Probability-based Scoring Function as a Software Tool Used in the Genome-based Identification of Proteins from Spirulina Platensis

One of the major goals of proteomic research is the identification of proteins, a goal that often requires various software tools and databases. These tools have to be able to handle large amounts of data, such as those generated by PMF (Peptide Mass Fingerprinting), a high throughput technique. A newly sequenced organism, Spirulina platensis, was recently used to generate an in silico database, and thus an in-house tool designed for compatibility with this database and its inputs (PMF) was constructed in the present study. With a probability based scoring function, this tool effectively ranked ambiguous protein identification results by using five criteria: score, number of matched peptides, % coverage, pI and molecular weight. As a result, the protein identification step of Spirulina proteomic studies can be achieved precisely. Moreover, a very useful function of this tool is its capability for batch processing, in which the system can handle protein-identification searches of a hundred of proteins automatically, from a single user's input. Therefore, the tool not only gives accurate protein identification results but also saves the user time in processing a large amount of data.


INTRODUCTION
The identification of differentially expressed proteins by employing experimental techniques and tools is a goal of proteomics.After protein isolation by various techniques, including two dimensional-gel electrophoresis (2D-PAGE), a protein of interest is subjected to protein identification by using mass spectrometry techniques coupled with bioinformatics.Mass spectrometry (MS) technology is widely used to identify proteins by generating either the peptide mass fingerprinting (PMF) or the peptide fragmentation fingerprinting (PFF) of proteins of interest.Then, these PMFs and PFFs are searched against PMFs and PFFs in available databases, to identify the proteins.PMFs are analyzed by comparing an experimental mass list with theoretical mass lists in databases.This identification step requires both efficient software tools and appropriate databases to obtain reliable protein identification results.
At present, several software tools for protein identification using PMF have been constructed, including MASCOT, the MS-Fit tool, the ALDENTE tool, etc.These tools use several algorithms to calculate the scores of matched proteins, e.g. the Bayesian theory of probability-based scoring methods, the genetic algorithm method [1] and HMMs (Hidden Markov models) [2].However, the tool constructed in the present study was designed not only to serve the needs of *Address correspondence to this author at the BEC Unit, KMUTT, 83 Moo8, Thakham, Bangkhuntien, Bangkok 10150, Thailand; Tel: 662-470-7509; Fax: 662-452-3455; E-mails: apiradee@biotec.or.th, apiradee@pdti.kmutt.ac.th users but also to assist in a precise and time-saving protein identification process by accurate scoring plus ranking function, pI-filtering and PMF-batch processing.Therefore, an in silico database and an in-house software tool were constructed for S. platensis protein identification by using PMFs as inputs.In our previous study, simple ranking methods were employed, by counting the number of matched peptides and calculating the coverage percentage of the matched peptides compared to the whole proteins, in order to rank ambiguous protein identification results.However, an effective and accurate scoring method is required to get rid of ambiguous data and make protein identification more reliable.Thus, in the present study, an effective scoring method was developed using a probability-based scoring function.Moreover, the isoelectric point (pI) and molecular weight values of a protein were also used as criteria to pick out a target protein among redundant protein identification results, which may contain proteins with very close scores.For effective use of the tool, a batch processing module was also developed to handle hundreds of inputs simultaneously in a single run.

Programming Software
In this study, the PHP programming language was used to write code to calculate protein isoelectric point values and develop a probability-based scoring function, including a connection to the database.In the case of this tool, Apache 2.5.10 was employed to view all data from a database, which was installed with Navicat MySQL version 5.0.5.Moreover, phpMyAdmin version 2.10.3 was used to manage the database and perform tasks such as creating tables.To manage the web interface, the input and output interfaces of this tool were created as Graphic User Interfaces (GUIs) in the HTML language employed in the Macromedia Dreamweaver MX program.

pI Calculation Method
The pI of an individual protein was calculated by employing the Bisection method [3,4] as shown in Fig. (1).The pH values, which are related to the total charge of proteins, were divided into two sections, pH 0-7 and 8-14, in order to determine total charge.If the total charge of a protein is near zero at a certain pH value, then the pH value at this state will be the pI.However, the total charge of each protein was calculated from the summation of the charges of its constituent amino acids.These macromolecules can be categorized into two major groups: (i) the group of positively charged amino acids, which consists of histidine (H), lysine (K) and arginine (R), and (ii) the group of negatively charged amino acids, which consists of aspartate (D), glutamate (E), cysteine (C) and tyrosine (Y).In order to reduce the time required for the pI calculation, the pKs values of the seven charged amino acids [4] were collected in a MySQL database.Then, the PHP programming language was used to calculate the pI values of an individual protein from the S. platensis database.Finally, the theoretical pI was shown in the 'Hits' protein results on the GUI.

