Methylation of CpG dinucleotides is the major epigenetic modification of mammalian genomes, critical for regulating chromatin structure and gene activity. The mixed‐lineage leukaemia (MLL) CXXC domain selectively binds nonmethyl‐CpG DNA, and is required for transformation by MLL fusion proteins that commonly arise from recurrent chromosomal translocations in infant and secondary treatment‐related acute leukaemias. To elucidate the molecular basis of nonmethyl‐CpG DNA recognition, we determined the structure of the human MLL CXXC domain by multidimensional NMR spectroscopy. The CXXC domain has a novel fold in which two zinc ions are each coordinated tetrahedrally by four conserved cysteine ligands provided by two CGXCXXC motifs and two distal cysteine residues. We have identified the CXXC domain DNA binding interface by means of chemical shift perturbation analysis, cross‐saturation transfer and site‐directed mutagenesis. In particular, we have shown that residues in an extended surface loop are in close contact with the DNA. These data provide a template for the design of specifically targeted therapeutics for poor prognosis MLL‐associated leukaemias.
In human leukaemia, the mixed‐lineage leukaemia (MLL) gene is a frequent target for recurrent specific chromosomal translocations (Djabali et al, 1992; Gu et al, 1992; Tkachuk et al, 1992; Corral et al, 1993; Domer et al, 1993; Thirman et al, 1993) that result in the generation of novel chimaeric fusions between MLL and over 30 different partner genes (Daser and Rabbitts, 2005). MLL is the human homologue of the Drosophila trithorax gene, and is required for the maintenance of Hox gene expression during mammalian development for the establishment of body segment identity (Yu et al, 1995, 1998). MLL is required for Hox‐dependent expansion of normal haematopoietic progenitors (Hess et al, 1997; Yagi et al, 1998; Ernst et al, 2004a, 2004b) and transformation of myeloid progenitors by MLL fusion proteins is dependent on specific Hoxa genes (Nakamura et al, 2002; Ayton and Cleary, 2003; Kumar et al, 2004; So et al, 2004; Zeisig et al, 2004; Wang et al, 2005).
The MLL protein is an SET domain‐dependent histone H3 lysine 4 (K4)‐specific methyltransferase that exists as part of a multiprotein supercomplex of at least 29 proteins (Milne et al, 2002; Nakamura et al, 2002). H3‐K4 methylation status correlates with an active transcriptional state (Strahl et al, 1999; Noma et al, 2001), and provides the molecular basis whereby Hox gene expression is maintained by the MLL protein (Yu et al, 1998). However, the mechanisms by which wild type or oncogenic MLL fusion proteins are recruited to specific target genes in a chromatin context are poorly understood. The amino‐terminal region of MLL contains a cysteine‐rich CXXC domain (zf‐CXXC; Pfam PF02008), characterised by two CGXCXXC repeats, which is also present in a number of other chromatin‐associated proteins. These include the methyl‐CpG binding domain protein (MBD1) (Cross et al, 1997), DNA methyltransferase 1 (DNMT1) (Bestor and Verdine, 1994), the major DNA maintenance DNA methyltransferase, CpG binding protein (CGBP), a component of the mammalian Set1 H3‐K4 methyltransferase complex (Lee and Skalnik, 2005) and FBXL11, recently characterised as a histone demethylase that specifically demethylates histone H3 at lysine 36 (Tsukada et al, 2006). The CXXC domain is retained in all MLL fusion proteins and is essential for target gene recognition, transactivation and myeloid transformation (Ayton et al, 2004). The CXXC domain of several proteins, including MLL, has been shown to bind to nonmethyl‐CpG dinucleotides (Lee et al, 2001; Birke et al, 2002; Ayton et al, 2004; Jorgensen et al, 2004). Cytosine methylation is the major epigenetic DNA modification in eukaryotes, and in vertebrates is found almost exclusively in a 5′ CpG context where it functions to maintain stable gene silencing through mitotic cell divisions. DNA methylated at the cytosine of CpG dinucleotides is found in transcriptionally inactive genes, whereas actively expressed genes are generally hypomethylated (Cross and Bird, 1995). The CXXC domain may therefore play an important role in directing MLL to transcriptionally active genes. To understand the molecular basis of nonmethyl‐CpG DNA recognition, we have determined the solution structure of the MLL CXXC domain. By combining NMR spectroscopy with chemical shift perturbation analysis, cross‐saturation transfer, site‐directed mutagenesis and mass spectrometry, we have identified the DNA binding interface and revealed residues that are critical for DNA binding and maintaining the fold of the MLL CXXC domain. These studies provide a structural basis for understanding how vertebrates interpret the methylation status of CpG dinucleotides and provide a framework for the development of novel therapeutics for the treatment of the poor prognosis MLL‐related leukaemias.
