Structure of the human activating natural cytotoxicity receptor NKp30 bound to its tumor cell ligand B7-H6

We produced recombinant NKp30 by in vitro folding from bacterial inclusion bodies and B7-H6 by secretion from baculovirus-infected insect cells. Surface plasmon resonance (SPR) was used to demonstrate specific binding of NKp30 to B7-H6, which is characterized by very fast on and off rates (Fig. 1 A). Both rates were too fast to measure reliably. Under equilibrium binding conditions, a dissociation constant (K_D) of 1.0 ± 0.2 × 10⁻⁶ M was obtained, comparable to the affinity of PD-1 for PD-L1 (K_D = 6 × 10⁻⁶ M; Lázár-Molnár et al., 2008) or of CTLA-4 for B7-1 (4 × 10⁻⁷ M; van der Merwe et al., 1997). We determined the structure of unbound B7-H6 to 2.0 Å resolution by molecular replacement using the V-like and C-like domains of human PD-L1 (Lin et al., 2008) as separate search models. The partially refined B7-H6 structure and the human CTLA-4 structure (Schwartz et al., 2001) were used as search models for solving the structure of the NKp30–B7-H6 complex to 2.3 Å resolution (Table I). We first describe the structures of NKp30 and B7-H6 individually and then proceed to the NKp30–B7-H6 complex.

NKp30 is a type I transmembrane glycoprotein composed of a single Ig-like extracellular domain connected by a short stalk region of six residues to a transmembrane segment and cytoplasmic domain (Pende et al., 1999). NKp30 is monomeric in the crystal, in agreement with its behavior as a monomer in size exclusion chromatography (unpublished data) and on the NK cell surface (Pende et al., 1999). The Ig-like domain of NKp30 is a two-layer β-sandwich that exhibits the chain topology found in IgC domains (C1 set; Williams and Barclay, 1988; Fig. 2). The front and back sheets, composed of β-strands GFC and ABED, respectively, are linked by a canonical intersheet disulfide bond (Cys39–Cys107). In addition, NKp30 contains two α-helices (α1 and α2) preceding β-strands D and E, respectively. One of these helices (α1) is unique to NKp30 among Ig-like domains.

A Dali structure homologue search (http://www2.ebi.ac.uk/dali/fssp/) identified the closest homologue of NKp30 as the V-like domain of PD-L1, a ligand for the T cell inhibitory receptor PD-1, with a Z-score of 12.8 (24% sequence identity). Superposition of NKp30 onto PD-L1 resulted in a root mean squared (r.m.s.) difference of 1.8 Å for 101 α-carbon atoms. The structural similarity between NKp30 and PD-1 is lower (Z-score = 10.8; 21% sequence identity; r.m.s. difference = 3.1 Å for 101 α-carbon atoms), indicating that NKp30 is more structurally akin to the PD-L1 ligand than to its receptor. NKp30 is considered a member of the CD28 family, which includes PD-1 (CD279), CTLA-4 (CD152), ICOS (CD278), and BTLA (CD272; Brandt et al., 2009). Superposition of NKp30 onto CTLA-4 (Z-score = 10.1; 28% sequence identity) gave an r.m.s. difference of 2.6 Å (95 α-carbon atoms). Thus, NKp30 is clearly more structurally related to PD-1 and CTLA-4 than to adhesion molecules, such as CD2 (Z-score = 9.2 for domain 1; 16% sequence identity) and CD58 (9.2 for domain 1; 16%; Wang et al., 1999), or to the NCRs NKp44 (9.1; 14%; Cantoni et al., 2003) and NKp46 (8.6 for domain 1; 15%; Foster et al., 2003).

Although β-strands GFC in the front sheet and ABE in the back sheet of NKp30 align well with the corresponding strands of PD-1 and CTLA-4, there are large differences on one side of the β-sandwich. Whereas the back sheets of PD-1 and CTLA-4 also include strands C′C′′ or C′, NKp30 lacks these strands and instead contains helices α1 and α2 connected by a short D strand (Fig. 3, A and B). One of these helices (α2) participates in binding B7-H6, as discussed in the next section.

