Research ArticleCancer

Exploiting an Asp-Glu “switch” in glycogen synthase kinase 3 to design paralog-selective inhibitors for use in acute myeloid leukemia

See allHide authors and affiliations

Science Translational Medicine  07 Mar 2018:
Vol. 10, Issue 431, eaam8460
DOI: 10.1126/scitranslmed.aam8460

The right GSK for the job

The enzyme glycogen synthase kinase 3 (GSK3) has been proposed as a potential therapeutic target in many diseases, but none of the drugs targeting this enzyme so far have translated to the clinic. A major reason for this failure to translate is the existence of two closely related paralogs of GSK3 such that most drugs target both forms at once and cause unacceptable toxicity. By applying a rational structure-based approach, Wagner et al. were able to design selective inhibitors for the α and β isoforms of this enzyme and then show that selective GSK3α inhibitors specifically show promising activity against acute myeloid leukemia with no detectable effects on healthy hematopoietic cells.


Glycogen synthase kinase 3 (GSK3), a key regulatory kinase in the wingless-type MMTV integration site family (WNT) pathway, is a therapeutic target of interest in many diseases. Although dual GSK3α/β inhibitors have entered clinical trials, none has successfully translated to clinical application. Mechanism-based toxicities, driven in part by the inhibition of both GSK3 paralogs and subsequent β-catenin stabilization, are a concern in the translation of this target class because mutations and overexpression of β-catenin are associated with many cancers. Knockdown of GSK3α or GSK3β individually does not increase β-catenin and offers a conceptual resolution to targeting GSK3: paralog-selective inhibition. However, inadequate chemical tools exist. The design of selective adenosine triphosphate (ATP)–competitive inhibitors poses a drug discovery challenge due to the high homology (95% identity and 100% similarity) in this binding domain. Taking advantage of an Asp133→Glu196 “switch” in their kinase hinge, we present a rational design strategy toward the discovery of paralog-selective GSK3 inhibitors. These GSK3α- and GSK3β-selective inhibitors provide insights into GSK3 targeting in acute myeloid leukemia (AML), where GSK3α was identified as a therapeutic target using genetic approaches. The GSK3α-selective compound BRD0705 inhibits kinase function and does not stabilize β-catenin, mitigating potential neoplastic concerns. BRD0705 induces myeloid differentiation and impairs colony formation in AML cells, with no apparent effect on normal hematopoietic cells. Moreover, BRD0705 impairs leukemia initiation and prolongs survival in AML mouse models. These studies demonstrate feasibility of paralog-selective GSK3α inhibition, offering a promising therapeutic approach in AML.


The development of selective, targeted small molecules to understand the mechanisms of human diseases and improve their treatment remains a challenge. This challenge is pronounced when selective inhibitors of highly similar and ubiquitously expressed proteins are desirable. One exemplary target class, relevant to numerous biological pathways and human diseases, is glycogen synthase kinase 3 (GSK3).

Multiple eukaryotic kinases have undergone gene duplication events resulting in two or more paralog genes. Duplicate genes are thought to diverge in either their regulation or their biochemical functions (1). GSK3 represents an example of such an evolutionary selection system, resulting in two paralog genes (1): GSK3A on chromosome 19 and GSK3B on chromosome 3. The encoded paralog proteins, GSK3α and GSK3β, have a high degree of similarity [67% overall amino acid identity and 95% identity in the adenosine triphosphate (ATP)–binding site] (Fig. 1A) (2). Although some functional overlap is clear, the GSK3α and GSK3β paralogs also have distinct, albeit uncharacterized, functions (3, 4). Both GSK3 paralogs are ubiquitously expressed and implicated in the pathophysiology of a number of human disorders (5, 6), including non–insulin-dependent diabetes mellitus (7), cardiac hypertrophy (8, 9), neurological and neurodevelopmental disorders such as bipolar disorder and Alzheimer’s disease (10), and cancer (1116). The precise mechanistic role of each paralog in disease pathogenesis, however, has been difficult to determine given the lack of paralog-selective chemical probes.

Fig. 1 A single amino acid difference in the ATP-binding domain of GSK3α and GSK3β results in structural and topological differences.

(A) Aligned primary amino acid sequences of GSK3α and GSK3β. Residue differences between the two paralogs are in red. The main regions of the kinases are color-coded: hinge (yellow), backend region (green), P-loop (pink), DFG motif (blue), and activation loop (orange). Key amino acids for autoregulation, S21/9 and Y279/216, are underlined. Amino acids within 6 Å of the ligand are marked by an asterisk. (B) X-ray crystal structure (2.4 Å resolution) of hGSK3β bound to BRD0209, a dual GSK3α/β inhibitor. The inset shows a hydrogen bond network centered on the Asp133 side chain, on the backend region of the hinge connecting the N- and C-lobes of GSK3. The main regions of the kinases are again color-coded. (C) A comparison between the hinge backend and ATP-binding site of apo GSK3β and an MD simulation of apo GSK3α highlight the size difference between their hydrophobic pockets.

A large number of dual GSK3α/β inhibitors from varied chemotypes have been reported: the natural product–based indirubins, paullones, bisarylmaleimides, indolocarbazoles, synthetically derived aminopyrimidines, ruthenium complexes, thiadiazolidinones, and imino-thiadiazoles (1720). Few of these inhibitors have good selectivity against the human kinome, a property paramount for the unambiguous interpretation of in vitro and in vivo effects as solely driven by GSK3 inhibition (21). Furthermore, paralog-selective inhibitors of GSK3α or GSK3β with demonstrated functional selectivity in vitro and activity in vivo are nonexistent. Oxadiazole-based selective GSK3α inhibitors have been described, but their functional selectivity or activity in mammalian models in vivo has not been characterized (2224). In addition, although a small molecule preferentially inhibiting GSK3β has been recently reported, its degree of paralog selectivity and off-target inhibition against other kinases complicates the use of these molecules to study the biology of GSK3 paralogs (25).

One concern in the clinical application of dual GSK3α/β inhibitors is the predicted stabilization and nuclear translocation of β-catenin (1921), a direct GSK3 phosphorylation substrate in the APC/AXIN/GSK3 complex. This is likely a problem in the translation of dual GSK3α/β inhibitors for the treatment of acute myeloid leukemia (AML), a poorly treated form of cancer in need of new therapies. Numerous studies support the targeting of GSK3 in AML (2630); however, the well-established oncogenic role of β-catenin activation in the context of myeloid leukemia results in a clear biological challenge to targeting GSK3 in this disease (3133). We recently suggested a more parsimonious approach to targeting GSK3 in AML. We found that selective genetic suppression of the GSK3α paralog promotes AML differentiation and impairs leukemia progression in mouse models of AML without increasing β-catenin (34). In addition, Doble et al. (22) reported that selective genetic suppression of either GSK3 paralog in embryonic stem cells and whole-body knockout (KO) of GSK3α in mice do not increase β-catenin. A gene dosage effect resulting in β-catenin pathway activation was observed only in cells lacking three or all four of the GSK3 alleles. In an adult hematopoietic stem cell (HSC) context, GSK3β paralog KO or dual GSK3α/β KO both induced a preneoplastic disease and overt AML, whereas HSC GSK3α KO did not result in myeloid transformation (31). Therefore, paralog-selective inhibition of GSK3α should avoid this mechanism-based safety concern.

