Computational Design of an Allosteric Antibody Switch by Deletion and Rescue of a Complex Structural Constellation.

Therapeutic monoclonal antibodies have transformed medicine, especially with regards to treating cancers and disorders of the immune system. More than 50 antibody-derived drugs have already reached the clinic, the majority of which target cytokines or cell-surface receptors. Unfortunately, many of these targets have pleiotropic functions: they serve multiple different roles, and often not all of these roles are disease-related. This can be problematic because antibodies act throughout the body, and systemic neutralization of such targets can lead to safety concerns. To address this, we have developed a strategy whereby an antibody's ability to recognize its antigen is modulated by a second layer of control, relying on addition of an exogenous small molecule. In previous studies, we began to explore this idea by introducing a deactivating tryptophan-to-glycine mutation in the domain-domain interface of a single-chain variable fragment (scFv), and then restoring activity by adding back indole to fit the designed cavity. Here, we now describe a novel computational strategy for enumerating larger cavities that can be formed by simultaneously introducing multiple adjacent large-to-small mutations; we then carry out a complementary virtual screen to identify druglike compounds to match each candidate cavity. We first demonstrate the utility of this strategy in a fluorescein-binding single-chain variable fragment (scFv) and experimentally characterize a triple mutant with reduced antigen-binding (Rip-3) that can be rescued using a complementary ligand (Stitch-3). Because our design is built upon conserved residues in the antibody framework, we then show that the same mutation/ligand pair can also be used to modulate antigen-binding in an scFv build from a completely unrelated framework. This set of residues is present in many therapeutic antibodies as well, suggesting that this mutation/ligand pair may serve as a general starting point for introducing ligand-dependence into many clinically relevant antibodies.

Introduction

Monoclonal antibodies have had a transformative impact on biology and medicine, both as tools for scientific discovery and as precisely targeted therapeutic agents. Their ability to precisely inhibit or activate some biological target of interest, coupled with dramatic engineering successes to allow antibody humanization and enhanced effector functions,[1] antibody-drug conjugates,[2] and bispecific antibodies,[3] together provide ample room for antibodies to continue growing as tools for therapeutic intervention and for enhancing understanding of complex biological systems.

Most antibody constructs approved as drugs or in current clinical trials address various indications in oncology or immunology by targeting cytokines or cell-surface receptors.[4] While aberrant signaling from these antigens is typically localized to a subset of tissue types, the biodistribution of antibody-derived constructs can be hard to precisely control.[5] This is particularly problematic because many of these potential targets—cytokines and cell-surface receptors—also serve important functions unrelated to the disease state, elsewhere in the body and in other biological processes. Accordingly, such pleiotropic activities can underlie dose-limiting toxicity and/or other adverse events associated with systemic antagonism of these targets.[6−9]

To address this, we envisioned a scenario in which “switchable” antibodies could be systemically administered, and then locally activated in a spatially regulated manner. As a first step toward this goal, we therefore sought to engineer ligand-dependent antigen recognition into an antibody framework. A number of approaches have been described for building small-molecule-dependent activity into proteins, most commonly by fusing a (pre-existing) responsive domain into the protein of interest[10,11] or by splitting the target protein into two separate pieces that are brought together upon assembly of fused ligand-dependent dimerization domains.[12,13] In order to avoid adding an additional domain onto the antibody as part of our design strategy, however, we instead sought to integrate the ligand-binding site directly into the antibody framework itself.

In the past, we have shown that introducing a tryptophan-to-glycine (W → G) substitution at a carefully selected position can lead to loss of protein activity via discrete conformational changes and/or altered protein stability or dynamics; the subsequent addition of indole—chosen to match the atoms removed by this mutation—can precisely revert this disruption and, thus, rescue the protein’s activity.[14−17] We have applied this “indole rescue” strategy to modulate activity of enzymes,[15,16] a fluorescent protein,[14] a transcription factor,[14] and most recently an antibody.[17] In each case, however, millimolar concentrations of indole were needed in order to recover meaningful protein activity: this strongly limits the potential applications of these switchable proteins and certainly precludes any in vivo applications.

In the course of these previous studies we explored rescue of a W → G substitution using a series of indole analogues and found that none of these rescued activity better than indole itself:[16] this underscored the need to use a ligand that precisely matched the designed cavity. At the same time, we speculated that the high concentration of indole needed to activate these designed switches is a fundamental limitation of the attainable binding affinity available with such a small ligand.[18]

To overcome this limitation, here, we report a computational strategy for enumerating larger and more complex cavities that can be introduced into proteins through multiple simultaneous large-to-small mutations at adjacent buried sites. We couple this approach with virtual screening to define which of these cavities can be complemented with a suitable ligand and, thus, can serve as the basis for a more effective protein switch. We have applied this strategy to screen for candidate mutant/ligand pairings in a model antibody system, a single chain variable fragment (scFv) that recognizes fluorescein as its antigen. Our overarching design goal was to identify a set of cavity-forming mutations that would “rip” apart (dissociate) the heavy and light chains of the scFv, prior to “stitching” them back together using a ligand that complements the cavity and restores the precise orientation of the domains needed for recognition of antigen (Figure a).

Figure 1

Design of switchable antibodies. (A) Summary of the design strategy. Introducing cavity-forming mutations in the interface between the heavy and light chains will lead to dissociation of this domain–domain interface, leading to loss of antigen-binding activity. Subsequent addition of a rescuing ligand will induce reassociation of the domain–domain interface and, thus, restore activity. (B) Computational design strategy. (I) All possible combinations of two/three-residue cavity-forming mutations at this domain/domain interface are exhaustively considered, to identify those yielding a suitable cavity for subsequent rescue by a druglike small molecule. (II) Energetically favorable three-dimensional conformations (“conformers”) are generated from each member of a large compound library. (III) For each “constellation of atoms” that can be deleted from the protein domain/domain interface, potential structural matches are identified from the library of three-dimensional compound conformations. (IV) The top-scoring structural matches are refined in the context of the (mutant) protein environment, and the best 5 resulting designs are selected for experimental characterization. Using anti-His and anti-kappa (light chain) Western blots, we find that only M2 is solubly expressed in its complete form (the uncropped Western blots are shown in Figure S1); we therefore focused further characterization on this design, which we refer to as Rip-3/Stitch-3. (C) Design model of the rescued Rip-3/Stitch-3 complex. The crystal structure of 4D5 (which harbors the same framework as 4D5Flu) was used as a starting point, and the fluorescein antigen was modeled from a separate antibody (4-4-20). The residues that comprise the Rip-3 constellation are indicated in blue sticks and are shown in superposition with the ligand predicted to rescue this mutation (Stitch-3, orange).

