FoxO1, which is up-regulated during early stages of diet-induced leptin resistance, directly interacts with STAT3 and prevents STAT3 from binding to specificity protein 1 (SP1)–pro-opiomelanocortin (POMC) promoter complex, and thereby inhibits STAT3-mediated regulation of POMC transcription. FoxO1 also binds directly to the POMC promoter and negatively regulates its transcription. The present study aims to understand the relative contribution of the two interactions in regulating POMC expression. We studied the structural requirement of FoxO1 for its interaction with STAT3 and POMC promoter, and tested the inhibitory action of FoxO1 mutants by using biochemical assays, molecular biology and computer modelling. FoxO1 mutant with deletion of residues Ala137–Leu160 failed to bind to STAT3 or inhibit STAT3-mediated POMC activation, although its binding to the POMC promoter was unaffected. Further analysis mapped Gly140–Leu160 to be critical for STAT3 binding. The identified region Gly140–Leu160 was conserved among mammalian FoxO1 proteins, and showed a high degree of sequence identity with FoxO3, but not FoxO4. Consistently, FoxO3 could interact with STAT3 and inhibit POMC promoter activity, whereas FoxO4 could not bind to STAT3 or affect POMC promoter activity. We further identified that five residues (Gln145, Arg147, Lys148, Arg153 and Arg154) in FoxO1 were necessary in FoxO1–STAT3 interaction, and mutation of these residues abolished its interaction with STAT3 and inhibition of POMC promoter activity. Finally, a FoxO1–STAT3 interaction interface model generated by computational docking simulations confirmed that the identified residues of FoxO1 were in close proximity to STAT3. These results show that FoxO1 inhibits STAT3-mediated leptin signalling through direct interaction with STAT3.
Leptin signalling in the hypothalamus is a key regulator in controlling food intake and energy homoeostasis [1,2]. Defective leptin signalling in animals and humans, such as in the case of ob/ob (no leptin production due to leptin gene mutation) and db/db (null mutation of leptin receptor) mice, leads to severe obesity, elevated blood circulating insulin levels, hyperglycaemia and infertility [3,4]. Leptin is primarily produced in and secreted from white adipocytes, whose level is proportional to body fat. High levels of circulating leptin in obese humans and animals fail to regulate food intake and energy expenditure effectively, resulting from a phenomenon called leptin resistance [5,6].
In pro-opiomelanocortin (POMC) neurons, leptin stimulates POMC expression through the Jak2/STAT3 pathway . POMC can be further cleaved by prohormone convertases into several biologically active peptides, including α-MSH and β-endorphin [8,9]. The α-MSH acts on melanocortin-4 receptors in the paraventricular nucleus of the hypothalamus to inhibit food intake. STAT3 is a key signalling node and transcription factor in the leptin signalling pathway. Upon leptin binding and activation of its receptor (OBRb), STAT3 is phosphorylated and translocated into the nucleus, which eventually promotes Pomc gene expression in POMC neurons . Leptin resistance may occur upstream of STAT3 activation based on the findings from a diet-induced obese (DIO) mouse model after 14 weeks of high-fat diet (HFD) feeding . However, leptin resistance could also happen after only 4–5 weeks of HFD feeding. We have shown previously that early leptin resistance arises because of a mechanism downstream of STAT3 phosphorylation, specifically, inhibition of STAT3 function by FoxO1 .
FoxO1 belongs to the forkhead box-containing protein O subfamily. Structurally, FoxO1 contains four domains: a forkhead DNA binding domain (DBD), a nuclear localization signal (NLS), a nuclear export sequence (NES) and a transactivation domain [12–15]. The forkhead domain (FHD) is highly conserved and contains three α-helices, three β-strands and two wing-like loops, and (the whole FHD domain, not just the wing-like loops?) is responsible for DNA recognition, binding and regulation [13,16,17]. NSL and NES control FoxO1's subcellular localization through their phosphorylation status [15,18,19]. It has been shown that FoxO1 can induce Agouti-related peptide (AgRP) expression in AgRP neurons . Activation of insulin signalling in AgRP neurons results in phosphorylation of FoxO1 and its nuclear exclusion, which leads to down-regulation of orexigenic Agrp expression and reduced food intake [20,21]. On the other hand, in POMC neurons, FoxO1 inhibits leptin signalling by blocking STAT3 from interacting with specificity protein 1 (SP1)–POMC promoter complex, thus reducing POMC expression . Interestingly, it has been reported that the STAT3 and FoxO1 binding sites in Agrp and Pomc promoters are adjacent to each other, and that they may compete with each other in promoter binding . Moreover, mice with specific depletion of FoxO1 in the POMC neurons appear resistant to diet-induced obesity . These studies indicate that FoxO1 in POMC neurons plays a negative role in leptin signalling and that inhibition of FoxO1 action may alleviate leptin resistance. However, the relative contribution to the negative effect of FoxO1 in leptin signalling between FoxO1–POMC promoter interaction and direct FoxO1–STAT3 binding remains unclear.
