Solvent accessibility based
iDEntification and Recognition
The SPPIDER protein interface recognition server can be used to:
(1) predict residues to be at the putative protein
interface(s) by considering single protein chain with resolved 3D
structure; (2) analyse protein-protein complex with given 3D structural
information and identify residues that are being in interchain
Training and control sets of proteins used for SPPIDER training and
S435 - Training set
S149 - Control set
S21(S21a) - Subset of
control set that has individually crystallized chains - homologs
API: For automated submissions a perl script can be used
The zip achive contains two files: the perl script and
example of a list file with PDB codes.
Don't forget to keep a delay between requests for at least
15 seconds or your IP address will be banned.
Types of queries
Currently, there are two available types of queries:
(1) Recognition of
putative interacting sites for protein-protein interaction based on
consensus classifier. It is able to determine the residues that are
potentially take part in some protein-protein interactions. Predicted
interacting sites may belong to the different protein interfaces.
Accuracy of recognition can be adjusted using Tradeoff option (see
below). Requests of this type go to the machine cluster queuing
system and are being processed in the order received. Thus, recognition
results may be delayed and can be optionally sent via e-mail.
(2) Identification of protein interface within protein-protein
complex, which one can submit in the
PDB format by either specifying registered PDB entry code or uploading custom
PDB file. Based on 3D structure analysis, server will issue a
precise information about what the actual residues are involved in a given
protein-protein interaction. But it won't be able to tell anything
beyond that, for example, if there are alternative active sites within
each participating protein chain. Automatic identification is
performed in real-time and usually one can get results immediately.
Note, that for each type of query a set of options is different.
- Interacting sites recognition within single protein chain
E-Mail address is an
optional setting to let user obtain results via e-mail without
waiting for on-line output, which might take a while. A link for
the visualization of results is provided in e-mail message.
PDB code and
PDB file are fields to
specify either existing PDB entry or custom protein structure.
Chain label is to select
a chain of interest if PDB file represents more than one protein
chain. If option is omitted, first defined chain is taken for its
interacting sites prediction.
Version is to choose the
strategy of prediction: to have higher sensitivity or better
specificity. What is the difference between versions I and II?
Briefly, SPPIDER I (February 2005) has higher recall
(sensitivity), while SPPIDER II (February 2006) shows higher
precision (specificity) and per residue accuracy. And both versions have
the same correlation coefficient (see table below).
In details, (i) feature space they use differs in one
parameter; (ii) SPPIDER II was trained on smaller but cleaner
and better annotated training set.
Below is a table with performances of both versions applied to the
same control set of 149 protein chains (19977 vectors), baseline
Q2=65% with no sequence homology to the training set.
|MCC||Q2, %||Recall, %||Precision, %
Tradeoff between Sensitivity
and Specificity is an option for adjusting results according
to the purpose of search. Within each SPPIDER version, one
can reach a high degree of recall for putatively interfacial
residues having in the same time low level of precision to the
knowledge about interfacial residues. Although, it should
not be discouraging since many negatives can be false just because
of the lack of information about protein-protein
interaction. Estimates for tradeoff between sensitivity and
specificity are represented in table below based on control set
consisting of 19977 vectors, baseline Q2=65% with
no sequence homology to the training set.
|MCC||Q2, %||R, %||P, %
||MCC||Q2, %||R, %||P, %
Where MCC - Matthew's correlation coefficient,
Q2 - percent of correctly classified vectors in
two-class problem, R - recall (sensitivity) and
P - precision (specificity).
Generate PDB file with
prediction encoded by B-factors setting supplys user with
downloadable file in the PDB format modified to contain target
chain only and
B-factors values changed to either two-state classification
output (0.00 - Negative class or 100.00 - Positive class) or
probabilities of residue to be at interface depending on option
selected. Thus, prediction results can be easily visualized
in 3D using this PDB file with any molecular modeling software that
supports coloring atoms by temperature or B-factors. In SPPIDER
output, it can be viewed by JMol java-applet.
Add information about known
interactions option provides user with possibility to retrieve
information about known interacting sites derived from PDB. This
data can be either attached to or contrasted with prediction
results as independent section in the output. System for known
protein interaction data analysis and retrieval is implemented as
independent service - SCORPPION. WARNING: use this option with
caution as it is still under active development and did not pass
the thorough testing. Thus, results obtained using this option are
not fully reliable.
Add solvent accessibility
prediction results to e-mail message option lets
user to get SABLE
prediction results for protein secondary structure and relative solvent
accessibility as a main data contributor to the feature space used
for interface recognition in our classifier. It might be helpful
for the results analysis.
- Interface identification within protein-protein complex
PDB code is to be the
registered entry code at the
Protein Data Bank
server and has to represent a protein-protein complex
with at least 2 protein chains.