Here, we aimed to develop small-molecule, paralog-selective inhibitors of GSK3. In addition, we intended to determine whether the pharmacological inhibition of the GSK3α paralog would mimic its paralog-selective genetic suppression in the context of AML differentiation and leukemia progression. Starting from BRD0209, a highly kinome-selective GSK3 inhibitor (21), a set of inhibitors stemming from the same single chemotype, or isochemogenic inhibitors, with paralog selectivity, was rationally designed, synthesized, and characterized. We applied this set of selective GSK3α and GSK3β inhibitors to a series of cellular assays to refine our understanding of the individual paralogs in driving mechanism-based β-catenin toxicities. BRD0705, a selective GSK3α inhibitor, induced differentiation, reduced transcriptional programs of stemness, and impaired colony formation in AML cell lines and primary patient samples without affecting normal hematopoietic cell growth. BRD0705 did not induce β-catenin stabilization or nuclear translocation at concentrations efficacious in multiple mouse models of AML, resulting in leukemia initiation impairment and prolonged survival. Together, our findings suggest that paralog-selective small-molecule inhibition of GSK3α offers a safe and unexplored therapeutic strategy for the treatment of AML.


Structure-based design considerations identify an Asp→Glu “switch” driving topological differences between the GSK3 paralogs

GSK3α and GSK3β, which are highly similar in sequence and domain architecture, are highly conserved across species (fig. S1). The primary sequence of human GSK3α is 67% identical to human GSK3β (2), with the least homology in the N- and C-terminal domains (difference in amino acid sequence in red; Fig. 1A). Inspection of the ATP-binding domain (asterisk in Fig. 1A) reveals a 95% amino acid sequence identity (100% similarity), with a single amino acid difference in the hinge-binding region (amino acids in yellow): a glutamic acid [Glu(E)196] in GSK3α to an aspartic acid [Asp(D)133] in GSK3β.

Our group recently published the discovery of kinome-selective dual inhibitors of the GSK3 kinases, such as BRD0209 (Fig. 1B) (21). A high-resolution (2.4 Å) x-ray crystal structure of hGSK3B bound to BRD0209 was obtained [Protein Data Bank (PDB): 5KPK; detailed experimental conditions and crystallographic parameters in Supplementary Materials and Methods, tables S1 and S2, and fig. S2] (3538). Consistent with previous x-ray crystal structures of hGSK3β (53 distinct structures with more than 20 amino acids in PDB based on a UniProt search of P49841), the general structural motifs consist of (i) an N-lobe characterized by six β sheets and a flexible P-loop (pink in Fig. 1, A and B); (ii) a kinase hinge region (D/E-Y-V; yellow in Fig. 1, A and B); and (iii) a C-lobe, consisting primarily of α helices and featuring the activation loop (orange in Fig. 1, A and B) and the DFG motif (blue). Within the ATP-binding domain (Fig. 1B, inset), the crystal structure revealed a tridentate hydrogen bond interaction (pink dotted lines) established by the constrained tricyclic pyrazolo-tetrahydroquinolinone structure of BRD0209. This hydrogen bond donor-acceptor-donor motif serves not only as a critical binding motif but also as a geometric anchor that orients substituents from the pyrazole ring toward the kinase hydrophobic pocket (39) adjacent to the gatekeeper residue [Leu (L); purple in Fig. 1A). Notably, an H-bond donor interaction through the pyrazolo-NH- engages the backbone carbonyl of Asp133, the unique residue difference between GSK3β and GSKα (Glu196), within the ATP-binding domain (Fig. 1, B and C). We also discovered a hydrogen bond network on the backend region of the hinge connecting the N- and C-lobes centered on the Asp133 side chain (Fig. 1B). The Asp133 carboxylate forms one H-bond with Arg113 in the N-lobe and a second with Lys197 in the C-lobe. Inspection of all GSK3β structures in the PDB confirmed that this backend hydrogen network was conserved across structures independent of the bound ligand. We wondered whether this hydrogen bond network was unique to the Asp133 in GSK3β or would be conserved in GSK3α, where the corresponding residue, Glu196, contains one additional side-chain carbon atom.

Although numerous structures of hGSK3β are reported in the PDB, no x-ray crystal structures of GSK3α have been reported, preventing a direct structural comparison. To understand the impact of a glutamic acid substitution at this residue on the backend hydrogen bond network, we performed a Basic Local Alignment Search Tool (BLAST) search of GSK3α to identify a kinase with high similarity to GSK3α, including a hinge glutamic acid. Our search identified a fungal GSK3 ortholog found in Ustilago maydis (PDB: 4E7W), which has 49 and 58% overall identity to GSK3α and GSK3β, respectively (40). Analysis of the fungal structure revealed a different Glu126 side-chain conformation compared to the analogous Asp133 side chain in hGSK3β (fig. S3). The GSK3β Asp133 hydrogen bonds to the side chain of Arg131 within the N-lobe, and Arg131 in turn hydrogen bonds with the side chain of Glu80 within the N-lobe, forming a network of hydrogen bonds (Fig. 1C). However, the Glu126 in the fungal structure adopts a downward conformation and does not engage the corresponding Lys106, and there is no Lys106 to Asp73 hydrogen bonding interaction, indicating that a single carbon difference at this hinge residue may affect the overall architecture of this backend hydrogen bond network (fig. S3).

This observation led us to investigate whether similar structural differences between the Asp-Glu switch in the human GSK3 paralogs were present by running long time-scale molecular dynamics (MD) simulations of the GSK3α and GSK3β apo forms (see Supplementary Materials and Methods for experimental conditions of MD simulations) (4144). Analyses of the two sets of kinase simulations were consistent with our initial observations and revealed two different geometrically preferred paralog conformations of the hinge Asp133/Glu196 side chains (in GSK3β and GSK3α, respectively) (Fig. 1C). The side chain of Asp133 in GSK3β hydrogen bonds to the side chain of Arg113 within the N-lobe (Fig. 1C, panel 1, and figs. S4, distance A, and S5), which in turn stabilizes an Arg113/Glu80 side chain–to–side chain hydrogen bonding interaction (Fig. 1C, panel 1, and figs. S4, distance B, and S5) (45). In addition, Asp133 also forms a hydrogen bond interaction to Lys197 within the C-lobe, effectively bridging the two major structural components within the kinase, the N- and C-lobes. Conversely, the corresponding amino acid at the hinge of GSK3α, Glu196, does not strongly interact with Arg176 (Fig. 1C, panel 3, and fig. S4, distance A) due to the extra side-chain methylene, which in turn destabilizes the corresponding Arg176/Glu143 side chain–to–side chain hydrogen bond interconnections within the N-lobe (Fig. 1C, panel 3, and fig. S4, distance B). This difference in the stability of the network of hydrogen bonds on the backend of the ATP-binding pocket for GSK3α and GSK3β was quantified by measuring these distances across time as well as plotting the free energy contour plots (figs. S4 and S5; see Supplementary Materials and Methods for interaction measurements and energy plot protocols). The ability of the carboxylate side chain of Asp133 in GSK3β to hydrogen bond with Arg113 stabilizes the backend region, driving it into a single preferred conformation. In comparison, the carboxylate side chain of Glu196 in GSK3α has a fleeting interaction with the analogous Arg176, which in turn does not favor the Arg176-to-Glu143 interaction because all three residues sample distances outside typical hydrogen bond distances (fig. S5, C and D). This subtle, but quantifiable, conformational difference between the two kinase hinges triggers an indirect effect on the size and topology of the ATP-binding domain and the adjacent hydrophobic pocket (Fig. 1C, panels 2 and 4).

