C9orf72 nucleotide repeat structures initiate molecular cascades of disease - PubMed (original) (raw)

. 2014 Mar 13;507(7491):195-200.

doi: 10.1038/nature13124. Epub 2014 Mar 5.

Christopher J Donnelly 2, Goran Periz 1, Eric A J Simko 1, Patrick G Shaw 3, Min-Sik Kim 3, Nicholas J Maragakis 4, Juan C Troncoso 5, Akhilesh Pandey 3, Rita Sattler 2, Jeffrey D Rothstein 6, Jiou Wang 1

Affiliations

C9orf72 nucleotide repeat structures initiate molecular cascades of disease

Aaron R Haeusler et al. Nature. 2014.

Abstract

A hexanucleotide repeat expansion (HRE), (GGGGCC)n, in C9orf72 is the most common genetic cause of the neurodegenerative diseases amyotrophic lateral sclerosis (ALS) and frontotemporal dementia (FTD). Here we identify a molecular mechanism by which structural polymorphism of the HRE leads to ALS/FTD pathology and defects. The HRE forms DNA and RNA G-quadruplexes with distinct structures and promotes RNA•DNA hybrids (R-loops). The structural polymorphism causes a repeat-length-dependent accumulation of transcripts aborted in the HRE region. These transcribed repeats bind to ribonucleoproteins in a conformation-dependent manner. Specifically, nucleolin, an essential nucleolar protein, preferentially binds the HRE G-quadruplex, and patient cells show evidence of nucleolar stress. Our results demonstrate that distinct C9orf72 HRE structural polymorphism at both DNA and RNA levels initiates molecular cascades leading to ALS/FTD pathologies, and provide the basis for a mechanistic model for repeat-associated neurodegenerative diseases.

PubMed Disclaimer

Figures

Figure 1

Figure 1. DNA and RNA of the C9orf72 HRE form G-quadruplexes

a) The CD absorptivity shows characteristic K+-dependent spectra for antiparallel DNA G-quadruplexes with (GGGGCC)4. b) The presence of K+ during annealing induces a conformational change that decreases the mobility of (GGGGCC)4 DNA in a gel mobility shift assay. c) A DMS footprinting assay on (GGGGCC)4 DNA shows protection of the N7 positions on all the guanines when a G-quadruplex is formed. d) The proposed topology for the antiparallel DNA G-quadruplex formed by (GGGGCC)4. Each gray plane represents a G-quartet, as shown in the lower corner. Four separate G-quartets are stacked 5´ to 3´ with two cytosines forming each loop region. e) The CD spectra identified the formation of a parallel G-quadruplex for the RNA (GGGGCC)4 in the presence of 100 mM KCl. f) (GGGGCC)4 RNA demonstrates a slower mobility in the presence of KCl compared to (CCCCGG)4 RNA. g) The RNase T1 protection assay identifies single-stranded guanine residues (denoted in black) not involved in the formation of the G-quadruplex. h) The proposed parallel G-quadruplex topology formed by the RNA of the C9orf72 HRE. Three stacks form the parallel G-quadruplex, with the RNase T1-sensitive guanines shown as black dots.

Figure 2

Figure 2. Abortive transcription in the C9orf72 HRE

a) Increasing lengths of GGGGCC repeats cause accumulation of abortive transcripts in a length-dependent manner in vitro. The transcriptional products were separated on a denaturing gel with a 500 nt ssRNA control (CTL). b) Transcripts levels shown in (a) were densitometrically quantified and then plotted as the ratio of full-length transcripts that contain regions 3´ of the repeat divided by all transcripts that contain 5´ regions. The curve was fit to a single exponential. Data are means ± s.d. n = 4. c) The C9orf72 HRE induces the formation of R-loops on C9orf72 HRE-containing plasmids with (GGGGCC)~70. Treatment of the in vitro transcription products with RNase A and H digests the RNA still hybridized with relaxed or supercoiled plasmid and reduces the smearing that was caused by the size heterogeneity of RNA•DNA hybrids. Genomic DNA (top band) serves as an internal loading control. d) Patients carrying the C9orf72 HRE have reduced pre-mRNA 3´/5´ratios relative to C9orf72 WT, consistent with the HRE-induced abortive transcription reducing full-length transcript levels. Data are means ± s.e.m. n = 5/6 (B lymphocytes), n = 12/10 (motor cortex), n = 8/5 (spinal cord) for C9orf72 WT/HRE samples, respectively. ***P < 0.001, **P < 0.01, *P < 0.05.

