Structures and membrane interactions of native serotonin transporter in complexes with psychostimulants

To isolate porcine SERT (pSERT), we exploited the 15B8 Fab (46), an antibody fragment that binds to a tertiary epitope of human SERT (hSERT), yet does not hinder the binding of ligands or the transport activity. We hypothesized that because pSERT and hSERT are closely related (SI Appendix, Fig. S1), the 15B8 Fab would also bind to pSERT and could serve as a powerful tool for immunoaffinity isolation of the transporter. We thus engineered the 15B8 Fab with a carboxy terminal mCherry fluorophore and an affinity tag (Fig. 1A).

To isolate pSERT from brain membranes, we explored different membrane protein solubilization conditions, aiming to extract the transporter under the mildest of conditions while retaining as much surrounding native lipids as possible. Solubilization in the presence of styrene-maleic acid (SMA) copolymer (47) or the recently developed amphipols (48, 49) did not yield a measurable amount of native pSERT (Fig. 1B). We then examined the mild nonionic detergents, digitonin or n-dodecyl-β-D-maltoside (DDM) together with cholesterol hemisuccinate (CHS), based on their utility in the extraction of native (50) or recombinant hSERT (51), respectively. Surprisingly, a peak in the FSEC trace for the pSERT–15B8 Fab–mCherry complex was only observed with the DDM/CHS mixture (Fig. 1B), at a position that is appropriate for the pSERT–Fab complex, while no similar peaks were observed upon solubilization under the additional conditions. We therefore utilized DDM/CHS in all subsequent studies. To isolate pSERT from brain tissue, we incubated the solubilized membranes with an excess of the 15B8 Fab-mCherry protein, as well as with saturating concentrations of either methamphetamine or cocaine. The transporter was purified by affinity chromatography, followed by size-exclusion chromatography, collecting fractions manually. Shown in Fig. 1C are FSEC traces of the solubilized material throughout the purification work flow. We also compare the elution position of the pSERT–15B8 Fab–mCherry complex with the monomeric ts2 variant of hSERT in complex with the 15B8 Fab (46), and the coelution of the two peaks supports the conclusion that pSERT, solubilized in the presence of DDM/CHS, behaves as a monomeric entity (Fig. 1C). Analysis of the isolated material by western blot revealed a band migrating with an apparent mass of 75 kDa (Fig. 1D), consistent with the estimated mass of pSERT. The well-resolved and symmetrical FSEC peak indicated that the purified pSERT–15B8 Fab–mCherry complex was monodisperse and best described as a single species. The elution volume with FSEC of the pSERT–15B8 Fab complex was consistent with the recombinant ts2–15B8 complex (Fig. 1C), indicating that pSERT purified in the presence of DDM/CHS is best described as a monomer.

To explore the function of pSERT, we then carried out saturation binding experiments using the high-affinity SSRI [³H]paroxetine for which we determined a dissociation constant (K_d) of 6.5 ± 1.3 nM (Fig. 1E), in agreement with previous measurements (52). To characterize methamphetamine and cocaine binding with pSERT, competition experiments were performed, similarly employing [³H]paroxetine, and measuring inhibitory constants (K_i) of 199 ± 103.4 µM (Fig. 1F) and 179 ± 68 nM (Fig. 1G), respectively, thus indicating that both methamphetamine and cocaine compete for [³H]paroxetine binding, consistent with binding of the psychostimulants to the central site. The potencies of methamphetamine and cocaine on pSERT differ by a factor of ~1,000, in accord with previous studies, whereas on DAT, the difference is only a factor of ~10 (53), underscoring the differences in residue composition and plasticity of the central binding pockets of SERT and DAT.

From one porcine brain, we obtained ~20 μg of purified protein in a volume of 200 μL, which was sufficient for visualizing particles on continuous carbon-coated grids under cryogenic conditions. We then collected single-particle cryo-EM datasets and carried out reconstructions of the methamphetamine- or cocaine-bound SERT complexes, obtaining density maps that extended to approximately 2.9- and 3.3-Å resolution, respectively (SI Appendix, Figs. S2–S6, S7 A and B, and Table S1). Particle picking was employed with an elliptical “blob” of sufficient size to capture pSERT oligomers. Thorough 2D and 3D classifications yielded only a single class for each dataset in which pSERT was found as a monomeric entity, with no evidence of dimers or higher-ordered oligomers (SI Appendix, Figs. S2 and S3), consistent with the molecular size of SERT estimated by FSEC. Overall, the density maps are of sufficient quality to assign most of the amino acid side chains, to identify additional nonprotein density within the central binding site as bound ligands, and to suggest the presence of cholesterol or CHS molecules surrounding the transporter transmembrane domains (Fig. 2).