Results

In earlier studies of indole rescue, it was possible to simply test one at a time each potential (single) W → G substitution, and compare activity of the mutant in the presence and absence of indole.[14,15,17] If mutations are allowed at residues other than tryptophan, and multiple mutations can be introduced at once, the number of potential combinations grows rapidly. Further, if each variant harbors a different and more complex cavity, one cannot expect to test activity with a single preselected rescuing ligand. For this reason, we developed a new computational pipeline to screen for promising mutant/ligand pairings (Figure b) and applied it to the crystal structure of 4D5Flu, a fluorescein-binding scFv. The overarching strategy of this pipeline and its application to 4D5Flu are described below, with implementation details provided in the Methods section.

Computational Strategy: Enumerating Candidate Cavity-Forming Mutations

At the outset, we defined the following criteria for candidate mutations:

We will consider only double and triple mutants.

We will consider all possible substitutions in which the original residues are replaced by a smaller one, provided that the atoms of the new side chain are an exact subset of the original side chain (i.e., Thr can be replaced with Ser, Ala, or Gly).

The mutations must together remove at least 12 non-hydrogen atoms from the protein.

At least one mutated residue must be aromatic.

The mutated side chains must be in close proximity to one other, so that the resulting cavity will be contiguous.

These criteria were primarily selected such that they would define cavities approximately the size of druglike chemical matter (i.e., molecular weight 200–500 Da). The requirement that substitutions must leave behind only atoms from the original side chain (rule 2) circumvents the need for carefully modeling the new cavity explicitly: instead, we model the protein variant by simply removing atoms deleted by the substitution of interest.

We implemented these rules in a new program that is part of the Rosetta macromolecular software suite.[19] The program loops over all possible 2- and 3-residue cavity-forming mutations and evaluates adherence to these rules. For each valid combination, our program determines which atoms would be removed from the mutated side chains. It then exports the corresponding “constellation”: the three-dimensional arrangement of deleted atoms in the original protein structure, which in turn defines the shape of the cavity and will thus serve as a template for identifying the complementary ligand (Figure b, part I).

In addition to these, we also required that candidate mutations be positioned in the antibody framework region (i.e., in the domain–domain interface of the scFv), and that none of the target residues could be located in the CDR loops. Applying this exhaustive search to 4D5Flu yielded a collection of 862 unique constellations to serve as templates for virtual screening: each of these corresponding to the cavity produced by a distinct double or triple mutant at the scFv’s domain–domain interface.

Computational Strategy: Identifying the Complementary Rescuing Ligand

To search for ligands that might complement individual constellations from our collection, we began by assembling a (virtual) chemical library. At the time of our study the ZINC12 database[20] was composed of ∼20 million commercially available compounds: we then curated this collection by filtering for compounds based on Lipinski’s criteria[21] (to ensure drug-likeness of the compounds we would ultimately test) and also required molecular weight ≥180 Da (to avoid fragments like indole not expected to provide sufficient binding affinity) and maximum 4 rotatable bonds (to avoid compounds with long aliphatic chains). We further excluded any compounds likely to exhibit pan-assay interference properties (PAINS)[22−24] or containing a chemically reactive group.[25,26] Collectively these filters reduced the size of our collection to ∼3 million entries. In order to fit back into the cavity provided by the mutant protein, a given compound must match the three-dimensional shape and properties defined by the constellation. For each compound, we therefore built a set of ∼100 low-energy conformations and collected these into a large 3D “conformer” library comprised of ∼300 million entries (Figure b, part II).

In essence, the challenge of identifying compounds to match a given constellation is analogous to “scaffold hopping” in medicinal chemistry:[27] our constellation corresponds to a core structure (albeit without a single bond-connected structure), and we seek an alternate chemical scaffold to engage the same “receptor” (the cavity from which the constellation was generated). Inspired by this analogy, we used the ROCS scaffold hopping software[28,29] to rapidly compare every one of the 862 constellations against each of the ∼300 million conformers (Figure b, part III). Each of these ∼2.5 billion comparisons (overlays) entailed aligning the conformer onto the constellation, and then evaluating their similarity on the basis of overall shape and conservation of chemical features (e.g., position of hydrogen bond donor/acceptors, aromatic rings, etc.).

Finally, to account for the low-resolution nature of these pharmacophore comparisons, we explicitly built models of the protein–ligand complexes from the top-scoring constellation/ligand pairings. While the rescuing ligand is intended to mimic the atoms in the constellation, inevitably some differences arise because the rescuing ligand is a single entity: in particular, there are often additional atoms linking together the functional groups that correspond to parts of the constellation. The extra atoms comprising these linkers, if not chosen carefully, might clash with the walls of the designed cavity.

For a given conformer, the intended alignment to the constellation was already known from the previous step. Because the constellation was taken directly from the protein structure, the ligand–constellation alignment thus provided a starting point for orienting the ligand relative to the protein. Modeling the designed mutations into the protein was also trivial, given our requirement that each side chain could only be mutated to a smaller residue with atoms corresponding to a subset of the original side chain. For each of the top-scoring constellation/ligand pairings, then, we rapidly generated an initial model of the protein/ligand complex (Figure b, part IV). This model of the complex was then subjected to refinement using the Rosetta energy function, and the resulting models were filtered on the basis of a series of structural and energetic criteria (see the Methods section).

At this point, we selected for experimental characterization the top-scoring five designs, which we denoted M1–M5. We began by evaluating the soluble expression of each construct in Escherichia coli and found M2 to be much more readily produced than the other four (Figure b); while M5 was also expressed, it appeared to lack the His-tag we planned to use for purification. In addition to the full-length construct, we found partial fragments from our expression of the wild-type (WT) construct, as well as for M1, M2, and M5 (Figure S1). While examination of the other four top-scoring models provides insight into the types of matches that emerge from our computational pipeline (Figure S2), on the basis of protein production considerations we chose to focus our efforts fully on characterization of the M2 design. We shall heretofore refer to this particular designed pair as Rip-3/Stitch-3 (Figure c).

