• We are sorry, but NCBI web applications do not support your browser and may not function properly. More information
Logo of narLink to Publisher's site
Nucleic Acids Res. Nov 2006; 34(20): 5839–5851.
Published online Oct 24, 2006. doi:  10.1093/nar/gkl732
PMCID: PMC1635311

Phage_Finder: Automated identification and classification of prophage regions in complete bacterial genome sequences


Phage_Finder, a heuristic computer program, was created to identify prophage regions in completed bacterial genomes. Using a test dataset of 42 bacterial genomes whose prophages have been manually identified, Phage_Finder found 91% of the regions, resulting in 7% false positive and 9% false negative prophages. A search of 302 complete bacterial genomes predicted 403 putative prophage regions, accounting for 2.7% of the total bacterial DNA. Analysis of the 285 putative attachment sites revealed tRNAs are targets for integration slightly more frequently (33%) than intergenic (31%) or intragenic (28%) regions, while tmRNAs were targeted in 8% of the regions. The most popular tRNA targets were Arg, Leu, Ser and Thr. Mapping of the insertion point on a consensus tRNA molecule revealed novel insertion points on the 5′ side of the D loop, the 3′ side of the anticodon loop and the anticodon. A novel method of constructing phylogenetic trees of phages and prophages was developed based on the mean of the BLAST score ratio (BSR) of the phage/prophage proteomes. This method verified many known bacteriophage groups, making this a useful tool for predicting the relationships of prophages from bacterial genomes.


Bacteriophages (phages) are viruses that infect bacteria. Phages not only play important roles in the biology of their hosts, but they also have a major influence on the ecology of the oceans by cycling limiting nutrients (1) and boosting photosynthesis (2). Temperate phages (phages that integrate into the host genome) can provide essential virulence and fitness factors, affecting metabolism, bacterial adhesion, colonization, invasion, spread, resistance to immune responses, exotoxin production, serum resistance, destruction of competing bacteria and resistance to antibiotics (3,4). These capabilities can be generated by the introduction of novel genes or by altering expression of existing genes. Phages have also contributed significantly to our understanding of many cellular processes and have been a source of countless enzymes used routinely in molecular biology and biotechnology. Furthermore, phages themselves (5) or proteins produced by them (6,7) may be used as antimicrobials to cure bacterial infections in humans and are being developed as household disinfectants.

More than 5000 phages have been classified since 1959 by electron microscopy (8), yet <3% of these have been completely sequenced and deposited in public databanks. Considering the estimate of ~1030 phages in the world (9), the diversity of phages is likely to be great which means more complete phage genome sequences will be needed to fully comprehend the true extent of genetic diversity, the capacity for genetic mobilization/exchange and the evolution of bacteriophages. Until this happens, there is an enormous, poorly explored, publicly available resource of bacteriophage genomes within the complete sequence of bacterial genomes. The phage genomes that can be recovered in this way are the double-stranded DNA tailed phages in the order Caudovirales and single-stranded DNA filamentous phages in the order Inoviridae that are known to integrate into the genome of their host.

Current bacterial genome sequencing projects poorly identify and annotate regions of bacteriophage origin and there are no standard definitions for the classification of these regions. There are multiple reasons for this. Many groups identify phage regions by manual inspection (10). Some consider any region with matches to phage sequence a prophage, while others have a more stringent definition and require complete or nearly complete sets of phage genes within the region to be considered a prophage. The annotation of the genes within these regions is even more variable because each group has different criteria for gene annotation. One of the greatest problems with the identification of prophage regions is the enormous diversity within the phage population, which can be observed in the sequence of many of the complete phage genomes where >50% of open reading frames (ORFs) have no database match. Another problem is that some regions are not complete phages, but just contain mainly tail fibers and lytic enzymes that have been hijacked by the host and used as weaponry to fight off competitors. These regions would be considered bacteriocins and have been called by many different names, depending on the species of origin (i.e. monocin for Listeria monocytogenes, pyocin for Pseudomonas aeruginosa) (11). Some investigators try to rely on altered G + C content of the region or the disruption of genes (particularly, the targeting or tRNAs) as methods of defining the boundaries of prophage regions. We have shown that prophage regions do not always have atypical G + C nucleotide composition (12), so this is not a reliable method. Phages do not always integrate into coding regions and do not exclusively use tRNAs as the target site for integration, making scanning for disrupted genes or tRNAs as a standalone method inadequate for finding prophage regions.

Phage_Finder was developed as a heuristic computer program to identify prophage regions within bacterial genomes and is freely available (http://www.tigr.org/software/ or http://phage-finder.sourceforge.net). It uses tab-delimited results from NCBI BLASTALL (13) or WU BLASTP 2.0 (http://blast.wustl.edu) (14) searches against a collection of bacteriophage sequences and results from HMMSEARCH (15) analysis of 441 phage-specific hidden Markov models (HMMs) to locate prophage regions. By using FASTA33 (16), MUMMER (17) or BLASTN (14), it can find potential attachment (att) sites of the phage region(s). Data from tRNAscan-SE (18) and Aragorn (19) are used to determine whether a tRNA or tmRNA served as the putative target for integration. Additionally, by looking for the presence or absence of specific proteins using HMMs, Phage_Finder can predict whether the region is most likely prophage and which type (Mu, P2, or retron R73), an integrated element, a plasmid, or a degenerate phage region. The pipeline was tested against a set of manually-defined prophage regions (20). Phage_Finder found 91% of these regions, resulting in 7% false positives and 9% false negatives with a test dataset using default settings (Table 2). To test the robustness of the pipeline, 302 complete bacterial genomes were processed through the pipeline. A total of 403 putative prophage regions were identified, which accounted for 1.7% of the total bacterial DNA. In addition to finding prophage regions, Phage_Finder has found integrated elements and integrated plasmids. A novel method for constructing phylogenetic trees of phages and prophages was developed based on the mean of the BLAST score ratio (BSR) of bacteriophage proteins.