Recognizing that these topological differences are driven via the Asp→Glu switch within the hinge domain and that our pyrazolo-tetrahydroquinolinone–based hinge binders formed a direct H-bond to the analogous hinge position (Asp133 or Glu196 residue of GSK3β or GSK3α, respectively), we envisioned a rational approach to exploit this hypothesis and generate paralog-selective inhibitors. The tridentate binding mode of our core scaffold provided a rigid molecular platform within the ATP-binding domain well suited to explore apparent differences within the hydrophobic selectivity pockets. Therefore, we set out to systematically probe these differences by designing inhibitors predicted to be preferential binders for either GSK3α or GSK3β.

Paralog-selective GSK3 inhibitors were rationally designed by exploiting an Asp (D)–to–Glu (E) switch

To map the topography of the selectivity pocket adjacent to the hinge region, our initial structure-activity relationship (SAR) exploration probed substituents from the pyrazole ring at R1, which vectors into the hydrophobic selectivity pocket (see V in Fig. 2A). Structural modifications to the pyrazolo-tetrahydroquinolinone scaffolds are easily obtained via an efficient three-component, two-step reaction sequence (Fig. 2A). An electrophilic aromatic substitution between aminopyrazoles (I) and acetophenones (II) generates C-substituted pyrazolo intermediates such as III. Intermediate III reacts with dimedones (IV) to form vinylogous amides that cyclize in situ to provide the fully assembled tricycles, such as (V) [see Supplementary Materials and Methods for detailed synthetic protocols (fig. S6) and compounds’ characterization and spectra (fig. S7)].

Fig. 2 Paralog-selective inhibitors of GSK3α and GSK3β were designed and characterized.

(A) General synthetic scheme for synthesis of the pyrazolo-tetrahydroquinolinone scaffold. TFA, trifluoroacetic acid; TFE, 2,2,2-trifluoroethanol. (B) IC50 values for the inhibition of GSK3α and GSK3β were determined at Km of ATP in a motility shift microfluidic assay (Caliper) measuring phosphorylation of a synthetic substrate. Values are average of three or more experiments. Data are shown as IC50 values in μM ± SD. Compounds were tested in duplicate in a 12-point dose curve with threefold dilution starting at 33.3 μM. (C) Kinome-wide selectivity for BRD0705 and BRD3731 represented on a kinome phylogenetic tree. Each inhibitor was screened against 311 kinases at a 10 μM concentration. Kinases with >50% inhibition are depicted (percentage inhibition proportional to size of red dot).

Starting with minimal structural requirements at R1, we installed a hydrogen, leaving the selectivity pocket unoccupied (Fig. 2B). Compound 1 displayed fivefold selectivity for GSK3α versus GSK3β [IC50 (median inhibitory concentration), 42 and 225 nM, respectively; measured using a mobility shift microfluidics assay (Caliper; assays were run at Km,ATP of each enzyme; see Supplementary Materials and Methods for assay details and protocol)]. However, the chloro-substituted (~50% increase in atomic radius versus hydrogen) compound 2 displayed equipotency for two paralogs (IC50, 4 and 9 nM, respectively). A similar increase in potency for both paralogs was observed with electronically neutral but increasingly bulkier substituents: a methyl in compound 3 and a cyclopropyl in BRD0209, implicating hydrophobic rather than electronic effects. These modifications and increased binding affinities were tolerant of simple substitutions on the phenyl ring (see BRD0320). Increasing the ring size at R1 to cyclobutyl in compound 4, we observed our first hints of preferential binding (fourfold selectivity) for GSK3β (IC50, 25 nM) over GSK3α (IC50, 100 nM). To further exploit these differential interactions and steric requirements in this highly hydrophobic selectivity pocket, we synthesized the difluorocyclobutyl compound 5, which increased preferential binding toward GSK3β, exhibiting an eightfold selectivity window. Further structure-based design aimed toward maximizing steric requirement at the R1 position and the corresponding differences within the selectivity pockets of GSK3α and GSK3β led to the neopentyl-substituted compound BRD3731, which displayed 14-fold selectivity for GSK3β (GSK3β IC50, 15 nM; GSK3α IC50, 215 nM).

With the successful achievement of >10-fold selectivity for GSK3β using large substituents at R1, we turned our attention toward the development of more selective GSK3α inhibitors by revisiting the smallest substituent at R1, hydrogen. Recognizing subtle differences between the ATP-binding domains of GSK3α and GSK3β, in addition to the now validated differences in the hydrophobic selectivity pockets, we explored combinations between R1 = H and the quaternary center R2. Whereas branched alkyl (compound 6) and long-chain alkyl substituents at R2 resulted in losses in potency, an ethyl substitution (BRD0705) maintained potency and displayed increased selectivity for GSK3α (eightfold) versus GSK3β (GSK3α IC50, 66 nM; GSK3β IC50, 515 nM). The inverted quaternary center stereochemistry, the enantiomeric BRD5648, was relatively inactive against the GSK3s, demonstrating the limited three-dimensional space in the ATP-binding domain.

With the successful design of a chemically matched set of GSK3α (BRD0705)– and GSK3β (BRD3731)–selective inhibitors, as well as a corresponding negative control compound (BRD5648), we turned our attention toward the larger kinome as a test of specificity. Although minor chemical modifications provided a robust selectivity and SAR within the GSK3 kinases, our compounds were untested against the larger human kinome. Testing all four compounds against a panel of human kinases (311 kinases at 10 μM compound concentration) revealed exquisite overall kinase selectivity (Fig. 2C; see Supplementary Materials and Methods for assay details and protocol and tables S3 to S6). For BRD0705, the cyclin-dependent kinase (CDK) family of kinases (CDK2, CDK3, and CDK5) were next most potently inhibited at values of 6.87, 9.74, and 9.20 μM (87-, 123-, and 116-fold selectivity relative to GSK3α, respectively; table S5). This set of small-molecule inhibitors is a selective and fully characterized set of chemical tools to probe the biological function of the GSK3 kinases.

