Human antibody-based chemically induced dimerizers for cell therapeutic applications

Chemically induced dimerizers (CIDs) have emerged as one of the most powerful tools for artificially regulating signaling pathways in cells; however, currently available CID systems lack the properties desired for use in regulating cellular thera- pies. Here, we report the development of human antibody-based chemically induced dimerizers (AbCIDs) from known small- molecule–protein complexes by selecting for synthetic antibodies that recognize the chemical epitope created by the bound small molecule. We demonstrate this concept by generating three antibodies that are highly selective for the BCL-xL–ABT-737 complex compared to BCL-xL alone. We show the potential of AbCIDs for application in regulating human cell therapies by using them to induce CRISPRa-mediated gene expression and to regulate CAR T-cell activation. We believe that the AbCIDs gener- ated in this study will find application in regulating cell therapies and that the general method of AbCID development may lead to the creation of many new and orthogonal CIDs.
IDs are powerful tools for dose and temporal control of protein–protein interactions1–3. CIDs have been used in a myriad of applications, including development of artificial cellular circuits4, activation of split-enzyme activity5,6, and control of protein localization. Recently, there has been a growing inter- est in using CIDs to regulate the activity of cell therapies after they have been administered to a patient7,8. Of particular interest has been the utilization of CIDs as safety switches for chimeric anti- gen receptor T-cell (CAR T-cell) therapies, as several patient deaths have occurred in CAR T-cell clinical trials9. Although a number of homo- and hetero-CIDs have been developed, they generally lack the properties required for use in human cell therapies1,3,10–16. For example, the classical FKBP–FRB CID system utilizes the small molecule rapamycin, which is both toxic and immunosuppressant2. Orthogonal ‘rapalogs’ show reduced toxicity, but have undesirable pharmacokinetic (PK) properties2.

Several plant-based CID systems have been developed, but the nonhuman nature of these proteins makes them prone to immunogenicity issues when incorporated into a cell therapy17. For CIDs to reach their full potential for use in cell therapies, it is critical that new human-protein-based CIDs be developed that utilize small molecules with drug-like properties. Ideally, the small molecules should have favorable PK properties and be bioorthogonal or well-tolerated. Additionally, new CIDs should exhibit dose dependence and be easily incorporated into different cellular signaling pathways. To date, the vast majority of CID sys- tems have been based on naturally occurring CIDs, and the ability to engineer and optimize both the small molecule and binding part- ners has been limited. Though chemically linking two pharmacoph- ores together has been employed to rationally design heteromeric CIDs not found in nature, the resulting small molecules almost universally lack drug-like properties. For these reasons, a general method to design novel CIDs with desirable properties for use in regulating human cell therapies would be of great utility.Here, we demonstrate a strategy to generate chemical-epitope- selective antibodies that has the potential to turn many known small-molecule–protein complexes into AbCIDs (Fig. 1a). We demonstrate this approach by engineering AbCIDs using the BCL- xL–ABT-737 complex. Furthermore, we show that AbCIDs can be used to regulate cellular processes; including CRISPR activation(CRISPRa)-mediated gene expression and CAR T-cell activity. We believe there is broad applicability for this approach, as it enables rapid generation of CIDs from human protein–small-molecule complexes using proteins and small molecules that meet the criteria for application in regulating human cell therapies.

RESULTS
We reasoned that the ideal complexes to generate selective anti- bodies against would be those in which a large portion of the small molecule remains solvent exposed when bound. Nature has employed a similar principle in the rapamycin–FKBP12–FRB CID system, in which rapamycin first binds FKBP12, generating a new binding surface, which is then recognized by FRB. Several other natural products use a similar approach for artificial protein recruitment2. Additional design principles include that the target protein should be a small monomeric domain and that the small molecule inducer be commercially available with desirable phar- macokinetic properties and low toxicity, making it amenable for animal model applications.After a survey of small-molecule-bound structures in the Protein Data Bank ((PDB; http://www.rcsb.org/pdb/home/home. do), we turned our attention to the human BCL-xL–ABT-737 complex (PDB: 2YXJ)18. BCL-xL is a member of the antiapoptotic BCL-2 family of proteins19. This small monomeric protein (~26 kDa) is located on the outer membrane of the mitochondria, where it sequesters proapoptotic members of the BCL-2 family. Because of its antiapoptotic role, a number of animal and clinically active small-molecule inhibitors have been developed against BCL-xL for the treatment of cancers20. The crystal structure of our candidate ligand, ABT-737 (1)21, bound to BCL-xL shows that a large portion of ABT-737 is solvent exposed (308 Å2), providing a potential chem- ical epitope for antibody binding. In comparison, an analysis of 866 small-molecule-bound structures in the PDB (Supplementary Results, Supplementary Fig. 1) revealed a mean solvent-exposed surface area of 125 Å2, with rapamycin bound to FKBP12 being an outlier at 528 Å2 (PDB: 1FKB)22. Thus, we felt that the BCL-xL– ABT-737 complex would be a good first target for the development of an AbCID.