System and software requirements

Phage_Finder was written in PERL (http://www.perl.org) and tested using PERL version 5.8.5+ on Linux and Mac OS X 10.3/10.4 operating systems, but should work in most Unix environments if the following helper programs are installed and functional: NCBI BLASTALL (13) or WUBLAST 2.0 (http://blast.wustl.edu) (14) for BLAST searching, HMMSEARCH (15) to find HMM matches, tRNAscan-SE (18) to find the location of tRNAs, Aragorn (19) to locate tmRNA sequences, and FASTA33 (16), MUMMER (17), or BLASTN (13) to find att sites. Phage_Finder.pl utilizes the Math::Round PERL module that was written by Geoffrey Rommel to round numbers by defined multiples and is freely available (http://search.cpan.org/~grommel/Math-Round-0.05/Round.pm).

Input requirements

If running the included Phage_Finder.sh BASH script, the input requirements are as follows: name of file containing the protein sequences of the bacterial genome to be searched in FASTA format, name of BLAST-formatted phage protein sequences, the name of the file containing the DNA sequence of the entire bacterial genome to be searched (not the coding sequences), and the Phage_Finder information file (tab-delimited: contig_ID, size_of_contig, feat_name, end5, end3, annotation) or a GenBank .ptt file.

If invoking Phage_Finder.pl directly, then the following tab-delimited files are required for full functionality: NCBI or WU BLASTP data, HMMSEARCH data, tRNAscan-SE data, Aragorn data and either a Phage_Finder information file or GenBank .ptt file. If the HMM data is not provided then, the search for att sites will not be performed. If the data from tRNAscan-SE/Aragorn is not provided, then the Phage_Finder.pl will not associate any putative target-site duplications with tRNA/tmRNA genes.

Identification of prophage regions

One of the original intentions of Phage_Finder was to have a program that can distinguish between largely intact, possibly functional prophages versus small regions or clusters of prophage remnants and other mobile elements. It takes advantage of several features of functional prophages to filter out unwanted fragmented regions. Since functional temperate phages integrate as linear molecules in a size range of 18–150 kb, good candidate prophage regions should have clusters of phage-like genes in this size range. Functional phages also have a large fraction of hypothetical or conserved hypothetical proteins. These stretches of phage-like and unknown genes are consecutive, not broken up by operons of house-keeping genes, although an occasional metabolic enzyme can be encoded on a phage. Tailed phages will have a conserved late gene operon that is responsible for packaging and head morphogenesis (2123). This conserved region includes a small and large terminase subunit to recognize pac or cos sites and to cleave phage genome concatemers for packaging of the phage genome into capsids (24); a portal protein to form a hole for passage of the phage genome during packaging and release (25); a prohead protease to generate mature capsids (23,26); and the major capsid protein that forms the bacteriophage capsid or head (23). A functional prophage region will also lack ribosomal RNA sequences.

The boundaries of functional prophages that integrated into specific locations can be determined by locating a site-specific recombinase at one of the ends of the phage region. Phages can integrate into tRNA/tmRNA genes, other conserved genes or intragenic regions. Since many phages and genetic elements tend to have an affinity for tRNA genes as the target for integration, any tRNA/tmRNA gene present within the putative phage region is checked first as a target for integration. The integrase can integrate into the anticodon-loop, the T-loop, or the 3′ end of the tRNA/tmRNA gene (27). The phage genome will contain sequence near the integrase gene that is homologous to the 3′ part of its target to avoid inactivating the gene after insertion (28). Following integration, the target gene will be a fusion with the 5′ end being of bacterial origin and the 3′ end of phage origin and the original bacterial-derived 3′ end on the other side of the inserted phage genome. By searching the other end of the putative phage region with the sequence of the tRNA/tmRNA gene, including extra sequence in case of miscalculation of the boundaries, one can identify what looks like a target-site duplication, the sequence of the replaced 3′ end of the tRNA/tmRNA gene. This homologous sequence, flanking the genome of the integrated phage genome, is referred to as the core attachment site (attcore) and the two half sites are attL (the phage-derived sequence) and attR (the original bacterial sequence). Sequences that are 3′ of the tRNA/tmRNA gene can also be part of the attcore (27).

Phage_Finder overview

The first analysis that Phage_Finder does is to count the number of valid BLASTP phage matches within a user-defined window size, sliding by a user-defined step size until the size of the genome is reached. I have determined that a window size of 10 000 bp and a step size of 5000 bp are optimal settings for defining clusters of phage hits with the least amount of noise. These setting are the default window and step size settings. The center of each prophage region is then defined as the window with the greatest number of phage database matches in a region that begins with a window having at least the user-defined number of hits per window (default is four hits per window).

If at least one region is found within the minimum number of hits per window, then the 5′ and 3′ boundaries of each region is roughly determined. Beginning with the previously defined center of each region, Phage_Finder.pl slides gene by gene toward the 5′ and 3′ ends of the region, making a decision about inclusion. The decision to include a gene within a particular phage region is made in the following order: (i) if the protein has a phage HMM hit, then include; (ii) if it has a phage database BLASTP valid match, then include; (iii) if the gene is a tRNA or tmRNA, then include; (iv) if either of the next three genes are tRNAs or tmRNAs, then include; (v) if the gene has annotation that has been observed in known phages and there are at least three valid BLASTP phage database matches in the current window, then include (this ‘ok annotation’ is described below); (vi) if the gene has annotation that has been observed in known phages and there are at least two valid BLASTP phage database matches in the next window, then include; (vii) if there are at least three valid BLASTP phage database matches in the next window, then include; lastly, (viii) if there are at least two valid BLASTP matches in the current step and the current gene is before the matching gene, then include every gene up to the gene with the database match. If a putative prophage region contained valid HMM matches to XerC/D (TIGR02224 and TIGR02225) or integron (TIGR02249) integrases, then these regions would be excluded from further consideration.