Site-directed mutagenesis and crystallographic studies reveal a key role for the hinge D-to-E switch

With the successful demonstration of GSK3α- and GSK3β-selective inhibitors and a working hypothesis based on computational and structural data for the observed paralog selectivity of BRD0705 and BRD3731, we set out to further validate our structural findings and computational predictions through additional crystallographic structure-based approaches. Because our efforts to obtain ligand-bound or apo structures of GSK3α were not successful, presumably due to its unorganized N terminus, we chose to pursue mutagenesis studies in GSK3β to test our Asp→Glu switch hypothesis. We surmised that if this one amino acid residue change (Asp133→Glu196) within the hinge—the net result being the addition of a single carbon atom in GSK3α—were the driver of the observed structure activity relationships, then mutating Asp133 to Glu133 in GSK3β would abrogate our observed selectivity patterns [see Fig. 3A for construct primary sequence, GSK3β (D133E)]. In addition, this design directly tests the importance of the hinge residue on the nature of the backend hydrogen bond network and the corresponding topological effects on the ATP-binding domain and adjacent hydrophobic selectivity pocket. Toward this end, a single-point GSK3β (D133E) mutant protein was designed, expressed, purified, and fully characterized (PDB: 5T31; detailed experimental conditions and crystallographic parameters in Supplementary Materials and Methods and tables S7 and S8). This GSK3β (D133E) mutant exhibited comparable kinase activity when compared to recombinant GSK3α [wild-type (WT)] or GSK3β (WT) with Kcat/Km [GSK3β (D133E)] of 2.3 × 104 M−1 s−1, Kcat/Km [GSK3α] of 1.6 × 104 M−1 s−1, and Kcat/Km [GSK3β] of 7 × 103 M−1 s−1. For comparison, we also obtained additional high-resolution x-ray GSK3β (WT) ligand-bound structures with our GSK3β-selective inhibitor BRD3731 (PDB: 5KPM; Fig. 3B and fig. S8, panels 1 to 3) and our GSK3α-selective inhibitor BRD0705 (PDB: 5KPL; Fig. 3B and fig. S8, panels 4 to 6).

Fig. 3 An Asp (D)–to–Glu (E) switch in the enzymatic hinge backend underlies paralog selectivity of GSK3α and GSK3β inhibitors.

(A) Primary sequence comparison of WT GSK3α, WT GSK3β, and GSK3β (D133E) and GSK3α (E196D) mutants. (B) X-ray structures of WT hGSK3β bound to BRD3731 (panels 1 to 3) and BRD0705 (panels 4 to 6) and hGSK3β (D133E) bound to BRD0705 (panels 7 to 9). The side chain of Asp133 in GSK3β forms a complex hydrogen bond network at the back of the kinase hinge and controls the shape and size of the adjacent selectivity pocket. (C and D) GSK3β (D133E) mutation “switches” compound selectivity. IC50 values for the inhibition of GSK3β (D133E) were determined at Km of ATP in a motility shift microfluidic assay (Caliper) measuring phosphorylation of a synthetic substrate. Representative IC50 curves show the comparison in enzymatic assays using GSK3α (WT, blue), GSK3β (WT, red), and GSK3β (D133E, black) protein constructs (C). Values are the average of three or more experiments. Data are shown as IC50 values in μM ± SD (D). Compounds were tested in duplicate in a 12-point dose curve with threefold dilution starting at 33.3 μM. (E) FLAG-tagged WT and E196D mutant GSK3α were overexpressed in HEK 293T cells. Western immunoblot for p-GSK3α (Tyr279) and total GSK3α after FLAG-GSK3α immunoprecipitation (IP) and treatment with dimethyl sulfoxide (DMSO) or BRD0705 at 10 μM.

As in previous ligand-bound GSK3β (WT) structures, for both GSK3 (WT) BRD3731- and BRD0705-bound crystal structures, we observed the hydrogen bond network on the backside of the hinge-binding domain centered on Asp133 (Fig. 3B, panels 1 and 4). In both GSK3β (WT) structures, Asp133 formed a hydrogen bond with Arg113 in the N-lobe and Lys197 in the C-lobe, as described above. However, in the corresponding BRD0705 ligand-bound structure with the GSK3β (D133E) construct, the increase of a single carbon unit in the GSK3β mutant protein (Glu133) disrupted the described H-bond interaction. This observation was predicted from our earlier comparison to the fungal GSK3β ortholog with a Glu at this position and from our MD simulation studies of GSK3α. In addition, this Asp133→Glu133 switch on the kinase hinge has indirect topological effects on the hinge-binding domain and adjacent selectivity pocket (Fig. 3B and fig. S8, panels 7 to 9). The volume of the hydrophobic selectivity pocket adjacent to the hinge, specifically pocket “c,” is much larger in the BRD0705-GSK3β (WT) structure than in the BRD0705-mutant GSK3β (D133E) structure (Fig. 3B, blue contour in panel 5 versus panel 8 and pocket c in panel 6 versus panel 9, and fig. S8). This observation is consistent with our observed SAR that the larger R1 neopentyl substituent in BRD3731 provides the most preferential binding for GSK3β. The smaller predicted pocket topology in regions “b” and “c” in GSK3α is less able to accommodate such large sterically demanding substituents. Conversely, smaller R1 substituents, such as a hydrogen in compound 1 and BRD0705, do not benefit from hydrophobic interactions in the hydrophobic selectivity pocket of GSK3β and showed preferred binding to GSK3α. In addition, pocket “a” in GSK3β (D133E) was slightly larger (Fig. 3B, orange contour in panel 6 versus panel 9) and tightly bound the R2 ethyl substituent of BRD0705, providing additional selectivity for GSK3α (overall eightfold selectivity).

As further validation of our Asp→Glu switch hypothesis and SAR, we tested our GSK3α-selective (BRD0705) and GSK3β-selective (BRD3731) inhibitors in the same enzymatic assay using the mutant GSK3β (D133E) construct. Both compounds displayed “GSK3α-like” activity against this GSK3β mutant construct (Fig. 3C). BRD0705 displayed similar potency toward the GSK3β (D133E) and GSK3α (WT) constructs, with an IC50 of 110 and 66 nM, respectively (Fig. 3D), whereas BRD3731 displayed reduced potency toward the GSK3β (D133E) with an IC50 of 53 nM compared to GSK3β (WT) (IC50, 15 nM).

Finally, we sought to validate in cellulo the suggested biochemical explanation for the GSK3α paralog selectivity of BRD0705 using GSK3α Tyr279 autophosphorylation as a surrogate for GSK3α enzymatic activity (46, 47). As predicted, overexpression of the E196D mutant form of GSK3α in human embryonic kidney (HEK) 293Tcells attenuated the repression of GSK3α active site phosphorylation after BRD0705 treatment, compared to overexpression of GSK3α (WT) in this same cellular context (Fig. 3E).

Through a variety of structural and functional approaches, we concluded that the observed inhibitory profiles for our GSK3α (BRD0705)– and GSK3β (BRD3731)–selective inhibitors are driven in large part through their direct ligand interactions with Asp133 in GSK3β and the corresponding Glu196 in GSK3α combined with difference in the overall topology of the ATP-binding domain and the adjacent hydrophobic selectivity pocket in each of these kinase paralogs. We next explored the biological impact of both GSK3α- and GSK3β-selective inhibition in a variety of cell-based systems and extended our investigation into the downstream effects of paralog inhibition on β-catenin stabilization.