The NMR spectra of residues V1146 to K1214 of MLL were assigned and the solution structure determined using standard techniques (Wüthrich, 1986; Bax, 1994). Residues R1150–P1201 adopt a well‐defined tertiary structure with an r.m.s. deviation of 0.4 Å for backbone atoms. The N‐terminal residues V1146–G1149 and C‐terminal residues S1202–K1214 are unstructured as judged by a lack of long‐range NOEs and negative heteronuclear NOE values (data not shown) and were excluded from the statistical analysis. Experimental restraints and structural statistics for the 20 accepted lowest energy structures are summarised in Table I. The coordinates for the structure are available from the Protein Data Bank (entry code 2j2s).
The CXXC domain adopts an extended crescent‐like structure that incorporates two Zn ions (Figure 1A–C). The presence of zinc and the metal binding stoichiometry (zinc:protein 2:1) was established by induction coupled plasma mass spectrometry (data not shown) and mass spectrometry under native and denaturing conditions (Table II). Each of three cysteine residues in the two CGXCXXC motifs provides a ligand for the coordination of a Zn ion (Figure 1C). Both motifs adopt a very similar conformation in which the second and third cysteine residues lie within a small helix (residues T1171–L1174) or form part of a small helix‐like turn (residues P1159–Q1162). The residue that follows the first cysteine has a positive phi angle, which accounts for the strong preference for glycine at this position. After the second motif the main chain changes direction by 180° to enable C1189 and C1194 to provide the additional fourth ligand for coordinating Zn, together with the three cysteines from the second and first CGXCXXC motifs, respectively.
The topology of the fold is primarily dictated by the pattern of Zn coordination and contains little regular secondary structure. Residues R1151–R1154 and L1197–M1200 form a short two‐stranded antiparallel β‐sheet that places the N and C termini close together in an arrangement seen in many protein modules. Following the small helix in the second CGXCXXC motif is a 310 helix (residues P1177–F1179), in which F1179 packs onto G1168 of the second CGXCXXC motif (Figure 2). The gamma and delta carbon atoms of the K1178 pack onto the aromatic ring of F1179, while the zeta nitrogen is close to the carboxyl group of D1166. The packing of these helices dictates the overall structure of the turn and acts as the scaffold for an extended surface loop. This loop begins with the break in structure caused by the sequential glycines G1180 and G1181 and ends with the distal Zn ligand residue, C1189. There are several charge–charge interactions, the most notable being a salt bridge between D1166 and R1192. D1166 is situated between the two CGXCXXC motifs while R1192 lies in the helix‐like turn (residues 1190–1192) between the two cysteine residues that provide the fourth zinc ligands for each motif. Hydrogen exchange experiments showed that amide protons are protected only in the CGXCXXC motifs and the helix‐like turn (residues K1190–R1192) between the fourth zinc ligands. Amide protons in more peripheral parts of the structure, such as the helix‐like turn (residues P1159–Q1162) and the β‐sheet, exchange rapidly with the solvent.
The structure of the CXXC domain differs from that of other Zn‐dependent binding motifs (Krishna et al, 2003) and a search for structurally similar proteins using the program DALI produced no hits (Holm and Sander, 1995). However, this type of Zn ligation has been seen previously in the CGXCXXC motif of the RecQ family of helicases (Bernstein et al, 2003). However, in RecQ, a cysteine residue N‐terminal to the motif provides the fourth Zn ligand.
