Information analysis of Fis binding sitesPaul N. Hengen1, Stacy L. Bartram1,2,+, Lisa E. Stewart1 and Thomas D. Schneider1,*
1Laboratory of Mathematical Biology, National Cancer Institute, Frederick Cancer Research and Development Center, PO Box B, Building 469, Room 144, Frederick, MD 21702-1201, USA and 2Middletown High School, 200 High Street, Middletown, MD 21769, USA
Received August 25, 1997;Revised and Accepted October 30, 1997
ABSTRACT
Originally discovered in the bacteriophage Mu DNA inversion system gin, Fis (Factor for Inversion Stimulation) regulates many genetic systems. To determine the base frequency conservation required for Fis to locate its binding sites, we collected a set of 60 experimentally defined wild-type Fis DNA binding sequences. The sequence logo for Fis binding sites showed the significance and likely kinds of base contacts, and these are consistent with available experimental data. Scanning with an information theory based weight matrix within fis, nrd, tgt/sec and gin revealed Fis sites not previously identified, but for which there are published footprinting and biochemical data. DNA mobility shift experiments showed that a site predicted to be 11 bases from the proximal Salmonella typhimuriumhin site and a site predicted to be 7 bases from the proximal P1 cin site are bound by Fis in vitro. Two predicted sites separated by 11 bp found within the nrd promoter region, and one in the tgt/sec promoter, were also confirmed by gel shift analysis. A sequence in aldB previously reported to be a Fis site, for which information theory predicts no site, did not shift. These results demonstrate that information analysis is useful for predicting Fis DNA binding.
Fis is a pleiotropic DNA-bending protein that enhances site-specific recombination, controls DNA replication, and regulates transcription of a number of genes in Escherichia coli and Salmonella typhimurium (1 -4 ). Fis is composed of two 98 amino acid polypeptides, with each polypeptide having four [alpha] helices, A-D. Homodimers of Fis bind to and deflect DNA from 40° to 90° (3 ,5 -7 ). Mutational analyses suggest that the N-terminal portion of the Fis monomer containing the A helix is necessary for recombination, while the C-terminal portion containing the D helix is thought to be involved in DNA binding (8 ,9 ). However, X-ray crystal structures of Fis reveal that the D helices appear to be too close together for Fis to fit into two successive major grooves on straight B-form DNA, suggesting that the DNA bends to accommodate Fis (6 ,10 ,11 ), that Fis is flexible, or that Fis binds in a completely unanticipated manner.
Although Fis binds to precise sequences according to footprint data, it is often noted that the Fis binding site has a poorly defined consensus sequence (3 ,6 ,8 -25 ). Consensus sequences are often used to locate binding sites, but it is widely known that this does not work well, especially in the case of Fis (24 -26 ). Since a consensus is constructed by selecting the most frequent base or bases at every position across a binding site, creating a consensus sequence throws out important information about the observed frequency of bases in the binding site.
In contrast to consensus sequences, information theory provides quantitative models for binding sites. These models are represented by the sequence logo, a graphical method that retains most of the subtleties in sequence data (27 -30 ). Even a glance at a sequence logo often reveals the possible nature of specific base contacts, which side of a base pair is likely to face the protein, and whether or not the DNA is distorted away from B-form (30 ,31 ).
Because of its importance in a variety of genetic systems and because many binding sites were already well defined, interaction of Fis with DNA was an attractive candidate for a thorough information analysis. In addition to the information content measure (32 ) and the sequence logo (29 ), we used a new method, `individual information' (Ri), that defines the information content of individual binding sites (33 ) and displays the results graphically as a `sequence walker' (34 ). These methods have an advantage over other methods in that training is not required to obtain a quantitative binding site model, and only examples of functional sites are used to construct the model. Using well-defined biochemical data helps to ensure that the models are realistic. Our analysis of Fis binding sites and their surrounding sequences revealed many previously unidentified sites adjacent to known ones, and experiments demonstrated that some of the predicted sequences are indeed bound by Fis in vitro.
Fis sites identified by footprinting, gel shift or mutational data were gathered from the GenBank accession numbers, coordinates and orientations shown in Figure 1 . The exact alignment of the sites was confirmed by maximizing the information content (54 ).