To identify unique chemical-epitope-selective antibodies, we used a C-terminally truncated form of BCL-xL (residues 2–215) lacking the mitochondrial transmembrane domain. Biotinylated BCL-xL was immobilized on streptavidin resin and used for phage selections with a previously developed synthetic antibody-fragment library and selection stategy23. During each round of selection, the phage library was first subjected to stringent counterselection against BCL-xL in the absence of small molecule, thereby removing fragment antigen- binding (Fab) phage that were not selective for the ABT-737-bound form. Positive selections were then performed in the presence of saturating amounts of ABT-737 (1 M), ensuring that the majority of BCL-xL was bound to ABT-737 (Fig. 1b). A total of four rounds of selection were performed. Encouragingly, we observed substan- tial enrichment of phage titers for selections against BCL-xL in the presence of ABT-737 (Supplementary Fig. 2). After round four, individual Fab-phage clones were isolated and sequenced. Ten Fab phage with unique sequences in the complementarity-determining regions (CDRs) were identified (Supplementary Table 1).The unique Fabs were subcloned into a bacterial expression vector, expressed, and purified23. Gratifyingly, enzyme-linked immuno- sorbent assays (ELISA) with BCL-xL in the presence or absence of ABT-737 showed that all ten Fabs had enhanced binding to BCL-xL in the presence of the drug. Several Fabs showed excellent potency and extremely strong selectivity for binding in the presence of ABT-737 (Supplementary Fig. 3). To further profile the three best Fabs, we characterized the kinetics of BCL-xL binding in the pres- ence or absence of ABT-737 by biolayer interferometry (Fig. 1c and Supplementary Fig. 4)24. All three of the Fabs (AZ1, AZ2, and AZ3) were very potent binders of BCL-xL in the presence of ABT-737 (KD < 10 nM) and showed almost no detectable binding in the absence of ABT-737 at concentrations up to 5,000 nM of Fab (Supplementary Table 2). Our most selective Fab (AZ2) showed >2,000-fold selectiv- ity for the ABT-737-bound form of BCL-xL compared to the apo form. Formation of the AbCID ternary complex was reversible, either through washout of the Fab (Fig. 1c and Supplementary Fig. 4) or the small molecule (Supplementary Fig. 5).

We hypothesized that the exquisite selectivity of our Fabs was the result of direct interactions of the Fab CDRs with parts of ABT- 737. We reasoned that if this were the case, the Fab would bind less potently to other BCL-xL–ligand complexes. ABT-263 (2) is an analog that binds to the same conformation of BCL-xL with potency similar to that of ABT-737 (r.m.s. deviation = 0.8; Fig. 2a and Supplementary Fig. 6a)25. To test our hypothesis, we measured the ability of AZ1 to discriminate between ABT-737, ABT-263, and the native-ligand-derived Bak-peptide26 bound BCL-xL (Fig. 2b). As predicted, we observed dramatically weaker binding of Fab to the BCL-xL–ABT-263 complex and no detectable binding of Fab to the Bak peptide complex. Although we do not have a crystal struc- ture of the AbCID complex, these data strongly suggest that AZ1 binds near, and possibly covers, the small-molecule binding site.Although AZ1 is able to discriminate between ABT-737 and its close analog ABT-263, suggesting that ABT-737 comprises a por- tion of the epitope recognized by AZ1, we hypothesized that AZ1 also makes contacts with BCL-xL. To test this, we measured the ability of AZ1 to bind to ABT-737-bound BCL-W and BCL-2. BCL-W and BCL-2 are both homologs of BCL-xL that are known to bind ABT-737 with potency similar to that of BCL-xL21. Importantly, BCL-xL, BCL-W, and BCL-2 all have similar folds (Supplementary Fig. 6b). In the presence of saturating concentrations of ABT-737, AZ1 showed reduced binding to BCL-W and almost no detectable binding to BCL-2, suggesting that the epitope recognized by AZ1 encompasses specific residues on the surface of BCL-xL in addi- tion to specific chemical epitopes on ABT-737 (Fig. 2c). This data supports the hypothesis that AZ1 makes direct contact with both the small-molecule and the protein portion of the BCL-xL–ABT- 737 complex.