Before further defining the boundaries of each putative prophage region, the program attempts to define whether the regions are type prophage or type bacteriocin and whether regions can be further classified as Mu-like, retron phage R73-like, P2-like or P4-like. Degenerate regions are determined after analysis of putative attachment sites. A region is defined as type prophage if: (i) it has a core HMM match or (ii) it has a lytic HMM match and a tail HMM match and an integrase HMM match. If the region has a lytic HMM match and a tail HMM match, but no integrase HMM match, then the region is defined as type bacteriocin, which is analogous to the phage-like bacteriocins (pyocins) in P.aeruginosa or monocins in L.monocytogenes (11). The R-type and F-type pyocins are probably the best studied, being defined phenotypically and genetically (29). These pyocins encode headless phage tails, regulatory proteins and lysis proteins for the production and release of phage tails that are specific to closely related Pseudomonads, resulting in destruction of the target cell through membrane disruption (30). The region is sub classified as Mu-like if there are proteins within the region that match the following Mu-specific HMMs: PF02316 (Mu DNA-binding domain), PF02914 (Bacteriophage Mu transposase), PF06074 [labeled as protein of unknown function (DUF935), but matches the Mu portal, gp29] and PF07030 [phage conserved hypothetical protein (DUF1320), matching Mu gp36]. PF06074 and PF07030 were determined to be Mu-specific by searching the phage database with these models and via mapping to single-linkage clusters of the phage database searched against itself by BLASTP (data not shown). Only Mu-like phages were hit by these HMMs. A combination of specific HMM matches and region length were used to distinguish between the retron phage R73, P2 and P4. The retron phage R73 is a P4-like cryptic prophage from a clinical Escherichia coli isolate, containing a retroelement (31). Bacteriophage P4 is a satellite phage that can use the head and tail genes of coliphage P2 to package inself into infectious viral particles (32). All three of these phages can be identified as having HMM hits to PF04606 (Phage transcriptional activator, Ogr/Delta) and TIGR01613 (phage/plasmid primase, P4 family, C-terminal domain) and an integrase match. The region is classified as retron R73-like if there is a match to PF00078 [Reverse transcriptase (RNA-dependent DNA polymerase)]. If the length of the region is >25 000 bp, then it is considered P2-like and P4-like if the size is <15 000 bp.

Provided that a file containing the DNA sequence of the genome and HMM or tRNA/tmRNA data was provided and the region is not Mu-like, a search for putative phage attachment sites is conducted. The user can specify BLASTN, FASTA33 or MUMMER to do the nucleotide similarity searches. BLASTN and FASTA33 have the advantage in that imperfect direct repeats can be identified, while MUMMER only looks for exact matches. BLASTN is the default nucleotide similarity search tool because it appears to do a better job of finding more significant matches. Only the top two matches are processed.

If tRNA/tmRNA gene(s) are within a putative prophage region, the sequence of the inmost tRNA gene is used as the query to search the remaining 20% of the region plus 15 000 additional nucleotides for a similar direct repeat. If a direct repeat is identified, then Phage_Finder attempts to extend the region of homology by searching against the tRNA/tmRNA sequence plus 200 additional nucleotides at the 3′ end of the tRNA/tmRNA gene.

If no homologies are found, the program will identify the outermost integrase gene. The sequence beginning 1 nt outside the outermost end of the integrase gene and extending outside of the phage region by 400 nt is used as the query to search the remaining 20% of the region plus 15 000 additional nucleotides for a similar direct repeat. If a putative att site is identified, the coordinates are used to identify a putative target gene. Analysis of the distance between integrase and att sites of known phages was conducted to determine that 400 nt are sufficient sequence to include the att site. The coordinates of tRNA/tmRNA genes are used to determine whether a putative att site targeted a tRNA/tmRNA gene, in case the incorrect tRNA was chosen from a series of multiple tRNAs.

The final step is to format the data for printing a summary to the terminal and multiple output files to disk. If there are more contigs/assemblies to analyze, the program begins by finding clusters of phage database BLASTP matches on the next contig. This is repeated until there are no further contigs to search.


Phage_Finder begins by checking for the two required data files: NCBI or WUBLAST BLASTP tab-delimited data where the query bacterial genome protein sequences were searched against a BLAST-formatted database of bacteriophage protein sequences and the Phage_Finder information file or GenBank .ptt file (described above). If both files are given and present, then the data from the Phage_Finder information file or GenBank .ptt file is processed, otherwise, the program terminates with a help menu. A flowchart of Phage_Finder logic is presented in Figure 1. If provided, data from tRNAscan-SE and Aragorn are processed and stored for future use. Other files, that change infrequently, are stored in the Phage_Finder home directory and are hard-coded into the program.

Figure 1
Flow chart of Phage_Finder pipeline (A) and Phage_Finder.pl script (B) logic. Standard symbols for constructing flow charts were used.

A list of satisfactory gene annotations is read from a file so the program can differentiate between ‘house-keeping’ genes and those genes that have been previously associated with functional phages or can be associated with phages (i.e. hypothetical proteins, conserved proteins and regulatory proteins). This list of ‘ok annotation’ was generated by parsing the annotation associated with every phage sequence in the phage database used in BLAST searches, removing spaces and making non-redundant. This also allows those genes that are actually phage-derived, but not matching anything in the current, limited phage database, to be included in a phage region. If orthologs of every phage protein were already in the database, this step would not be necessary.

Because orthologs of some proteins that are found in the genomes of functional phages are also present in bacterial genomes in non-phage regions, matches to these proteins are not specific to phage regions and must be excluded from analysis. Examples of proteins that fall into this category are transposases (non-Mu-like), ABC transporters, ribonucleotide reductase and certain other enzymes commonly encoded in the genomes of lytic phages. The accession numbers of these proteins are stored in the phage_exclude.lst file, which is used by Phage_Finder.pl to exclude these protein database matches from analysis.