Delila system programs were used for handling sequences and information calculations (29 ,32 -34 ,43 ,57 ,58 ). Figures were generated automatically from raw GenBank data using Delila and UNIX script programs. Further information is available on the World Wide Web at http://www-lmmb.ncifcrf.gov/~toms/ .
In designing the sequences of Figure 6 we chose the hin site of S.typhimurium (12 ) because it is well characterized and the binding site prediction is clear. We chose 32 bases of the hin sequence because according to the information-theory based search (Fig. 5 ) this region contains two overlapping Fis sites, one of which is the Fis site proximal to the recombination junction hixL (12 ). We added five bases of natural DNA sequence on each end-half a twist of DNA-to be sure we were not missing important components, although this region does not show up significantly in the sequence logo. Beyond these ends we added EcoRI and HindIII overhangs. We created three other sequences using the anti-consensus of the Fis weight matrix to destroy the proximal site, the newly identified `medial' site, or both sites. The anti-consensus sequence is the sequence that should bind Fis the worst (33 ). It is predicted from the number of bases at each position or the Riw(b,l) matrix by noting which bases appear least frequently at each position of the site or which give the lowest weight. In ambiguous cases we chose C or G when possible because these appear rarely in the logo (Fig. 2 ). (Note: the anti-consensus sequence in the early model we used had C at -5 rather than G.)
To create a model for Fis binding, 60 Fis binding sites for which DNA footprinting or mutational data are available were collected (Fig. 1 ). Each footprint was compared to the prediction from information scans and walkers to be sure that the location of the binding sites matched the final model. By this method, the locations of several sites given in previous compilations were corrected (see Materials and Methods).
Since E.coli and S.typhimurium Fis proteins have identical amino acid sequences (3 ,24 ) and no modifications that we are aware of, we used binding sites from both species. This would be inappropriate if in vivo DNA binding conditions are different between the two species. Because Fis binds as a dimer (6 ,10 ), the sequences and their complements were aligned to produce a sequence logo (29 ). Analysis of such logos can reveal interaction details that are otherwise obscured by a consensus (31 ). From the sequences, the sequence logo (Fig. 2 a) the structure of DNA base pairs (Fig. 2 b), and molecular modeling, the following observations were made:
(i) The correlation between sequence conservation peaks at ±7 and ±3 and a 10.6 base spacing (shown by the sine wave in the figure) suggests that Fis makes contacts in two consecutive major grooves (31 ). Further, the information content at ±7 is above one bit, which also suggests major groove binding (30 ,35 ). This is consistent with protection data showing that the methylation of the major groove N7 of G at ±7 interferes with Fis binding (12 ), with Fis binding that protects against DMS methylation at ±7 (36 ), and with hydroxyl radical footprints (37 ).
(ii) As seen on the logo, if an A is substituted for the majority G at -7 (or the complementary G at +7), a possible G-O6 contact would be lost while a G-N7 contact could be retained as A-N7. On the other hand, if T is substituted, a G-N7 contact would be lost but a G-O6 contact could be replaced by T-O4. This is consistent with the observed frequency of bases at position -7 (and it's complement at +7) for which G>A~T>C. That the frequency of As and Ts are nearly the same at this position suggests this A-N7 contact is energetically equivalent to a T-O4 contact. Similar contacts appear at ±15 in OxyR binding sites (31 ) and at +6/-7 in CRP (30 ). The conservation can be explained by direct contacts or indirect through-water bonds.
(iii) At position ±6 there is no observed sequence conservation, yet methylation of a G in the major groove at that position interferes with Fis binding (12 ). This suggests that Fis passes close to the base in that region but does not make a specific contact.
(iv) At -4 and +3 the logo shows conservation of Cs or Ts, while at positions -3 and +4 the logo shows the complementary As or Gs. Because N7 is the only contact common to both A and G in the major groove (Fig. 2 b), this observation suggests that all four positions have N7 contacts (38 ). These contacts are consistent with DMS interference experiments (12 ). The relative heights of the letters reveal a 4.8-fold preference for A over G at -3 (T over C at +3), suggesting other direct contacts in the major groove or DNA bending effects.
(v) At positions 0, ±1 and ±2 there is an A-T region where Fis most likely faces the minor groove. Since A is as frequent as T but C and G are allowed at low frequency, this preference could be caused by a series of protein probes that sterically interfere with the N2 of G in the minor groove (Fig. 2 b) (30 ). Consistent with this, methylation of A at N3 in the minor groove at positions 0 and ±1 interferes with Fis binding (12 ,39 ).