Probability-Based Scoring Function Method
The scores of peptides that matched theoretical peptides from the Spirulina-PMF database were calculated in the form of probabilities.In the first step, the total charge of each protein was obtained from the summation of the negative and positive charges of each macromolecule, as shown in equations 1 and 2 [4], where pK Ni and pK pi represent the pK of negative and positive macromolecules of size i.
Charge of negative macromolecules: Charge of positive macromolecules: In the case of the probability-based scoring function, the input peptides were matched with proteins in the database, and probabilities and scores were calculated, as shown in equation 3 and 4, respectively, where represents a constant between zero and one, m ij is the number of the matched peptide, and M j is the number of all peptides obtained from digestion of the theoretical protein.Score = log Pr(Pk) Equation ( 4) If the input-peptide does not match with a theoreticalpeptide in the database, the probability of this input-peptide is zero.Thus, the score has values between zero and one, which requires visualizing the results in the form of decimal numbers.For ease of comparison, these probabilities are converted into logarithmic form.For example, if the probabilities are 0.0001 and 0.002, these scores in logarithmic form will be -4 and -3.6989 which are simpler for users to analyze.Finally, the scores of the 'Hits' protein results are shown on the GUI.

Tool Validation
For tool validation, cross-species proteins were used to check an accuracy of the current tool, in order to compare the 'Hits' results with online PMF tools such as the Mascot tool, by using two main methods.First, cross-species proteins such as the photosystem II D1 protein found in cyanobactria (Synechocystis sp.PCC6803 (slr0752)), and the pyruvate kinase of Escherichia coli K-12 MG1655 were digested, in order to collect their PMF data using miss cleavage values of zero to three.Second, the PMF data of known proteins, obtained from the 2-DE technique in the NCBI database (http://www.ncbi.nlm.nih.gov/Ftp/) were searched for with the tool.Finally, the E. coli genome was downloaded to the database and then PMFs of the known proteins of E. coli were searched with this tool.

RESULTS & DISCUSSION
The input interface of the current version of the in-house software tool was designed as shown in Fig. (2).By using this GUI, users could fill experimental PMF data into 'Query' and 'Autosearch' sections on the input interface (Fig. 2), together with protein mass and pI obtained from 2-DE experiments [5].The source code of the tool is available at http://spirulina.biotec.or.th/~spirulina/proteome/index.htm.
Moreover, users could select a pI-database to represent the "Hits" protein results from EMBOSS, DTASelect, Solomon, Sillero, Rodwell and Wikipedia databases as shown in the dashed box in Fig. (2).In order to rank the 'Hits' protein results, five criteria were considered consecutively, (i) probability-based score, (ii) number of matched proteins, (iii) pI, (vi) protein molecular mass and (v) % coverage.On the output interface, all results were represented in the form of tables and lists of the amino acid sequences of theoretical protein 'Hits,' which were obtained from the Spirulina database under the input criteria setting (Fig. 3).Criteria for searching are presented in a table in Fig. (3), such as database name (only two versions, 677 and 847), allowed missed cleavage, ion mode (MH+ or Mr Modes), type of filtering, protein mass, protein tolerance, pI and pI tolerance.
A limitation of this tool underlying the search algorithm is the required filtering of the 'Hits' results by the protein mass and pI, because of differences in their distribution patterns.However, in both versions of the database, the same pattern of protein distributions were represented in Fig. (4a) and Fig. (4b), for database versions 677 and 847, respectively.The protein mass distribution of both versions showed that the most abundance protein masses were in the range of 5 kDa to 99 kDa.Thus, if a user selected a wide gap of protein tolerance for the experimental protein within this range, the execution time would be very long.Therefore, for the experimental proteins, which have protein masses within the range of 5 kDa to 99 kDa, a user should use a protein tolerance of 0.1 kDa when searching the 'Hits' protein results.On   the other hand, if the experimental proteins have protein masses of more than 100 kDa, a value for protein tolerance could be selected from 10 kDa to 30 kDa, resulting in a the searching time of 30 minutes.
In the case of pI distribution, the pI values of both databases differed in their pKs values from the pKs database, as shown in Fig. (5a) and (b) for versions 677 and 847, respectively.In these patterns, the most abundant pI values were found within the range of 3 to 7 and 8 to 11, for database versions 677 and 847, respectively.These ranges are very wide.Thus, if a user selected only pI filtering for protein identification, the process would take around 30 min to execute.Therefore, users are recommended to use protein mass filtering coupled with pI filtering.
For tool validation, two known proteins were searched for using the current Spirulina tool in order to compare these results to the 'Hits' results of Mascot.The 'Hits' results for each protein are shown in     first place in the 'five Hits' results (Table 1) by using the Mascot tool and our current tool.For Mascot, this protein was found with a score of 160, expected values of 1.5E-10, a matched number of eleven, a protein mass of 39.696 kDa, and coverage of 0.68% (Table 1a).On the other hand, this protein was found in first place, using the current tool, with a score of 15.7398, three out of thirteen matched peptides, a pI of 5.386, a protein mass of 39635.9211Da, and matched peptide coverage of 10.8635 % (Table 1b).The input setting for Mascot was: the MSDB database (Mascot database), Taxonomy of bacteria, allowed up to one missed cleavage, and a peptide tolerance of 1.2 Da.In the current version of the tool, six different input parameters were used: the Spirulina database version 847, allowed zero missed cleavages, protein mass filtering at 39 kDa, a protein tolerance of 0.7 kDa, a peptide tolerance of 1.2, pI filtering of five, and a pI tolerance of one.
In the case of the other known protein used for tool validation, pyruvate kinase from E. coli, the search results from both tools show pyruvate kinase I second on the list.The E. coli database and the Spirulina databases were used for Mascot and the current tool, respectively, for the protein identification process.Using the Mascot tool, this protein was found in second place with a score of 62, expected values of 0.021, a match number of seven, a protein mass of 50.697 kDa, and coverage of 7% (Table 2a).According to the current version of the PMF-Spirulina tool, the protein was also found in second place, with a score of 8.4036, a match number of four from seven peptide masses, a pI of 5.7986, a protein mass of 63.339 kDa, coverage of 21.19 %, and a fragment number of six, as shown in Table 2b.
Consequently, a second set of tool validation experiments were carried out.The complete E. coli genomes were down-   Note: Gray color is the correct theoretical protein of each AC number.Black boxes represent the results of each AC number that the correct theoretical protein was found within the third order.loaded to test our in-house tool and also to compare with Mascot by using the PMF data of E. coli proteins (crossspecies proteins), P0C0V0-protease Do, P61889-malatede hydrogenase, P0A9B2-glyceraldehyde-3-phosphate dehydrogenase A, P0AFL3-peptidyl-prolyl cis-trans isomerase A, P0A6P9-enolase and P00811-beta-lactamase as the input PMFs.The best 'Hits' results from Mascot and our current tool are shown in Table 3a and Table 3b, respectively.
According to the search results from Mascot, these proteins were found first in the best 'Hits' results for the first three proteins, and third for the last three, as shown in gray boxes in Table 3a.The input parameters, MSDB (the Mascot database), taxonomy of E. coli, an allowed miss cleavage of one, and a peptide tolerance of 1.2 Da were used.
The results from our current tool illustrated that five out of six proteins were shown first in the 'Hits' results.The protein with accession number P0A6P9 was identified second on the list as enolase (protein ID 2688) with a score of 6.0782, which is less than that of the 3-deoxy-D-arabinoheptulosonate-7-phosphate synthase.
In conclusion, the current tool has an accurate ability to identify proteins using the scoring function and appropriate input parameters.This version of the tool with a probabilitybased scoring function has high accuracy in protein identification by using peptide mass fingerprints.To the best of our knowledge, this is the first time that a tool used as search engine for protein identification contains pI-filtering, pIcalculating and a batch processing module.When the limitation of batch processing is its long execution time, this problem can be overcome by automatically obtaining the protein mass and pI values from each PMF file of the batch process.
Improvements to the batch process are in progress.Thus, the current tool obtained in this study has a high impact on the protein identification step in our proteomic work due to its accurate scoring function and ranking criteria, including pI.

