This document describes the process by which operons are predicted for genes within the BioHealthBase database.

Transcription

1 1. Purpose This document describes the process by which operons are predicted for genes within the BioHealthBase database. 2. Methods Description An operon is a coexpressed set of genes, transcribed onto a single mrna. Operon finding software (OFS) can identify operons using only the genetic code. This is an important goal because gene coexpression controls intracellular enzyme concentrations, influences the operation of metabolic pathways, impacts drug discovery, and is essential in understanding microarray and other gene regulation data associated both with the target micro-organism and the host it infects. Operons are predicted for the BHB prokaryotes: Francisella and Mycobacterium. Operons in BHB are generated from three sources. Two in-house software methods are used predictively: the SRI Pathway Tools operon finding algorithm ( Ptools-OFS ) and the Washington University-St Louis operon finding software package ( WUStL-OFS ). These computational methods first gather information about genetic open reading frames (ORFs), determine the genes directionality (plus or minus strand), calculates the base pair distance between flanking open reading frames, detects the presence of those genes in the same metabolic pathway (PTools only), and determine the presence of common GeneOntology (GO) terms within the two protein Genbank records (WUStL only). The data is then compiled and used to make a prediction about which genes are present in operon units. A third source of data is the operon prediction set from TBDB.org. These operons are for Mycobacterium tuberculosis H37Rv only. TBDB shares this data, and to make it they use a modification of the methods described in Moreno-Hagelseib et al (2003) and in Salgado (2001). The TBDB operons are predicted without using pathway data, however the method is tuned to the most common intergenic distances (IGD) of Mycobacteria and E. coli (Salgado et al), the -1 or -4bp IGD (overlap) of ATGA or TAATG/TGATG (Figure 1 Right, Salgado Figure 2). Functional annotations of E. coli genes developed by Riley (1996) were also used to tune the method so the appropriate IGD was used as cutoff. For this prediction method, known operons from E.coli and B subtilis were used as truth standards to tune the algorithm. The odds of being in an operon (log-

2 odds value) for every intergenic distance was plotted, using the within-operon (WO) and transcription-unit boundary (TUB) distances of all gene pairs. Taking the log of the fraction of WO over TUB, according to equation 1: The log-likelihood (log-odds, LogL) of a pair of neighboring genes being in the same operon as a function of distance was calculated with the formula: where N op and N nop are pairs of genes in operons and at transcriptional boundaries, where TN op and TN nop are the total number of pairs of genes in operons and at the transcription unit boundaries, respectively. The most accurate intergenic distance value for predicting the occurrence in an operon was demonstrated to be a universal constant among tested prokaryotes (logl threshold=0.05) Figure 2 Right (Salgado 5b). At this peak accuracy, over 90% of the gene pairs are present in operons, where Accuracy is the sum of true positives and true negative, divided by the sum of all predictions true, false, positive, and negative (trues over totals). Thus given the intergenic distance in base pairs, the method reads the log-odds value at that intergenic distance, from the E. coli/b. subtilis training set standard curve, using the boundary cutoff of 0.05 for WO/TUB pairs. Operation of the prediction tools Actual operation of the two prediction tools is dramatically different (below). While the WUStL tool uses precompiled protein table data files (PTT), it also uses different algorithmic weighting parameters and a novel genbank file mining step, so that the operon prediction decisions made by WUStL are sometimes different than PTools-OFS. The WUStL-OFS tool is operated as a shell script in which informant genomes are concatenated, run through formatdb; and then BLASTed against the target genomes. The algorithm next calculates orthologs, base pair distances between ORFs, directionality of ORFs, and clusters the output into pairs and operon groups. The PTools-OFS tool is run graphically and provided with genome data, generates a pathway map for the organism, finds holes in the built pathway using BLAST, compares the pathway maps of target and informant organisms, and predicts operons.