Like other members of the B7 family, the extracellular portion of B7-H6 is composed of a membrane distal V-like domain and a membrane proximal C-like domain (Fig. 2). B7-H6 behaved as a monomer in solution (unpublished data) and crystallized as a monomer in both free and NKp30-bound forms. B7-H6 and PD-L1 share significant structural similarity, with a Z-score of 16.9 (20% sequence identity) and r.m.s. difference of 2.8 Å (191 α-carbon atoms). The structural similarity between NKp30 and B7-1 is lower but still notable (Z-score = 14.1; 17% sequence identity; r.m.s. difference = 5.3 Å for 177 α-carbon atoms). The two antiparallel β-sheets of the B7-H6 V-like domain are formed by strands GFCC′C′′ and ABE, whereas the β-sandwich of the C-like domain is composed of strands GFCC′ and ABED. B7-H6 contains six potential N-linked glycosylation sites, two in the V-like and four in the C-like domain. In both the free and NKp30-bound B7-H6 structures, clear electron density was visible for carbohydrate chains attached to Asn43 and Asn57 in the V-like domain and to Asn208 in the C-like domain (Fig. S1). The V-like domain of B7-H6 terminates at Val140, which is followed by a three-residue coil (Ala141Ser142Pro143) before the start of β-strand A of the C-like domain (Fig. S2). However, this coil is integral to the C-like domain rather than forming a flexible interdomain linker, as in antibody molecules, which probably helps stabilize the extended, upright stature of B7-H6.

A structure-based sequence alignment of B7-H6 and NKp30 showed several shared secondary structural elements, notably strands ABCEFG and helix α2, which is unique to these two proteins among Ig-like domains (Fig. 3, C and D). Additionally, superposition of B7-H6 onto NKp30 resulted in an r.m.s. difference of 1.7 Å for 81 α-carbon atoms, indicating close similarity. The main differences between these molecules reside on one side of the β-sandwich, where B7-H6 has strands C′C′′ and NKp30 instead has strand D, helix α1, and a long loop connecting strand C and helix α1.

The structure of the NKp30–B7-H6 complex revealed an assembly with a 1:1 receptor–ligand stoichiometry and a binding interface formed by the front and back β-sheets of the NKp30 C-like domain and the front β-sheet of the B7-H6 V-like domain (Fig. 2). Residues from strands CF and loops BC and FG (front sheet) of NKp30, along with helix α2 (back sheet), juxtapose residues from strands GFCC′ and loops BC, C′C′′, and FG (front sheet) of B7-H6. The overall organization of the NKp30–B7-H6 complex differs considerably from those observed in the PD-1–PDL1 (or PD1–PDL2; Lázár-Molnár et al., 2008; Lin et al., 2008) and CTLA-4–B7-1 (or CTLA-4–B7-2) structures (Schwartz et al., 2001; Stamper et al., 2001). These differences are best appreciated by superposing NKp30 in the NKp30–B7-H6 complex onto PD-1 and CTLA-4 in the PD-1–PD-L and CTLA-4–B7 complexes (Fig. 4, A and B). Compared with NKp30, PD-1 and CTLA-4 engage their ligands at angles of ∼90° and ∼60°, respectively.

The difference in overall complex architecture is explained by two main differences in the binding mode of NKp30. First, whereas PD-1 and CTLA-4 use only their front β-sheets to engage the corresponding β-sheets of their ligands in a face-to-face manner (Fig. 4, D and E), the binding site of NKp30 comprises elements from both its front and back β-sheets, resulting in engagement of B7-H6 via the side, as well as face, of the NKp30 β-sandwich (Fig. 4 C). Second, B7-H6 contacts NKp30 in an antibody-like manner (see the next section), with greater involvement by the loops of the V-like domain of the ligand than is the case for PD-L or B7. In the PD-1–PD-L complex, whose interface is dominated by strand-to-strand interactions (Fig. 4 D), the result is a relatively compact structure with an end-to-end distance spanning ∼75 Å, compared with ∼100 Å for the NKp30–B7-H6 complex (Fig. 4 A). However, in the CTLA-4–B7 complex, the involvement by the loops of CTLA-4 in binding ligand, which is greater than in the case of NKp30 (Fig. 4, C and E), results in as extended a structure as the NKp30–B7-H6 complex, with a similar end-to-end distance of ∼100 Å (Fig. 4 B).

