The PredictProtein server

Burkhard Rost ^{1, 2, 3, *} & Jinfeng Liu^{1, 3, 4}

This article is published in (Nucleic Acids Research, issue, 2003 and pages) © copyright Oxford University Press (2003). OUP is the only authorised source. All copying of this article including placing on another website requires the written permission of the copyright owner.

Abstract

PredictProtein (PP, http://cubic.bioc.columbia.edu/pp/) is an internet service for sequence analysis and the prediction of aspects of protein structure and function. Users submit protein sequence or alignment. The server returns a multiple sequence alignment, PROSITE sequence motifs, low-complexity regions (SEG), ProDom domain assignments, nuclear localisation signals, regions lacking regular structure and predictions of secondary structure, solvent accessibility, globular regions, transmembrane helices, coiled-coil regions, structural switch regions, and disulfide-bonds. Upon request fold recognition by prediction-based threading is available. For all services, users can submit their query either by electronic mail, or interactively from World Wide Web.

Key words: sequence analysis, prediction of protein structure and function, globularity, protein domains, secondary structure, solvent accessibility, multiple alignments, transmembrane helices.

Overview

10 years of sustained service for protein structure prediction. PredictProtein (PP) is an automatic service that searches up-to-date public sequence databases, creates alignments, and predicts aspects of structure and function ( Fig. 1 ). Users send a protein sequence and receive a single file with results from database comparisons and prediction methods. PP went online in 1992 at the European Molecular Biology Laboratory (EMBL, Heidelberg) [1] ; it was the first internet server for protein structure prediction, and belonged to a group of five pioneering internet sites for molecular biology [2] . Originally, PP handled all requests through email. Since summer 1993, users can also query the server through a web interface and opt for an HTML output format. When the server moved from EMBL to Columbia University (1999), this output format was extended significantly and we added the option of retrieving the results through http download to account for the continuously growing sizes of alignments. With the explosion of the web in the mid 90's, many other servers have implemented particular aspects covered by PP. However, PP remains the most widely used public server for structure prediction: over one million requests from users in 95 countries have been handled over the first decade of PP ( Fig. 2 >

Fig. 1: Data flow for PredictProtein (PP). Users query PP with their protein and receive a single file with results from database comparisons and prediction methods (Fig. 3 for details). Users can query PP through email and web interfaces.

Fig. 2: Usage of PredictProtein. The number of requests to PP continues to rise (A); over one million requests came from 95 countries. Almost half of all the requests originated from North America (B).

Ultimate goal: collect all predictions for a protein in one image. Many of the servers launched after PP are either more specialised (e.g. only secondary structure prediction), or list all available tools irrespectively of performance accuracy. From the beginning, the ultimate goal of PP was to concatenate the results from all valuable and sustained prediction methods into one single image. This objective has driven us to build one of the most comprehensive public servers for sequence analysis and structure prediction. We hope to make the next leap into that direction in the near future.

Attempt to simplify output by incorporating hierarchy of thresholds. The attempt to 'pre-digest' as much information as possible to simplify the ease of interpreting the results is another unique pillar of PP. For example, by default PP returns only those proteins found in the database that are very likely to have a similar structure as the query protein [3] . Particular predictions such as those for membrane helices, coiled-coil regions, signal peptides, nuclear localization signals are not returned if found below given probability thresholds. Over the years, we have added so many methods into the output of PP that our original goal 'easy-to-interpret' is challenged. We hope that a variety of improvements in the near future will reduce this problem.