In the present study, we have identified that the amino acid sequence Ala137–Leu160 in FoxO1 is necessary for its interaction with STAT3, among which Gln145, Arg147, Lys148, Arg153 and Arg154 are the key residues. Deletion of this region in FoxO1 abolishes its interaction with STAT3 and suppression of leptin-stimulated STAT3-dependent POMC activation, without affecting FoxO1–POMC promoter binding. This indicates that direct FoxO1–STAT3 interaction is essential for FoxO1 inhibition of POMC promoter activation. The identified STAT3-interaction site on FoxO1 may potentially be used for designing competitive inhibitors to reverse leptin resistance that is attributed to FoxO1 inhibition.
MATERIALS AND METHODS
FoxO1 truncation mutants FoxO11–494 (Mutant 1), FoxO11–370 (Mutant 2), FoxO11–240 (Mutant 3), FoxO11–167 (Mutant 4) and FoxO11–123 (Mutant 5) were made by inserting the following PCR fragments into pCMV5-Myc vector: 1–1482, 1–1110, 1–720, 1–501 and 1–369 of the mouse FoxO1 coding region, respectively. The fragment between 409 and 480 of FoxO1 coding region was deleted from pCMV5-Myc-FoxO1 to generate the deletion mutant FoxO1∆137–160 (Mutant 6). Multiple rounds of site-directed mutagenesis were done using pCMV-Myc-FoxO1 as the template to generate FoxO1Q145A+R147A+K148A+R153A+R154A (Mutant 7). pXJ40-Flag-FoxO3a was a generous gift from Dr. Fukamizu (University of Tsukuba). pBR332-Flag-FoxO4 was purchased from Addgene. The coding regions of FoxO3a and FoxO4 were sub-cloned into pCMV5-Myc vector to generate pCMV5-myc-FoxO3a and pCMV5-myc-FoxO4, respectively.
Cell lines and culture
293-OBRb, a HEK293 cell line with stable OBRb expression , was cultured in Dulbecco's minimal essential medium (Invitrogen) supplemented by 10% FBS in a 37°C incubator with 5% CO2.
293-OBRb cells were transfected with pXJ40-Flag-STAT3 plus either pCMV5-Myc-FoxO1 or one of the mutants by Lipofectamine 2000 (Invitrogen) and harvested 48 h after transfection by using whole cell lysis buffer containing 20 mM Hepes (pH 7.9), 280 mM KCl, 1 mM EDTA, 0.1 mM Na3VO3, 10% glycerol, 0.5% NP40, 1 mM DTT, 1 mM PMSF and proteinase inhibitor cocktail (Roche Applied Sciences). One milligram of cell lysate was incubated with anti-Flag antibody (Sigma) or control mouse IgG overnight, followed by an additional 1 h incubation with protein A/G-Sepharose beads (Invitrogen). The beads were washed with lysis buffer four times and resuspended in 2× SDS/PAGE sample buffer. The protein samples were resolved by SDS/PAGE, and STAT3 was detected with anti-Myc antibody (Santa Cruz Biotechnology). Ten micrograms of cell lysate was loaded as input.
293-OBRb cells were transfected with pGL3-POMC-Luciferase, pXJ40-Flag-STAT3, pRenilla plus either pCMV5-Myc-FoxO1 or one of its mutants by using FuGENE 6 (Roche). Sixteen hours after transfection, the cells were starved in DMEM with 1% FBS for 5 h, followed by treatment with leptin (Invitrogen) or vehicle for 20 h. The cells were washed with PBS and harvested in passive lysis buffer provided in the Dual Luciferase Reporter Assay System (Promega). Luciferase activities of the cell lysate were measured by using an LMax II Luminometer (Molecular Devices). POMC promoter activity was associated with firefly luciferase activity and normalized against Renilla luciferase. A 40% or more increase in luciferase activity by leptin treatment in comparison with vehicle treatment is considered as activation of POMC expression.