A list of core phage HMMs as well as lists of HMMs that are specific for phage lysis genes and tail proteins are read in from separate files (Table 1). Matches to these HMMs as well as the integrase HMMs (PF00239, PF00589, or PF02899) are used to characterize the putative prophage regions as either prophage or bacteriocin. Matches to other specific phage HMMs and the length of the region are used to distinguish between Mu-like regions, retron phage R73, P2 and P4.

Table 1
List of HMMs used to categorize putative prophage regions

Tab-delimited data from BLASTP searches against a database of bacteriophage protein sequences is read in and processed. Data from NCBI BLASTP option −m 8 or WUBLASTP that has been processed with the BTAB BLAST output parser (33) are acceptable formats. Only valid matches are considered for further analysis. A valid match is the top or best match whose subject accession number is not in the exclude list and has an E-value less than or equal to the specified cut-off (default is 10−6).

The results of phage-specific HMM searches are then processed. Protein sequences are searched with HMMSEARCH against 441 total models [295 glocal-mode (built with hmmls mode) and 146 fragment-mode (built with hmmfs mode)]. For an explanation of the glocal/hmmls and fragment/hmmfs alignment modes, please refer to the HMMER User's Guide (http://hmmer.wustl.edu/). Eight Pfam fragment-mode models were removed due to frequent matches to non-phage proteins (PF05442, PF07352, PF05012, PF06992, PF06806, PF07068, PF05037 and PF07230). A valid HMM match is recorded if the total score is greater than the noise cut-off (for glocal-mode) or is greater than the trusted cut-off for the fragment-mode). Unfortunately, due to a lack of sequence diversity that was included in the HMM seed for several glocal-mode model Pfams, several reasonable matches were not used by Phage_Finder.pl because the total score was between the noise and trusted cut-offs. There were five Pfam glocal-mode models (PF00589, PF02316, PF02914, PF06074 and PF07030) where the noise cut-off had to be set to a lower value within the Phage_Finder.pl program to increase the number of valid matches to these models. This was not the case for the fragment-mode models, where only scores above the trusted cut-off appear reliable.


Phage_Finder will generate at least eight different filetypes as output if a phage-like region is identified. These include the following: (i) a log file that gives useful information about how Phage_Finder processed the data and a summary of the findings; (ii) a file that can be imported into XGRAPH (http://www.xgraph.org/), which plots the number of phage matches to the database per window and step size; (iii) a tab-delimited report file that shows (coordinate incremented by the step size, # hits per window, and the feat_name or locus name of the hits); (iv) a file containing the 5′ end of each gene, tRNA or att site within each region, the name of the feature and the annotation/database match/HMM match as well as the G + C% content of each region, a best guess for the type of region and the coordinates of each region with or without att site adjustments. There are three different names for this file, depending on the size of the regions (1–10 000, 10 001–18 000 and >18 001 bp); (v) a tab-delimited file containing (contig_id, size of the genome, G + C% content of the genome, 5′ end of the phage region, 3′ end of the phage region, size of region in bp, label (small, medium, large), region type (prophage, integrated element, degenerate), sequence of attR, sequence of attL, name of integration target, G + C% of region, 5′ feat_name or locus name, 3′ feat_name or locus name, # integrase HMM hits, # core_HMM hits, # above noise core_HMM hits, # lytic gene HMM hits, # tail HMM hits, # Mu HMM hits, orientation of the prophage based on orientation of the target or the position of the integrase, the distance from att site to integrase, and the number of genes in the region; (vi) a file in FASTA format containing the DNA sequence of the phage region; (vii) a file in FASTA format containing the DNA sequence of each gene within the phage region; and (viii) a file in FASTA format containing the protein sequence of each gene within the phage region.

Calculation of distance from BSR

The BSR has been used to compare three genomes at a time (34). This approach can be expanded to compare n number of genomes by computing the average of BSRs between each genome. For the purpose of tree building, a Phylip-style distance matrix was required (35,36). BLASTP was used to identify bidirectional matching protein sequences as described previously (37). The BSR was calculated on all protein matches that met the prerequisites as described previously (37). Those proteins that failed to meet the prerequisites were given a BSR value of zero. Because the Phylip-style distance matrix uses a different numerical scale than the BSR, a simple calculation (Dab = −99.999999 × BSR +99.999999) was used to convert the BSR (1 = exact match, 0 = no match) into a phylip distance (0 = exact match, 99.999999 = no match). To ensure that Dab equals Dba for all the proteins between two genomes (a and b), the following calculation was used to compute the mean of the distance between genomes a and b: D = (Dabi+Dbai)/N, where N is the total number of proteins in the subject and query genomes a and b. This number, D, for every genome pairwise combination was formatted as a Phylip-style distance matrix file that was used as input for the Phylip NEIGHBOR program.

Sources of genome sequence data and phage database

The majority of phage sequences that were included in the phage database were obtained from NCBI's phage page (http://www.ncbi.nlm.nih.gov/genomes/static/phg.html). Additional phage sequences were obtained from GenBank and some prophages were obtained from whole genome projects at TIGR. Complete bacterial genome sequences were obtained from NCBI (http://www.ncbi.nlm.nih.gov/genomes/static/eub_g.html).