GSK3 paralog-selective inhibitors demonstrate cell-based functional selectivity and cell context–dependent effects on β-catenin stabilization

With the goal of using this set of compounds in a cellular and in vivo context, it was crucial to demonstrate that a ~10-fold selectivity in a biochemical assay using recombinant enzymes would translate to sufficient selectivity in a human cellular context. Paralog-selective target engagement for GSK3α and GSK3β was quantitatively assessed using a bioluminescence resonance energy transfer (BRET) reporter system (NanoBRET) in live, nonpermeabilized HEK 293 cells to match physiological drug delivery conditions (48). N-terminal NanoLuc fusion constructs of the individual GSK3α and GSK3β paralogs were designed to serve as luminescent energy transfer donors and used a fluorescent nonselective kinase BRET tracer serving as an energy acceptor. Compound binding is evident from a loss of BRET signal owing to competitive displacement of the BRET tracer (detailed assay conditions in Supplementary Materials and Methods). In this functional cell-based assay, the 10-fold selectivity for each paralog-selective GSK3 inhibitor translated to a more than four- to sixfold selectivity in intact live HEK 293 cells. BRD0705 inhibits GSK3α with a Kd of 4.8 μM, and BRD3731 impairs GSK3β with a Kd of 3.3 μM in this cellular context (Fig. 4A and sample curves in fig. S9).

Fig. 4 GSK3α- and GSK3β-selective inhibitors play a differential role on β-catenin stabilization in a context-dependent manner.

(A) Biophysical measurement of GSK3 target engagement in HEK 293 cells by BRET signal between a NanoLuc-fused protein target and a small molecule labeled with the NanoBRET acceptor dye. Kd values for each inhibitor are reported as means ± SEM of two replicates. (B) β-Catenin immunofluorescence staining in HEK 293T after treatment with the indicated inhibitor. β-Catenin is shown in red, and 4′,6-diamidino-2-phenylindole is shown in blue. (C) β-Catenin TCF/LEF luciferase reporter assay in HEK 293T after treatment with the indicated inhibitor. P < 0.05 calculated using a Mann-Whitney test in comparison with control conditions. Data are means ± SEM of 10 replicates. (D and E) Western immunoblot for β-catenin and vinculin after treatment with BRD0705 and BRD0320 (D) or BRD3731 and BRD0320 (E) in HL-60 cells. (F) Western immunoblot for β-catenin, phospho–β-catenin (S675), phospho–β-catenin (S33/37/T41), and actin after treatment with the indicated inhibitors in TF-1 cells. (G) β-Catenin TCF/LEF green fluorescent protein (GFP) reporter assay in TF-1 after 24 hours of treatment with the indicated inhibitor. P < 0.05 is calculated using a Welch’s t test in comparison with control conditions. Data are means ± SEM of three replicates. (H and I) Western immunoblots for β-catenin and actin after treatment with the indicated inhibitors in MV4-11 (H) and MLL-AF9 murine leukemic cells (I). BM, bone marrow.

Because genetic KO or knockdown of the GSK3α or GSK3β individually does not induce β-catenin accumulation in an embryonic stem cell context (22), we set out to define whether a ~10 and >4- to 6-fold selectivity in a biochemical and cellular context would translate to sufficient selectivity to mitigate the effect of dual inhibition on total β-catenin protein stabilization and nuclear translocation. Immunofluorescence staining of β-catenin was performed in HEK 293T cells (Fig. 4B), demonstrating an absence of nuclear β-catenin accumulation after selective inhibition of either GSK3α (BRD0705) or GSK3β (BRD3731) for 24 hours (20 μM). In contrast, dual paralog inhibition, with BRD0320 or CHIR99021, a commercially available pan-GSK3 inhibitor (49), resulted in total β-catenin stabilization and nuclear translocation. Using a β-catenin–dependent T cell factor (TCF)/lymphoid enhancer-binding factor (LEF) luciferase reporter assay, we confirmed the absence of β-catenin–induced target activation after treatment with BRD0705 or BRD3731 (Fig. 4C). In contrast, the dual GSK3 inhibitor BRD0320 or CHIR99021 activated the reporter.

Given the increasing interest in targeting GSK3 in cancer and the potential oncogenic risk of β-catenin stabilization, we set out to validate the safety of our GSK3 inhibitors in a neoplastic context using multiple AML models. The effect of paralog GSK3 inhibition using our selective inhibitors was evaluated in three human AML cell lines and one murine primary AML model. As observed by β-catenin Western immunoblotting in Fig. 4 (D and E), total protein amounts remained largely below dual GSK3 inhibition–induced β-catenin stabilization up to 80 μM of BRD0705 treatment, whereas BRD3731 induced β-catenin stabilization starting at 20 μM in the HL-60 AML cell line. To further assess the threshold of β-catenin stabilization and nuclear translocation in AML, we transduced the TF-1 AML cell line with a TCF/LEF reporter construct. We first evaluated total protein amounts and phosphorylation of β-catenin after GSK3 inhibition (Fig. 4F). Consistent with our observation in the HL-60 cell line, GSK3β inhibition with BRD3731 at 20 μM decreased β-catenin S33/37/T41 phosphorylation and induced β-catenin S675 phosphorylation, resulting in increased β-catenin. GSK3α inhibition with BRD0705 at 20 μM did not change β-catenin total protein and phosphorylation. TCF/LEF induction was only observed at 60 and 20 μM of BRD0705 and BRD3731 treatment, respectively, whereas 10 μM of BRD0320 was sufficient to promote β-catenin nuclear translocation and subsequent transcriptional effects in the TF-1 context (Fig. 4G). Because the effects of the genetic suppression of GSK3β on β-catenin stabilization have been inconsistent (22, 31, 34), we decided to further evaluate the effects of BRD3731 in two additional AML models. As observed in Fig. 4 (H and I), no β-catenin stabilization was observed after GSK3β inhibition in both the MV4-11 AML cell line and an MLL-AF9–induced murine AML at 20 μM, in contrast to our observations in TF-1 and HL-60 cells.

Together, these data demonstrate an absence of β-catenin stabilization after GSK3α inhibition in all AML cell lines tested, whereas GSK3β inhibition appears to drive β-catenin stabilization in a context-dependent manner, consistent with previous data obtained with genetic suppression of GSK3β (22, 31, 34). From a toxicity perspective, we demonstrated that ~10-fold selectivity for GSK3α versus GSK3β is sufficient to achieve potent GSK3α inhibition with no effect on β-catenin stabilization. With BRD0705, we have successfully decoupled GSK3α kinase inhibitory effects from WNT/β-catenin pathway activation, mitigating a potential mechanism-based toxicity. In light of this finding and previous studies supporting the targeting of GSK3α in AML and raising concerns over targeting GSK3β (22, 31, 34), we have focused our attention on the GSK3α-selective compound BRD0705 in the AML cellular context.