Figure 3

Figure 3. Identification of conformation-dependent C9orf72 HRE RNA-binding proteins

a) Western blotting analyses of the sequential fractions eluted from a biotinylated-RNA pulldown of HEK293T cells with increasing KCl concentrations. The RNA-binding proteins identified using LC-MS were differentiated among those that recognize different structural motifs, including the RNA antisense (AS) hairpin, (CCCCGG)4; the RNA sense (S) hairpin (GGGGCC)4; and the sense G-quadruplex (G). NCL and hnRNP U have binding preferences for the G motif of the sense RNA. hnRNP F and RPL7 bind both guanine-rich sense sequences, regardless of the underlying RNA structure. hnRNP K prefers binding to the cytosine-rich AS. b) A representative spectrum from SILAC analysis is shown for the preferential binding of NCL to the G motif formed by (GGGGCC)4 compared to the hairpin motifs formed by the sense or antisense sequences. c) An RNA pulldown performed with GST-NCL demonstrates that NCL directly binds (GGGGCC)4 RNA in vitro with highest affinity for the G-quadruplex. Data are means ± s.d. n = 3.

Figure 4

Figure 4. Nucleolar stress is a result of repeat-containing RNA transcripts from the C9orf72 HRE

a) In the control C9orf72 WT B lymphocytes, NCL (green) is localized to the condensed nucleolus. In contrast, the cells of patients with the C9orf72 HRE show an increased NCL diffusion and fractured nucleoli in the nucleus (Hoechst, blue). A heat map of NCL intensities marks the difference between cells. b) iPS motor neurons derived from patients carrying the C9orf72 HRE also demonstrate NCL mislocalization. β-III tubulin (Tuj1) (red) was used to identify neurons. c) NCL colocalizes with RNA foci (red) formed in motor cortex tissue from patients carrying the C9orf72 HRE. A (CCCCGG)2.5 probe was used to detect the (GGGGCC)n RNA foci, a previously identified pathological feature of the C9orf72 HRE tissues. d) Transfection of (GGGGCC)21 abortive transcripts (Figure 2a) recapitulates NCL pathological features observed in patients cells with the C9orf72 HRE.

Figure 5

Figure 5. A model for the molecular cascade resulting from the C9orf72 HRE structural polymorphism

The DNA and RNA•DNA structures formed in the GGGGCC repeat region impede RNA polymerase transcription, which results in transcriptional pausing and abortion. This leads to a loss of full-length products and an accumulation of abortive transcripts. Abortive transcripts that contain the hexanucleotide repeats form G-quadruplexes and hairpins and bind essential proteins in a conformation-dependent manner. Sequestration of these proteins leads to nucleolar stress and other downstream defects. The repeat-containing transcripts can also escape the nucleus and be bound by ribosomal complexes, thereby increasing repeat-associated non-ATG-dependent translation that results in aggregative polydipeptides.

Extended Data Figure 1

Extended Data Figure 1. Spectroscopic characterization of DNA/RNA structural polymorphism

