Bioinformatics Advance Access originally published online on July 16, 2008
Bioinformatics 2008 24(18):2094-2095; doi:10.1093/bioinformatics/btn371
© The Author 2008. Published by Oxford University Press. All rights reserved. For Permissions, please email: journals.permissions@oxfordjournals.org
MetalDetector: a web server for predicting metal-binding sites and disulfide bridges in proteins from sequence
Marco Lippi 1,
Andrea Passerini 1,
Marco Punta 2,3,4,
Burkhard Rost 2,3,4 and
Paolo Frasconi 1,*
1Machine Learning and Neural Networks Group, Dipartimento di Sistemi e Informatica, Università degli Studi di Firenze, Via di Santa Marta 3, 50139 Firenze, Italy, 2Department of Biochemistry and Molecular Biophysics, Columbia University, 630 West 168th Street, 3Columbia University Center for Computational Biology and Bioinformatics (C2B2), 1130 St Nicholas Ave. and 4Northeast Structural Genomics Consortium (NESG), Columbia University, 1130 St Nicholas Ave. Rm. 802, New York, NY 10032, USA
*To whom correspondence should be addressed.
 |
ABSTRACT
|
|---|
Summary: The web server MetalDetector classifies histidine residues in proteins into one of two states (free or metal bound) and cysteines into one of three states (free, metal bound or disulfide bridged). A decision tree integrates predictions from two previously developed methods (DISULFIND and Metal Ligand Predictor). Cross-validated performance assessment indicates that our server predicts disulfide bonding state at 88.6% precision and 85.1% recall, while it identifies cysteines and histidines in transition metal-binding sites at 79.9% precision and 76.8% recall, and at 60.8% precision and 40.7% recall, respectively.
Availability: Freely available at http://metaldetector.dsi.unifi.it
Contact: metaldetector{at}dsi.unifi.it
Supplementary Information: Details and data can be found at http://metaldetector.dsi.unifi.it/help.php
|
1 INTRODUCTION
|
|---|
Metal-binding proteins play critical catalytic, regulatory and
structural roles in the cell. They are implicated in heavy metal
toxicity, in processes such as apoptosis (Formigari
et al.,
2007) and aging (Mocchegiani
et al.,
2006), as well as in numerous
diseases, including Alzheimer (Crouch
et al.,
2007), Parkinson
(Santamaria
et al.,
2007) and AIDS (Diamond and Bushman,
2006).
Their identification and characterization can contribute toward
a better understanding of these phenomena. Here, we introduce
a web server that takes the protein sequence as input and outputs
predictions of transition-metal binding for cysteine and histidine
residues; for cysteines it also predicts disulfide bonding bridges.
|
2 METALDETECTOR: INTEGRATING METAL LIGAND PREDICTOR AND DISULFIND
|
|---|
We previously developed a method, Metal Ligand Predictor (MLP;
Passerini
et al.,
2006), which predicts transition-metal binding
for cysteines and histidines from sequence information alone.
The method classifies cysteines into one of three states: free
(F), disulfide bridged (D) metal bound (M) and histidines into
one of two states (F or M). The main purpose of MetalDetector
is to make the predictor available online as a web application.
When in the process of developing a server for MLP, however,
we observed some inconsistencies with DISULFIND (Ceroni
et al.,
2006), a server we previously made available for predicting
the disulfide bonding state of cysteines and their disulfide
connectivity. In particular, on the same test set used in (Passerini
et al.,
2006), conflicting cysteine classifications by the two
predictors involved 761 out of 9187 cases (i.e. 8.3%). Two types
of inconsistency may arise: (1) MLP predicts D and DISULFIND
predicts F (554 cases), and (2) MLP predicts F or M and DISULFIND
predicts D (207 cases). MetalDetector integrates MLP and DISULFIND
and tries to resolve their inconsistencies.
|
3 CONCEPT
|
|---|
When a protein sequence is submitted to MetalDetector, both
constituent methods, MLP and DISULFIND, are queried. For histidines,
the results are just read off MLP. For cysteines, the output
of MetalDetector is determined by a decision tree architecture
(
Fig. 1). We start with the output of DISULFIND that classifies
all cysteines as either F or D. For the same residues, MLP provides
probabilities for classes F, D and M (
PF,
PD,
PM). For a given
cysteine, if DISULFIND predicts class F, we apply a simple threshold
TD to the
PD output of MLP. If
PD>
TD, MetalDetector will
predict class D, else the cysteine will be predicted to be either
in class F (if
PF>
PM), or M (if
PF<
PM). We apply a similar
threshold
TM when DISULFIND predicts D. If the output
PM of
MLP exceeds
TM, the cysteine will be assigned to class M, otherwise
to class D. Changing the thresholds
TD and
TM enables the user
to decide how much trust to put in each of the constituent predictors.