Electrophoretic mobility shift assay
The oligonucleotides coding for the FoxO1 binding site of the Pomc promoter (5′-TAG TGA TAT TTA CCT CCA AAT GCC AGG AAG GCA G-3′)  were annealed and labelled with 32P. Wild-type, Mutant 6 and 7 of FoxO1 were expressed and extracted from HEK293 cells. After binding reaction, the samples were separated by using a natural poly-acrylamide gel and transferred to filter paper. The interactions between wild-type FoxO1, Mutant 6 and Mutant 7 with the radioactive-labelled probe were detected by Phospho Screen using a Typhoon FLA7000 scanner (GE Healthcare).
Computer simulation of FoxO1 and STAT3 interaction
The program Modeller [23,24] was used to build and/or complete the 3D structures of FoxO1 and STAT3. The stretch of residues between 150 and 250 of FoxO1 was modelled by using chain A of the DBD of the human fork head transcription factor AFX (FoxO4, pdb id:1e17) as the template, and in the models of STAT3, the sequence of the unphosphorylated STAT1 (pdb id:1yvl) was used as template. The two alignments used to build the model were obtained by using a profile–profile alignment and fold recognition algorithm (FFAS) extracted from the meta-server (http://meta.bioinfo.pl). The sequence identity for the models to the pdb templates was 87% for FoxO1 and 54% for STAT3. We were unable to generate coordinates in a defined secondary structure motif for the residues 120–160 of FoxO1, as all the prediction programs showed a tendency of this region to form a random coil, indicating an absence of regular secondary structure (Supplementary Figure S1).
The above models were used as starting structures in the docking analysis. Docking algorithms are computational methods that generate models of putative biomolecular complexes from structural information of its starting constituents. All the models for the complexes in the present study were constructed by using default parameters incorporated in the software employing a data-driven docking approach called HADDOCK (High Ambiguity Driven protein–protein DOCKing) . The advantage of such an approach is the possibility of incorporating different types of available information on protein interfaces and encodes it into ambiguous interaction restraints (AIRs) to conduct the docking. This program permits a flexible range of input data to drive the docking [26,27]. Based on a combined conservation and structure-based criterion, we used multiple sequence alignments  to identify the active residues involved in the binding. The predicted residues were filtered and only those having a relative solvent accessibility of >50% as defined by NACCESS  were kept to drive the docking (see Supplemental Table S1). Following the general protocol of HADDOCK, the residues that are exposed to solvent and those that flank the active residues were also included in the definition of the binding interface, as passive residues. The AIRs are defined as ambiguous distance restraints between active residues of one protein, and all active and passive residues on the other protein and vice versa. In the HADDOCK algorithm, flexibility is accounted for in different ways during the docking, which allows modelling of subtle conformational changes upon binding.
The HADDOCK protocol incorporates the driving information to perform the docking in a qualitative manner, and thus does not provide precise distances or relative orientations of the molecules, or allow for the unambiguous identification of pairs of atoms that are in contact in the complex. Given the ambiguous nature of the restraints in this protocol and the addition of flexibility, the output is a cluster of possible modelled biomolecular complexes from which statistically relevant residue pair interaction information can be extracted, rather than relying on a well-defined and static conformer. This is highly relevant in the case of complexes with a high inherent flexibility, like the one analysed in the present study, as it provides a more dynamic picture more likely to mimic the biological solution ensemble.
To optimize the interactions at the interface, the best model obtained from the bioinformatics-driven docking run was further refined by incorporating the experimental information obtained from the various FoxO1 mutants in the stretch Gln145-Arg154 as restraints in the water refined phase of the HADDOCK protocol; this also allows for this loop region to move freely during this refinement step.
Data are presented as means ± S.E.M. All statistical tests were performed using Student's two-tailed t test for independent data sets. The significance limit was set at P<0.05.
