resistance factors into a genome [4][5][6][7][8].
Because of the development of resistance to conventional antibiotic therapeutics, the use of phage therapy as a tool to combat bacterial infections has gathered renewed interest from the scientific community and public health decision makers as a promising alternative to antibiotics [9]. Cataloging the repertoire of prophages and their integration sites is an important initial step in understanding their clinical implications, either as a contributor to bacterial virulence, or as a therapeutic tool. While a handful of prophage identification tools have been designed [4,10,11], successful prediction of prophages from bacterial genomes is hampered by the extraordinary diversity, low similarity and extensive genome size variation (which can range from 3 to 497 kb) across phage families [8,12]. Among the few publicly available prophage prediction tools, Prophinder [13], PHAST [14], PHASTER [15] and PhySpy [16] are the most well-known. Prophinder, PHAST and PHASTER are only available as web applications, which limits their throughput, and routinely burdens users with long response times during periods of high demand. This limits applicability for analysis of large datasets consisting of thousands of bacterial genomes, which are increasingly commonplace. Thus, we sought to design a stand-alone prophage prediction tool that does not have the constraints associated with a webbased tool. Herein, we present ProphET (Prophage Estimation Tool) a fast, easy to use, scalable stand-alone application that identifies prophages in bacterial genomes. In benchmarking, ProphET performed comparably or with higher precision than Prophinder, PHAST, PHA-STER and PhySpy, and allowed prophage predictions in a shorter or comparable time compared to currently used tools.

ProphET repository
ProphET is a Linux stand-alone application that can be freely downloaded at https://github. com/jaumlrc/ProphET. Instructions regarding installation, input files formats, tutorials on how to run ProphET and a test dataset can be found at this link.

Phage sequence database
ProphET predictions rely on similarity searches against a database of phage genes, which is generated automatically during ProphET's installation. The ProphET database is built from the predicted proteome sequences of phage genomes available in GenBank, belonging to 18 families listed in Krupovic 2011 [17]. We excluded protein sequences annotated as ABC transporters, as these are also highly conserved in non-prophage containing regions of both prokaryotes and eukaryotes genomes, and thus could lead to false-positive phage predictions [18]. The current version contains 142,575 protein sequences representing the proteome of 1,435 phages (from bacterial and archaeal, DNA and RNA bacteriophages, with linear, circular or segmented genomes). However, as novel bacteriophage genomic sequences are deposited in GenBank, the database size will increase. The database can be manually updated by executing the update command from ProphET's home directory. A file reporting the database download date is copied to the results directory of every ProphET execution, and all previous database instances, including the download date and the number of coding sequences per phage family are stored, allowing for reproducibility of results.

Input files
ProphET takes two input files: a bacterial genome sequence in FASTA format and its corresponding gene annotation in General Feature Format (GFF). The GFF file must be formatted as specified by the Sequence Ontology Consortium (https://github.com/The-Sequence-Ontology/Specifications/blob/master/gff3.md). However, ProphET also provides a GFF converter, as part of GFFLib (package in ProphET setup), allowing the user to convert other GFFs to a format that can be used by ProphET. Instructions on how to run the converter are provided in ProphET GitHub (https://github.com/jaumlrc/ProphET).

Phage prediction
ProphET identifies prophages based on the density of phage-like genes within the bacterial genome, a process that can be divided into three steps (Fig 1).
1. Similarity search. The coding sequence of each gene is extracted from the bacterial genomic sequence according to the coordinates in the user provided GFF file, and then translated into its corresponding protein sequence. Each protein sequence is then searched against the ProphET reference database to identify phage-like matches (e-value < 10 −5 ).
2. Prophage density calculation. The number of phage-like genes is calculated within a sliding window of length 10 kb, moving in increments of 1 kb across the genome sequence.
Instances with at least eight phage-like genes and an overall length longer than half of the sliding window length (5 kb) are selected. Overlapping windows are merged and assigned as preliminary prophage predictions. The length of the sliding window and the minimum number of phage-like genes in each window were defined according to the distribution of these traits for the bacteriophage genomes in the ProphET database: only 6.2% of phage genomes are shorter than 10 kb, and even those have more than 8 genes (S1 Fig). 3. Edge refinement. Finally, ProphET trims the borders of preliminary prophage predictions at the last phage-like gene. As tRNA genes are hotspots of prophage integration [19], ProphET searches 3 kb upstream and downstream of the putative prophage borders for tRNA genes. If a tRNA gene is found, it extends or contracts the prophage predicted border accordingly, resulting in the final prophage prediction (Fig 2).

ProphET output files
ProphET outputs i) a file listing the bacterial genome coordinates of each predicted prophage in BED format, ii) an image depicting the predicted prophage positions in the bacterial genome, and iii) FASTA files with the nucleotide sequence of each predicted prophage. ProphET also outputs a folder with intermediate files used in prophage predictions, as the BLAST results.

