Cataloguing proteins in cell cycle control

Kazimierz O. Wrzeszczynski ^{1, *}, & Burkhard Rost ^{1, 2, 3, *}

Abstract

Key words:  cell cycle control, genome sequence analysis, protein function prediction, multiple alignments.

Abbreviations used

Introduction

No direct path from sequence similarity to 'biological' similarity. How can bioinformatics tools help to identify particular types of proteins? In general, the answer depends on the type of protein. Alignment methods can identify similarities between two proteins. However, while database search tools are optimised to finding the best possible superposition between two proteins, they fail in answering questions such as: Does the query protein Q perform the same function as the protein in the database H for which we have some experimental data about function? In fact, alignment methods typically provide some statistical score evaluating the probability that the similarity between Q and H happened by chance [1, 3] . The precise function relating such a statistical score for sequence similarity to the actual 'biological' similarity of two proteins, i.e. similarity in terms of their three-dimensional (3D) structure and/or function depends on the problem. For example, if the PSI-BLAST expectation value for the similarity between Q and H is below 10^-5, then this typically implies that H and Q have similar local 3D structure [6] . However, less than 70% of all pairs of enzymes that have this level of sequence similarity have exactly the same enzymatic activity [7] and over 90% of all pairs with so similar sequences are observed in the same sub-cellular compartment [8] . Establishing these estimates typically requires solving three different tasks: (1) define biological similarity (3D, enzyme activity, sub-cellular localization), (2) build unbiased data sets of experimentally reliable information, (3) and establish thresholds that relate sequence to biologically similarity. These steps have been completed for a variety of biological features such as structure [9, 10, 11, 12, 6] , enzymatic activity [13, 14, 15, 16, 17, 7] , active sites [14] , binding sites [14] , functional keywords [14] , functional classes [14, 16] , sub-cellular localization [8, 18] . However, there is no way to infer from these results at which level of sequence similarity we can conclude that two homologous proteins play the same role in processes such as cell cycle control.

The field of proteomics has evolved into various levels of biological and computational techniques that identify and classify proteins in the context of entire genomes and proteomes. These techniques include a broad spectrum of approaches; from detailed literature searches [19, 20] or text analysis of database annotations [21] , database mining [22, 23, 24, 25, 26, 27, 28, 20] , multiple sequence alignments [29, 30, 31, 3, 32, 33] , protein family clustering [34, 35, 36, 37, 38, 39] , methods predicting aspects of protein function and structure [40, 41, 42, 43, 44, 45, 46] and computational modelling of the cell cycle [47] to gene microarray or 'chip' expression techniques [48, 49, 50, 51, 52] , yeast two-hybrid systems [53, 54] and recently mass spectroscopy of protein complexes [55, 56] . The process of unifying these techniques from an assortment of cataloguing tools into a more eloquent analysis of the cell cycle and specifically cell cycle control proteins is only beginning to take shape. Here, we present a first step for this process using database mining and literature searches to evaluate the current status of cell cycle control proteins present in various databases, combined with sequence alignment evaluation to identify cell cycle control proteins in various proteomes. We began by archiving proteins known to be involved in cell cycle control through database and literature searches. Then, we established levels of sequence similarity that imply similarity in function. Finally, we attempted identifying cell cycle control proteins through homology in entirely sequenced eukaryotic proteomes.

Materials

Public databases

Curated, well-formatted and annotated databases comprise one of the most important resources for bioinformatics. A few public databases contain information about cell cycle proteins ( Table 1 ); from these we built a resource that identifies the general register of cell cycle information currently available. To create this repository, we collected about 3811 records from MEDLINE [57] . Using SRS [4] , we retrieved about 364 proteins from SWISS-PROT [5] , and 98 proteins of known structure from PDB [2] . Only seven of these 98 were classified as 'cell cycle control' proteins. A closer inspection of the SWISS-PROT dataset revealed 534 proteins with the keyword 'cell cycle', and 940 with the keyword 'cell division'. ProtoNet [36, 58] is a tool that clusters all proteins from SWISS-PROT into somehow related families. ProtoNet identified 1476 clusters with a total of 512 proteins for the SWISS-PROT keyword 'cell cycle' and 887 proteins in 1983 clusters with the keyword 'cell division'. The obvious next task was to peel out a catalogue of unique families of proteins related to cell cycle (Methods).