PDB file is a custom
file in the PDB format. It should contain at least 2 protein chains.
RSA change threshold is
an option to adjust identification
system for the user's need. In literature, different research
groups use various relative solvent accessibility change as a
threshold for assigning residue to be in the interface or
non-interacting. Option provides user with the possibility to set
up RSA change threshold in both absolute (Å2) and
relative (% of maximal surface exposed area for given amino
acid) surface scale.
Output setting specifies
how protein interface analysis is supposed to be presented: as a
summary of all interacting residues within each chain, or
as a pairwise interfaces for each pair of chains separately, which
makes further analysis easier in case of complexes with multiple chains
involved in the interaction.
For both types of query, plain text output is available, which can be
used for further parsing. Beside the text, server provides a graphical
visualization of the results. Default view for query type 1 is made by
POLYVIEW, whereas for query type 2 the default option is JMol (which
requires Java to be installed on computer and java-applets to be
enabled in web-browser). But user can also follow the links provided
to receive enchanced views either by POLYVIEW-3D or by Protein Explorer.
Examples of prediction
Apart from 2D output generated by POLYVIEW visualizer (for details, refer to
one can get a 3D animated image of the whole query protein as
well as publication quality static slides (see
Examples of SPPIDER I prediction (click on picture to see animated
image - big files!).
- Colored ribbons (if any) - chains beyond prediction
- White CPK - target chain
- Red CPK - atoms in residues predicted to be at interface
Query protein is GROEL/GROES complex (PDB code
target chain A.
Query protein is Deoxyhemoglobin Rothschild
1hba) with target chain A.
Examples of SPPIDER II prediction (front and rear view). Target proteins were
not included in the training set and have no sequence redundancy to
any of the training proteins.
- Red - true positives (residues correctly predicted to be at interface)
- White - true negatives (residues with no functional annotation)
- Yellow - false positives (residues wrongly predicted to be at interface)
- Blue - false negatives (known but not recognized interfacial residues)
Human erythrocyte catalase as a part of the oxidoreductase complex (PDB entry 1f4j, chain A).
Cyclin-dependent kinase 6 (CDK6) as a part of the p18(ink4c)-cdk6-k-cyclin ternary complex (1g3n:A).
Von Hippel-Lindau disease tumor suppressor from the pvhl/elongin-c/elongin-b complex (1lqb:C).
Information about protein binding sites was derived from complexes
containing target protein chains and available at PDB and further mapped to
the corresponding structures specified above
(1f4j:A -- 1f4j, 1dgb, 1dgf, 1dgg, 1dgh, 1qqw, 1tgu, 1th2, 1th3, 1th4, 4blc, 8cat; 1g3n:A -- 1g3n, 1bi7, 1bi8, 1blx, 1xo2, 1jow; 1lqb:C -- 1lqb, 1lm8, 1vcb).
Comparison with other predictors
We have recently performed a large scale evaluation of SPPIDER in
comparison with other structure-based methods for
prediction of protein-protein interaction sites available
online as web-servers. Evaluated servers include
Evolutionary Trace method, ConSurf, PROMATE, Cons-PPISP,
WHISCY, PIER, and SPPIDER. Three benchmark sets were used:
the SPPIDER control set (Porollo and Meller, 2007,
Proteins), protein docking benchmarks containing both
bound and unbound proteins (Albou et al., 2008, Proteins;
Hwang et al., 2008, Proteins).
The detailed survey can be found in the book chapter
Porollo A, Meller J. Computational
Methods for Prediction of Protein-Protein Interaction
Sites. In: Protein-Protein Interactions - Computational
and Experimental Tools; W. Cai and H. Hong, Eds. InTech
2012; 472: pp. 3-26.
The book chapter discusses the
effects of protein interaction site definition and mapping
from homologous complexes. Among different angles of
assessment, the survey also evaluates the methods
using bound and unbound forms of proteins in the benchmark
are with open access.
Reference to the SPPIDER method:
A. Porollo, J. Meller
Prediction-based Fingerprints of Protein-Protein Interactions
Proteins: Structure, Function and Bioinformatics (2007) 66: 630-45.
Reference to the survey of methods for prediction of
protein-protein interaction sites:
A. Porollo, J. Meller
Computational Methods for Prediction of Protein-Protein Interaction
In: Protein-Protein Interactions - Computational
and Experimental Tools;
W. Cai and H. Hong, Eds. InTech
2012; 472: pp. 3-26.
This work was supported by the
University of Cincinnati College of Medicine,
Cincinnati Children's Hospital Research Foundation, and
NIH through grants: AI055338, R01 AR050688, 5R01GM067823-02.
Back to the SPPIDER server home page