a) The decanucleotide GGGGCCGGGG from the C9orf72 HRE adopts a parallel G-quadruplex conformation indicated by the CD spectra in the presence of KCl (left). The proposed intermolecular parallel G-quadruplex topology (right) of two decanucleotide sequences that form three stacks of G-quartets 5´ to 3´ with three or four nucleotides, GCCG, forming a loop region. b) The complementary sequence strand for the C9orf72 HRE, (CCCCGG)4, shows no apparent structural differences ± KCl. c) The variability in CD spectra increases with the number of repeats (dotted lines), but the CD spectra can be recapitulated (lines) by fitting a linear combination of three component spectra for hairpin, antiparallel Gquadruplex, and parallel G-quadruplexes from the spectra corresponding to 0 mM KCl (GGGGCC)4, 100 mM KCl (GGGGCC)4, and 100 mM KCl GGGGCCGGGG, respectively. The residuals of the fit are plotted below. d) CD spectral analysis of the thermal stability of the DNA antiparallel G-quadruplex. CD spectra were obtained as described in Methods with temperatures ranging from 25–95°C in 10 mM Tris-HCl pH 7.4 with 50 mM KCl. This two-state transition was evaluated by singular value decomposition (SVD) of the CD melting spectra (Matlab). The two basis component spectra, shown as left coefficients, represent an antiparallel G-quadruplex and an unfolded single-stranded DNA oligo. The residuals show the difference between the original CD spectrum and the basis component spectra fit. e) Increasing KCl concentrations increase the abundance and stability of antiparallel DNA G-quadruplexes, as shown by UV-VIS melting curves. CD absorbance was normalized to the equation (Abst – min)*(max – min)−1, where Abst is the absorbance at a given temperature, max is the observed maximum absorbance at 295 nm at that [KCl], and min is the minimum value obtained for all [KCl]s. Data was fit to a Boltzmann distribution (Prism) with a lower boundary constraint set to min. f) The decamer HRE sequence forms a parallel G-quadruplex that is less stable than the antiparallel conformation. The data obtained were normalized as described above, except the CD absorptivity was measured at 260 nm. g) CD melting spectrum shows a two-state transition from a folded parallel RNA G-quadruplex state to an unfolded linear state. The CD absorptivity data measured at 260 nm were normalized as described above and fit to a Boltzmann distribution without constraints. h) The (GGGGCC)4 forms a G-quadruplex that is unperturbed by a 3ʹ biotin label. CD spectra were obtained ± KCl as in Figure 1 to identify the stabilization of K+ in the formation of a G-quadruplex that had been chemically modified with the biotin tag. All calculated melting temperatures are provide in Extended Data Table 1 and results detailed further in Supplementary Results.

Extended Data Figure 2

Extended Data Figure 2. Depurination of the HRE-containing plasmid template contributes minimally to abortive transcription

a) Schematic digestion map for analyzing depurination levels present on the template plasmid. b) The same patterns are shown for the (CCCCGG)~70 repeat template and a GFP control template with piperidine treatment and prior acid treatment, confirming that the repeat tract in the repeat-containing template is depurinated basally at nonspecific/spontaneous levels similar to those of the GFP template. The ~135 nucleotide 5´-end-radiolabeled cleavage fragment in the pCR8-70 samples serves as a size marker for the lower boundary of the repeat region corresponding to the region where the repeats would begin to appear on the gel. More details concerning these findings and the experimental designs are described in Supplementary Results and Methods. c) Densitometric plot profiles of the gel bands from each lane in (b) are shown to illustrate the depurination results. The left panel shows the small depurination differences between the piperidine-treated template without prior treatment and those that went through the denaturing/annealing treatments ± KCl. The middle and right panels show that treatment with piperidine causes similar patterns of minimal cleavage for the pCR8-70 and control GFP templates, respectively, as compared to the acid-induced depurination controls, consistent with basal levels of nonspecifically/spontaneously cleaved products. The densitometry analysis was done using ImageJ and plotted in Prism.

Extended Data Figure 3

Extended Data Figure 3. G-quadruplexes increase abortive transcripts within the C9orf72 HRE region in vitro

a) A newly developed BG4 nanobody directly binds the K+-dependent Gquadruplex formed by DNA (GGGGCC)4. BG4 was purified from E. coli, and an ELISA experiment using biotinylated DNA (GGGGCC)4, with or without 100 mM KCl during annealing, was used to determine the specificity of BG4, essentially as previously described. Data show the mean ± S.D. n = 3. b) The ΔCt graph for 100 mM KCl vs 0 mM KCl shows that BG4 pulled down >50-fold more plasmid when the plasmid containing (GGGGCC) repeats was annealed in the presence of 100 mM KCl, directly demonstrating the presence of Gquadruplexes formed by the sequence. The DNA pulldown with BG4 is described in the Supplementary Methods. n = 1. c) Formation of G-quadruplexes in the coding region prior to transcription causes a shift to an earlier termination of abortive transcripts and further loss of the full-length products. The shift is represented below the densitometric plot profile by a significant change in the center of mass for the repeat region (n = 5). The further loss of full-length transcriptional product is shown on the graph to the right of the transcription profile (n = 3). _P_-values were calculated using a paired fit for each trial assuming a parametric distribution. Data are means ± s.e.m. *P < 0.05; **P < 0.01. d) Accumulation of the abortive transcripts indicates that the repeat-dependent impairment of transcription is not transient but persists over time ± KCl. Aliquots were removed from the in vitro transcription reactions at different time points and then visualized and analyzed as in (c).