Each request triggers the application of over 10 different methods.Currently, users receive a single output file with the following results (some of these are optional, Table 1 ). (1) Database searches: similar sequences are reported and aligned by a standard, pairwise BLAST [4] , an iterated PSI-BLAST search [5] , and by the dynamic-programming method MaxHom [6] . While the pairwise BLAST searches are identical to those obtainable from the NCBI site, the iterated PSI-BLAST is performed on a carefully filtered database to avoid accumulating false positives during the iteration [7, 8] . The dynamic-programming method MaxHom is only available through PP. Additionally database searches comprise a standard BLAST-based search through ProDom [9] and a standard search for functional motifs in the PROSITE database [10] . Optionally, users can request searches for remotely similar proteins by the prediction-based threading method TOPITS [11] . (2) Structure prediction methods: secondary structure, solvent accessibility, and membrane helices predicted by the PHD and PROF programs [12, 13] , coiled-coil regions by COILS [14] , and bonded cysteine residues by CYSPRED [15] . Putative structural switching regions are detected by the program ASP [16, 17] , low-complexity regions are marked by SEG [18] , and long regions with no regular secondary structure are identified by NORSp [19] . The PHD/PROF programs and TOPITS are only available through PP. The particular way in which PP automatically iterates PSI-BLAST searches and the way in which we decide what to include into sequence families is also unique to PP.

 

Method	Task	MainAuthor(s)	Quote
Database
SWISS-PROT *	annotated protein sequences	ABairoch (SIB) & R Appweiler (EBI)	[28]
TrEMBL*	rawprotein sequences	RAppweiler (EBI)	[28]
PDB*	proteinstructures	PBourne (UCSD)	[29]
BIG	non-redundantcombination of SWISS-PROT, TrEMBL, PDB	DPrzybylski (Columbia)	[8]
Alignment
MaxHom	dynamicprogramming, multiple alignment	RSchneider (LION) and C Sander (Sloan Kettering)	[6]
BLASTP*	pairwisealignment	SKarlin & S F Altschul (NCBI)	[4]
PSI-BLAST*	profilebased alignment	S FAltschul (NCBI)	[5]
TOPITS	prediction-basedthreading	BRost	[11, 30, 31]
Proteindomains and unusual regions
ProDom*	structuraldomain-like regions	FCorpet, F Servant, J Gouzy & D Kahn (Toulouse)	[32]
SEG*	low-complexityregions	J CWootton & S Federhen (NCBI)	[18]
NORSp	floppyregions	JLiu & B Rost	[19, 33]
Proteinstructure
PHDsec	secondarystructure	BRost	[34, 35, 12]
PHDacc	solventaccessibility	BRost	[36, 12]
PHDhtm	membranehelices	BRost	[37, 12, 38]
PROFsec	secondarystructure	BRost	[13]
PROFacc	solventaccessibility	BRost	unpublished
GLOBE	globularity	BRost	unpublished
COILS	coiled-coiledregions	ALupas (TŸbingen)	[39]
CYSPRED*	disulfide-bonds	PFariselli & R Casadio (Bologna)	[15]
ASP	structuralswitches	MYoung & S Highsmith (Sandia)	[17]
Proteinfunction
PredictNLS	nuclearlocalisation signals	RNair, M Cokol & B Rost (Columbia)	[40, 41]
PROSITE*	functionalsequence motifs	K Hofmann,P Bucher & A Bairoch (SIB)	[10]
Toolsintegrated into PP
MView*	HTMLalignment viewer	NBrown	[42]
ESPript*	Ready-to-publishoutput for sequence alignments and secondary structure	PGouet & E Courcelle (IPS Toulouse)	[43]

OriginalURLs
SWISS-PROT: http://expasy.cbr.nrc.ca/sprot/
TrEMBL	http://www.ebi.ac.uk
PDB:	http://www.rcsb.org/pdb/
BLASTP/PSI-BLAST:	http://www.ncbi.nlm.nih.gov/BLAST/
PROSITE:	http://www.expasy.ch/prosite/
ProDom:	http://protein.toulouse.inra.fr/prodom.html
SEG:	http://trex.musc.edu/manuals/unix/seg.html
CYSPRED:	http://prion.biocomp.unibo.it/cyspred.html
MView:	http://mathbio.nimr.mrc.ac.uk/~nbrown/mview/
ESPript:	http://prodes.toulouse.inra.fr/ESPript