ProphET performance evaluation
The performance of ProphET was evaluated and compared to four other published prophage prediction tools, Prophinder, PHAST, PHASTER and PhySpy, using 54 well-studied bacterial genomes containing 267 prophages, manually annotated by Casjens [5], serving as a gold standard (S1 Table). This was the same dataset previously used as the gold standard in publications reporting on Prophinder v0.4 and PHAST. All of these programs were run with default parameters, with exception of PhySpy which was run with both a generic (default) training set as well as a taxonomically-informed choice of training set. PhySpy version 3.2 currently has 31 training sets available, comprising 30 bacterial species from 21 different families and one "generic" default training set (S2 Table). If available, a training set from the same species as the evaluated ProphET, prophage estimation tool genome was selected (this was the case for 38 of the 54 genomes in our gold standard set). Otherwise, a training set from the same family was selected (this was the case for 7 of the 54 genomes). For the 9 remaining genomes in our gold standard set, neither of these conditions was met, and the test set was evaluated using the generic test set (S3 Table).
In order to evaluate the results of our benchmarking, we considered as true positives (TP) the overall length of the overlap between prophage predictions and the gold standard; false positives (FP) as the overall length of predictions that did not overlap with the gold standard; and false negatives (FN) as the overall length of regions of the gold standard that were not included in the predictions (Fig 3). The sensitivity (TP/TP+FN) and positive predictive value (PPV = TP/TP+FP) were computed and compared to those obtained by Prophinder, PHAST, PHASTER and PhySpy. To compare scalability, we benchmarked run times for the five prophage prediction tools using the same set of 54 genomes, measuring the time elapsed between the submission of the first genome and the receipt of results for all prophage predictions.

Results and discussion
In order to efficiently predict phages from large datasets of bacterial genomes, we designed ProphET, a fast, easy to use, and highly scalable tool that can accurately identify bacteriophages in complete and draft genome sequences. ProphET searches the predicted proteome from an input genome sequence against a database of phage genes automatically downloaded from NCBI-GenBank (see Materials and Methods). ProphET requires that the user provide precomputed gene annotations as input, permitting the user to use their preferred annotation tool as well as to remove known genes that could lead to incorrect predictions of prophages, if applicable. Since prophage predictions are based on the density of phage genes, fragmented assemblies and partial gene annotation files could result in missed prophage predictions. We recommend the use of complete genomes and GFF annotation files rather than draft assemblies and partial gene annotation files; however, ProphET can identify prophages wholly contained within a contig from a fragmented assembly.
We compared the performance of ProphET to four phage prediction tools, Prophinder v0.4 [13], PHAST [14], PHASTER [15], and PhySpy [16] (see Materials and Methods). ProphET had the highest overall performance among the four programs ( Table 1). The slightly higher sensitivity obtained by PHAST, PHASTER and PhySpy (77.7% for PHAST, 78.1% for PHA-STER and 79% for PhySpy, versus 73.3% for ProphET) came at the cost of a loss in PPV (69.2% for PHAST, 74.6% for PHASTER and 55.8% for PhySpy, versus 84.2% for ProphET). The higher PPV obtained by ProphET is a consequence of its lower false positive rate when compared to the other predictors. The majority of the prophages missed by ProphET (false negatives) were also not detected by the other predictors. These mainly consisted of short prophages without phage-like genes, according to our reference database. This limitation may be addressed as more phages are annotated and included in reference databases.
Finally, to compare scalability, we benchmarked run times for Prophinder, PHASTER, Phy-Spy and ProphET using the same set of 54 genomes. As PHAST is an older version of PHA-STER with lower sensitivity and specificity, it was not evaluated. We measured the time elapsed between the submission of the first genome and the receipt of results for all prophage   ) for Prophinder, PHASTER, ProphET, PhySpy (generic training set) and PhySpy � (taxonomically optimized training set) to perform prophage identification for our test set of 54 well-studied bacterial genomes (S1 Table). Three separate evaluations were performed on different days. predictions. In our comparison, ProphET was four times faster than PHASTER, over 15 times faster than Prophinder, and comparable in speed to PhySpy (Fig 4, S4 Table). The benchmarking of the standalone applications ProphET and PhySpy was performed using 50 processors; however, their run times could be further reduced if more powerful servers or cloud computing were used. Using a single processor to analyze the 2 MB genomic sequence of Lactobacillus johnsonii NCC 533 (NC_005362.1), ProphET took 12 minutes, with a peak memory usage of 70 MB of RAM on a desktop computer with 1.4 GHz Intel Core i5 processor and 8 GB 1600 MHz DDR3 RAM. In addition, in our benchmarking ProphET ran in an automated fashion without user intervention, whereas we had to manually re-submit around 10% of the genomes to PHASTER due to dropped server connections.

Conclusion
ProphET is a fast and scalable, easy to install and to use program to identify prophage sequences in bacterial genomes. In our benchmarking, ProphET ran faster or with comparable speed with current used prophage predictors, with comparable sensitivity and substantially better precision. Unlike the current most widely-used prophage predictors which are webbased, ProphET runs as a stand-alone application and scales well with large datasets, a very desirable characteristic as projects requiring the sequence and annotation of thousands of bacterial genomes are becoming routine. Therefore, we believe that ProphET will be a great asset for the large-scale identification of prophage within rapidly expanding genomic datasets.   Table. Run-time evaluation of prophage prediction programs. The total elapsed time to perform prophage predictions in the 54 bacterial genomes was measured on three different days, represented by the "Round" columns. PhySpy: Predictions using the generic training set. PhySpy � : Predictions using a taxonomically optimized training set (Materials and Methods; S3 Table).