Stitch-3 Rescues Antigen-Binding in 4D5Flu Rip-3

4D5Flu is an anti-fluorescein scFv that was designed by grafting the CDR loops from 4-4-20 (an anti-fluorescein Fab)[30,31] onto the humanized anti-HER2 scFv scaffold 4D5.[32] The two halves of 4D5Flu are connected with a 30-residue linker (Gly4Ser)6. 4D5Flu is a convenient model system because this construct is readily expressed in E. coli and because the intrinsic fluorescence of 4D5Flu’s antigen, fluorescein, allows for straightforward monitoring of binding. The Rip-3/Stitch-3 pairing from our computational design pipeline entails incorporating a triple mutation (Rip-3 = VLF98G, VHV37A, and VHW110G) into 4D5Flu to reduce activity, and then adding its complementary ligand (Stitch-3 = 6-(benzyloxy)indazole) to restore activity (Figure c).

Binding by 4D5Flu quenches the fluorescence signal from fluorescein; we took advantage of this property to monitor the quenching activity. In the absence of rescuing ligand, Rip-3 showed diminished quenching relative to WT 4D5Flu; upon addition of 100 μM Stitch-3, activity of rescued Rip-3 is restored nearly to that of WT (Figure a). We further found that rescue of fluorescence quenching exhibits dose-dependence with respect to Stitch-3, yielding an EC50 value of 16 μM (Figure b).

Figure 2

Stitch-3 rescues antigen-binding activity in 4D5Flu Rip-3. (A) The designed Rip-3 variant of 4D5Flu quenches fluorescein fluorescence less than WT 4D5Flu, implying reduced antigen-binding. Upon addition of Stitch-3, increased quenching is observed. (B) Addition of Stitch-3 rescues fluorescence quenching of Rip-3 in a dose-dependent manner. (C) Protein titrations in the presence or absence of 100 μM Stitch-3. All data are presented as mean ± SEM, n = 8.

To fully rule out the possibility of a direct effect of Stitch-3 on fluorescein fluorescence, we collected dose–response curves for both WT 4D5Flu and Rip-3, in the absence and presence of Stitch-3 (Figure c). It is evident that presence or absence of Stitch-3 does not affect the WT binding curve, whereas addition of Stitch-3 shifts the binding curve of Rip-3 closer to that of the WT (i.e., Rip-3’s binding affinity for fluorescein is enhanced in the presence of Stitch-3). Though the midpoint of the curves is shifted, the end points of the curves match in all cases: this confirms that the difference in fluorescence quenching is not because Rip-3 binds fluorescein in an altered state that is more/less quenched, but rather that mutation and rescue simply modulate binding affinity of Rip-3 for fluorescein.

In summary, then, we have shown that collectively introducing the Rip-3 triple mutation into 4D5Flu leads to partial loss of antigen-binding activity, and that activity can be rescued by addition of Stitch-3. These observations are consistent with the original model underlying our design strategy (Figure a).

Stitch-3 Stabilizes Rip-3 in the Presence of Antigen

In certain previous cases in which indole was used to complement tryptophan-to-glycine (W → G) substitutions, we observed that inactivation and rescue could be mediated through loss and rescue of protein stability.[14,17] We therefore used differential scanning fluorimetry (DSF, aka ThermoFluor) to probe Rip-3’s thermostability.[33]

For both WT 4D5Flu and the Rip-3 variant, we began by using DSF to monitor the thermal unfolding transition that results from increasing temperature (Figure S3) and found apparent melting temperatures (Tm) of 56.2 and 44.0 °C, respectively (Figure a). Upon addition of 5 μM fluorescein (equimolar to the protein concentration in this experiment, and far above the dissociation constant for this interaction), both the WT and the Rip-3 show increased Tm indicating that fluorescein binding stabilizes the folded protein (Figure b).

Figure 3

Stitch-3 rescues thermal stability of Rip-3 4D5Flu and induces its compaction and linker ordering. (A) Apparent thermal unfolding temperature for 4D5Flu WT and Rip-3, as measured via differential scanning fluorimetry. (B) Upon addition of 5 μM fluorescein, both WT and Rip-3 show increased thermal stability. (C) Upon addition of 100 μM Stitch-3 (without fluorescein), both WT and Rip-3 show decreased thermal stability. (D) In the presence of both fluorescein and Stitch-3, both WT and Rip-3 show increased thermal stability. The increase in stability of Rip-3 exceeds that observed upon addition of fluorescein alone, whereas the increase in stability of the WT is less than that observed upon addition of fluorescein alone. All data are presented as mean ± SEM, n = 8. (E) 1D NMR spectra of T1 relaxation measurements at 0.02 s relaxation delay, for 4D5Flu Rip-3 alone (blue) and upon addition of 100 μM Stitch-3 (red). The spectrum shows three dominant peaks in the 8.1–8.5 ppm range: based on their intensities and their positions in the 1H-15N HSQC spectrum, we infer that these peaks correspond to the glycine and serine residues in the scFv’s synthetic linker. (F) Portion of the 2D 1H-15N HSQC spectrum of Rip-3 4D5Flu in the presence of 100 μM Stitch-3.

We next added 100 μM Stitch-3 (without fluorescein) and found this to destabilize both WT and Rip-3 (Figure c). The change in Tm for the WT scFv (which lacks a designed binding site for Stitch-3) suggests that this compound is destabilizing to the protein through a nonspecific mechanism, presumably by making the solvent more nonpolar and thus stabilizing exposure of the protein’s hydrophobic side chains. The fact that Rip-3 is similarly destabilized suggests that the binding site for Stitch-3 is not present, presumably because the protein does not adopt the necessary conformation in the absence of fluorescein (at least, under these conditions).

To test this, we next monitored unfolding in the presence of both fluorescein (5 μM) and Stitch-3 (100 μM). Under these conditions the WT construct was slightly stabilized, but less than observed with fluorescein alone (Figure d). In essence, the two ligands appear to act independently on the WT construct, in competing directions. By contrast, addition of the two ligands together stabilizes Rip-3 more than fluorescein alone. Put another way, under these conditions, Stitch-3 stabilizes Rip-3, but only if it has already adopted the conformation appropriate for antigen-binding—which is also the conformation needed for Stitch-3 binding. Thus, addition of fluorescein (at a concentration far above the dissociation constant) reverses the conformational change resulting from introducing the triple mutation and, thus, preorders the designed binding site for Stitch-3. Whereas Stitch-3 is destabilizing for the WT construct in the presence of fluorescein, this explains why it is stabilizing for Rip-3 in the presence of fluorescein.