A structure‐based sequence alignment of CXXC domains is shown in Figure 3. The CXXC domain is highly conserved in MLL proteins from different species. For example, there is only one amino‐acid change between the Homo sapiens and Fugu rubripes proteins (Caldas et al, 1998). The CXXC domain of the human MLL paralogue MLL4 is less well conserved (FitzGerald and Diaz, 1999; Huntsman et al, 1999). Other CXXC domains are more diverse in sequence. The residues that provide the ligands for the Zn ions are, however, strictly conserved, and it is likely that all the other CXXC domains have a similar overall fold to that of MLL. Residue R1192 is invariant among all CXXC domains, and D1166 is highly conserved or shows conservative substitution to glutamate. These residues form the salt bridge described above and their strict conservation would seem to indicate importance to the structure. Comparison of the three CXXC domains within the MBD1 protein (only one of which, CXXC‐3, shown as MBD1c in Figure 3, binds DNA) suggests residues that may be critical for CpG recognition (Jorgensen et al, 2004). Residues K1178–G1181, discussed above, comprise a KFGG motif that is conserved in other CXXC domains known to bind DNA. Q1187 follows an identical pattern of conservation to that of the KFGG motif suggesting functional importance. This is particularly apparent in the MBD1 protein in which the only CXXC domain to maintain both the KFGG motif and residue Q1187 is the one that binds DNA.
A series of isothermal titration calorimetry (ITC) experiments was conducted to measure binding of the CXXC domain to DNA palindromes of 12, 16 and 20 base pairs (bp), all with a unique and centrally positioned CpG dinucleotide. No significant difference in binding affinity was observed for different length of DNA (data not shown) and subsequent work was carried out using 12 bp duplexes. The CXXC domain binds CpG 12‐mer DNA with a Kd of 4.3 μM (standard error of 0.4 μM from three independent determinations) and an enthalpy of complex formation of 1.4 kcal mol−1 at 22°C under the conditions of the ITC buffer (Figure 4). Binding to CpG 12‐mer DNA was measured at a series of temperatures to determine the ΔCp for binding. A value of −0.3 kcal mol−1 K−1 was determined (data not shown), consistent with other known protein–DNA interactions (Peters et al, 2004). There was no evidence of binding to the same DNA containing a central methyl‐CpG under these conditions. A 12 bp DNA containing a central GpC showed only minor heat effects above the baseline, possibly indicating significantly weaker binding for this sequence, but this was too small to be analysed. These findings are consistent with the results obtained in previous studies on DNA binding by this domain using other techniques (Birke et al, 2002; Ayton et al, 2004).
The DNA binding site was localised by monitoring the changes in the 2D 1H‐15N‐HSQC spectra of the MLL CXXC domain upon the addition of a 12‐bp DNA duplex containing a central CpG dinucleotide. DNA binding significantly alters the NMR spectrum of the domain (Figure 5A) with many residues undergoing large changes in chemical shift (Figure 5B). The majority of these residues are located on one face of the CXXC domain (Figure 6A). This region of the protein contains many positively charged amino acids (Figure 7A) consistent with it being the DNA binding site. Cross‐saturation transfer experiments, which provide direct information on through space interactions (Ramos et al, 2000; Lane et al, 2001), were also employed to precisely identify residues at the DNA binding surface. Specific saturation of the imino protons of the DNA resulted in an attenuation of peak intensity in the 1H‐15N HSQC spectrum for residues R1182–C1188 (Ramos et al, 2000). These residues lie within an extended surface loop (compare Figures 2 and 6B).
A series of mutations was devised to identify the role that individual residues play in the function and stability of the domain. DNA binding was tested by gel‐shift assays (Figure 7B). Mutations that disrupted the native fold of the CXXC domain were detected by performing mass spectrometry under native conditions (Table II).
Mutation of any of the cysteine residues involved in Zn ligation led to unfolding of the domain. Furthermore, disruption of the conserved salt bridge between D1166 and R1192 also unfolded the protein. Mutations R1153A, Q1162A, N1172A, Q1195A and N1196A had no effect on either stability or DNA binding. These residues are on the opposite face of the CXXC domain to that implicated in DNA binding. Mutation of residues R1151, R1154, D1175, K1176, K1178, F1179, K1185, K1186, Q1187 and K1193 abolished or significantly decreased DNA binding, but had no effect on the global fold. For most of these mutations, the lack of DNA binding activity is likely to be solely the result of the removal of a functionally important side chain. Furthermore, most of these residues localise to the binding face of the domain as discussed above (see Figure 7A). Other mutations may, however, perturb local structure that is important for binding without the residues themselves playing a direct role. For example, as discussed above, the conserved K1178 and F1179 are packed in such a manner that they significantly contribute to the local structure (see Figure 2) and mutation of either residue is found to impair DNA binding. Furthermore, mutation of the KFGG motif (residues K1178–G1181) to alanine destabilises the protein sufficiently to cause it to unfold (see Table II).