Critical FoxO1 sequence for its interaction with STAT3
Previous studies indicate that FoxO1 negatively regulates STAT3-mediated leptin regulation of Pomc transcription in hypothalamic POMC neurons [11,20,29]. To examine whether the inhibitory effect is due to direct FoxO1–STAT3 interaction, we first mapped the STAT3 binding site in FoxO1. We generated a series of FoxO1 truncation mutants, namely, FoxO11–494 (Mutant 1), FoxO11–370 (Mutant 2), FoxO11–240 (Mutant 3), FoxO11–167 (Mutant 4) and FoxO11–123 (Mutant 5) (Figure 1A), and performed co-immunoprecipitation experiments to assess their interaction with STAT3 (Figure 1B). The truncation mutants FoxO11–494, FoxO11–370, FoxO11–240 and FoxO11–167 retained the ability to bind to STAT3, whereas FoxO11–123 failed to interact with STAT3. We then performed luciferase assays to study the effects of FoxO1 mutants in the regulation of STAT3-mediated leptin-dependent POMC expression. As previously reported, expression of wild-type FoxO1 inhibited STAT3-induced POMC promoter activation (Figure 1C) . Similarly, the truncation mutants that retained the STAT3 binding (FoxO11–494, FoxO11–370, FoxO11–240 and FoxO11–167) also inhibited leptin-regulated POMC promoter activation. In contrast, FoxO11–123, which failed to interact with STAT3, was unable to inhibit leptin activation of POMC transcription (Figure 1). To further define the interaction sequence on FoxO1, we generated a FoxO1 deletion mutant, FoxO1∆137–160 (Mutant 6) (Figure 2A), and examined its STAT3 binding by co-immunoprecipitation (Figure 2B). Deletion of amino acids 137–160 abolished FoxO1 binding to STAT3 and Mutant 6 showed no inhibitory effect on POMC promoter activity (Figure 2B and 2C). These results indicate that the sequence between amino acids 137 and 160 in FoxO1 is essential for its interaction with STAT3, and suggest that FoxO1–STAT3 interaction is necessary for FoxO1 to inhibit STAT3-mediated POMC promoter activation.
Identification of essential FoxO1 fragments for its interaction with STAT3
FoxO1 fragment between Ala137 and Leu160 is essential for its interaction with STAT3
Sequence conservation in the FoxO family for STAT3 interaction
Considering that the first few amino acids are alanine (Ala137 to Ala139), which are unlikely to be involved in STAT3 interaction, we postulated that the region between Gly140 and Leu160 is the critical sequence for FoxO1–STAT3 binding. The putative STAT3-interaction site in FoxO1 contains 10 residues (Gly140–Thr149) to the N-terminal start of the FHD and 11 residues (Ser150–Leu160) of the FHD. To assess the homology of the identified STAT3-interaction region of FoxO1, we performed multiple sequence alignment analysis of FoxO1 protein sequences of several mammalian species. The multiple alignment data (Figure 3A) showed that the identified region (Gly140–Leu160 in mouse FoxO1) was highly conserved in these species. The conservation was not limited to the FHD but also included the linker region, suggesting that this region may be functionally important.
The homology of the FoxO1–STAT3 interaction site
We further performed multiple sequence alignment for mouse FoxO proteins, and found that the postulated STAT3-interaction site in FoxO1 shared >70% identity with FoxO3 but ~50% with FoxO4 (when the highly conserved FHD sequence was taken into consideration; Figure 3B). Mouse and human Foxo4 are identical in the sequence containing the STAT3-interaction site and FHD (Figure 3B). As we were not able to obtain the mouse FoxO4 cDNA, we used human FoxO4 instead in the following experiments. We first examined FoxO3 and FoxO4 interaction with STAT3 by co-immunoprecipitation and found that FoxO3, but not FoxO4, could interact with STAT3 (Figure 3C). We then tested the effect of FoxO3 and FoxO4 in the regulation of STAT3-mediated POMC transcription by the luciferase assay. FoxO3 effectively inhibited STAT3-mediated leptin-dependent POMC activation, whereas FoxO4 had no inhibitory effect in the leptin regulation of POMC transcription (Figure 3D). These results further support the hypothesis that the sequence Gly140–Leu160 is essential for STAT3 interaction, and that FoxO1 inhibition of STAT3-mediated leptin action is through direct FoxO1–STAT3 interaction.
Characterization of FoxO1–STAT3 interfaces in the in silico model
To understand the structural relevance of the identified sequence, we generated a model for the interaction between FoxO1 and STAT3 by initially using the information extracted from the sequence conservation in the alignment of the protein families. The model was subsequently refined using the experimental information obtained from the mutational data arising in the current study (see Materials and Methods for a more detailed protocol).