Methamphetamine binds to the central site of the pSERT complex, adopting a similar binding pose to that observed in DAT (16), lodged between the aromatic groups of Tyr213 and Tyr132. The amine groups of methamphetamine interact with Ser475 and form hydrogen bonds with the carboxylate of Asp135 and the main chain carbonyl of Phe372, as seen with the hydrogen bonds formed between the amine group of methamphetamine and the equivalent Asp46 and Phe319 in DAT (16) (PDB code: 4XP6). The side chain of Phe378 forms edge-to-face aromatic interactions with the phenyl group of methamphetamine (Fig. 3A and SI Appendix, Fig. S7C). Further studies, at higher resolution, and complexes of methamphetamine with DAT, will be required to better explain the weaker binding of methamphetamine to SERT in comparison to DAT. Comparison of the positions of TM1, TM6, and the extracellular gate to the equivalent elements of serotonin-bound outward-open complex (54) [Protein Data Bank (PDB) code: 7LIA, RMSD: 0.606] indicates that the pSERT–15B8 Fab–methamphetamine complex adopts an outward-open conformation (Fig. 3B and SI Appendix, Fig. S7D), which is consistent with previous structural studies of dDAT in complex with methamphetamine (16).

The structure of the pSERT–cocaine complex reveals an outward-open conformation, with cocaine occupying the entire central binding pocket with an overall similar pose to the experimentally derived dDAT–cocaine complex (16) and reminiscent of the hDAT–cocaine complex determined using computational methods (55). The nearly complete filling of the volume of the central site by cocaine perhaps accounts for the increased affinity of SERT for cocaine compared to methamphetamine, reminiscent of how increasing the volume of serotonin via methylation can enhance ligand binding (54). The benzoyl moiety of cocaine is accommodated between TM3 and TM8, where it forms van der Waals interactions with Ile209, Tyr213, Phe378, and Thr476. The methyl ester group protrudes into the base of the extracellular vestibule and the tropane rings are bordered by Tyr132, Ala133, Asp135, Phe372, and Ser475. Interestingly, the side chain of Phe372 undergoes substantial displacement and moves further into the central site than seen in the dDAT complex (16). This reorganization translates into the formation of a “thin” extracellular gate (56), blocking the release of cocaine from the central site and ultimately occluding the binding pocket, a conformation that was not visualized in the cocaine-bound dDAT structure (Fig. 3C and SI Appendix, Fig. S7E). Nevertheless, the overall structure of the pSERT–15B8 Fab–cocaine complex is similar to 5-HT-bound recombinant hSERT in its outward-open conformation (PDB code: 7LIA, RMSD: 0.6 Å), with residue pairs that define the extracellular gate, Arg141–Glu530 and Tyr213–Phe372, 11.1 Å and 15.0 Å apart, measured from Cα to Cα, respectively (Fig. 3D and SI Appendix, Fig. S7F).

SERT is an integral membrane protein embedded in a complex membrane composed of phospholipids, sphingolipids, and cholesterol (57, 58). SERT (58–60), NET (61, 62), DAT (63–65), and GlyT (66, 67), as well as some excitatory amino acid transporters (68), associate with cholesterol in brain tissues or in transfected cell lines. Cholesterol is implicated in a variety of biological processes, including membrane protein organization and compartmentalization within the membrane. It is also known to play a key indirect role in modulating neurotransmission via its effects on the activities of DAT (69) and SERT (58). Indeed, depletion of cholesterol from membranes affects the function of neurotransmitter transporters (69, 70). Previous molecular dynamics (MD) studies revealed six potential cholesterol-binding sites of SERT, defined as CHOL1-6 (71). Bound cholesterol has been observed at the CHOL1-binding site in dDAT structures (16). The cholesterol analog, CHS, has been shown to bind at the CHOL2-binding site in dDAT (16, 72) and hSERT (46), as well as at the CHOL3-binding site in hSERT (52). In investigating the interactions of cholesterol with pSERT, we carefully examined the density maps of methamphetamine- and cocaine-bound SERT complexes, and the quality of the density maps enabled the identification of CHS, or perhaps cholesterol, at the CHOL1- and CHOL2-binding sites in both structures (Fig. 4 A, B, D, and E), consistent with previous observations.