In the naturally occurring rapamycin–FKBP12–FRB CID, it is known that rapamycin potently binds FKBP12 (KD in the subnano- molar range) and has only weak affinity for FRB on its own (KD in the micromolar range)2,10. However, FRB is able to potently bind the FKBP–rapamycin complex (KD in the low nanomolar range)2,10. As we generated our AZ1 AbCID by selecting for Fabs against the pre- viously known BCL-xL–ABT-737 complex18, we hypothesized that our assembled CID uses a mechanism similar to that of rapamycin –FKBP12–FRB. To test this, we used differential scanning fluorim- etry to look for changes to the melting temperature (Tm) of AZ1 in the presence of ABT-737. As suspected, ABT-737 seemed to have no effect on the Tm of AZ1, suggesting that AZ1 does not bind ABT-737 on its own at the concentrations used in our dimerization assays (Supplementary Table 3). In comparison, BCL-xL, which is known to potently bind ABT-737, showed a ~10 °C increase in Tm in the presence of ABT-737. Together, this data support a mechanism in which ABT-737 first binds to BCL-xL, creating a new epitope that is then potently recognized by AZ1.To develop AbCIDs, we used design principles relevant to regu- lating cellular therapies. For that reason, when choosing cellular applications to demonstrate AbCIDs, we focused on cellular models of two main cell therapy modalities: regulation of gene expression and activation of immune cells. We incorporated our AbCIDs into the important technologies of CRISPRa and CAR T-cell engineering to show the great promise of applying AbCIDs to next-generation cell therapies.

Current CID technologies are often used for controlling intracellular signaling pathways2,3. Because of the disulfide bond linking the heavy and light chains of Fabs and the reducing envi- ronment inside the cell, it is generally believed that intracellular expression of Fabs in mammalian cells would lead to an inactive species. Recently, we reported a single-chain Fab (scFab) con- struct in which the light and heavy chains are genetically fused as a single polypeptide27. The scFab scaffold has a very high melting temperature (Tm = ~81 °C), so that once formed, it is very stable27. We hypothesized that conversion of our ABT-737-inducible Fabs into a scFab format may allow them to be used in living cells. Indeed, transfection of the AZ1 gene in scFab format (scAZ1) into HEK293T cells resulted in robust expression as measured by immunoblotting (Supplementary Fig. 7). To test whether scAZ1 was active in living cells, we constructed a genetic circuit in which scAZ1 is fused to the VPR transcriptional activation domain28 and BCL-xL is fused to dCas9 to generate a two-part CRISPRa construct (ref. 29; Fig. 3a). Both constructs contain a nuclear localization sequence, which reduces the possibility of interaction with endogenous BCL-xL while simultaneously priming the sys- tem for activation in the nucleus. The dCas9–BCL-xL fusion can be targeted by addition of a specific single guide RNA (sgRNA) to a promoter that drives a luciferase reporter. If the AbCID is functional in cells, addition of ABT-737 should lead to localiza- tion of AZ1-VPR to the luciferase reporter, promoting expression of luciferase, which can be readily detected. For comparison, we generated an identical two-part CRISPRa circuit except with a conventional CID based on the rapamycin–FKBP12–FRB sys- tem10, as recently reported30. Indeed, addition of ABT-737 to our engineered cells resulted in robust expression of luciferase, indi- cating that AZ1 and BCL-xL function as an ABT-737-inducible AbCID in living cells (Fig. 3b). The level of activation observed using the AbCID was comparable to that observed with the con- ventional CID. The induction of luciferase expression was dose dependent, with an EC50 of 8.7  1.1 nM (Fig. 3c). This value was consistent with the EC50 measured by in vitro characterization of the AZ1–ABT-737–BCL-xL complex using biolayer interferome- try, suggesting little barrier to cellular activation (Supplementary Fig. 8). Importantly, addition of ABT-737 to an AbCID-gated sys- tem with a negative sgRNA resulted in no increase in luciferase expression (Supplementary Fig. 9). Together, these results support that our AbCID can be used for tunable control of bio- logical systems in living cells.