(vi) The -4 to +4 central region of the Fis logo can be interpreted in a different way. We constructed a three-dimensional model of Fis-DNA binding predicted from the logo and probable contact points (see http://www-lmmb.ncifcrf.gov/~toms/fismodels/ for details). Mutations at Arg 85 and Lys 91 of Fis alter its ability to bind DNA (8 ,9 ), and molecular dynamics docking of Fis with DNA supports the notion that these residues contact the DNA (40 ). When we compressed a Fis binding site in the minor groove from -2 to +2 to account for the A-T region, kinked the DNA at ±3.5 and ±7.5 to create a bend at pyrimidine-purine pairs, and aligned Fis so that Arg 85 contacts G ±7 and Lys 91 contacts phosphate ±1.5, an unavoidable gap appeared that prevents direct contact between Fis and bases -4 to +4. Because our detailed model incorporates all the features observed in the sequence logo, and shows the same gap observed by others (5 ,6 ,10 ,11 ,40 ), the entire conservation from -4 to +4 might be accounted for by indirect contacts instead of direct contacts. Direct contacts represent physical contact between the protein and the DNA, while indirect `contacts' are those in which there is no direct contact but instead the structure of the DNA is distorted, indirectly leading to sequence conservation. Since molecular modeling is not entirely reliable, both the direct and the indirect binding modes are plausible and further experimental work would be required to distinguish between them. However, these two binding modes are not exclusive since it is possible that Fis can flex enough to bind to straight DNA using direct contacts to all of the bases. Subsequently, the DNA could bend using the bending properties of the central bases.
(vii) The sequence logo shows that Fis sites easily accommodate the Dam methylase site 5'-GATC-3' at ±4 through ±7, suggesting that under some circumstances Fis binding may be controlled by methylation. A Fis site in Tn5 is only bound when overlapping GATCs are unmethylated (41 ). The only other occurrence of GATC within a Fis site in our list (Fig. 1 ) is at 0 to +3 of oriC Fis 283, so there may be a connection between Fis and this feature of the origin of DNA replication, as suggested previously (42 ).
(viii) It is possible that the bases at one position of a Fis site are correlated to those in another position. For example, an A at -3 might only appear when there is an A at -2, but not when there is a T at -2. This would make the sequence logo an incomplete model because these are not displayed. The Diana program (43 ) shows only faint correlations between -20 and -19 (0.14 bits; P < 1 × 10-7 given the background of correlations from -20 to +20 of -0.02 ± 0.03 bits) and between -2 and -1 (0.12 bits; P < 1 × 10-6) and their complements. As these values are within the error of the total sequence conservation (±0.27 bits), there is little or no missing sequence conservation in the sequence logo model.
To model the base preferences of Fis, we computed a weight matrix from:
Riw(b,l) = 2 + log2f(b,l) - e(n) (bits per base)
1
where f(b,l) is the frequency of each base b at position l in the aligned binding site sequences and their complements, and e(n) is a sample size correction factor for the n = 120 sequences used to create f(b,l) (33 ). Riw(b,l) values range between -[infinity] and 2 bits. To evaluate a DNA sequence, the bases of the sequence are aligned with the matrix entries and the Riw(b,l) values corresponding to each base are added together to produce the total Ri value. This measure has several advantages over other methods. First, the scale is in bits, which are easy units to think about and which allow direct comparison to many other systems. Second, by adding the weights together for various positions in a particular binding site, we get the total `individual information' (Ri) for that site. Third, the average Ri for all of the binding sites used to create the Riw(b,l) matrix is the average information content, Rsequence (32 ). This is the same as the area under the sequence logo. Fourth, unlike a neural network that needs to be cyclically trained and requires both sites and non-sites, the matrix can be created immediately using only proven sites as examples. This avoids the danger of training against unknown functional sites, and therefore was critical for obtaining the results presented here. Fifth, functional binding sites have positive Ri values, within the error of the method, allowing one to make predictions. Finally, unlike consensus sequences which destroy the available sequence data by arbitrarily rounding the frequencies up or down, the individual information method uses the base frequences directly and so it preserves subtleties in the data.