BRD0705 selectively targets GSK3α and induces differentiation in AML cell lines and primary patient samples

To further assess GSK3α-selective target engagement in AML, we performed time-response (Fig. 5A) and concentration-response (Fig. 5B) immunoblotting studies of GSK3α/β Tyr279/216 phosphorylation after treatment with BRD0705 or BRD0320. GSK3α/β Tyr279/216 autophosphorylation sites are used as a surrogate for GSK3α/β enzymatic activity and are reflective of proximal-based substrate inhibition (46, 47). As expected, the GSK3α-selective inhibitor BRD0705 impaired GSK3α Tyr279 phosphorylation in a time- and concentration-dependent manner without affecting GSK3β Tyr216 phosphorylation. In contrast, the dual inhibitor BRD0320 decreased active site phosphorylation for both GSK3α and GSK3β. A decrease in the phosphorylation of glycogen synthase, a direct target of GSK3 (50), was also observed (Fig. 5B). As expected, BRD5648, the inactive enantiomer of BRD0705, did not induce changes in enzyme phosphorylation or total β-catenin protein stabilization, supporting the on-target GSK3 activity of these selective inhibitors (Fig. 5B). Finally, to further assess BRD0705’s clinical translational potential, we evaluated the functional selectivity of BRD0705 toward direct selective inhibition of GSK3α (phospho-GSK3α, Y279) and confirmed the absence of downstream stabilization of β-catenin in primary blasts from five independent patients with AML (Fig. 5C, data shown for two representative patient samples).

Fig. 5 BRD0705 induces differentiation in AML cell lines and primary patient samples through GSK3α-selective inhibition.

(A and B) Time-response (A) and dose-response (B) Western immunoblots for β-catenin, phospho-GSK3α/β (Tyr279/216), total GSK3α/β, glycogen synthase (GYS), phospho-GYS (S641), and vinculin after treatment with the indicated inhibitors in U937 cells. (C) Dose-response Western immunoblot for β-catenin, phospho-GSK3α/β (Tyr279/216), total GSK3α/β, and vinculin after treatment with the indicated inhibitors in primary AML patient samples. (D) May-Grunwald-Giemsa staining after BRD0705 treatment for 6 days. Scale bar, 10 μm. (E) Fluorescence-activated cell sorting (FACS) analysis of the expression of CD11b, CD11c, and CD14 cell surface markers after BRD0705 treatment. *P < 0.05 calculated using a Welch’s t test in comparison with the control condition. Data are means ± SEM of three replicates. (F) FACS analysis of the expression of CD14 and CD117 cell surface markers in five primary AML samples treated with BRD0705. *P < 0.05 calculated using nonparametric Mann-Whitney test in comparison with control condition. Data are means ± SEM of five primary AML samples. (G) Western immunoblot for total GSK3α/β and vinculin in U937 after GSK3α CRISPR KO. (H) FACS analysis of the expression of CD11b cell surface marker in U937 GSK3α-WT, GSK3α-KO#1, and GSK3α-KO#2 treated with the indicated inhibitor. P < 0.05 calculated using a Welch’s t test in comparison with the DMSO condition for each clone (#) or the DMSO condition in the GSK3α-WT clone (*). Data are means ± SEM of three replicates.

Together, these data demonstrate the functional selectivity of BRD0705 in targeting endogenous GSK3α kinase activity in multiple cell-based systems, including patient-derived AML cell lines and primary AML cells. This functional selectivity translated to proximal substrates including GSK3α, glycogen synthase, and β-catenin.

Given the implication of GSK3α as a differentiation target in AML (34), we treated a panel of six AML cell lines (HL-60, NB-4, U937, TF-1, MOLM13, and MV4-11) with the GSK3α-selective inhibitor BRD0705 to evaluate potential phenotypic changes. Selective GSK3α inhibition induced changes in morphology (Fig. 5D) and cell surface markers (Fig. 5E, CD11b, CD11c, and CD14), consistent with AML differentiation. We then tested whether these findings were relevant to primary patient AML cells through cell surface marker staining (CD14 and CD117) in a panel of five primary AML patient blast samples (Fig. 5F). Increased differentiation marker (CD14) expression and decreased immature marker (CD117) expression were confirmed in all five patient samples in a dose-responsive manner after BRD0705 treatment. Conversely, BRD3731 effects were inconsistent, with a modest differentiation phenotype in some cell lines but a decrease in cell surface differentiation markers in a subset of the AML cell lines (fig. S10, A to C). These results support the GSK3α genetic suppression–induced phenotype in AML (34) and confirm that enzymatic inhibition recapitulates the phenotype observed with genetic suppression.

To provide more definitive evidence supporting the selectivity of BRD0705 for GSK3α, we generated GSK3α isogenic KO U937 clones (Fig. 5, G and H), characterized the phenotypic effect of KO on differentiation, and tested BRD0705 and BRD3731 treatment of the two different KO clones. GSK3α KO recapitulated the GSK3α inhibition phenotype, as evidenced by increased expression of differentiation cell surface markers with dimethyl sulfoxide (DMSO) treatment. These cells showed no additional increase in expression of the differentiation cell surface marker CD11b upon treatment with BRD0705, in contrast to the GSK3α-expressing clones (Fig. 5H), providing validation of in vitro paralog selectivity and on-target activity. Moreover, GSK3β inhibition using BRD3731 produced a similar increase in CD11b marker on GSK3α WT and GSK3α KO clones (Fig. 5H).

GSK3α inhibition triggers differentiation and reduced stemness transcriptional programs without increasing β-catenin–related signatures

To characterize the effects of GSK3 inhibitor treatment on genome-wide transcriptional programs, we used RNA sequencing (RNA-seq) to profile the U937 AML cell line after 24 hours of treatment with BRD0705, BRD3731, or BRD0320 at 10 μM. Dual GSK3 inhibition induced a greater change in transcriptional programs (556 up-regulated and 975 down-regulated genes) compared with selective GSK3α inhibition (55 up-regulated and 193 down-regulated genes) or selective GSK3β inhibition (187 up-regulated and 203 down-regulated genes), as illustrated in Fig. 6 (A and B). The top 30 differentially up-regulated and down-regulated genes upon selective GSK3α, selective GSK3β, or dual GSK3 inhibition are depicted in Fig. 6C. All differentially expressed genes based on a permutation P < 0.05 and false discovery rate (FDR) < 0.05 are listed in tables S9 to S11.

Fig. 6 GSK3α- and GSK3β-selective inhibitors trigger differential transcriptional programs.

(A and B) Venn diagrams (A) and scatterplots (B) of genes modulated after BRD0705, BRD3731, or BRD0320 treatment. Significantly depleted or enriched genes after BRD3731 or BRD0705 treatment are highlighted in blue and red, respectively (log2FC < −0.5 or > 0.5 and P ≤ 0.05). (C) Heatmaps of the top 30 genes found to be differentially expressed by genomic profiling upon BRD0705, BRD3731, or BRD0320 treatment in the U937 cell line. Depleted and enriched genes are in blue and red, respectively. Data are presented as row-normalized. Each column corresponds to one of three replicates for each treatment condition. The BRD0705, BRD3731, and BRD0320 signatures identified by RNA-seq were then interrogated in a functional enrichment analysis across the MSigDB database. (D to F) GSEA plots for β-catenin signaling pathway induction after treatment with BRD0705 (D), BRD3731 (E), or BRD0320 (F). NES, normalized enrichment score.