In the complex, 12 NKp30 residues contact 11 B7-H6 residues through predominantly hydrophobic interactions, with no bound water molecules in the interface (Fig. 5 A; Table II). The three potential N-linked glycosylation sites of NKp30 are located outside the interface with B7-H6 and so should not interfere with binding. The NKp30–B7-H6 complex buries a total solvent-accessible surface of only 1,130 Ų, of which 578 Ų is contributed by NKp30 and 552 Ų by B7-H6. This small interface, which is similar in size to that of the CTLA-4–B7-1 complex (1,250 Ų) but substantially less than the interface of the PD-1–PDL1 complex (1,900 Ų), is at the lower limit of the mean value of 1,600 (±400) Ų for stable protein–protein complexes (Lo Conte et al., 1999). However, the interface is characterized by high shape complementarity, based on a calculated shape correlation statistic (S_c; Lawrence and Colman, 1993) of 0.77 (S_c = 1.0 for interfaces with geometrically perfect fits), which is at the upper end of the range for protein–protein complexes. This high S_c value is readily explained by the topologies of the contacting surfaces, whereby the FG loop of B7-H6 forms a prominent bulge on the surface of the ligand which fits snugly into a deep groove between strands FG and helix α2 of the NKp30 receptor that is unique to this CD28 family receptor (Fig. 5 B). The high shape complementarity of the NKp30–B7-H6 interface, which is also characterized by the complete exclusion of bulk solvent, probably compensates for its small size to achieve sufficient affinity for NK cell activation.

Residues Thr127 and Pro128 of the FG loop of B7-H6 each make 16 van der Waals contacts with NKp30, which together account for nearly half (44%) the total number of contacts (73) in the complex. These residues bind at the base of groove between strands FG and helix α2 of NKp30 (Fig. 5 B). The only two hydrogen bonds in the interface are formed by these two FG loop residues: B7-H6 Thr127 Oγ–N NKp30 Gly51 and B7-H6 Pro128 O–N NKp30 Val53 (Fig. 5 A; Table II). Notably, the FG loop corresponds to complementarity-determining region (CDR) 3 of antibodies and T cell receptors. Additional contacts are provided by residues from the BC (CDR1-like) and C′C′′ (CDR2-like) loops of B7-H6, underscoring an antibody-like interaction with NKp30 which is not evident in the interaction of PD-L or B7 with their receptors (Fig. 4, C–E). In particular, B7-H6 Thr59 and Ser60 (BC loop) engage NKp30 Ile50 (BC loop) and Leu86 (α2 helix) through multiple main chain–main contacts (Fig. 5 A). Overall, the CDR-like loops of B7-H6 contribute 64% of the total contacts to the receptor, compared with 13 and 18% for CDR-like loops of PD-L1 and B7-1, respectively. The remaining interactions are contributed by strands C′, F, and G, including a salt bridge linking B7-H6 Lys130 (G stand) with NKp30 Glu111 (Fig. 5 A; Table II). On the NKp30 side of the interface, the CDR-like loops of the receptor account for 46% of the total contacts to B7-H6, with helix α2 making a substantial contribution (16%).

To functionally validate the NKp30–B7-H6 interface observed in the crystal structure, we mutated contacting residues of NKp30 in loop BC (Ile50), β-strand C (Ser52), or helix α2 (Phe85, Leu86) and measured binding to B7-H6 by SPR. Mutation of NKp30 Phe85 and Leu86 to alanine reduced affinity for B7-H6 by 53-fold (K_D = 5.3 × 10⁻⁵ M versus 1.0 × 10⁻⁶ M for wild-type NKp30; Fig. 1 B). A less pronounced (14-fold), though still significant, effect resulted from alanine substitution of NKp30 Ile50 and Ser52 (Fig. 1 C). Much more disruptive was the replacement of NKp30 Ser52 by arginine (Fig. 1 D), which reduced binding to B7-H6 by 150-fold, thereby confirming the interface identified in the structure.

Superposition of the V-like domain of free B7-H6 onto the V-like domain of B7-H6 in the complex with NKp30 gave an r.m.s. difference in α-carbon positions of 1.1 Å, indicating close overall similarity (Fig. 6 A). However, strands C′C′′ and their associated C′C′′ loop have undergone a movement toward NKp30 to optimize interactions with the receptor (Fig. 6 B). In loop C′C′′, the side chain of B7-H6 Phe82 is rotated by ∼90° around the Cα–Cβ axis to avoid steric clashes with NKp30 Leu113, and the α-carbon atom of B7-H6 Gly83 is shifted by 2.7 Å toward the receptor to allow formation of van der Waals contacts with NKp30 Ile50 and Leu113. The main chains of loops BC and FG of B7-H6 are each displaced ∼1.6 Å away from NKp30 to avoid steric clashes with the receptor.