Residues important for DNA binding
Taken together, our NMR binding and mutagenesis data clearly delineate the DNA binding interface of the MLL CXXC domain (Figures 6A, B, 7A and B). In particular, there are two distinctive features of the domain with potential relevance to binding. A positively charged groove runs along the DNA binding face of the domain consisting of residues shown to abolish or significantly decrease DNA binding upon mutation (R1154, K1176, K1178, K1186, K1193) (Figure 7A and B). There is also a surface patch at the tip of the domain corresponding to residues R1182–C1188 of the extended loop that are all shown to make direct contact with protons in the DNA by cross‐saturation experiments. Methylation of cytosine at the 5′ position places a methyl group in the major groove of the DNA. One could envisage a model whereby the positively charged groove on the binding surface interacts with the DNA phosphate backbone, while the residues of the extended loop insert into the major groove to probe the methylation state of the CpG dinucleotide.
The MBD1 transcriptional repressor is unique in that it contains both a methyl‐CpG binding domain and a CXXC domain that binds specifically to nonmethyl‐CpG. This allows MBD1 to interpret the CpG dinucleotide as a repressive signal in vivo regardless of its methylation status (Jorgensen et al, 2004). Taken together, the architectures of the MLL CXXC domain and the MBD domain of the MBD1 protein provide a structural basis for understanding how vertebrates interpret the methylation status of DNA, the major epigenetic DNA modification in eukaryotes. Although the CXXC domain of MLL is known to be required for the transforming activity of MLL fusion proteins, the biochemical role it plays in this process has not been fully defined. The mutation of several residues in the CXXC domain has been shown to abolish both DNA binding and prevent myeloid transformation (Ayton et al, 2004). To date, however, these mutations have either removed zinc ligands, which would perturb the structure of the domain, or have involved changes in multiple residues that we have shown result in unfolding of the protein (Table II). It is also possible that these types of mutation could affect other activities of the domain. In addition to DNA binding, the CXXC domain of MLL also recruits the polycomb repressor proteins HPC2 and BMI‐1, and the corepressor CTBP (Xia et al, 2003) and forms part of a low‐affinity binding site for the menin tumour suppressor oncogenic cofactor (Yokoyama et al, 2005). With the structure of the CXXC domain now available, it will be possible to design nondisruptive mutations that can help to define the role of individual residues in these activities, and thus promoting a deeper understanding of the role of the CXXC domain in transformation by MLL fusion proteins. The CXXC domain is retained in all forms of leukaemogenic MLL fusion proteins, including partial tandem duplications (Lochner et al, 1996) and internal PHD finger 1 deletions (Chaplin et al, 2001; von Bergh et al, 2001; Deveney et al, 2003; Morel et al, 2003). Thus, novel approaches to the treatment of MLL‐associated leukaemias might involve addressing the continued occupancy of key target genes by MLL fusion proteins by disrupting the interaction between the CXXC domain and nonmethyl‐CpG DNA. Our structure provides a potential template for the development of such novel reagents for the treatment of poor prognosis MLL‐related leukaemias.
Materials and methods
Preparation of the protein and the DNA
The CXXC domain of the MLL protein used for NMR and calorimetry, corresponding to residues V1146–K1214, was cloned into a modified pET24a plasmid (Novagen) that expresses proteins fused to the lipoyl domain of Bacillus stearothermophilus dihydrolipoamide acetyltransferase. The fusion protein was expressed in the E. coli strain Tuner [DE3] (Novagen). For isotope labelling, K‐MOPS minimal medium containing 15N‐NH4Cl and/or 13C‐glucose was used. The fusion protein was initially purified by Ni2+‐chelating sepharose affinity chromatography. Subsequent TEV protease digestion and Ni2+‐chelating sepharose affinity chromatography removed the lipoyl domain fusion‐tag. The CXXC domain was concentrated and then gel‐filtered through a Superdex 75 (Amersham) column and the fractions containing the CXXC domain pooled. All DNA was supplied as an HPLC purified powder by Operon Biotechnologies, Inc. DNA was dissolved in buffer (20 mM MES pH 6.5, 250 mM NaCl, 5 mM β‐mercaptoethanol) before being annealed by heating to 90°C for 10 min and cooled slowly to room temperature.