To assess more specifically BRD0705’s functional effects in AML, we interrogated our RNA-seq data with published gene signatures for statistically significant enrichment by gene set enrichment analysis (GSEA). We generated a functional network illustrating GSK3α inhibition’s transcriptional effects in AML (fig. S11A and table S12). This analysis first confirmed that GSK3 inhibition gene sets were statistically enriched among genes down-regulated by BRD0705 treatment (fig. S11B). Similarly, GSK3α inhibition by BRD0705 induced differentiation transcriptional programs and down-regulated stemness signatures (fig. S11, C and D). As reported by Guezguez et al. (31), GSK3α inhibition up-regulated multiple mitochondria metabolism transcriptional programs (fig. S11E). Overall, this genomic profiling allows independent validation of the previously established on-target, phenotypic effects of BRD0705 through an unbiased approach. Finally, a similar enrichment analysis performed on BRD0320-induced transcriptional signatures revealed a strong induction of β-catenin signaling pathway genes, whereas BRD0705 treatment did not induce enrichment for β-catenin–related gene sets. BRD3731 did not show significant induction of β-catenin transcriptional programs (Fig. 6, D to F).

BRD0705 impairs colony formation in AML cell lines and patient samples and demonstrates in vivo efficacy in multiple xenograft and syngeneic AML mouse models

To further characterize the phenotypic consequences of BRD0705 treatment, we tested its ability to impair AML colony formation in methylcellulose. We found that BRD0705 impaired AML colony formation in all six tested cell lines (MOLM13, TF-1, U937, MV4-11, HL-60, and NB-4) in a concentration-dependent manner (Fig. 7A), as opposed to BRD3731, which impaired colony formation in TF-1 but increased colony-forming ability in the MV4-11 cell line (fig. S12). We then tested whether these findings were relevant to primary AML patient samples through methylcellulose colony-forming assays in AML samples from five patients previously selected for colony formation ability. Reduced ability to form colonies (Fig. 7B) was confirmed in all five patient AML samples, again in a concentration-dependent manner, after BRD0705 treatment. No effect of BRD0705 treatment was observed on normal CD34 cell colony formation, thus offering a potential therapeutic window for selectively targeting leukemic cells (Fig. 7C). Detailed patient characteristics are provided in table S13. In accordance with these findings and the RNA-seq data suggesting down-regulation of stemness-related gene expression signatures with BRD0705 treatment, we evaluated the capacity of our selective GSK3α inhibitor to eradicate leukemia-initiating cells (LICs) in a syngeneic mouse model of AML driven by the oncogene MLL-AF9. BRD0705 treatment significantly impaired development of MLL-AF9 leukemia in secondary recipient mice in comparison to the vehicle-treated group (Fig. 7D; P < 0.05). Dual GSK3 inhibition with BRD0320 only induced a mild delay in leukemia onset, whereas GSK3β-selective inhibition with BRD3731 did not affect leukemia development. A limiting dilution experiment performed in this mouse model revealed a 3.79-fold decrease in LIC frequency after BRD0705 pretreatment (Fig. 7E). No significant difference in LIC frequency was observed after BRD3731 or BRD0320 pretreatments.

Fig. 7 BRD0705 impairs colony formation in AML cell lines and patient cells and shows in vivo efficacy in multiple AML mouse models.

(A to C) Colony formation assay of the indicated AML cell lines (A), five primary AML samples (B), and human CD34 cells (C) after BRD0705 treatment. Data are represented as means ± SEM of three replicates. (D) MLL-AF9 cells (250,000) injected into secondary recipient mice after 24 hours of pretreatment with DMSO, BRD0705, BRD3731, or BRD0320. Kaplan-Meier curves are shown (n = 5 per group). (E) MLL-AF9 cells injected into secondary recipient mice in a serially diluted manner (500,000, 250,000, 100,000, or 50,000 cells). Frequency of LICs calculated using ELDA software. (F) MLL-AF9 AML cells injected into secondary recipient mice 3 days before treatment with vehicle (black) or BRD0705 at 30 mg/kg (red). Kaplan-Meier curves for each group of mice (n = 5 per group). (G) HL-60–Luc cells injected into recipient mice and treated with vehicle (black) or BRD0705 at 15 mg/kg (red) or 30 mg/kg (blue). Bioluminescence was quantified. (H) Kaplan-Meier curves for each group of mice treated as in (G) (n = 5 per group). (I) MV4-11–Luc cells injected into recipient mice and treated with vehicle (black), BRD0705 (red), BRD3731 (blue), or BRD0320 (dashed blue). Bioluminescence was quantified. Data are means ± SEM. *P value calculated for the latest time point. (J) Kaplan-Meier curves for each group of mice (n = 10 per group). P < 0.05 determined by log-rank (Mantel-Cox) test in comparison with vehicle.

To better enable in vivo studies, we next characterized the peripheral exposure of BRD0705 in mice. After a single oral dose administration of BRD0705, plasma concentrations were quantifiable for up to 24 hours, with Tmax of 0.25 hours and area under the curve (AUC) of 67.6 μmol/L.h (fig. S13A). To determine whether selective inhibition of GSK3α would translate to a decreased AML burden in vivo, we tested the efficacy of BRD0705 in the MLL-AF9 syngeneic mouse model. Mice were separated into two groups and treated with either vehicle or BRD0705 at 30 mg/kg through oral gavage twice daily. Mouse survival was significantly improved under BRD0705 treatment (Fig. 7F; P < 0.05).

We then evaluated BRD0705 efficacy in an orthotopic HL-60 xenograft AML model generated through injection of luciferized HL-60 cells into the tail veins of NSG mice. Mice were separated into three groups and received either vehicle or BRD0705 at 15 or 30 mg/kg by oral gavage once a day. BRD0705 treatment decreased leukemia progression in a dose-dependent manner, as evidenced by the decrease in bioluminescence measurements of disease burden (Fig. 7G; P < 0.05 and P < 0.01 at 15 and 30 mg/kg, respectively) and circulating human CD45-expressing AML cells (fig. S13B), resulting in a prolonged overall survival (Fig. 7H). No major weight loss and no complete blood count changes were observed upon in vivo administration of BRD0705 (fig. S13, C to G) at this dose and schedule. GSK3 inhibition has been previously shown to decrease glycogen synthase phosphorylation in peripheral blood mononuclear cells, resulting in an increased cellular glycogen content (25). Thus, we first validated glycogen accumulation (periodic acid–Schiff staining and quantitative measurement) as an on-target effect of GSK3α inhibition after BRD0705 treatment in vitro in AML cell lines, including HL-60 (fig. S14, A and B). We then confirmed the presence of increased glycogen in peripheral mononuclear cells obtained from BRD0705-treated mice as compared to vehicle-treated animals (fig. S14C). Glycogen measurement may therefore serve as a potential marker of BRD0705 on-target activity.