Collectively, these findings point to positive cooperativity between antigen-binding and Stitch-3 binding in the Rip-3 construct. Our fluorescence quenching experiment demonstrated that Stitch-3 enhances Rip-3’s binding affinity for antigen, and here, we demonstrate the converse: the presence of antigen enhances binding of Stitch-3.

Stitch-3 Enhances the Interaction between the Two scFv Domains

Next, we sought to directly test the structural basis for rescue of activity, by characterizing protein conformational changes in Rip-3 occurring in response to Stitch-3. To do so, we produced uniformly 15N labeled Rip-3 protein and measured longitudinal (T1) and transverse (T2) 15N relaxation times at 20 °C. In a stacked plot of the first spectra of T1 relaxation (delay of 0.02 s) of Rip-3 with and without the presence of equimolar Stitch-3 (100 μM), recorded under precisely the same conditions, the spectra are dominated by three large peaks around 8.1–8.4 ppm (Figure e). These peaks mainly arise from glycine and serine residues in the linker, as made evident by their large intensities and their positions in the 15N HSQC spectrum (Figure f). Addition of Stitch-3 results in a 3.4-fold increase in peak intensity (signal-to-noise ratio is 31 for the protein alone, and 104 upon addition of Stitch-3). While longitudinal T115N relaxation times for Rip-3 remain nearly unchanged with and without Stitch-3, the calculated transverse T215N relaxation times are significantly increased after the addition of Stitch-3; this observation is consistent for T2 values obtained both from the calculations based on peak intensity of the glycine residues in the linker (largest glycine peak around 8.3 ppm) and from calculations that integrate values between 7 and 10 ppm (Figure S4 and Table S1).

For rigid protein molecules, in the limit of slow molecular motion (τc ≫ 0.5 ns) in high magnetic field (500 MHz or greater), there exists a closed form solution for τc as a function of T1/T215N relaxation times:[34−36]According to the Stokes–Einstein equation:where νN is nuclear frequency, η is viscosity, k is the Boltzmann constant, T is temperature, and α is hydrodynamic radius. Thus, decreasing T1/T2 reflects a decrease in τc, and a decrease in τc reflects a decrease in α.

Thus, the fact that addition of Stitch-3 leads to decreased T1/T2 suggests that, upon binding, the rescuing ligand reduces the Rip-3 protein’s overall hydrodynamic radius and rotational correlation time. Addition of Stitch-3 also leads to narrowing of peaks in the glycine and serine region of the HSQC spectrum (Figure f), suggesting increased ordering of the linker as Stitch-3 induces the protein to become more compact.

These results may seem at odds with those of the preceding thermal unfolding analysis. In the thermal unfolding experiment, we did not observe evidence of Stitch-3 binding to the Rip-3 protein until antigen was present; by NMR, however, we observe a clear structural response to Stitch-3 even in the absence of antigen. However, we note that the conditions used for these two experiments are somewhat different from one another. Most notably, the thermal unfolding transition requires elevated temperature, which in turn may explain the apparent lack of binding in the earlier setting.

Overall then, these results are consistent with a model in which the two domains comprising the scFv are dissociated from one another in Rip-3 and then reassociate in response to addition of Stitch-3. Because residues involved in antigen-recognition are distributed across the CDR loops of both domains, this provides a mechanistic explanation for the loss and then recovery of antigen-binding upon mutation and subsequent rescue (Figure a).

Structure–Activity Relationship of the Rescuing Ligand

From our previous studies of (W → G) substitutions, we observed that indole (corresponding exactly to the tryptophan side chain) provided the most effective rescue of activity; however, we also observed that other ligands yielded partial rescue, with rank order commensurate with their similarity to indole.[16,17] These studies motivated our selection of Stitch-3 as the complement for this particular Rip-3 mutant: we anticipated that an optimal rescuing ligand would closely recapitulate the three-dimensional shape and chemical features of the deleted atoms in the original protein structure (i.e., the template “constellation”), and this formed the basis for our virtual screening approach.

To test the hypothesis that Stitch-3 is indeed a near-optimal rescuing ligand for Rip-3, we used a series of chemical analogues to explore how varying the structure of the rescuing ligand affects recovery of antigen-binding activity. In future, we do anticipate that analogues of Stitch-3 can be identified with improved potency (i.e., such analogues would activate Rip-3 at lower concentrations than those needed for Stitch-3). At this point, however, our primary goal was to test whether compounds with similar size and physicochemical properties would also activate Rip-3: if so, this would demonstrate that rescue could be occurring through nonspecific binding rather than precise molecular recognition.

The two ring systems in Stitch-3 are intended to separately mimic a deleted phenyl ring (from VLF98G) and an indole ring (from VHW110G); for this reason, we split the Stitch-3 structure into two parts: the left side consists of the phenyl ring and the ether linker, while the right side is composed solely of the indole ring. We then searched for Stitch-3 analogues aiming at probing changes to the left side of the compound (benzyloxy group), the right side (indazole), or both at once; we specifically sought analogues with shared chemical structure but different three-dimensional features, to probe the importance of precise complementarity in driving rescue of antigen-binding. For expediency, we restricted our search to commercially available compounds, and purchased 12 such analogues (Figure a).

Figure 4

Stitch-3 rescues activity more effectively than chemical analogues which do not share its complementarity for Rip-3. (A) Chemical structure of Stitch-3 (black). Several analogues were selected for characterization by varying the left side/linker of Stitch-3 (green), the right side of Stitch-3 (blue), or both sides at once (orange). (B) At a concentration of 10 μM, Stitch-3 rescues Rip-3’s antigen-binding activity (fluorescence quenching) more effectively than any of these chemical analogues tested. All data are presented as mean ± SEM, n = 8.

We tested the degree to which each of these compounds rescues activity of Rip-3 4D5Flu, using the fluorescence quenching assay described earlier. At a concentration of 10 μM, we find that none of these analogues rescue fluorescein binding to the extent of Stitch-3 (Figure b). This observation is consistent with our model (Figure a), whereby rescue requires that the rescuing ligand must bind precisely to the cavity left by the designed mutations in order to restore the structure to exactly its wild-type conformation. This further observation supports the utility of detailed structure-based modeling in designing such switches: our underlying hypothesis was that an optimal ligand would be the one that precisely matches the three-dimensional arrangement of the deleted atoms, and the activity of Stitch-3 (relative to these close analogues) supports this hypothesis. We expect this observation to guide future medicinal chemistry optimization studies aimed at improved potency of Stitch-3: it is unlikely that drastic changes to the chemical structure will allow derivatives to complement the designed cavity in Rip-3, and thus analoging should focus on relatively modest substitutions.