Extended Data Figure 4

Extended Data Figure 4. R-loops, and not G-quadruplex formation on nascent RNA transcripts, increase abortive transcription within the C9orf72 HRE region in vitro

a) RNA transcripts containing many GGGGCC repeats form G-quadruplexes under physiologically relevant KCl concentrations. A colorimetric assay was performed to identify the formation of RNA G-quadruplexes utilizing the enzyme-like peroxidase activity of G-quadruplex•hemin complexes. b) Workflow considerations for the transcriptional assay. The linear plasmid was first annealed ± 100 mM KCl or 100 mM NaCl in 10 mM Tris-HCl, pH 7.4. To prevent salt concentration-dependent effects on the in vitro transcriptional assay, a second adjustment was made to adjust the salts to a final 50 mM concentration in the assay. Reducing the effects on RNA polymerase allowed us to disambiguate the effects of salt on the conformation of DNA versus the nascent RNA. Comparison of DNA annealed in 0 mM KCl and 100 mM NaCl shows similar reduced polymerase processivity, suggesting that possible formation of G-quadruplexes on the nascent RNA transcripts makes a negligible contribution, but does not exclude intrinsic RNA hairpin-induced termination. c) RNase H treatment during transcription reduces the periodic accumulation of abortive transcripts and increases full-length transcripts. Treatment with RNase H, which specifically cleaves RNA•DNA hybrids, during transcription of the C9orf72 repeats causes a shift from truncated transcripts to full-length transcripts, suggestive of the formation of alternative secondary structures, known as R-loops, caused by increased nonduplex DNA during transcription. d) R-loop formation is observed for the plasmid containing the HRE insert, (GGGGCC), but not for a GFP insert in a control. Treatment with RNase A removes ssRNA but does not affect R-loops. As shown in Figure 2c, the HREs induce the formation of R-loops that cause a decrease in the plasmid mobility but can be relieved with RNase H treatment, which specifically cleaves RNA•DNA hybrids. The addition of RNase A and RNase H has little effect on the mobility of the plasmid containing the GFP insert. Radiolabeling the transcriptional products during in vitro transcription confirms the formation of R-loops, demonstrated by the shift consistent with the supercoiled and relaxed plasmids having altered mobility, which is relieved by treatment with both RNases. The transcripts were bodylabeled by including 20 μCi of α-[32P]UTP (Perkin Elmer) and then performing the in vitro transcription as previously described. e) There is a significant increase in the R-loop-induced plasmid mobility shift for the (GGGGCC) and (GGGGCC)~70 containing plasmids when compared to the GFP control insert. The shifted supercoiled plasmid bands were quantified by densitometrically measuring each band intensity after treatment with RNase A versus those after treatment with both RNase A and Rnase H (ImageJ, NIH). The overlapping densitometric signal of the supercoiled R-loop smear with the circular plasmid band (Figure 2b) prevented accurate quantification and was excluded. Data are means ± s.e.m. n = 3. *P < 0.05; **P < 0.01.

Extended Data Figure 5

Extended Data Figure 5. Workflow for the SILAC methodology used for identification and quantification of proteins that bind the hexanucleotide repeats in a structurally dependent manner

HEK293T cells that had complete incorporation of either heavy-, medium-, or lightlabeled isotopes were incubated with either the (GGGGCC)4 G-quadruplex, the (GGGGCC)4 hairpin, or the (CCCCGG)4 hairpin, respectively. The samples were then processed as described in Methods and presented in Figures 3. The final eluted samples were combined and prepared for trypsin digestion using filter-aided sample preparation (FASP). The buffer was exchanged to remove salts and SDS. The sample was then subject to LC-MS/MS analysis as described in the Methods.

Extended Data Figure 6

Extended Data Figure 6. Example of mass spectrometry spectra using SILAC obtained from the eluted fractions from the RNA pulldown

Four sets of spectra representing unique peptide sequences identified for hnRNP U, hnRNP F, RPL7, and hnRNP K show the quantitative RNA structural preferences that can be identified by using the SILAC methodology. In agreement with the work presented in Figure 3, each protein represented by the unique peptide spectrum here shows an individual preference for either RNA structure dependence, such as NCL (Figure 3) and hnRNP U, or sequence dependence, as in the cases of hnRNP K, hnRNP F, and RPL7, as shown by the abundance of the different isotope labels provided by the SILAC method. In Supplementary Table S1, the quantified differences are provided.