In addition to structural adjustments in regions of B7-H6 that contact NKp30, helix α2 of B7-H6 is displaced toward the body of the ligand by a mean of 1.7 Å in the position of its α-carbon atoms (Fig. 6 A). This helix is located on the back sheet of the B7-H6 β-sandwich and opposite the receptor binding site. However, this movement does not appear to be attributable to NKp30 binding but rather to lattice contacts with a neighboring NKp30 molecule in the crystal. Free and bound B7-H6 exhibit nearly identical hinge angles between the V-like and C-like domains, suggesting limited conformational flexibility in the interdomain region that is unrelated to receptor binding. In contrast, the hinge angles of free and bound PD-L1 were found to differ by ∼40° (Lin et al., 2008), indicating substantially greater flexibility that is likely attributable to structural differences in the interdomain regions of PD-L1 and B7-H6.

The NKp30–B7-H6 structure revealed that this NK cell activating complex is distinct from the CTLA4–B7 and PD-1–PD-L T cell inhibitory complexes in both overall organization and detailed atomic interactions that mediate binding and specificity. The structure provides the first information on the molecular basis for tumor cell recognition by any member of the NCR family, which includes NKp30, NKp44, and NKp46. In addition, the NKp30–B7-H6 structure establishes certain constraints that must be satisfied by any mechanism proposed to describe signaling through the NKp30 pathway and the integration of these signals into the overall NK cell response.

Upon NK cell–target cell interaction, a highly organized complex of molecules, termed the NK cell immune synapse, is formed at the contact area of both cells (Orange, 2008). These molecules include inhibitory NK receptors that monitor MHC class I expression on target cells (KIRs and Ly49s in human and mouse, respectively), activating NK receptors that recognize MHC-like or non-MHC ligands (NKG2D, DNAM-1, and 2B4), cytoskeletal proteins (F-actin and tailin), and integrins (CD11a and CD11b; Vyas et al., 2002; Orange, 2008). It has been proposed that the unique compartmentalization of molecules observed at immune synapses is regulated by a sorting mechanism based on molecular dimensions, with relatively small signaling molecules (e.g., NK receptors) localizing to a central zone and larger adhesion molecules (e.g., integrins) residing in the periphery (van der Merwe et al., 2000; Bromley et al., 2001; Orange, 2008). Accordingly, we compared the dimensions of the NKp30–B7-H6 complex with those of the KIR–MHC and NKG2D–MICA complexes (Boyington et al., 2000; Li et al., 2001), both of which accumulate at the NK cell synapse (Davis et al., 1999; McCann et al., 2007; Orange, 2008; Fig. 7). In the NKp30–B7-H6 complex, the distance spanned by the molecules is ∼100 Å, which is approximately the distances bridged by the KIR–HLA-C and NKG2D–MICA complexes. This suggests that these receptor–ligand pairs could potentially colocalize with NKp30–B7-H6 in the central zone of the NK cell immune synapse. In addition, tumor cells expressing both MICA and B7-H6 could possibly trigger synergistic activation of NK cells.

An important feature of NKp30 and B7-H6 is their monomeric nature. Whereas CTLA-4 exists as a disulfide-linked homodimer on the T cell surface, NKp30, like PD-1, exists as a monomer in solution, in the crystal, and on the NK cell surface (Pende et al., 1999). In this regard, it should be noted that NKp30 and PD-1 lack the extracellular equivalent of Cys122 in CTLA-4 that forms the intermolecular disulfide in the CTLA-4 homodimer. Similarly, whereas B7, the ligand for CTLA-4, forms noncovalent dimers (Ikemizu et al., 2000), B7-H6 is monomeric both in solution and the crystal. The monomeric state of NKp30 and B7-H6 precludes formation of the type of alternating, periodic assembly observed in CTLA-4–B7 crystal structures, in which bivalent CTLA-4 homodimers bridge bivalent B7 homodimers in the immune synapse (Schwartz et al., 2001; Stamper et al., 2001). Instead, simple diffuse processes probably mediate the engagement NKp30 and B7-H6 in interacting cells and the subsequent accumulation of NKp30–B7-H6 complexes in the central zone of the immune synapse, as proposed for PD-1–PD-L complexes, which are also monomeric (Lázár-Molnár et al., 2008).