Transferability of the Rip-3/Stitch-3 Switch

A further important consequence of this hypothesis is that Stitch-3 should also serve as a suitable rescuing ligand of the same triple mutant in other antibodies—or at least, other scFv’s—provided they share the three deleted side chains in precisely the same arrangement as found in 4D5Flu. Intrigued by this possibility, we searched for occurrences of this constellation in antibody databases.

Starting with AbYsis,[37] a database of antibody sequences, we find that the three residues comprising this constellation are present in 63% of the 150 697 total entries (Figure a). Restricting our search to antibodies with available crystal structures, as collected in the SAbDab database,[38] we find that 78% of the 3179 entries have this constellation (Figure a), and in each case the geometry precisely mimics that observed in 4D5Flu (Figure b). The fact that this constellation is so strongly conserved among antibodies—in sequence, but more importantly in structure—implies that this Rip-3/Stitch-3 pairing may serve as a very general means to modulate activity of many other antibodies (and especially scFv’s), without the need to repeat the computational design approach presented above.

Figure 5

The Rip-3 constellation is structurally conserved, allowing the Rip-3/Stitch-3 pairing to be transferred into other unrelated antibody frameworks. (A) In the antibody sequence database AbYsis, 63% of all entries harbor the three residues comprising the designed constellation (VLF98, VHV37, and VHW110). In the antibody structure database SAbDab, 78% of all entries harbor these three residues in precisely the same conformation as 4D5Flu’s parent, 4D5. Using alternate numbering schemes, these residues correspond to VLF98, VHV37, and VHW103 (Chothia), or VLF118, VHV42, and VHW118 (IMGT). (B) Despite their overall sequence divergence in the framework region, all five representative antibodies shown here [Adalimumab (anti-TNF-α), Ipilimumab (anti-CTLA-4), Atezolizumab (anti-PD-L1), Nivolumab (anti-PD-1), and 8B10 (anti-MPTS)] share the precise arrangement of these three side chains found in 4D5. (C) Taking as a model system the MPTS-binding antibody 8B10, and using MPTS quenching as a readout, we find that introducing the Rip-3 triple mutation leads to an scFv with antigen-binding EC50 of 1.0 μM. Upon addition of 100 μM Stitch-3, the EC50 improves 10-fold to a value of 100 nM. All data are presented as mean ± SEM, n = 8.

To test the transferability of our design, we selected from this group another antibody with readily assayed activity. The antibody 8B10 was affinity matured to recognize the fluorescent dye 8-methoxypyrene-1,3,6-trisulfonic acid (MPTS) as its antigen,[39,40] and it quenches MPTS fluorescence upon binding. The chemical structures of fluorescein and MPTS are quite dissimilar from one another (Figure S5a), and more importantly the framework regions of 4D5Flu and 8B10 are very different: they share only 56% sequence identity (Figure S5b). Unsurprisingly, given their different antigens and origins, the CDR loops in these two antibodies differ in both their lengths and their conformations. Despite these differences, however, both share the precise arrangement of the residues comprising the Rip-3 constellation (Figure b), suggesting that 8B10 could also be modulated using precisely the same Rip-3/Stitch-3 pairing.

We first designed an scFv version of 8B10, by directly transferring to it the serine/glycine linker from 4D5Flu. We then incorporated the Rip-3 triple mutation into the 8B10 scFv and used MPTS quenching as a measure of antigen-binding. In the absence of Stitch-3, this mutant scFv quenches MPTS with an EC50 value of 1.0 μM; upon addition of 100 μM Stitch-3, the EC50 becomes 10-fold tighter, with a value of 100 nM (Figure c). Thus, our characterization of the Rip-3/Stitch-3 pair in the context of 4D5Flu allowed us to rapidly generate an analogous switch in an unrelated antibody framework, by virtue of the structural conservation of the mutated constellation. These results confirm that this mutation/ligand pairing is indeed transferrable, provided that the target antibody framework harbors these conserved residues in precisely the same geometry as in 4D5Flu.

Discussion

The broad utility of antibodies—as therapeutics and as research tools—has motivated the desire for variants with antigen-binding activity that can be externally controlled. Applications drawn from cancer therapeutics exemplify this need: both chimeric antigen receptor engineered T cells (CAR-Ts) and checkpoint inhibitors are susceptible to adverse effects from off-tumor, on-target side effects from cross-reactivity with healthy tissues.[6−9] Some approaches to this problem seek to simply localize the antibody or CAR-T cell using tumor-selective markers: for example, building in targeting using bispecific formats of antibodies[41−44] or adaptor modules in CAR-T cells[8,45−47] that recognize tumor markers. In contrast to these “passive” localization strategies, other approaches instead seek to control antibody activity through direct responsiveness to external stimuli.

Such approaches require that antigen-binding be directly dependent on some external stimulus. To date, there have been two broad strategies to achieve this. The first strategy entails redesigning an scFv linker to include a conformationally responsive binding site drawn from some other protein, such that a conformational change in the fused protein pulls apart the folded scFv. By incorporating either an elastin-like polypeptide[48,49] or calmodulin[50] into the scFv linker, for example, antigen-binding could be coupled to changes in ionic strength or to a peptide. However, it is unclear to what extent re-engineering the scFv linker might be extensible to other antibody formats.

The second strategy to achieve direct responsiveness to external stimuli involves introducing a temporary “protecting-group” onto the antibody CDR loops to prevent antigen-recognition: this has been achieved either through phosphorylation (reversed when the antibody encounters a phosphatase)[51] or by fusing protein-based “masking domains” using a protease-sensitive linker (reversed when the cognate protease is encountered).[52,53] The use of such masking domains has proven especially promising, with an anti-PD-L1 “probody” activated by tumor-associated proteases recently having begun a Phase I/II clinical trial.[54,55] A structure-based computational method was also implemented to achieve the converse, namely, building a CDR loop that only adopts a conformation suitable for antigen-recognition upon binding Zn2+.[56] That said, a fundamental limitation of building responsiveness into the CDR loops is that the antibody’s structural diversity is focused here: designs that involve carefully altering the CDR loops are not expected to be robust or directly transferrable onto antibodies with different CDR loops (i.e., those that recognize different antigens).