Testing the program

Phage_Finder was written for the purpose of automated identification of prophage regions in bacterial genomes. As with any new piece of software, one would like to have a measure of performance—a benchmark. Because there are no other publicly available programs that perform a similar function, we were unable to benchmark Phage_Finder against other programs, as was done with tRNA-finding programs (19,38). Instead, the number of false positive and false negative prophage regions were calculated based on specific criteria and based on comparison to a manually curated set of prophages. Perhaps the best list of manually curated prophages was compiled by Sherwood Casjens (20). Rob Edwards provided the sequences and putative att sites of these prophages in Genbank format (http://phage.sdsu.edu/~rob/phage/). Using the two resources, Phage_Finder.pl was run on 42 bacterial genomes that had manually curated prophages with putative att sites. Of the 118 manually curated prophages, Phage_Finder found 107 (91%). This translates into 9% false negatives (11/118, Table 2). If the E.coli genomes were omitted, which have tandom (piggy-back) prophages and fragmented prophage regions, Phage_Finder found 82 of 88 (93%) of the manually curated prophage regions. Of the remaining six false negative prophages not reported by Phage_Finder using the default −S (strict mode) option, five of these regions were found when omitting the strict option and one was piggy-back with another phage. The piggy-back phages are either fused into one large phage region or one of the two phages is not reported. The number of false positive prophage regions was determined from the tab-delimited output file to be 7% (11/151), using default settings (Table 2). A region was considered false positive if the region was small and fragmented (<10 kb) or consisted mainly of transposons, restriction systems, transporters, plasmid-like genes or enzymes that could be host-derived in the absence of phage-specific genes.

Table 2
Testing Phage_Finder against a known dataset

Application of the program

After having demonstrated that a tool was produced that could identify prophage regions >90% of time, it was time to put the program to use on all completed bacterial genomes. At the time of writing, 302 complete bacterial genomes were available at NCBI. About half of all bacterial genomes processed contained a putative prophage region (154 out of 302). This accounted for 2.74% of the total bacterial chromosomal DNA or ~2.6 phage per genome. A little more than half of predicted prophage regions had putative att sites (285 out of 403). Mu-like phages accounted for roughly 8% of all predicted prophage regions.

Since there were so many putative attachment sites predicted, new questions could be asked that were not previously possible due to an insufficient amount of data. Do phage prefer tRNAs, tmRNAs, other genes, or intergenic regions as targets for integration? Put another way, what is the distribution of tRNA targets? What is the average distance from integrase to att site? Are certain tRNAs targeted more frequently? Which part of the tRNA do phages target? The distribution of 285 putative attachment sites is depicted as a pie chart in Figure 2A. It appears that tRNAs are targets for integration slightly more frequently (33%) than intergenic (31%) or intragenic (28%) regions, but because these numbers are so close, they each account for roughly a third of the total att sites. tmRNAs appear to be less prevalent as targets for prophage integration, accounting for 8% of the predicted prophage regions. As improved att site prediction algorithms are developed, these numbers will likely change. The most popular tRNA targets appear to be Arg, Leu, Ser and Thr (Figure 2B). The mapping of the insertion point on a consensus tRNA molecule revealed novel insertion points not previously characterized (27). In particular, the 5′ side of the D loop, the 3′ side of the anticodon loop and the anticodon were not previously known to be targeted (Figure 2C). When the frequency distribution was plotted linearly (Figure 2D), two distinct peaks were observed, spaced ~30 bp apart (or three turns of B-form DNA). These apparent ‘hot spots’ were confirmed independently (27).

Figure 2
Predicted prophage target-site distributions. The distribution of targets where Phage_Finder found putative attachment sites (A). The genetic code table indicates the distribution of tRNA targets (B). For each codon, the number of phages from Williams, ...

With the number of known phages and predicted prophages being well over 600, a novel way to display their relationships was needed. Rohwer and Edwards (39) developed a method of constructing phage trees using multiple sequence alignments, but with so many phages, a more rapid method of determining distance was required. The BSR has been used to compare up to three genomes at a time (34) and the mean of the BSR has been used to identify top phage matches (40). By converting the mean of the BSR of a phage proteome into a distance matrix, a tree can be generated. In-house PERL scripts were created to convert the BSR into a distance matrix. NEIGHBOR (35,36) was used to convert the distance matrix into a Neighbor-Joining tree (Figure 3). A test tree produced by this pipeline was compared against previously published Campylobacter 16S rRNA and concatenated protein trees and found to produce results similar to those produced using concatenated protein sequences (37). Given this benchmark, it is reasonable to believe that this method of tree building based on the BSR is a reliable method to infer genetic or evolutionary relationships.

Figure 3
Test phylogenetic tree generated by converting the BSR of BLASTP bidirectional matches into distance (A). Whole genome BLASTP data from Fouts et al., 2005 (37) was used to compute this tree. The previously published 16S rRNA tree (B) is shown for comparison. ...

A total of 679 phages, prophages and predicted prophages were processed with the mean BSR/distance matrix/Neighbor-Joining pipeline. PHYLODRAW (41) was used to generate an unrooted radial tree (Figure 4). Though this method most likely does not reflect true evolutionary decent, due to the mosaicism of phage genomes (42,43), some useful associations can be gained. There was clustering of known phages that do not integrate into the host chromosome and are known to be very similar, like the T4 and T5 phages and the Bam35 and PRD1 groups of Tectiviruses. As expected, there are no predicted or known prophages that cluster with these lytic phages. An exception to this is the Pseudomonas putida putative prophage Φ3, which clusters with the T7-like phages (Figure 4). It was previously shown that this P.putida prophage had best BLAST matches to T7-like phages (12). There are examples of clades or clusters (gold color, Figure 4) that are made up exclusively of predicted prophages from Phage_Finder, which suggests these prophage sequences are novel, lacking close relationship to known sequenced phages. There are a number of examples where the phage clusters agree with current knowledge, while there are also examples to the contrary. Specifically, most of the Mu-like phages cluster together, except for Deinococcus radiodurans RadMu and Campylobacter jejuni CMLP1 (Figure 4). The previously published c2-like and sk1-like relationships were confirmed (43); however, the Sfi21- and Sfi11-like phages were broken into separate groups. The K1-specific podophages K1F and K1E clustered with T7 and SP6, respectively, which agrees with previous observations (44). The Lambdoid phages λ, HK97, HK022 and N15 did not all cluster together. HK97 and HK022 clustered together as predicted (45), and E.coli N15 clustered with Klebsiella oxytoca ΦKO2 and Yersinia enterocolitica PY54 as previously shown (46). On the other hand, Lambda clustered with Coliphage 21 and with E.coli and Shigella predicted prophages instead of these other Lambdoid phages.