Sources of sequences for entire proteomes

All human sequences were extracted from SWISS-PROT and TrEMBL [5] . We retrieved all other proteome sequences from the respective public sites: Drosophila melanogaster: http://www.fruitfly.org/, Caenorhabditis elegans: ftp://ncbi.nlm.nih.gov/genbank/genomes/, Saccharomyces cerevisiae from the yeast genome directory [59] , Arabidopsis Thaliana: http://www.arabidopsis.org/, and Mus Musculus: http://www.ensembl.org.

^a Note1: we dropped the string 'http://' from the URL, e.g. to access KEGG you mayhave to type 'http://www.genome.ad.jp/kegg' in some browsers 
Note 2: we will make all our data along with a novel cell cycle specificdatabase available through our website cubic.bioc.columbia.edu

Methods

Cell cycle and cell cycle control proteins in public databases

Keyword search in SWISS-PROT. First, we searched for proteins of trusted experimental information about cell cycle control in SWISS-PROT. Most proteins retrieved thus control the g1/s and g2/m transitions, or are related to the m and s phases. In total, we found 361 proteins ( Table 2 ) that were distributed amongst various species. Next, we clustered these proteins into families.

Sequence-unique data sets. In order to reduce the bias from too similar sequences, we generated sequence-unique subsets for all types of proteins under consideration. 'Sequence-unique' was defined by that no pair in the set had more than 33% identical residues over more than 100 residues aligned (HSSP-threshold of 0 [6] ). Given an all-against-all pairwise alignment for the biased set, we simply used a greedy search to find the largest subset that fulfilled the above condition. This reduced the entire set of 361 to 42 unique proteins or protein families.

Extending simple keyword-based search. 42 unique proteins did not suffice to develop any statistical criteria for determining levels of significant sequence similarity and also implying similarity in the cell cycle process. We expanded our original data set by including searches for other cell cycle controlling factors such as ubiquitin, and those in the ras super-family, plus other proteins annotated for cell division control. This extensive search for cell cycle control proteins increased the list to a total of 595 proteins; 97 of these had multiple, conflicting annotations ( Table 2 ); 113 were sequence-unique, i.e. we increased the numbers of families from 42 to 113 through the extended keyword-based search. The entire dataset of cell cycle control proteins is in the preparation of being made available online at the CUBIC website: cubic.bioc.columbia.edu.

 1 Eukaryotic proteins presented, the remainder ofproteins in the set of 595 cell cycle proteins are involved in the prokaryoticcell cycle process.

Cell cycle control protein identification through sequence similarity

Establishing threshold for significant sequence similarity. If we want to find proteins that have similar roles in the cell cycle as the proteins for which we have experimental information in public databases, we have to first establish a threshold for 'significant sequence similarity', i.e. we have to address the question: at which level of sequence similarity can we infer similarity in the specific functional role of that protein. Obviously, such thresholds have to find a balance between accuracy and coverage, in other words, we have to navigate between the Skylla of 'high selectivity/low sensitivity', i.e. finding very few homologues all of which are right, and the Charibdis of 'low selectivity/high sensitivity', i.e. finding many putative homologues, most of which are wrong. Cumulative accuracy and coverage were calculated as:

  ; (1)

  ; (2)

with the thresholds for sequence similarity specified below.

Aligning proteins. We generated alignments for all sequences from the cell cycle unique dataset (595) against a set of non-nuclear (but including cytoplasmic) proteins of known function other than those functions in cell cycle control (total of 6728 proteins) using pairwise BLAST [1] . To refine the analysis, we also generated PSI-BLAST profiles using a filtered version of all currently known sequences with three iterations [60] . These profiles were then aligned against our 'cell cycle control plus all other proteins' dataset. Sequence similarity was defined by percentage identity, BLAST E-values, and the distance from the HSSP-threshold which relates percentage sequence identity to alignment length thus accounting for the fact that 80% pairwise identity is not significant when achieved over a stretch of 15 consecutive residues, however, it is highly informative when achieved over entire proteins [61] .