In the study presented here, we were determined to create a robust and transferrable allosteric antibody switch, that can be selectively activated using a bio-orthogonal ligand with druglike physicochemical properties. Already we have demonstrated that the Rip-3/Stitch-3 pairing can be directly transferred to an unrelated antibody framework, and we expect that this will prove true for many other antibodies that include this constellation of residues (Figure ). We note that the intrinsic strength of the VH/VL interaction can vary between antibody frameworks:[57] because of the mechanism of inactivation and rescue, and drawing from a theoretical groundwork developed using other biomolecular switches,[58−60] we expect that this intrinsic interaction strength may serve to “tune” the parameters of the switch. In particular, this “population-shift” model predicts that antibodies with intrinsically tighter VH/VL interactions may retain more background antigen-binding in the absence of Stitch-3 (but be activated at lower concentrations of Stitch-3), whereas those with weaker VH/VL interactions may require a higher concentration of Stitch-3 for activation (but have less background antigen-binding in the absence of Stitch-3). This thermodynamic model is particularly powerful because it provides a rational basis by which additional mutations might be incorporated into the VH/VL interface (remote from both the antigen-binding site and the Stitch-3 binding site), in order to arbitrarily tune the parameters of the switch.

In the present study we have not extended testing of the Rip-3/Stitch-3 pairing beyond the scFv format and, thus, cannot yet confirm applicability to a Fab (or IgG) context. We anticipate that, in such a context, the dissociated (mutant) variable domains may be more restricted by the presence of the constant regions: this, in turn, may lead to retention of higher background in the absence of Stitch-3. While further studies are undoubtedly necessary to define the applicability of the Rip-3/Stitch-3 pairing in this setting, we expect that this constraint will behave in a manner analogous to an scFv with intrinsically strong VH/VL interaction and, thus, may be rationally tuned via the same population-shift strategy described above.

Finally, we acknowledge that inducing antigen-recognition through binding of Stitch-3 is not itself sufficient for selectively activating an antibody in a specific microenvironment, such as in the vicinity of a tumor. In contrast to an antibody that responds directly to a selected stimulus (e.g., the presence of matrix metalloproteases found in the tumor microenvironment[52−55]), the modularity of our strategy allows for the future design of many diverse responsive derivatives. In particular, drawing from the fact that Stitch-3 is completely enclosed in the Rip-3 cavity, and the fact that our SAR studies demonstrate that rescue is sensitive to the exact complementarity of the ligand for this cavity, one can easily envision a broad set of labile handles that might be attached to Stitch-3 to prevent activity until triggered by the intended stimulus. We expect that the wealth of extant knowledge in designing prodrugs for activation in highly specific in vivo environments can now begin to be applied for controlling antibody activity.[61]

Methods

PDB Structures Used in Calculations

The crystal structure of 4D5Flu is not publicly available. Accordingly, the computational approach leading to the Rip-3/Stitch-3 design was deployed using the crystal structure of 4D5 (PDB ID 1FVC), which has the same framework as 4D5Flu. The location of the fluorescein antigen (Figure ) was modeled onto this structure using a related antibody, 4-4-20 (PDB ID 1FLR).

The Rip-3 triple mutation reported in our studies (VLF98G, VHV37A, and VHW110G) refers to sequential residue numbering. The corresponding residues in alternate numbering schemes are VLF98G, VHV37A, and VHW103G (Chothia[62]) or VLF118G, VHV42A, and VHW118G (IMGT[63]).

Comparisons to other antibodies harboring the Rip-3 constellation (Figure ) were based on the following crystal structures: Adalimumab (anti-TNF-α), PDB ID 3WD5; Ipilimumab (anti-CTLA-4), PDB ID 5XJ3; Atezolizumab (anti-PD-L1), PDB ID 5X8L; Nivolumab (anti-PD-1), PDB ID 5GGR; and 8B10 (anti-MPTS), PDB ID 4NJ9.

Enumerating Candidate Cavity-Forming Mutations

Candidate “constellations” were extracted from the protein structure using a new application in the Rosetta software suite[19] written for this purpose. This application is now distributed with the Rosetta source code; Rosetta is freely available for academic use (www.rosettacommons.org). For the studies reported here, we used the following Rosetta command-line options:

constel.macosclangrelease -s input_pdb -ignore_unrecognized_res -suppress_zero_occ_pdb_output -ignore_zero_occupancy false -pair_all_res -max_atom_sasa 10 -chain_interface -aromatic -cnl_exclude CDR_loop_resi.txt

Importantly, our previous studies have shown that the optimal rescuing ligand does not correspond exactly to the deleted atoms, but rather must account for the fact that a covalent bond is lost. Conceptually a W → G substitution would remove atoms corresponding to 3-methylindole, but addition of 3-methylindole would lead to a steric clash between the Gly Cα and the methyl group (in the wild-type protein, these are close to one another because they are covalently bonded). For this reason, we observe superior rescue with indole instead of 3-methylindole in these cases.[16] Consistent with this logic, we also find that indole rescues W → G better than W → A, presumably because the latter implies a steric clash with the Ala Cβ.[16] Accordingly, this Rosetta application does not export exactly the deleted atoms as the constellation for virtual screening but rather excludes the “bridging” atoms that covalently connect the deleted atoms to the wild-type protein (Figure S6).

Generating a Library of Low-Energy 3D Conformers

A library of ∼3 million commercially available small molecules, downloaded from ZINC12[20] and subsequently curated for druglike properties, was used as a starting point to create a 3D library. For each of the 3 million chemical structures, up to 100 low-energy conformers were generated with OMEGA (OpenEye Scientific Software)[64] using the following command-line:

omega2 -in input_file.smi -out output_file.sdf.gz -prefix ligand_name -warts -maxconfs 100

Screening for Complementary Ligands

To match constellations to complementary candidate ligands, we used shape-based pharmacophore matching as implemented in ROCS (OpenEye Scientific Software).[28,29] ROCS optimizes overlap and evaluates each match on the basis of overall shape (ShapeTanimoto) and a pharmacophoric score (ColorTanimoto) that captures the spatial overlap of features such as hydrogen bond donors, hydrogen bond acceptors, and aromatic rings. In our study, each constellation/conformer pairing was aligned and evaluated using the following command-line:

rocs -dbase database.sdf.gz -query query.pdb -prefix struct_name -oformat sdf -cutoff 1.0 -scdbase true -maxhits 4000 -maxconfs 4000 -outputquery false