MLL CXXC domain used for gel shift assays (residues T1136–K1208) was expressed with an N‐terminal His6‐tag in Escherichia coli strain C41 (DE3). The protein was purified using Ni2+‐NTA affinity resin (Qiagen) and resource S ion‐exchange (Amersham) and dialysed into 10 mM Tris, 150 mM NaCl, 1 mM DTT, pH 7.4.
Isothermal titration calorimetry
Experiments to determine the DNA binding characteristics of the CXXC domain utilised a number of palindromic DNA sequences of differing length and composition:
- CpG 12‐mer:
- CpG 16‐mer:
- CpG 20‐mer:
- GpC DNA:
- Methyl‐CpG DNA:
- 5′‐GTATC (5MeC)GGATAC‐3′
Samples of CXXC domain and DNA were dialysed extensively against ITC buffer (20 mM MES pH 6.5, 250 mM NaCl, 5 mM β‐mercaptoethanol) prior to the experiment. DNA was concentrated to ∼50 μM dsDNA prior to dialysis. Final DNA concentrations were determined spectroscopically after filtering and loading of the calorimetric cell by measuring absorbance at 260 nm, assuming 50 μg−1 ml−1 A260 unit−1. Final concentrations of DNA ranged from 47–54 μM except for methyl‐CpG DNA, which for reasons of solubility, was 30 μM. CXXC domain was concentrated to ∼1.2 mM prior to dialysis. After filtering it was loaded into the syringe and the final CXXC domain concentration was determined spectroscopically at 280 nm, given an extinction coefficient ε280 of 6990 cm−1 M−1. Final CXXC domain concentrations ranged from 1.0–1.3 mM. ITC experiments were performed at 22°C using a high‐precision VP‐ITC system (Microcal Inc.).
Experiments were conducted such that the heat change was measured over 250 s following a 10 μl injection for either 20 or 25 injections. Included was an initial preinjection of 3 μl, according to the manufacturer's recommendation to counter diffusion of samples during the thermal equilibration. Analysis was carried out using Microcal Origin Software. Individual injections were integrated following manual adjustment of the baselines. Heats of dilution and mixing were determined from separate control experiments or from the end point of the titration. This value was subtracted prior to curve fitting using a one‐site model.
The NMR spectra were recorded on Bruker Advance‐800, Advance‐600 and AMX‐500 spectrometers. 2D NOESY, TOCSY, DQF‐COSY, 15N‐HSQC, constant‐time 13C‐HSQC and 3D HNCACB, CBCACONH, HNCO, HNCACO HNHB, 15N‐NOESY, 15N‐TOCSY were recorded at 290 K. The mixing times chosen were 55 ms for TOCSY, and 120 ms for NOESY. Spectra were referenced relative to external sodium 2,2‐dimethyl‐2‐silapentane‐5‐sulfonate, for signals of proton and carbon, or liquid ammonium for that of nitrogen. Approximately half the Hβ resonances were assigned stereospecifically using a combination of HNHB and DQF‐COSY spectra. All the Val Hγ and Leu Hδ resonances were assigned stereospecifically using a 10% 13C‐labelled sample of CXXC domain (Neri et al, 1989). All the NMR spectra were analysed with ANSIG v3.3 (Kraulis et al, 1994).
All NMR sample concentrations were 1.0 mM and were prepared in 20 mM MES pH 6.5, 250 mM NaCl, 10% D2O. The CXXC domain was concentrated to 1.5 mM for spectroscopic measurements of the free form of the domain. For binding studies, protein–dsDNA complexes were made with a stoichiometry of 1:1 (protein: dsDNA).
For hydrogen exchange experiments, the 15N‐labelled CXXC domain was exchanged into NMR buffer containing 100% D2O using a NAP‐10 column (Amersham) and a series of 1H‐15N‐FHSQC spectra (Mori et al, 1995) were recorded over the course of 24 h.