The binding interface of the model for FoxO1–STAT3 complex was composed mainly of the loop region of FoxO1 comprising the region Arg147–Arg154 in the forkhead-DBD and a surface patch of STAT3 that belongs to the SH2 (Src homology 2) domain. The former, due to its large inherent flexibility, could adopt such a conformation that it wrapped this particular domain of STAT3. This created a large binding interface (2263.0 Å2; 1 Å = 0.1 nm) and an extensive set of inter-molecular salt bridges and hydrogen bonds that might confer specificity in the binding recognition between these two partners (Figure 4A). The binding interface of STAT3 could be divided into three different zones: two negatively charged regions sandwiching a positively charged one (Figure 4B). Consequently in the partner, the interactions were driven mainly by electrostatic forces along two positively charged patches. Thus, on one side, the complex was formed by salt-bridge interactions made by Arg153 of FoxO1 with both Glu530 and Glu592 of STAT3, and the second arginine in this stretch, Arg154, oriented to promote the formation of a hydrogen bond with Ser599 at the rim of this patch. At the other side of the interface, the stabilization of the complex arose from the interactions of residue Asp627 of STAT3 with both Arg147 and Gln145 of FoxO1. In the positively charged region of STAT3, the positively charged side chain of Lys148 pointed away from the interaction surface due to electrostatic repulsions, and only its backbone oxygen formed a hydrogen bond with another lysine at position 631.
Model of the FoxO1–STAT3 complex generated by HADDOCK
Residues Gln145, Arg147, Lys148, Arg153 and Arg154 are critical for FoxO1 to bind STAT3
Based on the in silico models and the biochemical analysis, we generated a FoxO1 mutant (Mutant 7) containing alanine substitutions of the following residues: Gln145, Arg147, Lys148, Arg153 and Arg154 (Figure 5A). Co-immunoprecipitation analysis revealed that Mutant 7 of FoxO1 completely lost its interaction with STAT3 (Figure 5B), and luciferase assay showed that the FoxO1 mutant was unable to inhibit STAT3-mediated leptin activation of POMC transcription (Figure 5C). These data show that the five residues (Gln145, Arg147, Lys148, Arg153 and Arg154) are critical for FoxO1 in its interaction with STAT3, and in its inhibition of STAT3-mediated POMC transcription, providing further support to the notion that FoxO1 inhibition of leptin–STAT3–POMC transcription is through direct FoxO1–STAT3 interaction.
Critical residues of FoxO1 necessary for its interaction with STAT3
FoxO1 inhibits POMC through direct interaction with STAT3
To examine whether FoxO1 binding to Pomc promoter plays any role in its inhibition of STAT3-mediated Pomc promoter activation, we performed EMSA to assess Pomc promoter binding by wild-type FoxO1, Mutant 6 and Mutant 7. Wild-type FoxO1, Mutant 6 and Mutant 7 were able to bind to the probe containing the FoxO1 binding site in the Pomc promoter (Figure 5D), indicating that deletion of amino acids 137–160 or mutation of the 5 key residues does not affect FoxO1 binding to the promoter. Although wild-type FoxO1, Mutant 6 and Mutant 7 could all bind to POMC promoter, only wild-type FoxO1, but not Mutant 6 or Mutant 7 inhibited STAT3-mediated POMC promoter activation (Figures 2, 5 and Supplementary Figure S2). As expected, when wild-type FoxO1 was co-expressed with Mutant 6 or Mutant 7, STAT3-mediated POMC promoter activation was impaired (Supplemental Figure S2). These findings indicate that FoxO1-promoter binding does not play a significant role in FoxO1 inhibition of STAT3-mediated promoter activation, consistent with our previous findings . Moreover, the above finding that Mutant 6 and Mutant 7 lost both STAT3-interaction and their inhibitory effect on POMC promoter activity without changes in their binding to the POMC promoter supports that FoxO1 inhibits POMC expression through its direct interaction with STAT3.