Final Refinement of Candidate Models

The refinement and final reranking of the resulting protein/ligand complexes were carried out using the Rosetta software suite.[19] As noted earlier, ROCS provides an alignment of a given conformer onto the template constellation. Since the constellation was extracted from the protein coordinates, this alignment is used to directly build a model of this ligand in complex with the mutant protein. The model was then subjected to energy minimization in Rosetta, using the following command line:

minimize.linuxgccrelease -s input_pdb −relax:fast -in:file:fullatom

After minimization, model complexes were filtered first based on similarity of the minimized complex to the WT structure (requiring backbone RMSD relative to WT ≤ 0.60 Å and ligand RMSD compared to constellation atoms ≤1.38 Å), and then on structural factors likely required for potent binding (at most one buried unsatisfied polar group, less than 0.13 Å2 solvent-accessible surface area of the ligand in the complex); the cutoffs for these filters were determined based on earlier experiments aiming to tune what fraction of designs were permitted to advance from each stage. To facilitate inactivation/reactivation in the mutant/rescued complex, we also required that our model of the apo mutant structure be worse than the WT structure by at least 5 Rosetta energy units (REUs), and that our model of the holo (rescued) mutant be within 5 REUs of the WT structure.[14]

After applying each of these filters, the remaining designs (cavity-forming mutations and their cognate ligands) were prioritized for experimental characterization on the basis of their protein–ligand interaction energies (as calculated using Rosetta).

Plasmid Construction

Because each of the two immunoglobulin domains in an scFv includes a disulfide bond, and we sought to use an E. coli system to express soluble scFv with the correct redox state, we elected to add a periplasmic signal sequence at the N-terminus of the scFv. This extra 21-residue signal sequence is auto-cleaved during the secretion of the polypeptide into the periplasm.[65,66] In addition, to facilitate purification and to improve expression of soluble fractions, we used a vector that includes a hexa-histidine tag (His-tag) and a maltose binding protein tag (MBP-tag).

The vector DNA selected for cloning was the pET PPL His6MBP LIC cloning vector (2K-T), obtained from QB3-Berkeley Core Research Facilities (Addgene plasmid 37183). In addition to the features described above, this vector also contains a cleavage site for tobacco etch virus (TEV) protease between the three tags and our gene of interest: thus, upon cleavage with TEV protease the desired scFv is produced with no remaining fusion tags. This vector was originally designed for Ligase Independent Cloning (LIC), but due to the emerging popularity and low failure rates of Gibson Assembly,[67] it has been adapted for cloning with the latter approach.

The empty vector was transformed into an NEB Turbo Competent E. coli (High Efficiency), plated in a LB agar plate with 100 μg/mL ampicillin, incubated overnight at 37 °C. A single colony from the plate was picked and grown in 50 mL of LB broth with 100 μg/mL carbenicillin, incubated with shaking at 250 rpm overnight at 37 °C. Plasmid DNA from liquid culture was isolated according to the protocol provided by the manufacturer (QIAprep Spin Miniprep Kit, QIAGEN). Next, the empty vector DNA was linearized using SspI-HF (NEB); the reaction was performed according to the protocol provided by the manufacturer, and the linearized vector was gel purified from a 0.8% agarose gel using a QIAquick gel extraction kit (QIAGEN).

As for insert DNA, WT and mutant genes were synthesized in a double-stranded linear DNA fragment format (GeneArt Strings DNA Fragments, ThermoFisher). The WT amino acid sequence was obtained from a previous published study.[32] These genes were prepared for Gibson Assembly by adding the following sequences to the 5′-end of the sense strand and antisense strand, respectively:

Sense: 5′- TACTTCCAATCCAATGCA-3′

Antisense: 5′- TAATAACATTGGAAGTGGATAA-3′

DNA strings were supplied in lyophilized vials and were reconstituted with molecular biology grade water.

Gibson assembly of the linearized 2K-T empty vector and prepared inserts was carried out by mixing 2 μL of vector (65 ng/μL) with 8 μL of insert (20 ng/μL), and then, 10 μL of Gibson Assembly Master Mix (NEB) was added to the insert/vector mixture. The Gibson assembly reaction was allowed to occur at 50 °C for 15 min, and then cooled down to 4 °C. Reaction mixture was used to transform into a NEB Turbo Competent E. coli (High Efficiency) and plated using the same protocol mentioned above. Three colonies from each transformation were picked, and plasmid DNA was prepared with the same protocol mentioned above; an aliquot of the plasmid DNA was sent out for successful cloning verification, using Sanger sequencing with T7 and T7-term primers.

Protein Expression

scFv genes in 2K-T plasmids were transformed into E. coli BL21 (DE3) competent cells (NEB) using the protocol suggested by the manufacturer and plated accordingly. A single colony was obtained from the plate and inoculated into a starter culture, 200 mL of TB broth, 0.4% glycerol, and 100 μg/mL carbenicillin and incubated with shaking at 250 rpm, overnight at 37 °C. Starter culture was inoculated into the main culture (using the same media as starter culture), with the ratio of 1:20 dilution. Main culture was grown with shaking at 200 rpm, at 37 °C, until O.D. reached 0.8 (measured at 600 nM); the temperature was then lowered to 15 °C, and isopropyl β-d-1-thiogalactopyranoside (IPTG) was added to a final concentration of 1 mM. Then, the culture was left shaking for 48 h. Cells were harvested by centrifugation at 3500g for 20 min. Used media was discarded, and the cell pellet’s surface was rinsed with sterile ddiH2O (past this point, samples were kept at 4 °C at all times). Cells were then resuspended in PBS buffer (6 mM phosphate buffer, 150 mM NaCl, pH 7.4) and passed through a French press for 3 rounds. Cell lysate was spun down at 15 000g for 30 min, and the supernatant (which contains soluble scFv) was filtered through a 0.22 μm PES syringe filter.

Western Blots

Induced cell culture (1 mL) was spun down at 2400g for 10 min. Used media was discarded, and the cell pellet was resuspended in 500 μL of 1× PBS; then, 500 μL of BPER-II (Thermo Fisher) detergent was added. The sample was agitated for 10 min at room temperature, and then spun down at 17 000g for 10 min. The supernatant was collected, mixed with 4X-LDS loading dye, and loaded onto a Bolt 4-12 Bis-Tris Plus gel (Invitrogen). After electrophoresis, gel was blotted onto a PVDF membrane using the iBlot 2 kit (Invitrogen).