Resonances of the bound form of the CXXC domain were reassigned using standard triple resonance techniques (Wüthrich, 1986; Bax, 1994). A cross‐saturation transfer period similar to that described by Ramos et al (2000) was incorporated immediately prior to the first 1H pulse of an FHSQC sequence (Mori et al, 1995). Saturation of the DNA imino proton resonances was achieved via a pulse train of 15 ms hyperbolic secant inversion pulses, which were centreed at δ1H=13 p.p.m. Saturation transfer periods were 0.360, 0.720 and 1.440 s, and the overall relaxation delay was kept constant at 1.94 s. To avoid sample heating effects, the pulse sequence contained an identical train of compensation pulses centreed at δ1H=−5 p.p.m., which were executed so that the overall number of pulses was kept constant. Attenuation was measured relative to control experiments executed at the beginning and end of the series of experiments with errors extracted from these controls.
The distance constraints derived from the NOESY spectra were classified into four categories corresponding to inter‐proton distance constraints of 1.8–2.8, 1.8–3.5, 1.8–4.75 and 1.8–6.0 Å, respectively. Hydrogen bond constraints of 1.8–2.1 Å were imposed on the distance between the hydrogen and the acceptor oxygen, while another constraint of 2.7–3.1 Å was imposed on the distance between the donor nitrogen and the acceptor oxygen. Artificial restraints were added to represent the constraints imposed by coordination of the zinc ions. Six sulphur–sulphur distance constraints of 3.55–3.95 Å and four zinc–sulphur distance constraints of 2.25–2.35 Å were incorporated for each of the zinc–cysteine clusters. Torsion angle constraints were obtained from stereo‐specific assignment of residue side chains and incorporated in the structure calculation, along with the backbone ϕ and ψ angle constraints determined with the program TALOS (Cornilescu et al, 1999). The structures were calculated using a standard torsion angle dynamics simulated annealing protocol with the program CNS (Brunger et al, 1998). Twenty structures were accepted where no distance violations were greater than 0.25 Å and no angle violations were greater than 5.0°.
Electrospray ionisation mass spectrometry (ESI‐MS)
Mass spectra of wild type or mutant proteins were generated on an LCT time‐of‐flight mass spectrometer with electrospray ionisation (ESI) (Micromass, Altrincham, UK). Before MS analysis, the protein samples were desalted by dialysis against water. For analysis in denaturing conditions, samples were diluted to 2 pmol ml−1 in 50% (v/v) methanol and 1% (v/v) formic acid. For analysis in native conditions, samples were diluted to 10 pmol ml−1 in 20 mM ammonium acetate buffer. The samples were infused into the ESI source at a flow rate of 10 ml min−1 using a Harvard Model 22 syringe infusion pump (Harvard Apparatus, Harvard, MA, USA) and calibration was performed in the positive ion mode using horse heart myoglobin. Typically, 60–80 scans were acquired and added to yield a mass spectrum. Molecular masses were obtained by deconvoluting the multiply charged protein mass spectra using the software package, MassLynxTM Version 4.0 (Micromass). Theoretical molecular masses of wild type and mutant proteins were calculated using Protparam (us.expasy.org/tools/protparam.html). The zinc content of each protein was derived from the difference in mass between the native and denatured proteins.
Palindromic olignucleotides PALCpG (5′‐GTATCCGGATAC‐3′), PALGpC (5′‐GTATGGCCATAC‐3′) and PALmeCpG (5′‐GTATCmCGGATAC‐3′) were annealed in 10 mM Tris pH 7.4, 1 mM EDTA, 100 mM NaCl and buffer exchanged into 10 mM Tris pH 7.4 using a Microspin G‐25 column (Amersham). DNA binding reactions were carried out in 20 μl of binding buffer (10 mM Tris pH 7.4, 1 mM DTT, 150 mM NaCl) for 30 min at room temperature. Binding reactions contained a final dsDNA concentration of 10 μM and a two‐fold molar excess of purified protein. Binding reaction mixtures were electrophoresed in 0.7% agarose in TB (89 mM Tris‐borate, pH 8.3) buffer at 4°C and DNA was visualised by ethidium bromide staining.
We are grateful to Sew Peak‐Chew for help with the ESI‐MS; Dr Keith Sinclair, Elsie Widdowson Laboratory, MRC Human Nutrition Research Laboratory for ICP‐MS studies, Dr A Andreeva for help with sequence alignment. This study was supported by the Leukaemia Research Fund, the Kay Kendall Leukaemia Fund and an MRC Senior Clinical Fellowship.
- Copyright © 2006 European Molecular Biology Organization