We used individual information (33 ) to study Fis binding sites throughout the E.coli genome and at several specific loci. Although the theoretical cutoff for distinguishing sites from non-sites is 0 bits, we often used a conservative 2 bit cutoff to define Fis sites because our previous experience showed that sites between 0 and 2 bits can bind Fis (data not shown). When comparing the output from the Scan, DNAplot, MakeWalker and Lister programs to previously reported footprinting data, we consistently found sites which were seen as DNase I protected regions.
For example, by using a degenerate consensus pattern, previous workers found five Fis binding sites upstream of the transcriptional start site of the nrd operon of E.coli (44 ). When we scanned for potential Fis binding sites, several more sites were identified (Fig. 3 ). These were confirmed to be bona fide sites since Cu-phenanthroline footprinting of this region had already been done by Augustin et al. Their data (Fig. 3 , lanes 4 and 5) correspond well with our predictions even though none of these additional sites were used in the Riw(b,l) model. In another case, in addition to the site found at position -58 of the tgt/sec promoter (45 ), an information scan shows a second strong site at -73 (34 ). Both tgt/sec sites were included in our model because they are supported by footprinting, gel shift and in vivo transcriptional assays. Although Fis has a poor consensus sequence, theoretically it can bind precisely (46 ,47 ), and indeed footprints reveal concise binding on well separated sites. Complex footprints appear to be imprecise binding if one uses a consensus sequence. Often the protected genetic regions can be dissected into their components by using individual information tools, so that the data is interpreted as representing overlapping sites.
The total number of Fis sites in the E.coli genome is not known, so the information needed to locate those sites (Rfrequency) cannot be calculated (32 ). However, the total sequence conservation at the binding sites is 7.86 ± 0.27 bits (Fig. 2 ), which suggests that there is one site roughly every 27.86±0.27 = 232 ± 43 bases or an average of 4.7 ± 0.9 sites at each of the ~4289 genes of the entire 4 638 858 bp E.coli genome (GenBank accession no. U00096, version of 16-JAN-1997). It also implies that ~20 000 ± 3700 Fis molecules would be needed to fill all Fis sites on a single chromosome. Using the method of individual information we scanned the genome and found 68 552 Fis sites with >2 bits of sequence conservation. These estimates are comparable to the number of Fis molecules per cell, which ranges from close to zero in stationary cells to between 25 000 and 50 000 Fis dimers per cell during the transition to exponential growth or an increase in nutrients (18 ). Thus, almost every Fis site could be filled by one Fis dimer under those growth conditions.
Fis is an autoregulatory protein with six strong binding sites and a number of lower-affinity sites near its promoter (18 ,19 ). A scan of the E.coli fis promoter shows up to 12 additional sites ( >= 2 bits) in the immediate region of the promoter, but few downstream (Fig. 4 a and b). Presumably the additional sites correspond to the weaker sites noted by Ball et al. (18 ).
Fis sites have been identified on recombinational enhancers (3 ,4 ,12 ,13 ,49 ,50 ). In the S.typhimuriumhin region there are two Fis sites that are proximal and distal to the hixL recombination site. An information scan of this region shows a third potential Fis site located 11 bp [~1 helical turn of DNA, 10.6 bp (51 ,52 )] to the right of and overlapping the proximal site (Fig. 5 , top left). We call this site the `medial' site; it is 37 bp (~3.5 helical turns) to the left of the distal site. The same structure is found in bacteriophage Mu gin and p15B min enhancers (Fig. 5 , left three graphs).
In the bacteriophage P1 cin, bacteriophage P7 cin, and E.coli e14pin enhancers, a potential overlapping site occurs 7 bp (~1/2 helical turn) to the left of the previously identified proximal site (Fig. 5 , right three graphs). Since this potential site is outside the region between the proximal and distal sites, we named it the `external' site.
At recombinational enhancer proximal Fis sites, when a potential new Fis site is found on the right, it is 11 bases away while when a potential new Fis site is found on the left, it is 7 bases away. It is not known whether this correlation is coincidental. We also observed that potential Fis sites corresponding in location to site III in gin (9 ) appear in all other enhancers scanned except hin and that in three cases a weaker potential site falls exactly between the distal site and the one corresponding to site III with spacings of 11-12 bp.