Performance of methods. A detailed review about the strengths, weaknesses, and pitfalls [20] of the many methods applied by PP is not possible, here. Hence, we give only a brief overview over trends in the following. (1) Alignment methods: while the dynamic programming method MaxHom still appears best in aligning two proteins, the iterated PSI-BLAST tends to be more sensitive in unravelling more distantly related proteins. Note, however, that PSI-BLAST tends to over-estimate the relevance of short matches, and that PSI-BLAST expectation values have to be viewed with extreme caution when inferring similarity in function [21, 22, 23] . (2) Protein domains and unusual regions: like for instance SMART [24] , ProDom tends to identify regions that are significantly shorter than structural domains [25] . Note that short regions of low-complexity (SEG) are fairly common and not necessarily informative. (3) Protein structure (see the EVA server [26] for an up-to-date evaluation of structure prediction): (3a) PROFsec secondary structure prediction: on average, 76% of all residues are correctly predicted by (only about 71% by PHDsec); (3b) PROFacc accessibility prediction: almost 80% of all residues are correctly predicted as either buried or exposed, and over 80% of the surface residues are correct; (3c) PHDhtm: about 80% of the membrane helices are correctly predicted, for about 66% of all tested proteins all membrane helices and the topology was correctly predicted [27] ; at the default threshold, membrane helices are incorrectly detected in about 2% of the tested globular proteins [27] ; about one fourth of all signal peptides (for secreted proteins) are mistaken for membrane helices [27] ; (3d) GLOBE: not accurate enough to identify domain boundaries, however, sufficient to capture trends like 'very unlike a globular protein'; (3e) COILS: perceived to be correct most of the time; (3f) CYSPRED: most disulfide-bonding residues are correctly identified, however, most predicted bonds are wrong; (3g) ASP: if the protein has a structural switching region, this is usually detected correctly. Note the PROF and PHD series and CYSPRED all are based on artificial neural network systems.

Input, output, and job options

Default output. The output format is self-documenting. The output contains: (1) a list of likely homologues found in the protein database (BIG) and the multiple sequence alignment of these sequence (by default in 'HTML' format from MView). (2) If found: a list of the putative PROSITE motifs. (3) If found: a list of ProDom domain assignments. (4) If found: a prediction of coiled-coil regions. (5) Information about the expected levels of accuracy of structure predictions. (We suggest that newcomers read this carefully.) (6) Prediction of aspects of protein structure. These are grouped in the following way: (i) Prediction of secondary structure for all residues, (ii) prediction of secondary structure for reliably scored residues only, with an expected three-state accuracy for these residues of > 85%, (iii) prediction of solvent accessibility for all residues, (iv) prediction of solvent accessibility for reliably scored residues only, with an expected correlation between experimental observation and prediction of 0.69, (v) prediction of transmembrane helices and their topology (if any detected). Note: for the prediction of transmembrane helices a conservative threshold is chosen. Thus, a membrane helix may not be detected.

Advanced input options. By default users submit proteins through its one-letter residue sequence. However, PP also accepts submissions in FASTA, PIR and SWISS-PROT format or through the SWISS-PROT identifier. Most predictions methods applied use the information from the multiple alignments created by PP; prediction accuracy increases with the quality of the alignment. PP's alignments are fully automated, thus may not be as accurate as the alignment that experts have hand-edited. Therefore, users may also submit their favourite alignment directly. PP accepts alignments as FASTA lists, PIR lists, as well as in SAF and MSF format. The fold recognition/prediction-based threading method TOPITS uses predictions of secondary structure and solvent accessibility to search through a library of proteins of known structure. Predictions can be submitted through a simple column-based format.