3 3. Input Data Preparation PTools-OFS and WUStL-OFS both require input files to predict operons. These data files are collected for each organism under study, in the format provided at the NCBI FTP site 1. The OFS algorithms use Genbank (*.gbk), fasta amino acid (*.faa), fasta nucleotide (*.fna), and pathway tools (*.ptt) files. a) WUStL-OFS The WUStL-OFS tool is run using only the PTT and FAA files as input. These files need to be generated for the target and informant organisms, and the shell script takes care of all downstream processing in an automated fashion. b) PTools-OFS The preparation for making PTools-OFS input files is more complicated. After the pathologic tools application is started up in its GUI, a new organism database is created as specified. A bacterial taxon number is entered using the NC**** identifier, and the phyla control is set to bacteria. Genbank (gbk) and nucleotide files (fna) are then loaded into the PTools system, and the software application is set to run overnight, thus building a pathway map. After a trial parse, a pathway hole filler is run, name matching is verified, consistency is checked, and an automated pathway genome database (PGDB) is finally built. This PDGB and the four input files above form the collection of data files needed to obtain operon predictions from PTools. The intermediate file produced in this process is as follows: Directon: (MT MT0001 MT0002 MT0003 MT0004 MT0005 MT0006 MT0007 TRNA-ILE-1 TRNA-ALA-1) Known TUs: None Predicted TU 1: (MT4041.2) MT hypothetical protein Predicted TU 2: (MT0001) MT0001 chromosomal DnaA Predicted TU 3: (MT0002 MT0003 MT0004) MT0002 DNA polymerase III, beta subunit MT0003 recf protein MT0004 conserved hypothetical protein (etc) 4. Output Data Post-Processing and Display Output of the operon predictions is a single tab-delimited file. The WUSTL output is a single file specifying the likelihood that juxtaposed, codirected genes are present on the same operon: Operon findings at probability cutoff ,MAV_0003,MAV_0004, ,MAV_0005,MAV_0006, ,MAV_0017,MAV_0019, ftp://ftp.ncbi.nlm.nih.gov/genomes/bacteria/

4 4,MAV_0020,MAV_0021, ,MAV_0021,MAV_0022, ,MAV_0029,MAV_0030,0.902 There is one line for each pair of genes belonging to the same operon. The leftmost number is the number of the operon, where all ORFs led by the same first column number are [physically present in the same operon. The 6 data rows in this example above represent 5 operons. The PTools output operon prediction report contains no scores, and is postparsed using a perl script, opr_parse_codeonly.pl. It looks like this: 1,MSMEG_0001a,MSMEG_0002 1,MSMEG_0002,MSMEG_0003 1,MSMEG_0003,MSMEG_0004 2,MSMEG_0005a,MSMEG_0006 3,MSMEG_0008a,MSMEG_0009 3,MSMEG_0009,MSMEG_0010 4,MSMEG_0011a,MSMEG_0012 Post processing of the TBDB data is a matter of loading the data into a data warehouse testbed, retrieving the insert errors, and then manually curating the data at fault. Most of the data that is not inserted into the warehouse is the result of two locus_tags or names being used for separate ORFS, pseudogenes being called differently (or not at all) between algorithms used at Genbank and TBDB, or redundant BHB table locations where it is not clear how to place the insertable data since two ORFS exist with the same name. In all cases the errors are expunged by submitting queries to the GBrowse BHB interface, to locate the correct gene name, and eliminating it if no ORF is called as a gene. As example from the June 2008 data load: TBDB Operons:~ 2525 Inserted without error: 2508 (99.3%) Load Errors: 17 (0.7%) Corrected in name: 12 (99.8%) No ORF/ pseudogene call (5 ; unloaded) Total: % In the instances where TBDB gene names or locus_tags did not match BHB, the names were compared. Where untranscribed pseudogenes, typos or obsolete names occurred, or where there was a redundant name (e.g. two 'erb1' homologs which were edited in Genbank after creation of the operon prediction, the new name was substituted to be consistent with BHB, and the entire file was reloaded to the BHB testbed without error.

5 5. References Ermolaeva, M.D., White, O., and Salzberg, S. (2001) Prediction of operons in microbial genomes. Nucleic Acid Res. 29, Karp, P.D., Paley, S., and Romero, P. (2002) The Pathway Tools software. Bioinformatics 18 Suppl 1:S Moreno-Hagelseib, G., and Collado-Vides, J. (2003) A powerful nonhomology method for the prediction of operons in prokaryotes. Bioinformatics18, S Riley, M. & Labedan, B. (1996) in: Escherichia coli and Salmonella: Cellular and Molecular Biology, eds. Neidhardt, F. N., Curtiss, R. I., Lin, E. C. C., Ingraham, J. L., Low, K. B., Magasanik, B., Resnikoff, W., Riley, M., Schaechter, M. & Umbarger, E. (Am. Soc. Microbiol., Washington, DC), pp H. Salgado, S. Gama-Castro, A. Martinez-Antonio, E. Diaz-Peredo, F. Sanchez-Solano, M. Peralta-Gil, D. Garcia-Alonso, V. Jimenez- JacinSalgado, H., Moreno-Hagelseib, G., Smith, T., and Collado-Vides, J. (2000) Operons in Escherichia coli: Genomic analyses and predictions. Nucleic Acid Res. 29,72. Westover, B.P., Buhler, J.D., Sonnenburg, J.L., Gordon, J.I. (2005) Operon prediction without a training set. Bioinformatics 21(7): 880. NCBI ftp site: URL ftp://ftp.ncbi.nlm.nih.gov/genomes/bacteria/