Accuracy and coverage of inferring cell cycle role by homology. When we aligned all trusted cell cycle proteins (595) against all true negatives (6116 non-cell cycle proteins), we found that at HSSP-distances of 15 (corresponding to 48% pairwise sequence identity for more than 100 aligned residues), we could seemingly infer the role in the cell cycle at an accuracy of 95% ( Fig. 1 ). However, when using the unbiased, sequence-unique subset of 113 cell cycle proteins to evaluate accuracy, we found levels of only 60% accuracy. In order to reach a level of 95% accuracy, we had to increase the HSSP-distance from 15 to 40 ( Fig. 2 ), i.e. have to require over 70% pairwise sequence identity). Replacing the HSSP-distance by the expectation values from BLAST or PSI-BLAST (E-values) did not yield a more accurate distinction between true and false positives. This finding confirmed our previous results on establishing thresholds for sequence similarity implying similarity in 3D structure and sub-cellular localisation [8, 7] .

Fig. 1. : Sequence conservation of all trusted cell cycle control proteins. We aligned all trusted cell cycle proteins (595) against all true negatives (6116 non-cell cycle proteins) using BLAST. Solid lines with filled squares describe cumulative accuracy (percentage of correctly identified cell cycle proteins at given threshold, Eqn. 1); dotted lines with open circles describe cumulative coverage (cell cycle proteins found at threshold/all cell-cycle proteins, Eqn. 2). We measured sequence similarity in three different ways: (A) by the percentage pairwise sequence identity (left graph), (B) the distance from the HSSP-threshold accounting for the length of the alignment (central graph), and (C) by the negative logarithm of the BLAST E-values (note: log to the base of 10) (right graph). For example, the accuracy exceeded 80% for levels > 60% pairwise sequence identity (left), HSSP-distances above 3 (centre), and BLAST expectation values below 10^-12 (right). At all levels of accuracy ³ 80, the HSSP-distance performed best in terms of coverage. Note that these estimates were based on large data sets, however, they constituted over-estimates, since the bias in the data sets was not removed.

Fig. 2. : Estimating accuracy and coverage for BLAST and PSI-BLAST. In order to correctly estimate the likely accuracy and coverage, we had to remove the bias from our initial data sets by aligning the subset of 113 sequence unique trusted cell-cycle proteins against all trusted cell-cycle proteins and against all true negatives. For this, we compared the performance of pairwise BLAST (open symbols) to that of PSI-BLAST (filled symbols). Accuracy (solid lines) and coverage (dashed lines with circles) were as in Fig 1. In general, PSI-BLAST clearly outperformed BLAST. For example, at HSSP-distances > 40 the accuracy of PSI-BLAST searches was above 95%. Note that these estimates were sufficiently lower than those that would have been obtained using the biased data (Fig. 1). Using only the E-values taken from PSI-BLAST and BLAST alignments required very high cut-off thresholds: even at levels of 10^-10, implying that only one in ten million hits occurred by chance, less than 70% of the inferences were correct. The residual problem with the data resulted from the small set sizes (rigged curves).

Identifying cell cycle control proteins from entirely sequenced proteomes. We used a variety of thresholds for inferring the role of cell cycle control proteins by homology as to confer the annotations about these roles from our trusted data set to homologues in entirely sequenced eukaryotes. In particular, we scanned the proteomes of human (Homo sapiens), mouse (Mus musculus), fly (Drosophila melanogaster), worm (Caenorhabditis elegans), weed (Arabidopsis thaliana), and yeast (Saccharomyces cerevisiae). At levels of around 95% accuracy, we could extend the number of proteins known to be involved in cell cycle control from 284 for the six completely sequenced organisms to about 747 ( Table 3 ). Our analysis also pulled out about 500-1300 additional proteins (difference between columns D=40 and D=25 and D=15 in Table 3) that may constitute candidates for unknown cell-cycle control proteins. On the other extreme end, our data illustrated that over 10000 proteins in any of these six proteomes have similar 3D structures to one of the known cell-cycle proteins. Supposedly most of these are not related to cell-cycle control, illustrating the variety of functions that can be adopted by proteins of similar structure.

 1 Distance from HSSP-Threshold chosen as seen inFig. 2 for various levels of percent accuracy using the PSI-BLAST curve. Levelsof accuracy are estimated according to Fig. 2, e.g. at a threshold of D=40 morethan 95% of the proteins for which we infer the involvement in cell cyclecontrol by homology are supposedly correctly inferred.2 The number of previously known annotated cellcycle control proteins represented in each specific proteome as used in ourtrusted data set is given for comparison.