Advanced prediction/job options. Not all methods are executed by default; some methods (like the prediction of membrane helices) use particular 'conservative' thresholds when included automatically and different thresholds when requested explicitly. In particular, the following methods can be toggled (switch on or off): MaxHom, BLASTP, PSI-BLAST, SEG, PHDsec, PHDacc, PHDhtm, PROFsec, PROFacc, COILS, CYSPRED, ASP, PROSITE, ProDom. Users can also explicitly request TOPITS or can evaluate the prediction accuracy of a secondary structure prediction method (EvalSec). Note that switching off methods has two advantages: it speeds up the execution and it reduces the size of the output. However, bear in mind that the database searches and their results are the limiting factor for speed and bytes produced.

Advanced output options. The default output now is an HTML formatted file, i.e. ready to display in any browser. Users can change this default to output in raw text in the following alignment formats: BLAST, no alignment, HSSP, HSSP profiles only, MSF, SAF, FASTA list. The results from the predictions are also available in a variety of machine-readable formats. (Developers: please do not write parsers for the human-readable PP output; in doubt, contact us, we can write almost any reasonable format if need be!) Due to the size of multiple alignments, we no longer email the results rather the output will be stored for a week on our web site (remember to download it in that period). Results can also be requested by email.

Interactive versus batch jobs. By default, the user submits requests to a batch queue and will be notified by email where to find the results (or will be sent these results). While PP also has an interactive mode that will write the results directly into the requesting web browser, this option comes with a restriction in the length of time for which the web connection is kept open: if PP has not completed a request within five minutes, we automatically switch the job to a batch mode and notify users by email. In practise, this implies that interactive jobs will only finish in time if (i) the PP queue is empty (works on a first-come-first-serve principle) and (ii) that the request does not require more than five minutes of CPU (typically the case if an alignment is submitted, and/or the query protein is short, and/or has few homologues in today's databases). We plan to upgrade the CPU resources for PP in the near future; this will increase the probability of successful interactive queries.

Acknowledgements

Making PredictProtein survive a decade was a major effort; many colleagues helped with hands and brains; thanks to all of them! Crucial contributions during the first years at EMBL: Antoine de Daruvar (Bordeaux Univ.) wrote most of the initial server software, Reinhard Schneider (LION Biosciences, Heidelberg) helped with software and ideas, Sean O'Donoghue (LION Biosciences, Heidelberg) helped with encouraged continuation at difficult moments, Chris Sander (Sloan Kettering, New York) contributed his invaluable support, and Hans Doebeling and Bjšrn Kindler provided sysadmin support. Thanks to Rolf Appweiler for his continued support at the European Bioinformatics Institute (EBI-EMBL, Hinxton, England), and to Volker Eyrich (Schroedinger, New York) for software support during the move to the USA. Further thanks to all who set up mirror pages and who consented to using their software in particular to Nigel Brown for MView, to Emmanuel Courcelle and Patrice Gouet (IPBS, Toulouse) for ESPript, to Florencio Pazos (Madrid) for Threadlize, to Andrei Lupas (Max Planck, TŸbingen) for COILS, to Piero Fariselli and Rita Casadio (Bologna Univ.) for CYSPRED, to Reinhard Schneider (LION Biosciences, Heidelberg) for MaxHom, to Malin Young (Sandia Labs, Albuquerque) for ASP, and to Rajesh Nair (Columbia Univ.) for NLSpred. Finally, thanks to all the users who supported the service, in particular to Roland Walker (NCBI, Washington), Timothy Springer (Harvard, Cambridge USA), Raphael Zidovetzki (UCI, Irvine), and Fernando Bazan (Stanford). Thanks to SGI, in particular to Juli Nash for crucial hardware support. JL and BR are supported by a grant to the Northeast Structural Genomics Consortium from the Protein Structure Initiative of National Institutes of Health (P50 GM62413). PredictProtein has attracted its first public support from the grant R01 LM07329-01 from the National Library of Medicine. Last, not least, thanks to Amos Bairoch (SIB, Geneva), Rolf Apweiler (EBI, Hinxton), Phil Bourne (San Diego Univ.), and their crews for maintaining excellent databases and to all experimentalists who enable computational biology!

References