Although NKp30 is structurally unrelated to NKG2D, which belongs to the C-type lectin-like superfamily, these two activating NK receptors present several intriguing parallels. Both recognize induced self-molecules expressed on tumor cells that alert innate immunity to cellular transformation: MICA and MICB for NKG2D (Raulet, 2003) and B7-H6 for NKp30 (Brandt et al., 2009). Moreover, both receptors bind multiple ligands. Besides MICA and MICB, human NKG2D interacts with ULBP1–4, whereas mouse NKG2D binds H-6, RAE-1, and MULT1 (Raulet, 2003). Likewise, in addition to B7-H6, NKp30 recognizes the nuclear factor BAT3 (Pogge von Strandmann et al., 2007) and pp65, a human CMV tegument protein (Arnon et al., 2005). However, whereas all NKG2D ligands are structurally related to MHC class I molecules, B7-H6, BAT3, and pp65 appear to be completely unrelated, both in terms of their structure (the proteins lack any sequence homology) and origin (B7-H6 is a cell surface protein; BAT3 is a nuclear protein released upon heat shock; pp65 is a structural protein of CMV found in the nucleus and cytoplasm of infected cells). In the case of NKG2D, crystallographic studies of NKG2D bound to MICA, ULBP3, and RAE-1 (Li et al., 2001, 2002; Radaev et al., 2001; McFarland et al., 2003) have shown that NKG2D recognizes these MHC-like ligands using a single binding site formed by the two identical subunits of this homodimeric receptor. In addition, only ∼60% of the solvent-accessible surface in the NKG2D–ligand complexes is nonpolar (McFarland et al., 2003), whereas the NKp30–B7-H6 interface is 69% hydrophobic. Whether NKp30 binds BAT3 and pp65 through the same site as B7-H6 is not clear, given the disparate nature of these ligands. It is conceivable that NKp30 may possess distinct, albeit overlapping, binding sites for these molecules, which may comprise residues from both its front and back β-sheets, as seen in the NKp30–B7-H6 complex, or be restricted to residues from the front β-sheet alone, in the manner of the CTLA4–B7 and PD-1–PD-L complexes (Schwartz et al., 2001; Stamper et al., 2001; Lázár-Molnár et al., 2008; Lin et al., 2008).

The availability of the NKp30–B7-H6 structure should accelerate the development of molecules to manipulate the NKp30 pathway for tumor immunotherapy. For example, soluble forms of B7-H6, or compounds designed to mimic the CDR-like loops of B7-H6 that bind NKp30, could be used to augment NK cell responses to tumor cells through induction of NKp30-mediated cytolytic activity. Alternatively, soluble forms of NKp30 fused to antibody Fc regions with effector functions may serve as reagents for eliminating B7-H6–expressing tumor cells via antibody-dependent cell-mediated cytotoxicity. Although such therapeutic applications will likely require structure-based engineering of variants of NKp30 that bind B7-H6 with high affinity (or vice versa), this has been achieved in the case of belatacept (Larsen et al., 2005), a high-affinity mutant of CTLA-4 in clinical trials for preventing rejection in kidney transplantation (Snanoudj et al., 2007).

The extracellular portion of NKp30 (residues Leu19–Gly135) was cloned into the expression vector pET-26b (Novagen) and expressed as inclusion bodies in BL21(DE3) Escherichia coli cells (EMD). The inclusion bodies were washed with 50 mM Tris-HCl, pH 8.0, containing 5% (vol/vol) Triton X-100, and then dissolved in 8 M urea, 50 mM Tris-HCl, pH 8.0, and 10 mM DTT. For in vitro folding, the inclusion bodies were diluted into ice-cold folding buffer containing 0.4 M l-arginine–HCl, 50 mM Tris-HCl, pH 8.0, 1 mM EDTA, 3 mM reduced glutathione, and 0.9 mM oxidized glutathione to a final protein concentration of 40 mg/liter. After 72 h at 4°C, folded proteins were separated from aggregates using a Superdex S-75 column (GE Healthcare). Further purification was performed by MonoS chromatography. Mutants of NKp30 were constructed by overlap PCR and expressed in E. coli in the same way as the wild-type receptor.