Table 1a . 'Hits' Results from the Mascot Tool Using PMF Data from the Photosystem II D1 Protein (Synechocystis sp. PCC6803: slr0752)
Note:The results obtained from the Mascot software tool were identified by using five parameters for searching; PMF data are 13 fragments obtained from the in silico digestion of the photosystem II D1 protein of Synechocystis sp.PCC6803: slr0752, Database name is MSDB (Mascot Database), Taxonomy is bacteria, Peptide tolerance set at 1.2 Da, and Allow missed cleavage up to one.

Table 1b . 'Hits' Results from the Current PMF Tool Using PMF Data from the Photosystem II D1 Protein (Synechocystis sp. PCC6803: slr0752) # Protein ID Orf Name Computer Annotation Human Curation Score Match pI Protein MW % Coverage
Note:The results obtained from the previous version of the PMF tool were identified by using eight parameters for searching; PMF data are 13 fragments obtained from the in silico digestion of the photosystem II D1 protein of Synechocystis sp.PCC6803: slr0752, Database name is SpiDB_v847 (Spirulina Database version 847), Peptide tolerance at 1.2 Da, Protein mass of 39 kDa, the protein tolerance of 0.7 kDa, pI of 5, pI Tolerance of 1, and Allow missed cleavage at zero.

Table 2a . 'Hits' Results from the Mascot Tool Using PMF Data from Pyruvate Kinase I
Note:The results obtained from the Mascot software tool were identified by using five parameters for searching; PMF data are 7 fragments obtained from in silico digestion of pyruvate kinase I of Escherichia coli K-12 MG1655, Database name is MSDB (Mascot Database), Taxonomy of other bacteria, peptide tolerance at 1.2 Da, and Allowed missed cleavage up to one.

Table 2b . 'Hits' Results from the Current PMF Tool Using PMF Data from Pyruvate Kinase I # Pro- tein ID Orf Name Computer Annotation Human Curation Score Match pI Protein MW %Coverage Fragments
Note:The results obtained from the previous version of the PMF tool were identified by using eight parameters for searching; PMF data are 7 fragments obtained from in silico digestion of pyruvate kinase I of Escherichia coli K-12 MG1655, Database name is SpiDB_v847 (Spirulina Database version 847), peptide tolerance at 1.2 Da, protein mass of 64 kDa, protein tolerance of 5 kDa, pI of 5, pI Tolerance of 1, and Allowed missed cleavage at zero.