The use of engineered T-cells for the treatment of malignancies has recently become an important paradigm in cancer therapeutics ovation domain. The scFv is specific for a tumor antigen and results in recruitment of the T-cell to the tumor and antigen-dependent activation of the T-cell. This technique has shown great responses in treating leukemia by targeting the CD19 antigen. However, hyperac- tivation of CAR T-cells has resulted in off-target cytotoxic effects, and in some cases death, limiting the utility of this promising modality31. For this reason, there has been great interest in developing remote control over the activity of these cells to tune the level of activation or shut off T-cell signaling should untoward toxicity develop32–35.Previous reports of small-molecule-activated CAR T-cells used intracellular splitting of the CAR activation domains32. An additional approach for controlling CAR T-cell activation uses uni- versal protein-based adaptor domains that confer antigen recog- nition and promote activation of the CAR T-cell when added33–35. We hypothesized that by taking advantage of the unique antibody nature of our AbCID system we could generate a hybrid of these two paradigms that has the universal nature of an adaptor strategy but the temporal control of a small-molecule-inducible system. To test this, we engineered Jurkat T-cells to express a CAR in which the scFv portion of the CAR is replaced by BCL-xL (Fig. 4a,b). This creates a T-cell that contains the machinery required for activation but no longer binds to the antigen-presenting cells. In parallel, we gener- ated a bispecific antibody by linking a clinically used CD19 scFv36 to Fab AZ1. We expected that upon addition of ABT-737, the bispe- cific antibody would be recruited to the CAR T-cell and mediate engagement with the CD19+ cell. Such a design permits both induc- ible and antigen-dependent CAR T-cell activation. To facilitate rapid quantitation of T-cell activation, we used a Jurkat T-cell line that had been engineered to express GFP upon activation of the nuclear fac- tor of activated T-cells (NFAT) pathway37. In the presence of CD19+ K562 cells and our bispecific antibody (AZ1-CD19), addition of ABT-737 resulted in dose-dependent activation of the CAR T-cells as measured by expression of GFP (Fig. 4c). Activation of the T-cells was further confirmed by expression of the canonical T-cell acti- vation markers CD69 and secreted interleukin-2 (Supplementary Fig. 10).

T-cell activation. (a) Schematic of AbCID-regulated CAR T-cell activation in which the CAR contains an extracellular BCL-xL domain in place of the typical scFv. Addition of an AZ1-CD19 bispecific antibody and various concentrations of ABT-737 results in recruitment to CD19+ cancer cells and tunable activation of the CAR T-cells. (b) Linear diagrams of the gene constructs used to produce the CARs and schematics of corresponding antibodies for this study. (c) Quantification of NFAT-dependent GFP reporter expression 20 h after initiation of co-culture with either CD19+ or CD19− K562 target cells and addition of antibody (5 nM) and varying concentrations of small molecule. Addition of ABT-737 in the presence of CD19+ K562 cells and bispecific antibody resulted in dose-dependent activation of the NFAT pathway, but no activation was observed in the absence of ABT-737 or with co-culture of CD19− K562 cells. Each data point represents the mean of three independent experiments  s.d.contain the CD19 scFv was used (Fig. 4c). In addition, ABT- 737 was not able to induce T-cell activation on its own. Though our T-cell system showed a ~65% activation level compared to the conventional CAR control, the reduced activity may actually be of benefit because of the hyperactivation and toxicity observed with conventional CARs. These data demonstrate that AbCIDs can be used for extracellular regulation of cellular signaling pathways and represent a novel paradigm for small-molecule control of CAR T-cell activation.ABT-737 is a soluble, cell-permeable, bioavailable, potent, and com- mercially available compound, making it an excellent molecule for use in a CID in cells and potentially in animals. However, it is known that ABT-737 induces apoptosis in some cells types, particu- larly hematopoietic cells that have high expression levels of BCL-2 family members21. We thus tested the concentration of ABT-737 necessary to induce apoptosis in Jurkat, K562, and HEK293T cells.