For example, if
TD=
TM=1, disulfide bridges are only predicted
by DISULFIND, while lowering both thresholds increases the weight
for MLP. Prior knowledge about the protein may therefore help
users to find a metal bound/disulfide bound/free cysteine. At
the end of the decision process, a finite state automation (Passerini
et al.,
2006) constrains the number of disulfide predictions
to be even (inter-chain bridges are ignored). In case of an
odd number of disulfide predictions, it relabels a single cysteine
from free or metal bonded to disulfide bonded or vice versa,
depending on which relabeling produces the least reduction in
likelihood. The probabilities used by the automaton come either
from DISULFIND, or from MLP, depending on which predictor has
made the final prediction on each residue. MetalDetector also
outputs predicted disulfide connectivity by calling the second
stage of DISULFIND.
The new method deals efficiently with inconsistencies: at the
default thresholds
TD=0.76 and
TM=0.65, there are 274 non-consistent
predictions, 191 of type (1) and 83 of type (2) (a reduction
from 8.3% inconsistencies to 3.0%). For these 274 residues,
the predictions of MetalDetector are identical to those of MLP
in 256 cases and better than those of DISULFIND 56 and 75% of
these cases, for inconsistencies of type (1) and type (2), respectively.
A paired
t-test revealed that MetalDetector is significantly
better than MLP in terms of accuracy (
P<0.01). MetalDetector
also significantly outperforms both DISULFIND and MLP on the
two-classes problem D versus M/F (
P<0.01), while there is
no significant difference between MLP and DISULFIND. Thus, the
new method provides better performance and succeeds in achieving
our stated goal, which was to make available a metal-binding
state predictor that would largely agree with DISULFIND on disulfide
bonding state. In
Tables 1 and
2, we report the best results
achieved by MetalDetector considering both cysteine and histidine
predictions using default thresholds. The corresponding protein-level
accuracy
Qp is 77% as in (Passerini
et al.,
2006). Sample predictions
are shown in
Figure 2.
View this table:
[in this window]
[in a new window]
|
Table 1. Comparison of precision (P), recall (R) and disulfide bonding state accuracy (A) on the test set used in (Passerini et al., 2006)
|
|
View larger version (15K):
[in this window]
[in a new window]
[Download PowerPoint slide]
|
Fig. 2. Sample predictions, inconsistencies highlighted in boldface. Top: MetalDetector (MD) corrects the first wrong D assignment of DISULFIND thanks to MLP prediction, but cannot correct MLP's missed metal. Bottom: MD corrects the wrong D assignments of MLP thanks to DISULFIND predictions. In all cases, where MLP predicts M and DISULFIND predicts F (highlighted in lowercase), MD picks the right choice from MLP.
|
|
|
4 SERVER
|
|---|
Three preset working points can be chosen from the web interface.
They correspond to high metal accuracy (default,
TD=0.76 and
TM=0.65), high metal–precision (
TD=0.5,
TM=1), and high
metal recall (
TD=1,
TM=0.5) for the metal class. In the case
of histidines, the decision threshold is 0.5. Precision/recall
for the disulfide class are 83.1/88.7 and 90.1/82.0 at the high
metal precision and high metal recall working points, respectively.
|
ACKNOWLEDGEMENTS
|
|---|
Funding: M.P. and B.R. were supported by the grants R01-GM079767,
R01-LM07329, and U54-GM75026 from the National Institutes of
Health (NIH) in the USA.
Conflict of Interest: none declared.
|
FOOTNOTES
|
|---|
Associate Editor: John Quackenbush
Received on April 17, 2008; revised on June 27, 2008; accepted on July 14, 2008
|
REFERENCES
|
|---|
Ceroni A, et al. Disulfind: a disulfide bonding state and cysteine connectivity prediction server. Nucleic Acids Res. (2006) 34(Web Server issue):W177–81.[Abstract/Free Full Text]
Crouch PJ, et al. The modulation of metal bio-availability as a therapeutic strategy for the treatment of alzheimer's disease. FEBS J. (2007) 3775–3783.
Diamond TL, Bushman FD. Role of metal ions in catalysis by hiv integrase analyzed using a quantitative pcr disintegration assay. Nucleic Acids Res. (2006) 6116–6125.
Formigari A, et al. Zinc, antioxidant systems and metallothionein in metal mediated-apoptosis: biochemical and cytochemical aspects. Comp. Biochem. Physiol. C Toxicol. Pharmacol. (2007) 443–459.
Mocchegiani E, et al. Zinc homeostasis in aging: two elusive faces of the same ‘metal’. Rejuvenation Res. (2006) 351–354.
Passerini A, et al. Identifying cysteines and histidines in transition-metal-binding sites using support vector machines and neural networks. Proteins (2006) 305–316.
Santamaria AB, et al. State-of-the-science review: Does manganese exposure during welding pose a neurological risk? J. Toxicol. Environ. Health B Crit. Rev. (2007) 417–465.
CiteULike 
Connotea 
Del.icio.us What's this?
TOP
ABSTRACT
1 INTRODUCTION