Scanning the Fis Riw(b,l) model across DNA inversion regions reveals pairs of Fis sites spaced either 7 or 11 bases apart (Fig. 5 ). Three of the sites are the footprinted proximal sites. In addition, the medial site for gin is supported by footprint data, but it was not identified as a Fis site (9 ). To test whether the proposed medial site exists at hin we performed gel shift experiments on DNAs in which we presumably had knocked out neither, one, or both of the sites. The DNA design is shown in Figure 6 using sequence walkers, a graphical representation of the individual information content at specific binding sites (34 ). Characters representing the sequence are either oriented normally and placed above a line indicating favorable contact, or upside-down and placed below the line indicating unfavorable contact. Functional sites therefore have most letters pointing upwards, while those we have destroyed have many letters pointing downwards. The walkers also show that we did not inadvertently create any other Fis sites. The results of shifting the hin sequences of Figure 6 are shown in Figure 7 .
Under our experimental conditions hin does have a second site as predicted, since the knockout of the stronger proximal site still allowed the DNA to shift (Fig. 7 , HL). However, more Fis protein was required to shift an equivalent amount of DNA than for the wild-type proximal site, indicating that Fis binds weakly to the medial site. This is consistent with the weaker sequence conservation of the medial site (Ri = 6.9 bits) compared to the proximal site (Ri = 8.3 bits).
To further investigate the predictive ability of our individual information model, we synthesized five oligonucleotides representing various interesting sites:
(i) We were curious as to whether the predicted cin external site (7 bases from the proximal site, Ri = 2.8 bits), could bind Fis, even though it is so close to the proximal site (Fig. 5 ).(ii) Four lines of evidence indicate that a site predicted to be at -73 of the tgt/sec promoter (Ri = 10.8 bits, Fig. 1 , # 23) should bind Fis (see fig. 2 of ref. 34). We decided to test this prediction directly.(iii) Footprinting data covers two overlapping sites spaced 11 bases apart within the nrd promoter at -283 (nrdF1, Ri = 14.6 bits) and -272 (nrdF2, Ri = 5.0 bits, Fig. 1 , #19) relative to the start of transcription (Fig. 3 ), but only one site had been identified (44 ). We decided to test both.(iv) The site previously identified as Fis site II at 238 in the aldB promoter (53 ) has a negative Ri value (-5.3 bits) and therefore should not bind Fis.
We thank Reid Johnson for generously supplying Fis protein, Denise Rubens for technical assistance and for sequencing, the Frederick Biomedical Supercomputing Center for access to computer resources, Peter Rogan, R. M. Stephens, Keith Robison and Dhruba Chattoraj for comments on the manuscript.
1 Craig,N.L. and Kleckner,N. (1987) In Neidhardt,F.C., Ingraham,J.L., Low,K.B., Magasanik,B., Schaechter,M. and Umbarger,H.E. (eds) Escherichia coli and Salmonella typhimurium: Cellular and Molecular Biology. Vol 2. American Society for Microbiology, Washington, DC. pp. 1054-1070.
2 Johnson,R.C. and Simon,M.I. (1987) Trends Genet., 3, 262-267.
3 Finkel,S.E. and Johnson,R.C. (1992) Mol. Microbiol., 6, 3257-3265.MEDLINE Abstract
4 Finkel,S.E. and Johnson,R.C. (1992) Mol. Microbiol., 6, 1023.
60 Sambrook,J., Fritsch,E.F. and Maniatis,T. (1989) Molecular Cloning, A Laboratory Manual. 2nd edition. Cold Spring Harbor Laboratory Press, Cold Spring Harbor, New York.
61 Hengen,P.N. and Iyer,V.N. (1992) BioTechniques, 13, 57-62.
62 Studier,F.W. and Moffatt,B.A. (1986) J. Mol. Biol., 189, 113-130.MEDLINE Abstract
68 Hengen,P.N. (1995) In Griffin,A.M. and Griffin,H.G. (eds) Molecular Biology: Current Innovations and Future Trends. volume Part 1. Horizon Scientific Press, Wymondham, UK. pp. 39-50.
79 Kahmann,R., Mertens,G., Klippel,A., Bräuer,B., Rudt,F. and Koch,C. (1987) In McMacken,R. and Kelly,T.J. (eds) DNA Replication and Recombination. Alan R. Liss, Inc, New York. pp. 681-690.
B. Galan, I. Manso, A. Kolb, J. L. Garcia, and M. A. Prieto The role of FIS protein in the physiological control of the expression of the Escherichia coli meta-hpa operon
Microbiology,
July 1, 2008;
154(7):
2151 - 2160.