Importantly, the concentration ranges used to induce AbCID CAR (<100 nM) and CRISPRa (<270 nM) activity were below the con- centrations at which cell death was observed (Jurkat IC50 ~2 M; K562 IC50 >10 M; and HEK293T IC50 ~10 M) (Supplementary Fig. 11). ABT-737 has been used extensively in mouse cancer mod- els and is generally well tolerated by mice except for causing plate- let toxicity21. However, the concentrations used to activate AbCID CARs in our cellular assays (<100 nM) are far below the concen- tration observed to be toxic to platelets (typically at low micro- molar concentrations)40. Additionally, others have also shown that ABT-737 can be applied to activate engineered proteins in live-cell experiments with little observed cytotoxicity41. Collectively these data support the feasibility of using ABT-737 activated AbCIDs in cellular and animal applications with minimal effect on the viability of these model organisms. Moreover, while ABT-737’s lack of com- plete bioorthogonality may be a caveat for some research applica- tions, it may actually be of benefit from a therapeutic perspective if the AbCID CAR approach described here were to be applied to the treatment of ABT-737-sensitive B-cell malignancies.

DISCUSSION
Here we describe a novel method to rapidly generate chemically induced dimerizers using known small-molecule–protein com- plexes and synthetic antibody libraries. We demonstrated this method by generating AbCIDs from the BCL-xL–ABT-737 com- plex. Additionally, we showed that these AbCIDs can be applied to regulate a diverse range of biological processes in living cells, includ- ing CRISPRa-mediated gene expression and CAR T-cell activation. Finally, we showed that the range of ABT-737 concentrations used to activate AbCIDs was far below the concentration at which toxic- ity was observed in cells.
One of the inspirations for developing AbCIDs came from previ- ous work showing that it is possible to use phage display to gener- ate antibodies that can specifically bind to protein conformations that are ‘trapped’ by binding of small molecules42–45. In these cases, the antibody shows an increased affinity for the small-molecule- bound form of the protein, similarly to a CID. However, the anti- body is often able to bind the protein in the trapped conformation independent of a small molecule. For this reason, the selectivity of conformation-selective antibodies for the bound form over the apo form is limited, reducing their utility as CIDs. With the develop- ment of AbCIDs, we specifically focused selections to generate anti- bodies that target a small-molecule–protein complex, and use both the small molecule and proteins as composite parts of the binding epitope. This provided higher selectivity for the bound form of the protein and, in turn, for the desired properties for use as CIDs. This solution is reminiscent of several naturally occurring CID systems, including the rapamycin–FKBP12–FRB system, in which binding of rapamycin to FKBP12 creates a novel binding surface necessary for recognition by FRB.

Although rapamycin has favorable PK properties in humans, its toxicity and immunosuppressant properties make it incompat- ible with regulating CAR T-cell therapy. In cellular CID assays, researchers typically use rapamycin in a concentration range of 10–100 nM despite rapamycin toxicity being observed in cell lines at concentrations of 100–300 nM, a difference of only three to ten-fold12,46,47. In comparison, the EC50 for activation of our AbCID CAR with ABT-737 is ~6 nM and the IC50 for cell kill- ing is ~2 M—a >330-fold difference. Though the commonly used rapamycin analog AP21967 lacks the toxicity and immuno- suppressive properties of rapamycin, its short half-life in mouse plasma (<4 h) greatly reduces its use for regulating cell therapies in vivo. In fact, previous studies demonstrating small-molecule activation of CAR T-cells in mice using the FKBP–FRB system have been limited because of the PK liabilities of AP21967 (ref. 32). Fortunately, ABT-737 has been shown to have a half-life in mouse plasma of 14–18 h, which should greatly facilitate the use of our AbCIDs to activate CAR T-cells in mouse models of cancer48.To our knowledge, AbCIDs represent the first demonstra- tion of a general strategy to engineer CIDs from existing small- molecule-protein interaction pairs. Though in this study we have used synthetic antibody fragment libraries, we envision that diver- sity libraries built upon alternative binding scaffolds will be able to be applied to this technique, including but not limited to, DARPins, FNIII, ubiquitin, knottins, and nucleic acid aptamer libraries49,50. We foresee that much of the power of our strategy will come from the ability to rapidly generate new AbCIDs from different small-molecule–protein pairs in which a substantial portion of the small molecule is solvent accessible. We believe that AbCIDs rep- resent a novel and promising approach to develop next-generation CID tools with the properties necessary for application in human cell ABT-737 therapies.