Figure 4
Phylogenetic analysis of Phage_Finder predicted prophages, known prophages and sequenced phage genomes. The radial tree was constructed with branch length extensions. The branches were colored as follows: sequenced phage genomes (black), known prophages ...


A software package has been described that searches complete bacterial genomes for the presence of bacteriophage-like regions and generates several output files. To test the accuracy of prediction, the program was ran against a set of 42 complete bacterial genomes that had manually curated prophages with putative attachment sites. Phage_Finder found all but 11 manually identified prophage regions. The missed regions fall into one of two categories (regions that did not meet the ‘strict’ definition and regions that have more than one prophage integrated into the same attachment site in tandom—so called ‘piggy-back’ prophages). The piggy-back phage regions were either fused into one large region or truncated. It is difficult for the program to tease these regions apart because there are multiple possible attachment sites (one pair marking the boundary of the first prophage and another pair marking the boundary of the entire tandem series). Since this happens infrequently (3% of the test dataset), future updates to the program will deal with this issue. There are still many putative prophage regions that do not have core HMM matches (large terminase, portal, major capsid), which could be the reason why some of the manually identified prophages were missed under the strict mode. As more HMMs are built to these important phage proteins, the robustness of phage detection will increase. The accuracy of boundary prediction was measured by calculating the difference in ORF counts (between the known and predict prophage regions) per bacterial genome (17 ORFs/genome, Table 2). Currently, the program chooses between one of the two top matches to either a tRNA sequence or the sequence extending 400 bp outside of the integrase gene. If the query sequence is a tRNA, then the longest att site is chosen, but if the query sequence was from near the integrase gene, then a different strategy is taken. The best att site is determined as one that is either in the 3′ end of a gene or has the longest length. Future updates will likely focus on a more sophisticated strategy for attachment site prediction, since the longest att site is not necessarily the one used by the phage integrase.

Never before the creation of this tool has it been possible to analyze such a large dataset of prophage sequences (447 prophage or prophage regions encoding 6171 proteins). Initial studies of the data in this report looked at the distribution of 285 putative attachment sites (Figure 2) and the distribution of 679 phages, prophages and predicted prophages (Figure 4). From data presented in Figure 2, it was concluded that tRNAs, intergenic and intragenic regions are targeted with about the same frequency (roughly a third of the time). tmRNAs were targeted infrequently at 8% of the time. The number of putative intergenic insertions may change over time as better attachment site prediction algorithms are developed. It is also possible that the number of intragenic insertions is artificially too high since very small regions of nucleotide similarity can result in inaccurate target site prediction. The region of the tRNA gene that is targeted is very specific (the 3′ end of the gene in most instances) and can be quite large (up to 121 bp), which increases the confidence level of prediction over intragenic regions. The 5′ side of the D loop, the 3′ side of the anticodon loop and the anticodon were not previously known to be targeted by site-specific integrases (27). It remains to be determined whether these integrases actually function by inserting into the predicted locations in the tRNA genes.

A new method for constructing phage trees was developed as a way to display the output of Phage_Finder and to determine whether such an approach might be useful as a way of grouping or classifying phages, prophages and predicted prophage regions from genomic sequences. Instead of using protein alignments (39), the mean of the BSR was used to generate a PHYLIP distance matrix. Since the test tree, consisting of Campylobacter sequences rooted with Desulfovibrio vulgaris (Figure 3) was able to generate a tree that was consistent with the published concatenated protein tree (37), the results of such analysis are credible. Further validity of this method came from the clustering of many known phages. One interesting result from Figure 4 was that RadMu did not cluster with other Mu-like phages. Perhaps RadMu is defective and has changed considerably such that it doesn't match any other phage very well. Alternatively, RadMu could represent a unique clade of Mu-like phage that has no other sequenced members. Though this method is potentially very exciting because it is less computationally intensive than creating whole protein alignments and because there is no universal phylogenetic marker for studying phage evolution, caution must be exercised since many phages are mosaics of other phages (9,42,45). However, this procedure does have utility for clustering related groups of phages and for classifying uncultured phages. Furthermore, in a number of examples (Campylobacter, Listeria, Bacillus, Mycobacterium, Staphylococcus and Streptococcus phages and prophages), there is evidence to suggest coevolution of phages with host bacteria (Figure 4). Future versions might include either bonuses or penalties for synteny or the lack of synteny, which may resolve mosaic phages.

Phage_Finder was initially developed to aid in the identification of prophage regions in complete bacterial genomes and to improve annotation of these genomes by associating some level of function to the many hypothetical and conserved hypothetical proteins that are encoded in bacterial genomes. It has been very successful in these goals and has even surfaced a few surprises. For example, 10% of all proteins (20% of the hypothetical proteins) in the genome of Enterococcus faecalis V285 were within Phage_Finder-predicted putative prophage regions (D. E. Fouts, unpublished data). Furthermore, even though E.coli O157:H7 Sakai had the greatest number of predicted prophage regions (13 under strict settings), Streptococcus pyogenes MGAS315, with six predicted prophage regions, had the highest percentage (13.6%) of its genome as prophage DNA. Another surprise was how well Phage_Finder could identify genomic islands (pathogenecity islands, mec regions, integrated plasmids, which lead to the implementation of the ‘strict’ (-S) option. This raises the question of whether phages evolved from these mobile elements or whether these mobile elements evolved from phages. Phage_Finder was run on all the complete archael genomes, but found no prophage regions under any setting. It is not clear whether these archael genomes do not have integrated phages or whether the prophage regions are not detected because phages of this type are lacking from the BLAST-formated database.