The extracellular portion of B7-H6 (residues Asp25–Ser262) was fused to the gp67 secretion signal sequence of baculovirus expression vector pAcGP67-B (BD) with a C-terminal His₆ tag and expressed in SF9 insect cells (Invitrogen). In a typical preparation, 500 ml Sf9 cells at 1.4 × 10⁶ cells/ml were inoculated with 12 ml of recombinant baculovirus at 1.0 × 10⁸ pfu/cell. Supernatants were harvested 4 d after infection and loaded onto a Ni²⁺-NTA column (GE Healthcare) for affinity purification. Recombinant B7-H6 was further purified by sequential Superdex S-75 and MonoQ columns.

Crystals of unbound B7-H6 (10 mg/ml) were grown in hanging drops at room temperature in 0.2 M LiCl₂ and 20% (wt/vol) polyethylene glycol (PEG) 3350. For crystallization of the NKp30–B7-H6 complex, 5 mg/ml NKp30 and 10 mg/ml B7-H6 were mixed in a 1:1 volume ratio. The complex crystallized in 20% (wt/vol) PEG 3350 and 0.2 M KF in hanging drops at room temperature.

For data collection, both crystals were transferred to a cryoprotectant solution of mother liquor containing 30% (vol/vol) glycerol, before flash cooling in liquid nitrogen. X-ray diffraction data were collected to 2.0 Å resolution for unbound B7-H6 and to 2.3 Å resolution for the NKp30–B7-H6 complex at beamline X29 of the Brookhaven National Synchrotron Light Source. The B7-H6 crystal belongs to space group P2₁ with one molecule per asymmetric unit; the NKp30–B7-H6 crystal belongs to space group P2₁2₁2₁ with one complex molecule per asymmetric unit. All data were indexed, integrated, and scaled with the program HKL2000 (Otwinowski and Minor, 1997). Data collection statistics are summarized in Table I.

The structure of B7-H6 was solved by molecular replacement with Phaser (Storoni et al., 2004) using the V-like and C-like domains of human PD-L1 (Protein Data Bank accession no. 3BIK; Lin et al., 2008) as separate search models. The partially refined B7-H6 structure and the human CTLA-4 structure (1I85; Schwartz et al., 2001) were used as search models for solving the structure of the NKp30–B7-H6 complex. All refinements were performed using the program CNS1.1 (Brünger et al., 1998), including iterative cycles of simulated annealing, positional refinement, and B factor refinement, interspersed with model rebuilding into σ_A-weighted F_o − F_c and 2F_o − F_c electron density maps using XtalView (McRee, 1999). Refinement statistics are summarized in Table I. Stereochemical parameters were evaluated with PROCHECK (Laskowski et al., 1993). Atomic coordinates and structure factors for the NKp30–B7-H6 complex and B7-H6 have been deposited in the Protein Data Bank under accession nos. 3PV6 and 3PV7, respectively.

The interaction of wild-type and mutant NKp30 receptors with B7-H6 was assessed by SPR using a BIAcore T100 biosensor (GE Healthcare). Typically, 200–1,000 resonance units of B7-H6 were immobilized on a CM5 sensor chip by random amine-coupling. Solutions containing different concentrations of NKp30 were injected sequentially over flow cells immobilized with B7-H6 or buffer as a blank. The data were analyzed using BIAevaluation 4.1 software (GE Healthcare). Dissociation constants (K_Ds) were determined from Scatchard analysis, after correction for nonspecific binding, by measuring the concentration of free reactants and complex at equilibrium. Standard deviations for two or more independent K_D determinations were typically <20%.

Fig. S1 shows electron density for two N-acetylglucosamine residues linked to Asn208 of B7-H6 in the NKp30–B7-H6 complex. Fig. S2 shows the interdomain region of B7-H6.

We thank H. Robinson (Brookhaven National Synchrotron Light Source) for x-ray data collection.

This study was supported by the National Institutes of Health grant AI047990. Support for beamline X29 comes from the Offices of Biological and Environmental Research and of Basic Energy Sciences of the US Department of Energy, and from the National Center for Research Resources of the National Institutes of Health.

The authors declare no competing financial interests.