Extended Data Figure 7

Extended Data Figure 7. Nucleolin defects observed in C9orf72 HRE-linked ALS patients are caused by HRE-containing transcripts

a) Quantification of nucleolin (NCL) in the nucleus shows significant NCL nuclear dispersion in B lymphocytes, iPS motor neurons, and fibroblasts from C9orf72 HRE patients. A single focal plane was obtained through the center of the nucleus (Hoechst staining; blue). To quantify the area differences in NCL relative to the size of the nucleus a single lower threshold setting in ImageJ (NIH), ranging from 25–75 (which provided quantification of both dispersed NCL and dense nucleolar NCL), was used to measure the pixel area of NCL relative to the area of the nucleus outlined by the Hoechst staining (depicted in the cartoon). The measurements for the C9orf72 HRE cells were then normalized to those for controls (Prism). Data are means ± s.e.m. n = 110 and n = 112 for 3 cell lines examined for the C9orf72 WT and C9orf72 HRE B lymphocytes, respectively. n = 29 for iPS motor neurons from 2 C9orf72 HRE patients and for the iPS motor neuron controls. A representative fibroblast line for C9orf72 WT, SOD1 D90A ALS, and C9orf72 HRE was quantified. n = 21. *P < 0.05; **P < 0.01; ****P < 0.0001. b) There are no significant changes in NCL protein levels between patients with and without the C9orf72 HRE, despite its mislocalization. The Western blotting analysis was performed as described in the Methods, following a typical protocol with the B lymphocyte lysates from patients and controls. The bands were visualized using a quantitative LI-COR imager, and the intensities for NCL and tubulin protein were quantified using Odyssey software. NCL levels were normalized against those of tubulin. Data are means ± s.e.m. n = 6. c) RNA foci are a unique feature of the C9orf72 HRE ALS pathology. Motor cortex tissue from C9orf72 WT (non-ALS) and non-C9orf72 sALS show no RNA foci or NCL colocalization, which is observed in motor cortex tissue from patients carrying the C9orf72 HRE (Figure 4c). All images were obtained as described in Methods. d) The protein hnRNP F, K, and U, identified from the pulldown (Figure 3a and Extended Data Figure 6), show no phenotypic differences in localization between patients carrying the C9orf72 HRE and controls, as previously observed in iPS motor neurons from FTD patients. n = 3. e) A representative HEK293T cell transfected with the control transcript shows the typical NCL localization (green), in contrast to the nucleolar dispersion observed in cells transfected with (GGGGCC) (Figure 4d). f) NCL is significantly mislocalized when HEK293T cells are transfected with (GGGGCC) transcripts, as compared to (GGGGCC)3 or control transcripts. RNA transfections were performed in duplicate, except for No RNA and (GGGGCC)3. Data are means ± s.e.m. n = 27, 50, 37, and 60 for the No RNA, CTL RNA, (GGGGCC)3, and (GGGGCC)21, respectively. ****P < 0.0001. g) There is a significantly decreased cell viability in HEK293T cells transfected with the (GGGGCC)21 abortive RNA transcripts when compared to a control RNA transcript. Cytotoxicity measurements were performed in a 96-well plate with HEK293T cells that had been transfected with 0–500 ng/mL of control RNA or (GGGGCC)21 repeat-containing RNAs generated from the in vitro transcription as described earlier and presented in Figure 2a. Cell viability was measured using the Cytotox Glo kit (Promega), with the fluorescence (viable cells) and luminescence (dead) measured on a Synergy H1 microplate reader. Data are means ± s.d. n = 3. *P < 0.05.