The Phage_Finder pipeline will be a valuable tool for the scientific community and has been made publicly available (http://www.tigr.org/software/ or http://phage-finder.sourceforge.net). Future plans are to integrate this pipleline into the existing infrastructure at TIGR and to make the results of Phage_Finder analysis publicly available. It may also be possible to modify the program to identify integrated viruses in eukaryotic genomes, which would greatly facilitate the identification of retroviruses.


The author wishes to thank Timothy D. Read for encouragement to pursue this project, Karen E. Nelson for support, and Daniel H. Haft for construction of some core phage HMMs, the XerC/XerD HMMs (TIGR02224 and TIGR02225) and the integron-specific HMM TIGR02249. Special thanks also go to Rob Edwards for providing the sequences and putative att sites of a manually curated prophage test dataset (http://phage.sdsu.edu/~rob/phage/) and to Mihai Pop for help with PERL theory and for reviewing the manuscript. I also thank Robert T. DeBoy, Pawel Gajer, Gagan Pandya, and Pratap Venepally for critical discussions on tRNA insertion frequencies and Claudia Haywood for legal assistance with the open source license. Funding to pay the Open Access publication charges for this article was provided by NSF-EF-0412091.

Conflict of interest statement. None declared.


1. Sullivan M.B., Waterbury J.B., Chisholm S.W. Cyanophages infecting the oceanic cyanobacterium Prochlorococcus. Nature. 2003;424:1047–1051. [PubMed]
2. Mann N.H., Cook A., Millard A., Bailey S., Clokie M. Bacterial photosynthesis genes in a virus. Nature. 2003;424:741. [PubMed]
3. Wagner P.L., Waldor M.K. Bacteriophage control of bacterial virulence. Infect. Immun. 2002;70:3985–3993. [PMC free article] [PubMed]
4. Schuch R., Fischetti V.A. Detailed genomic analysis of the Wbeta and Gamma phages infecting Bacillus anthracis: implications for evolution of environmental fitness and antibiotic resistance. J. Bacteriol. 2006;188:3037–3051. [PMC free article] [PubMed]
5. Stone R. Bacteriophage therapy. Stalin's forgotten cure. Science. 2002;298:728–731. [PubMed]
6. Schuch R., Nelson D., Fischetti V.A. A bacteriolytic agent that detects and kills Bacillus anthracis. Nature. 2002;418:884–889. [PubMed]
7. Liu J., Dehbi M., Moeck G., Arhin F., Bauda P., Bergeron D., Callejo M., Ferretti V., Ha N., Kwan T., et al. Antimicrobial drug discovery through bacteriophage genomics. Nat. Biotechnol. 2004;22:185–191. [PubMed]
8. Ackermann H.W. Frequency of morphological phage descriptions in the year 2000. Brief review. Arch. Virol. 2001;146:843–857. [PubMed]
9. Brüssow H., Hendrix R.W. Phage genomics: small is beautiful. Cell. 2002;108:13–16. [PubMed]
10. Fouts D.E. In: Microbial Genomes. Fraser C.M., Read T.D., Nelson K.E., editors. Totowa, NJ: Humana Press Inc.; 2004. pp. 71–91.
11. Daw M.A., Falkiner F.R. Bacteriocins: nature, function and structure. Micron. 1996;27:467–479. [PubMed]
12. Nelson K.E., Weinel C., Paulsen I.T., Dodson R.J., Hilbert H., Martins dos Santos V.A., Fouts D.E., Gill S.R., Pop M., Holmes M., et al. Complete genome sequence and comparative analysis of the metabolically versatile Pseudomonas putida KT2440. Environ. Microbiol. 2002;4:799–808. [PubMed]
13. Altschul S.F., Madden T.L., Schaffer A.A., Zhang J., Zhang Z., Miller W., Lipman D.J. Gapped BLAST and PSI-BLAST: a new generation of protein database search programs. Nucleic Acids Res. 1997;25:3389–3402. [PMC free article] [PubMed]
14. Altschul S.F., Gish W., Miller W., Myers E.W., Lipman D.J. Basic local alignment search tool. J. Mol. Biol. 1990;215:403–410. [PubMed]
15. Eddy S.R. Profile hidden Markov models. Bioinformatics. 1998;14:755–763. [PubMed]
16. Pearson W.R. Rapid and sensitive sequence comparison with FASTP and FASTA. Methods Enzymol. 1990;183:63–98. [PubMed]
17. Delcher A.L., Kasif S., Fleischmann R.D., Peterson J., White O., Salzberg S.L. Alignment of whole genomes. Nucleic Acids Res. 1999;27:2369–2376. [PMC free article] [PubMed]
18. Lowe T.M., Eddy S.R. tRNAscan-SE: a program for improved detection of transfer RNA genes in genomic sequence. Nucleic Acids Res. 1997;25:955–964. [PMC free article] [PubMed]
19. Laslett D., Canback B. ARAGORN, a program to detect tRNA genes and tmRNA genes in nucleotide sequences. Nucleic Acids Res. 2004;32:11–16. [PMC free article] [PubMed]
20. Casjens S. Prophages and bacterial genomics: what have we learned so far? Mol. Microbiol. 2003;49:277–300. [PubMed]
21. Desiere F., Lucchini S., Brüssow H. Comparative sequence analysis of the DNA packaging, head, and tail morphogenesis modules in the temperate cos-site Streptococcus thermophilus bacteriophage Sfi21. Virology. 1999;260:244–253. [PubMed]
22. Desiere F., McShan W.M., van Sinderen D., Ferretti J.J., Brüssow H. Comparative genomics reveals close genetic relationships between phages from dairy bacteria and pathogenic Streptococci: evolutionary implications for prophage-host interactions. Virology. 2001;288:325–341. [PubMed]