Notes

Limits of inferring function through homology

Everyday biologists are searching with their protein Q of interest by standard alignment methods to uncover putative homologies to their protein. Due to large-scale sequencing efforts, these database searches retrieve more and more often proteins without any annotation other than 'hypothetical protein'. To initiate hypotheses about function such results are obviously not very informative. More difficult are the 'helpful' cases when a protein with experimental annotation about function H is similar to Q. The number of pitfalls that can lead to incorrect hypotheses based on database searches are manifold [62, 63, 64, 14, 65, 28] . Nevertheless, an increasing number of publications in modern biology is based on some beneficial hints obtained from database searches. How can we separate the chaff from the wheat? Certainly, it is a sine qua non to establish thoroughly evaluated, statistically significant estimates for which level of sequence similarity implies what [13, 14, 66, 16, 17, 7] . In the context of cell cycle proteins, our approach aims at identifying commonalities in the evolutionary conservation of a selected group of functions. On the one hand, it appears evident that all proteins involved in cell cycle and cell cycle control have common evolutionary constraints. If true we can infer the involvement of a protein in the cell cycle process based on sequence similarity. On the other hand, we may suspect that two kinases such as pyruvate dehydrogenase kinase and Cdk1 are more similar than the two cell cycle proteins Cdk1 kinase and the E2F transcription factor. If true, we have to define all types of function related to cell cycle and have to establish thresholds for each functional type; in other words, our inference of cell cycle roles based on homology is rather limited. Arguably reality falls between these two extremes. Therefore, our ability to discover new proteins in cell cycle control through homology works to some extent, but is rather restricted.

Other tools targeting cell cycle proteins

Jones & Sgouros [67] studied cohesion complex proteins through sequence motifs and database searches. They used PSI-BLAST to identify all homologues of the SMC (Structural Maintenance of Chromosomes) and the SCC (Sister-Chromatid Cohesion) proteins from yeast (Smc1, Smc3, Scc1, Scc2, Scc3, and Scc4), as well as four proteins interacting with cohesion proteins (Trf4, Prp11, Tid3, Esp1). Next, the authors aligned the putative homologues identified by PSI-BLAST using the dynamic programming based method ClustalX [32, 33] , and constructed putative evolutionary trees from these ClustalX alignments using the program PHYLIP [68] . Finally, the study identified possible binding partners from the complete two-hybrid screens available through the Yeast Proteome Database and putative sequence motifs through the program Teiresias [69] . The study resulted in the establishment of five families of SMC proteins, a cohesion interaction network of 17 proteins and the identification of possible common sequence motifs for binding and a kinase active site.

Kel and colleagues [70] combined experimental and theoretical techniques in a comprehensive study identifying the 5' regulatory regions of cell-cycle related genes. First, the group developed a program that identifies context-specific binding sites for the E2F transcription factors. All these sites were identified in entirely sequenced genomes with the aim to identify new genes that play a role in controlling cell-proliferation, differentiation, and apoptosis. Finally, the predictions were verified by chromatin immunoprecipitation assays. The study resulted in a total of 313 new potential E2F targets found, 8 of which were verified through the in vivo experimentation.

Blaschke & Valencia [19] developed a text analysis system (SUISEKI: System for Information Extraction on Interactions) that automatically identifies cell-cycle related protein-protein interactions from scientific literature, i.e. from MEDLINE abstracts. At the heart of the system, text searches are defined into frames that capture the various language constructs used to convey protein interactions. The authors selected 5,283 abstracts that included the word Òcell-cycleÓ, the system detected 6,778 protein interactions from all of the abstracts, resulting finally in 4,657 distinct interactions from a total of 1,471 abstracts. The data is currently available at www.pdg.cnb.uam.es/suiseki/.

Acknowledgements

Thanks to Jinfeng Liu (Columbia) for computer assistance and the collection of genome data sets; to Jinfeng Liu, Dariusz Przybylski (Columbia), and Rajesh Nair (Columbia) for providing preliminary information and programs. Particular thanks to Volker Eyrich (Columbia) for programming and maintaining most of the immensely valuable software that runs the EVA and META-PredictProtein servers! The work of JL and BR was supported by the grants 1-P50-GM62413-01 and RO1-GM63029-01 from the National Institute of Health. Last, not least, thanks to all those who deposit their experimental data in public databases, and to those who maintain these databases.

References