Anti-His Western blots used 6X-His tag mouse monoclonal primary antibody (Rockland), with a 1:5000 dilution into the binding buffer of the Basic ONE-HOUR Western kit for mouse primary antibody (GenScript). The kit includes an HRP-secondary antibody conjugate, and the Western blot protocol followed instructions provided by the manufacturer. Next, 15 mL of ChromoSensor One-Solution TMB substrate (GenScript) was added to the blot after incubation with shaking (protected from light), and then, the reaction was stopped by rinsing the membrane in water. The membrane was allowed to dry and was then imaged.

For anti-kappa Western blots, the primary detection molecule was instead a protein L-HRP conjugate (GenScript). The protein L in this conjugate recognizes the kappa-VL of the scFv, and the HRP obviates the need for a secondary antibody. Thus, the same protocol was used as for the anti-His Western blot, except that no secondary antibody was used.

Protein Purification

First, cell lysate was purified with Capto L resin (GE Healthcare) by flowing lysate through the column (flow rate 1 mL/min) using an ÄKTA pure FPLC system (GE Healthcare). The column was washed using PBS until the UV signal reached a baseline level. Column was eluted using 15 mM NaOH with eluted fractions directly collected into a tube containing 1 M sodium citrate, pH 8.0 (5% of fraction volume). Fractions containing scFv with higher than 70% purity were pooled, concentrated down to 50 mL, and dialyzed against 50 mM sodium phosphate, 500 mM NaCl, 5 mM imidazole, pH 8.0 for 3 rounds.

Second, 5 mL of TEV protease 0.1 mg/mL was added to the dialyzed sample in the dialysis bag. The dialysis bag was then moved into a TEV reaction buffer 50 mM sodium phosphate, 500 mM NaCl, 5 mM imidazole, 2 mM DTT (freshly prepared) pH 8.0 and left to dialyze for 4 h before the bag was moved into a fresh batch of the same buffer and left to dialyze overnight. Samples were then dialyzed against 50 mM sodium phosphate, 500 mM NaCl, 5 mM imidazole, pH 8.0 for 3 rounds to remove DTT.

Third, samples were then applied through a Dextrin Sepharose (GE Healthcare) column (flow rate 1 mL/min) using an ÄKTA pure FPLC system (GE Healthcare). Flow-through fractions were collected and were then passed through a Ni Sepharose HisTrap FF (GE Healthcare) to remove TEV protease. Again, the flow-through fractions were collected; at this stage, scFv’s were about 90% pure.

Finally, samples were concentrated down to 2.5 mL and filtered through a 0.22 μm PES syringe filter. Next, the filtered samples were purified using a Superdex 75 size exclusion column (GE Healthcare), using 50 mM sodium phosphate, 500 mM NaCl, 5 mM imidazole, pH 8.0 buffer. Purified samples were then dialyzed into PBS, aliquoted into smaller volumes, and then flash frozen with liquid nitrogen and stored at −80 °C for further use. Samples of WT 4D5Flu and Rip-3 were analyzed using electrospray ionization mass spectrometry (ESI-MS) (Proteomics and Metabolomics Facility, Wistar), which confirmed that both constructs each contain two disulfide bonds as expected (Figure S7).

Fluorescence Quenching

For the protein titration experiment, samples were prepared by titrating varying concentrations of protein (WT or Rip-3) into PBS containing a final concentration of 10 nM fluorescein and 1% DMSO, with or without 100 μM Stitch-3. For the Stitch-3 titration experiment, samples were prepared by titrating varying concentrations of Stitch-3 into PBS with a final concentration of 10 nM fluorescein, 1% DMSO, and 500 nM protein (WT or Rip-3). For the SAR experiment (Figure ), samples were prepared by adding Stitch-3 analogues (to a final concentration of 10 μM) into PBS with a final concentration of 10 nM fluorescein, 1% DMSO, and 500 nM Rip-3.

Fluorescence quenching experiments were carried out in a SpectraMax i3x Multi-Mode microplate reader (Molecular Devices), using Costar 96-well black round-bottom plates, 200 μL reaction volume, excitation wavelength of 475 nm, and emission wavelength of 525 nm, at a temperature of 27 °C. Blank measurements containing all sample components except fluorescein were subtracted to remove background. All samples were normalized to a control of fluorescein alone (0% quenching) and buffer alone (100% quenching).

MPTS quenching was monitored in the same manner as fluorescein quenching except for using MPTS instead of fluorescein, 400 nM Rip-3 anti-MPTS antibody, and using an excitation wavelength of 375 nm, and an emission wavelength of 435 nm.

Differential Scanning Fluorimetry

SYPRO Orange dye has a higher fluorescence intensity when bound to exposed hydrophobic functional groups of proteins, and thus, it provides a probe for proteins’ thermal unfolding.[68−70] Differential scanning fluorimetry experiments were carried out in a QuantStudio 6 Flex real-time PCR system (ThermoFisher), using a 384-well white round-bottom plate, 20–30 μL reaction volume, excitation wavelength of 580 nm, and emission detected using the ROX filter setting (623 nm). Samples were prepared in PBS with 5 μM protein, 5× SYPRO Orange (ThermoFisher), and 1.1% DMSO with or without 100 μM Stitch-3. The scFv melting curve was collected from 25 to 95 °C (0.5 °C/min), and data were analyzed with PRISM 6 (GraphPad) using the Gibbs–Helmholtz equation[71] to determine the Tm values. Samples were run in 8 replicates.

NMR Studies

NMR spectra were recorded at 20 °C on a Bruker Avance II 600 MHz NMR instrument equipped with a TCI triple-resonance cryogenic probe. The protein sample was dissolved in PBS buffer (6.1 mM phosphate buffer, 154 mM NaCl, pH 7.4), at a concentration of 100 μM.

NMR data were collected using the Bruker standard pulse sequences hsqct1etf3gpsi3d and hsqct2etf3gpsi3d[34] and then were analyzed using TopSpin 3.2 (Bruker). T1 and T2 relaxation times were determined both by fitting peak intensities of the biggest peak at ∼8.3 ppm and by integrating between 7 and 10 ppm (covering most amide proton peaks) as a function of 11 relaxation delay intervals (T1, 0.02, 0.05, 0.1, 0.2, 0.3, 0.4, 0.6, 0.8, 1.0, 1.2, and 1.5 s; and T2, 1, 2, 3, 4, 5, 6, 8, 10, 12, 15, 20 s)[35] based on the following:Igor Pro software (WaveMetrics) was used to calculate the T1 and T2 relaxation times.

Safety Statement

No unexpected or unusually high safety hazards were encountered in the course of this study.