23. Duda R.L., Martincic K., Hendrix R.W. Genetic basis of bacteriophage HK97 prohead assembly. J. Mol. Biol. 1995;247:636–647. [PubMed]
24. Ackermann H.W. Tailed bacteriophages: the order Caudovirales. Adv. Virus Res. 1999;51:135–201. [PubMed]
25. Casjens S., Hendrix R. In: The Bacteriophages. Calendar R., editor. New York and London, Vol. 1: Plenum Press; 1988. pp. 15–91.
26. Black L.W., Showe M.K., Steven A.C. In: Molecular Biology of Bacteriophage T4. Karam J., editor. Washington, DC: American Society for Microbiology Press; 1994. pp. 518–558.
27. Williams K.P. Integration sites for genetic elements in prokaryotic tRNA and tmRNA genes: sublocation preference of integrase subfamilies. Nucleic Acids Res. 2002;30:866–875. [PMC free article] [PubMed]
28. Zhao S., Williams K.P. Integrative genetic element that reverses the usual target gene orientation. J. Bacteriol. 2002;184:859–860. [PMC free article] [PubMed]
29. Nakayama K., Takashima K., Ishihara H., Shinomiya T., Kageyama M., Kanaya S., Ohnishi M., Murata T., Mori H., Hayashi T. The R-type pyocin of Pseudomonas aeruginosa is related to P2 phage, and the F-type is related to lambda phage. Mol. Microbiol. 2000;38:213–231. [PubMed]
30. Uratani Y., Hoshino T. Pyocin R1 inhibits active transport in Pseudomonas aeruginosa and depolarizes membrane potential. J. Bacteriol. 1984;157:632–636. [PMC free article] [PubMed]
31. Sun J., Inouye M., Inouye S. Association of a retroelement with a P4-like cryptic prophage (retronphage ΦR73) integrated into the selenocystyl tRNA gene of Escherichia coli. J. Bacteriol. 1991;173:4171–4181. [PMC free article] [PubMed]
32. Dehò G., Ghisotti D. In: The Bacteriophages. Calendar R., Abedon S.T., editors. New York, Vol. 1: Oxford University Press, Inc.; 2006. pp. 391–408.
33. Dubnick M. Btab—a Blast output parser. Comput. Appl. Biosci. 1992;8:601–602. [PubMed]
34. Rasko D.A., Myers G.S., Ravel J. Visualization of comparative genomic analyses by BLAST score ratio. BMC Bioinformatics. 2005;6:2. [PMC free article] [PubMed]
35. Felsenstein J. PHYLIP—Phylogeny Inference Package (Version 3.2) Cladistics. 1989;5:164–166.
36. Felsenstein J. Seattle: Distributed by the author, Department of Genome Sciences, University of Washington; 2006. PHYLIP (Phylogeny Inference Package) version 3.65.
37. Fouts D.E., Mongodin E.F., Mandrell R.E., Miller W.G., Rasko D.A., Ravel J., Brinkac L.M., DeBoy R.T., Parker C.T., Daugherty S.C., et al. Major structural differences and novel potential virulence mechanisms from the genomes of multiple Campylobacter species. PLoS Biol. 2005;3:e15. [PMC free article] [PubMed]
38. Laslett D., Canback B., Andersson S. BRUCE: a program for the detection of transfer-messenger RNA genes in nucleotide sequences. Nucleic Acids Res. 2002;30:3449–3453. [PMC free article] [PubMed]
39. Rohwer F., Edwards R. The Phage Proteomic Tree: a genome-based taxonomy for phage. J. Bacteriol. 2002;184:4529–4535. [PMC free article] [PubMed]
40. Fouts D.E., Rasko D.A., Cer R.Z., Jiang L., Fedorova N.B., Shvartsbeyn A., Vamathevan J.J., Tallon L., Althoff R., Arbogast T.S., et al. Sequencing Bacillus anthracis typing phages gamma and cherry reveals a common ancestry. J. Bacteriol. 2006;188:3402–3408. [PMC free article] [PubMed]
41. Choi J.H., Jung H.Y., Kim H.S., Cho H.G. PhyloDraw: a phylogenetic tree drawing system. Bioinformatics. 2000;16:1056–1058. [PubMed]
42. Lawrence J.G., Hatfull G.F., Hendrix R.W. Imbroglios of viral taxonomy: genetic exchange and failings of phenetic approaches. J. Bacteriol. 2002;184:4891–4905. [PMC free article] [PubMed]
43. Proux C., van Sinderen D., Suarez J., Garcia P., Ladero V., Fitzgerald G.F., Desiere F., Brussow H. The dilemma of phage taxonomy illustrated by comparative genomics of Sfi21-like Siphoviridae in lactic acid bacteria. J. Bacteriol. 2002;184:6026–6036. [PMC free article] [PubMed]
44. Stummeyer K., Schwarzer D., Claus H., Vogel U., Gerardy-Schahn R., Muhlenhoff M. Evolution of bacteriophages infecting encapsulated bacteria: lessons from Escherichia coli K1-specific phages. Mol. Microbiol. 2006;60:1123–1135. [PubMed]
45. Juhala R.J., Ford M.E., Duda R.L., Youlton A., Hatfull G.F., Hendrix R.W. Genomic sequences of bacteriophages HK97 and HK022: pervasive genetic mosaicism in the lambdoid bacteriophages. J. Mol. Biol. 2000;299:27–51. [PubMed]
46. Casjens S.R., Gilcrease E.B., Huang W.M., Bunny K.L., Pedulla M.L., Ford M.E., Houtz J.M., Hatfull G.F., Hendrix R.W. The pKO2 linear plasmid prophage of Klebsiella oxytoca. J. Bacteriol. 2004;186:1818–1832. [PMC free article] [PubMed]

Articles from Nucleic Acids Research are provided here courtesy of Oxford University Press


Related citations in PubMed

See reviews...See all...


Recent Activity

Your browsing activity is empty.

Activity recording is turned off.

Turn recording back on

See more...