We identified a nonprotein density in the allosteric site for both complexes (Fig. 4 C and F), a binding site for a broad spectrum of ligands (52, 54, 73) (SI Appendix, Fig. S7 G–I). Previously, the maltose headgroup of a DDM molecule has been modeled into this site, based on a cryo-EM map from heterologously expressed hSERT (74). Because we isolated SERT from native tissue and the overall shape of the density also resembles the lipid molecule, we considered the possibility that the density feature might also be attributed to a lipid molecule. Because the local resolutions of the density maps within the allosteric site are not sufficient for unambiguous molecular identification, we used MD simulations to examine the binding of the most abundant lipid molecules, steroids, or fatty acids to the allosteric site. We performed all-atom MD simulations of the cocaine–pSERT complex with the allosteric site occupied by either docosahexaenoic acid (DHA), in charged or neutral forms (DHA^– or DHA⁰, respectively), cholesterol, CHS, or DDM, in triplicate simulation replicas for each molecular species (Fig. 4G and SI Appendix, Fig. S8A).

In all the three independent simulations, DDM remained bound stably at the site (center-of-mass displacement ≤ 3 Å). DHA^– remained bound in one simulation, but in the other two, transiently moved away and then returned to its initial location. By contrast, in all simulations with neutral DHA⁰, cholesterol, or CHS, the ligand unbinds from its binding site within nanoseconds, as highlighted by large center-of-mass displacement (>5 Å). We further performed bias-exchange umbrella sampling (BEUS) simulations to calculate binding free energy profiles for DHA^–, cholesterol, and DDM, verifying preferential binding of DDM, and DHA^– to a lesser extent, to the allosteric site, whereas cholesterol binding to this site appears to be accompanied with a large increase in free energy (Fig. 4H and SI Appendix, Fig. S8 B and C). These results suggest that the allosteric site is likely occupied by DDM under our experimental conditions, yet it can also accommodate fatty acid binding. However, cholesterol binding to this site is energetically unfavorable. The biological role of lipid binding within the allosteric site awaits further elucidation.

The 15B8 Fab construct (52) was cloned into the pFastBac-dual vector, including a GP64 signal sequence. A mCherry tag, followed by twin Strep [​Trp​Ser​His​Pro​Gln​Phe​Glu​Lys​(Gl​yGl​yGl​ySe​r)2​Gly​Gly​Ser​Ala​Trp​Ser​HisProGlnPheGluLys], and His₁₀ purification tags were fused to the C terminus of the heavy chain. Baculovirus was prepared according to standard methods. The Sf9 cells were infected by the recombinant baculovirus at a cell density of 2 × 10⁶ mL⁻¹ at 27 °C. The culture supernatant was then collected 96 h after infection by centrifugation at 5,000 rpm for 20 min using a JLA 8.1000 rotor at 4 °C. The 15B8 Fab was purified from Sf9 supernatant by metal ion affinity chromatography followed by size-exclusion chromatography.

One porcine brain (~150 g), obtained from a local slaughterhouse, was homogenized with a Dounce homogenizer in ice-cold Tris-buffered saline buffer (TBS: 20 mM Tris-HCl, pH 8.0, 150 mM NaCl) supplemented with 1 mM phenyl methylsulfonyl fluoride, 0.8 μM aprotinin, 2 μg mL⁻¹ leupeptin, and 2 μM pepstatin. The homogenized brain suspension was then sonicated using a sonicator equipped with a tip size of 1.27 cm, for 15 min with 3 s on and 5 s off, at medium power, on ice. The resulting solution was clarified by centrifugation for 20 min at 10,000 g at 4 °C, and the supernatant was collected and applied for further centrifugation at 40,000 rpm for 1 h at 4 °C (45 Ti fixed-angle rotor, Beckman) to pellet the membranes. The membranes were resuspended in 40 mL ice-cold TBS and further homogenized with a Dounce homogenizer. The membranes were solubilized in 100 mL ice-cold TBS containing 20 mM n-dodecyl-β-D-maltoside (DDM) and 2.5 mM cholesteryl hemisuccinate (CHS) in the presence of 1 mg 15B8 Fab, 100 μM methamphetamine, or 10 μM cocaine, for 1 h at 4 °C. The lysate was centrifuged at 40,000 rpm for 50 min at 4 °C (45 Ti fixed-angle rotor, Beckman) and the transporter–Fab complex was isolated by affinity chromatography using Strep-Tactin resin. The complex was further purified by FSEC (76) on a Superose 6 Increase 10/300 column in a buffer composed of 20 mM Tris-HCl (pH 8) supplemented with 100 mM NaCl, 1 mM DDM, 0.2 mM CHS, and 100 μM methamphetamine or 10 μM cocaine. The peak fraction containing the native pSERT–Fab complexes was manually collected and used for biochemical and single-particle cryo-EM analysis.