In the present study, we have investigated the relative contribution of FoxO1–STAT3 interaction and FoxO1–POMC promoter binding to the inhibitory role of FoxO1 in STAT3-mediated leptin activation of POMC expression. We find that FoxO1 inhibition of leptin–STAT3–POMC transcription is mediated by direct FoxO1–STAT3 interaction based on the following experimental evidence: first, by examining the interaction of a series of deletion and point mutants of FoxO1 with STAT3 using co-immunoprecipitation, we identify the critical sequence (Gly140–Leu160) and amino acid residues (Gln145, Arg147, Lys148, Arg153 and Arg154) in FoxO1 to be essential for its interaction with STAT3; secondly, by utilizing the POMC–luciferase assay system, which has been shown to be useful in analysing FoxO1 functions [6,11,18], we show that the FoxO1 mutants with deletion of the critical sequence (Gly140–Leu160) or mutations of the five residues (Gln145, Arg147, Lys148, Arg153 and Arg154) fail to inhibit STAT3-mediated leptin activation of POMC transcription; thirdly, by EMSA analysis, we find that the critical sequence (Gly140–Leu160) is not necessary for FoxO1's direct interaction with POMC promoter. Finally, FoxO3, which has high degree of sequence identity with FoxO1 and interacts with STAT3, is able to inhibit STAT3-regulation of POMC transcription; whereas FoxO4, which has low level of sequence identity with FoxO1 and does not interact with STAT3, has no inhibitory effect on STAT3-mediated POMC expression.
FoxO1 has been shown to be a key transcription factor in the hypothalamus to regulate energy homoeostasis: it promotes NPY and AgRP expression to increase food intake and reduce energy expenditure [5,18], while it inhibits POMC expression and contributes to leptin resistance [6,18]. In the present paper, we confirm that FoxO1 inhibits leptin-induced POMC activation and show that the FoxO1 action is mediated through direct interaction with STAT3. The finding suggests that targeting the FoxO1–STAT3 interaction may be an effective approach to reverse leptin resistance.
Structurally, STAT3 forms homodimers and possibly heterodimers with STAT1 . The full-length protein comprises a coiled-coil domain, a DBD, a connector domain, an SH2 domain and a carboxy terminus that is denoted as the transcriptional activation domain . The dimerization is driven by the phosphorylation of a conserved Tyr residue in the SH2 domain [32,33]. This phosphorylation event strengthens the dimer as well as the affinity for DNA binding. In a molecular dynamics simulation study, a large scissor-like motion of the coiled-coil domain has been observed in STAT3 dimerization . This large conformational change is driven by a tighter binding of the protein to DNA, thereby regulating gene expression. The same authors have also reported that the connector or linker domain is quite rigid in the dimer, whereas the highest fluctuations are seen in the SH2 domain, with the exception of certain residues (Lys591, Arg609, Ser611 and Ser613) involved in polar interactions with phosphotyrosine that are kept very rigid. In light of the biological function of STAT3 suggested in the above studies, we propose a model where the unstructured region of FoxO1 comprising residues 140–160 wraps the SH2 domain of STAT3, and such an interaction would abrogate the dimerization of the latter and consequently the binding to DNA (Figure 4C). The model obtained in the present study could open new avenues for possible drugs abrogating the interaction between these two proteins. The generation of a putative inhibitor peptide should expand to the whole interface covered by the unstructured region of FoxO1. In comparison with the wild-type sequence of this region in FoxO1 (-QPRKTSSSRR-), the mutation by positively charged residues at the threonine and serine positions in this peptide could enhance the binding due to the additional interactions with the positive surface of STAT3 (Figure 4B). Future efforts may be directed at developing and identifying compounds that compete for STAT3 binding, which could provide drug candidates for therapeutic use.
In summary, the region from Gly140 to Leu160 and the residues Gln145, Arg147, Lys148, Arg153 and Arg154 are critical for FoxO1 in its interaction with STAT3. The direct FoxO1–STAT3 interaction is necessary for the inhibitory effect of FoxO1 in leptin–STAT3–POMC promoter activation.
ambiguous interaction restraints
DNA binding domain
nuclear localization signal
nuclear export sequence
Src homology 2
specificity protein 1
Wei Ma, Gloria Fuentes and Chandra Verma performed the experiments and analysed the data. George Radda and Weiping Han designed the research. Wei Ma and Weiping Han wrote the paper. All the authors discussed and approved the paper.
The present study was supported by intramural funding from A*STAR Biomedical Research Council (W.H.). The authors declare that there is no conflict of interest associated with this manuscript.