[Abstract][Full Text][PDF]
S. I. Husnain and M. S. Thomas Downregulation of the Escherichia coli guaB promoter by FIS
Microbiology,
June 1, 2008;
154(6):
1729 - 1738.
[Abstract][Full Text][PDF]
B.-K. Cho, E. M. Knight, C. L. Barrett, and B. O. Palsson Genome-wide analysis of Fis binding in Escherichia coli indicates a causative role for A-/AT-tracts
Genome Res.,
June 1, 2008;
18(6):
900 - 910.
[Abstract][Full Text][PDF]
Z. Chen, K. A. Lewis, R. K. Shultzaberger, I. G. Lyakhov, M. Zheng, B. Doan, G. Storz, and T. D. Schneider Discovery of Fur binding site clusters in Escherichia coli by information theory models
Nucleic Acids Res.,
November 7, 2007;
35(20):
6762 - 6777.
[Abstract][Full Text][PDF]
M. D. Bradley, M. B. Beach, A. P. J. de Koning, T. S. Pratt, and R. Osuna Effects of Fis on Escherichia coli gene expression during different growth stages
Microbiology,
September 1, 2007;
153(9):
2922 - 2940.
[Abstract][Full Text][PDF]
R. K. Shultzaberger, L. R. Roberts, I. G. Lyakhov, I. A. Sidorov, A. G. Stephen, R. J. Fisher, and T. D. Schneider Correlation between binding rate constants and individual information of E. coli Fis binding sites
Nucleic Acids Res.,
August 13, 2007;
35(16):
5275 - 5283.
[Abstract][Full Text][PDF]
R. Quatrini, C. Lefimil, F. A. Veloso, I. Pedroso, D. S. Holmes, and E. Jedlicki Bioinformatic prediction and experimental verification of Fur-regulated genes in the extreme acidophile Acidithiobacillus ferrooxidans
Nucleic Acids Res.,
April 1, 2007;
35(7):
2153 - 2166.
[Abstract][Full Text][PDF]
R. K. Shultzaberger, Z. Chen, K. A. Lewis, and T. D. Schneider Anatomy of Escherichia coli {sigma}70 promoters
Nucleic Acids Res.,
February 16, 2007;
35(3):
771 - 788.
[Abstract][Full Text][PDF]
P. Bruscella, C. Appia-Ayme, G. Levican, J. Ratouchniak, E. Jedlicki, D. S. Holmes, and V. Bonnefoy Differential expression of two bc1 complexes in the strict acidophilic chemolithoautotrophic bacterium Acidithiobacillus ferrooxidans suggests a model for their respective roles in iron or sulfur oxidation
Microbiology,
January 1, 2007;
153(1):
102 - 110.
[Abstract][Full Text][PDF]
S. Hasan and M. Schreiber Recovering motifs from biased genomes: application of signal correction
Nucleic Acids Res.,
October 6, 2006;
(2006)
gkl676v3.
[Abstract][Full Text][PDF]
E. Bindewald, T. D. Schneider, and B. A. Shapiro CorreLogo: an online server for 3D sequence logos of RNA and DNA alignments.
Nucleic Acids Res.,
July 1, 2006;
34(Web Server issue):
W405 - W411.
[Abstract][Full Text][PDF]
L. S. Feldman-Cohen, Y. Shao, D. Meinhold, C. Miller, W. Colon, and R. Osuna Common and Variable Contributions of Fis Residues to High-Affinity Binding at Different DNA Sequences
J. Bacteriol.,
March 15, 2006;
188(6):
2081 - 2095.
[Abstract][Full Text][PDF]
S. Ferenczi, L. Orosz, and P. P. Papp Repressor of Phage 16-3 with Altered Binding Specificity Indicates Spatial Differences in Repressor-Operator Complexes
J. Bacteriol.,
February 15, 2006;
188(4):
1663 - 1666.
[Abstract][Full Text][PDF]
M. Hirsch and T. Elliott Fis Regulates Transcriptional Induction of RpoS in Salmonella enterica
J. Bacteriol.,
March 1, 2005;
187(5):
1568 - 1580.
[Abstract][Full Text][PDF]
CiteULike 
Connotea 
Del.icio.us What's this?