Porcine brain membranes were prepared as described above. The membrane fraction was resuspended and solubilized in ice-cold TBS containing buffer I (20 mM DDM, 2.5 mM CHS), buffer II (1% digitonin), buffer III (1% amphipol 17), or buffer IV (1% SMA-XIRAN30010). Mammalian cells expressing recombinant hSERT were solubilized with buffer I for comparison, as previously described (52). All the solubilizations were performed in the presence of 1 μM paroxetine for 1 h at 4 °C. The solubilized solutions were clarified by centrifuging at 40,000 rpm using a TLA-55 rotor for 20 min at 4 °C. The supernatant was then examined by FSEC on a Superose 6 Increase 10/300 column.

Purified native pSERT was run on a SDS-PAGE gel and subsequently transferred to a nitrocellulose membrane. Antibody used for detection was 10F2, a monoclonal antibody generated in-house. An IRDye 680RD anti-mouse secondary antibody (LI-COR) was used for visualization. Blots were developed from the secondary antibody at a ratio of 1:10,000 and imaged by Odyssey® DLx Imaging System.

A saturation binding experiment using [³H]paroxetine was performed via the scintillation proximity assay (SPA) (77) using the lysate of porcine brain membranes in SPA buffer (20 mM Tris-HCl, pH 8, 100 mM NaCl, 1 mM DDM, 0.2 mM CHS). The membrane lysates were mixed with Cu-YSi beads (0.5 mg mL⁻¹) in SPA buffer, and [³H]paroxetine at a concentration of 0.3 to 40 nM. Nonspecific binding was estimated by experiments that included 100 μM cold S-citalopram. Data were analyzed using a single-site binding function.

Methamphetamine and cocaine competition binding experiments were performed using SPA with Cu-YSi beads (0.5 mg mL⁻¹) in SPA buffer. For the methamphetamine competition assays, SPA was performed with Strep-purified native pSERT, 10 nM [³H]paroxetine, and 1 μM to 100 mM cold methamphetamine. For the cocaine competition assays, SPA was done with Strep-purified native pSERT and 10 nM [³H]paroxetine in the presence of 1 nM to 1 mM cold cocaine. Experiments were performed in triplicate. The error bars for each data point represent the SEM. K_i values were determined with the Cheng–Prusoff equation (78) in GraphPad Prism.

The purified native pSERT–15B8 Fab complex was concentrated to 0.1 mg mL⁻¹, after which either 10 mM methamphetamine or 1 mM cocaine, together with 100 μM fluorinated n-octyl-β-D-maltoside (final concentration), was added prior to grid preparation. A droplet of 2.5 μL of protein solution was applied to glow-discharged Quantifoil 200 or 300 mesh 2/1 or 1.2/1.3 gold grids covered by 2 nm of continuous carbon film. The grids were blotted for 2.0 s at 100% humidity at 20 °C, followed by plunging into liquid ethane cooled by liquid nitrogen, using a Vitrobot Mark IV. The native pSERT datasets were collected on a 300-keV FEI Titan Krios microscope located at the HHMI Janelia Research Campus, equipped with a K3 detector, at a nominal magnification of 105,000×, corresponding to a pixel size of 0.831 Å. The typical defocus values ranged from −1.0 to −2.5 μm. Each stack was exposed for 4.0 s and dose-fractionated into 60 frames, with a total dose of 60 e⁻ Å⁻². Images were recorded using the automated acquisition program SerialEM (79).