Extended Data Figure 8

Extended Data Figure 8. Patients carrying the C9orf72 HRE show phenotypes indicative of nucleolar stress

a) Another nucleolar component, nucleophosmin/B23 (green), shows a dispersed localization as seen for NCL in B lymphocytes from C9orf72 HRE patients. B23 (mouse monoclonal, B0556, Sigma) IF staining followed the manufacturer’s recommendations and the protocol described in Methods. B lymphocytes images were obtained as described in Methods. b) There is a significant increase in the total nuclear area occupied by B23 in B lymphocytes from a patient carrying the C9orf72 HRE when quantified similarly to NCL (Extended Data Figure 7a). The data are means ± s.e.m. n = 38 and n = 49 for representative C9orf72 WT and C9orf72 HRE B lymphocytes, respectively. *P < 0.05. c) The processing of the 45S pre-rRNA into the mature 28S, 18S, and 5.8S rRNAs that occurs in the nucleolus is depicted. d) Maturation of the 45S pre-rRNA to the 28S, 18S, and 5.8S rRNAs is impaired in patients carrying the C9orf72 HRE. B lymphocytes from patients (n = 6) show a decrease in 45S rRNA processing, but the processing is significantly decreased in the motor cortex tissues of patients carrying the C9orf72 HRE relative to controls (n = 4, 3, and 6 for C9orf72 WT, non-C9orf72 ALS, and C9orf72 HRE, respectively). Furthermore, ALS patients that do not carry the C9orf72 HRE show no significant impairment in 45s maturation (non-C9orf72 HRE ALS, n = 3). The primers for 45S, 28S, 18S, and 5.8S rRNAs (Extended Data Table 1) were previously described. The linear-fold change in mature rRNA levels was normalized relative to the parent 45S rRNA levels. Data are means ± s.e.m. *P < 0.05.

Extended Data Figure 9

Extended Data Figure 9. iPS motor neurons from patients carrying the C9orf72 HRE show an increased number of P bodies and increased sensitivity to tunicamycin

a) The number of processing bodies (P bodies), but not the size, is significantly increased in iPS motor neurons from patients carrying the C9orf72 HRE. There is an almost two-fold increase in the total number of P bodies per iPS neuron from C9orf72 HRE patients. The number and size of P bodies from 30 fields of view were quantified in two independent patient iPS motor neuron lines and two control lines. Data are means ± s.e.m. n = 43. *P < 0.05. b) The representative images of P bodies in iPS neurons were visualized and quantified using a DCP1A antibody (Abnova H00055802-M06, green), and the imaging and analysis are described in Supplementary Methods. c) The percent change in cell death with increasing doses of tunicamycin indicates that it takes lower concentrations of tunicamycin to induce significant toxicity in C9orf72 HRE iPS motor neurons than in control neurons. Moreover, there is a significant difference in tunicamycin responses at 0.3 μM when compared between C9orf72 HRE and C9orf72 WT iPS motor neurons. Each iPS line was differentiated twice and the tunicamycin sensitivity at varying concentrations (0.0, 0.1, 0.3, 0.6, 1.0 and 3.0 μM) was analyzed twice. The mean number of PIpositive iPS neurons was normalized to that of untreated iPS neurons from the same line. Data are means ± s.e.m. n = 69/58 (0.0), 62/69 (0.1), 75/55 (0.3), 54/49 (0.6), 55/52 (1.0), and 60/53 (3.0) for C9orf72 WT/HRE samples, respectively. **P < 0.01. d) Representative iPS motor neuron images showing the increased PI staining (white) in response to increasing tunicamycin concentrations (Supplementary Method).

Comment in

References

    1. Richard GF, Kerrest A, Dujon B. Comparative Genomics and Molecular Dynamics of DNA Repeats in Eukaryotes. Microbiology and Molecular Biology Reviews. 2008;72:686–727. - PMC - PubMed
    1. DeJesus-Hernandez M, et al. Expanded GGGGCC Hexanucleotide Repeat in Noncoding Region of C9orf72 Causes Chromosome 9p–Linked FTD and ALS. Neuron. 2011;72:245–256. - PMC - PubMed
    1. Renton AE, et al. A Hexanucleotide Repeat Expansion in C9orf72 Is the Cause of Chromosome 9p21-Linked ALS-FTD. Neuron. 2011;72:257–268. - PMC - PubMed
    1. Kiernan MC, et al. Amyotrophic lateral sclerosis. The Lancet. 2011;377:942–955. - PubMed
    1. Majounie E, et al. Frequency of the C9orf72 hexanucleotide repeat expansion in patients with amyotrophic lateral sclerosis and frontotemporal dementia: a cross-sectional study. Lancet Neurol. 2012;11:323–330. - PMC - PubMed

Publication types

MeSH terms

Substances

Grants and funding

LinkOut - more resources