The beam-induced motion was corrected by MotionCor2 (80). The defocus values were estimated by Gctf (81) and particles were picked by blob-picker in cryoSPARC (82). We explored blob picking with elliptical blobs of sufficient sizes to capture pSERT oligomers; yet after two rounds of 2D classification, we found only pSERT monomers, and proceeded to select 2D classes with clear secondary structures. An initial model was generated by cryoSPARC and employed for heterogeneous refinement. Next, a round of 3D classification without image alignment was performed in RELION-3.1 (83), with a soft mask excluding the constant domain of 15B8 Fab and micelle. The selected particles were imported back to cryoSPARC for homogeneous refinement, local contrast transfer function (CTF) refinement, and nonuniform refinement. The local resolution of the final map was estimated in cryoSPARC.

For the pSERT–15B8 Fab complex in the presence of (+)-methamphetamine, 7,794,907 particles were picked in cryoSPARC, which after rounds of 2D classification and heterogeneous refinement, left 348,745 particles (binned to a 200-pixel box, 1.662 Å pixel⁻¹). The particles were reextracted (360-pixel box, 0.831 Å pixel⁻¹) and subjected to homogeneous refinement and nonuniform refinement in cryoSPARC, and then subjected to 3D classification with 10 classes in RELION-3.1 without image alignment (360-pixel box, 0.831 Å pixel⁻¹). Particles from three classes with clear TM features were combined and subjected to homogeneous refinement, local CTF refinement, and nonuniform refinement in cryoSPARC (SI Appendix, Fig. S2).

For the pSERT–15B8 Fab complex with cocaine, a total of 7,560,137 particles were picked from 16,094 movies in cryoSPARC with a box size of 200 pixels (1.662 Å pixel⁻¹). After rounds of 2D classification and heterogeneous refinement, 338,343 particles were selected, reextracted (400-pixel box, 0.831 Å pixel⁻¹), and subjected to homogeneous refinement and nonuniform refinement in cryoSPARC and further subjected to 3D classification with 10 classes in RELION-3.1 without image alignment. Two well-resolved classes with 243,207 particles were combined and further refined in cryoSPARC with homogeneous refinement, local CTF refinement, and nonuniform refinement (SI Appendix, Fig. S3).

Interpretation of the cryo-EM maps exploited rigid-body fitting of the SERT–antibody complex models derived from previous cryo-EM studies. The outward-open ΔN72/C13 SERT–15B8 Fab complex with a 5-HT model (PDB code: 7LIA) was used as a reference. The initial model was generated via rigid-body fitting of the homology models to the density map in UCSF ChimeraX (84). The model was then manually adjusted in COOT (85). The model was further refined using real-space refinement in PHENIX (86). Strikingly, for both complexes, there was no density for the amino and carboxy terminal and thus the protein models begin at residue 116 and end at residue 654. Careful scrutiny of density maps showed evidence for N-linked glycosylation at Asn 245 but no evidence for phosphorylation or other post translational modification, at the present resolution. Figures were prepared in UCSF ChimeraX.

We performed MD simulations of the cocaine–pSERT complex in a hydrated lipid bilayer to explore the identity of the ligand in the allosteric site ligand. Triplicates of five simulation systems were performed, with the allosteric site occupied by either DHA in charged or neutral forms (DHA^– or DHA⁰, respectively), cholesterol (CHOL), CHS, or n-dodecyl-β-D-maltoside (DDM). The initial coordinates of CHS, DHA^–, and DDM were transferred from the experimental models, while CHOL was constructed into the CHS model, and DHA⁰ was constructed by protonating the DHA^– model using the PSFGEN plugin in VMD (87). The pSERT protein model contains residues 116-654. The missing side chains and hydrogen atoms in the protein were added using the PSFGEN plugin in VMD (87). The cocrystalized antibody fragment was removed. All bound Na⁺ and Cl^– ions, the cocaine molecule, and the two cholesterol molecules bound to transmembrane helices were retained. Glu173 was modeled as a protonated side chain according to pK_a calculations in PROPKA 3.0 (88). A disulfide bond was introduced between Cys 237 and Cys 246. Neutral N-terminal and C-terminal “caps” were added to the first and last residues of the protein segment, respectively. All protein models were internally hydrated using the DOWSER plugin (89, 90) of VMD and externally solvated using the SOLVATE program (https://www.mpinat.mpg.de/grubmueller/solvate). The models were then oriented according to the Orientations of Proteins in Membranes database (91) and embedded in a lipid bilayer consisting of 218 1-palmitoyl-2-oleoyl-sn-glycero-3-phosphocholine (POPC) and 94 CHOL molecules from CHARMM-GUI (92). The systems were next solvated and neutralized with a 150-mM NaCl aqueous solution in VMD (87), resulting in simulation systems of ~160,000 atoms, with approximate dimensions of 112 Å × 112 Å × 120 Å before equilibration. Each simulation system was replicated into three independent copies with lipid distribution randomized by shuffling lipid molecules within each leaflet using the VMD plugin Membrane Mixer (93).

All simulations were performed using NAMD2 (94, 95) and the CHARMM36m force fields (95) for proteins, CHARMM36 force fields for lipids (including CHS, CHOL, DHA^– and DHA⁰) and detergent DDM (96), and the TIP3P model for water (97), along with the non bonded fix modifications for nonbonded interactions (98, 99). The force field parameters for cocaine were obtained from the CGenFF server (100). All simulations were carried out as isothermal-isobaric ensembles under periodic boundary conditions and simulated in a flexible cell, whose dimensions could change independently while keeping a constant ratio in the xy (membrane) plane. A constant temperature of 310 K was maintained using Langevin dynamics with a 1.0-ps⁻¹ damping coefficient, and a constant pressure of 1.01325 bar was maintained with the Langevin piston Nosé–Hoover method (101). van der Waals nonbonded interactions were calculated with a 12-Å cutoff, and a switching function applied between 10 Å and 12 Å. Long-range, electrostatic nonbonded interactions were calculated with the particle mesh Ewald method (102). Bond lengths involving hydrogen atoms were constrained using the SHAKE (103) and SETTLE algorithms (104). Simulations were integrated in 2 fs time steps, and trajectories were recorded every 10 ps.

The five simulation systems, each replicated into three independent copies, were simulated following these steps: 1) 3,000 steps of energy minimization; 2) 15 ns of MD equilibration, during which Cα atoms of the protein, nonhydrogen atoms of ligands, and all bound ions were restrained by harmonic potentials with progressively decreasing force constants (k = 5, 2.5, 1 kcal mol⁻¹ Å⁻² for 5 ns each) to allow for protein side chain relaxation and protein hydration; 3) 150 ns MD run, during which harmonic potentials (k = 1 kcal mol⁻¹ Å⁻²) were applied to only Cα atoms to avoid undesired protein conformational deviation but allowing free diffusion of the allosteric site ligand; and 4) 150 ns production MD run without restraints.

BEUS method (105) was employed to characterize the binding energy profiles of CHOL, DHA^–, and DDM to the allosteric site. The ligand–TM1b/TM6a distance, measured as the center-of-mass distance between the nonhydrogen atoms in the ligand and Cα atoms in the extracellular ends of TM1b and TM6a (residues 145 to 148 and 361 to 364), was chosen as the reaction coordinate to sample ligand binding. The initial distances for the modeled CHOL, DHA^–, and DDM were 17.2, 15.2, and 13.9 Å, respectively. We chose reaction coordinates ranging from 14 to 21.5 Å for CHOL, 15 to 22.5 Å for DHA^–, and 13 to 20.5 Å for DDM, to sample their unbinding from the allosteric site. Each reaction coordinate was divided into 16 windows with a spacing of 0.5 Å. The initial conformations in each window (collectively shown in Fig. 4H as line representations) were captured from steered MD simulations using the COLVAR module⁹⁸ in NAMD, in which the ligand was pulled away from the extracellular ends TM1b and TM6a (residues 145 to 148 and 361 to 364) to the desired distances using a harmonic potential (k = 10 kcal mol⁻¹ Å⁻²) moving at a 0.5 Å ns⁻¹ rate. The BEUS simulations were performed for 60 ns in each window. The Hamiltonian replica exchange was attempted every 1 ps between neighboring windows. Weighted Histogram Analysis Method (WHAM) (106, 107) was used to construct the free energy profiles and perform error analysis using the Monte Carlo bootstrapping method.

D.Y. and E.G. designed research; D.Y., Z.Z., and E.T. performed research; Z.Z. and E.T. analyzed data; and D.Y., Z.Z., E.T., and E.G. wrote the paper.

The authors declare no competing interest.