Finally, to evaluate the differential effects of paralog-selective inhibition in vivo, we generated an MV4-11 orthotopic AML mouse model through tail vein injection of a luciferized MV4-11 cell line into irradiated NSG recipient mice. Mice were separated into four groups and treated with vehicle, BRD0705, BRD3731, or BRD0320 at 30 mg/kg once daily by oral gavage. BRD0705 treatment significantly decreased disease burden, as measured by bioluminescence and prolonged mouse survival (Fig. 7, I and J; P < 0.05). No significant disease burden or survival effect was detected under BRD3731 or BRD0320 treatments. Together, these results suggest a potential oncogenic effect of GSK3β-selective paralog inhibition, in part independent from β-catenin stabilization because BRD3731 did not increase β-catenin in MV4-11 cells. This undesirable effect could potentially counteract the beneficial effect of GSK3α inhibition when using a dual GSK3 inhibitor, thus offering an additional argument for selective GSK3α targeting in AML.


Reported GSK3 inhibitors generally demonstrate poor to moderate selectivity versus the broader human kinome. To date, only one reported GSK3 inhibitor has demonstrated preferential inhibition of GSK3α over GSK3β (23, 24, 51), but this compound has not been characterized for cell-based functional selectivity or studied in mammalian models in vivo. Dual GSK3α/β inhibitors, however, have entered clinical testing for both neurodegenerative disorders and cancer at tolerated doses, but the general lack of clinical activity was disappointing (5257).

From a cancer therapy perspective, the compound LY2090314, a dual GSK3 inhibitor, was initially tested in a phase 1 study in combination with pemetrexed and carboplatin in patients with advanced solid tumors, where the primary toxicities included visual disturbances and peri-infusional chest pain that was relieved with IV H2 blockers (53). Next, a phase 2 study was performed testing LY2090314 in patients with relapsed/refractory AML or patients with newly diagnosed AML who could not tolerate standard induction chemotherapy (56). Here, primary toxicities included decreased appetite and nausea; hematologic toxicities included rare febrile neutropenia, thrombocytopenia, and anemia. Although five patients had a decrease in peripheral blood blasts, there were no complete or partial remissions, despite evidence of target inhibition based on β-catenin stabilization in peripheral mononuclear cells and in patients’ leukemia blasts.

Although increased β-catenin provided confirmation of GSK3α/β target inhibition in the testing of LY2090314 in patients with AML, the notable lack of clinically meaningful responses raises the question of the potential role of β-catenin stabilization in attenuating response to this drug in patients with leukemia. The concurrent inhibition of GSK3α and GSK3β induces β-catenin stabilization and therefore increases self-renewal of LICs, raising potential efficacy concerns (22, 32, 33, 58). In support of these studies, more recently, Guezguez et al. (31) reported that dual GSK3α/β KO resulted in an overt, aggressive AML, and GSK3β KO drove a preneoplastic mixed myeloproliferative/myelodysplastic state. Paralog-selective GSK3α or GSK3β suppression with genetic approaches did not stabilize β-catenin in embryonic stem cell lines (22). In our own previous work, GSK3α- or GSK3β-directed short hairpin RNA (shRNA) did not stabilize β-catenin in the MOLM14 AML cell line, and GSK3α suppression promoted differentiation in AML cells (34). In the Guezguez et al. study (31), however, it was reported that GSK3β KO drives a β-catenin signature in the context of adult hematopoietic stem/progenitor cells as a mechanism of inducing the preneoplastic phenotype, lending caution to the application of GSK3β-selective inhibitors. These results are consistent with the data reported here, highlighting the importance of cellular context for β-catenin stabilization induced by GSK3β inhibition. In line with the data reported herein, where we did not see β-catenin stabilization with GSK3α inhibition, Guezguez et al. (31) had shown that GSK3α KO did not stabilize β-catenin and did not cause a leukemic or preleukemic phenotype, offering additional support for the exploration of GSK3α-selective inhibitors in AML treatment. Hence, in our studies of paralog-selective GSK3 molecules in AML, we have focused on GSK3α inhibitors, which display no functional effects on β-catenin.

Our GSK3 inhibitors were designed from a single-core pyrazolo-tetrahydroquinolinone scaffold, which presented a tridentate binding mode within the ATP-binding domain. To design these highly selective competitive inhibitors, we applied a rational, structure-based approach exploiting the hydrogen bond network on the backside of the hinge-binding domain centered on an Asp→Glu switch hypothesis. The difference of a single carbon atom between these two highly similar kinases (GSK3β-Asp133→GSK3α-Glu196) translated into topological differences within the ATP-binding pocket and the adjacent hydrophobic selectivity pocket. A secondary selectivity pocket specific to the ATP-binding site of GSK3α is also described. These differences were engaged by distinct small-molecule ligands to generate chemically matched, paralog-selective, and highly kinome-selective GSK3 inhibitors. A robust structural rationale for the observed selectivity was developed and subsequently validated through site-directed mutagenesis and crystallographic studies. Further medicinal chemistry rational design, taking full advantage of residue and volume differences in these selectivity pockets, should result in even greater paralog selectivity for second-generation inhibitors. We demonstrated that it is feasible to inhibit the kinase function of either GSK3α or GSK3β with sufficiently paralog-selective inhibitors (10-fold) that functionally decouple paralog-selective effects on β-catenin and further downstream activation of the WNT pathway in a cellular context. These paralog-selective inhibitors should provide an increased therapeutic index and allow for clinical doses and exposures that were not previously achievable.

Although we have pursued the study of an Asp→Glu switch within the GSK3 paralogs, it is tempting to speculate on the generality of this structural phenomenon in other paralog pairs of kinases. From an evolutionary perspective, studying the biological relevance of this backend hinge hydrogen bond network and single amino acid switch will be key to deciphering functional differences between such highly similar kinases and help identify preferential substrates of each paralog kinase. Future investigations should seek to determine whether this Asp→Glu switch is conserved in other kinases and whether the chemistry discovered here can be exploited toward the development of other paralog-selective compounds.

With an eye toward developing paralog-selective GSK3 inhibitors as a therapeutically tractable approach to cancers, such as AML, we explored the effect of selective GSK3α inhibition on differentiation, stemness, and colony formation. Small-molecule inhibition of GSK3α by BRD0705 induced myeloid differentiation through morphological and surface marker changes and impaired colony formation in AML cell lines and primary AML samples. Moreover, BRD0705 treatment reduced the frequency of LICs, resulting in delayed disease onset in an MLL-AF9 syngeneic mice model, supporting the in vivo relevance of the observed repression in stemness transcriptional programs observed in AML cells with BRD0705 treatment. BRD0705 showed a favorable tolerability profile in mice treated with 15 and 30 mg/kg, and BRD0705 decreased leukemia burden and prolonged survival in HL-60, MV4-11, and MLL-AF9 AML mouse models. No effect on disease burden or survival was detected with BRD3731 (GSK3β inhibition) treatment. Future studies of this paralog series will need to extend testing to additional mouse models of AML and seek to identify drug-drug combinations with GSK3α inhibitors in AML. This study provides a proof-of-concept roadmap toward the development of paralog-selective GSK3α inhibitors. However, additional optimization of theses molecular probes will be needed to develop this chemical series toward a drug candidate, setting the stage to test a paralog-selective GSK3α inhibitor in clinical trials for the treatment of AML. Beyond AML, this set of highly selective isochemogenic inhibitors will be useful chemical probes to refine our understanding of fundamental GSK3 biology and unravel substrate and functional selectivity of each GSK3 paralog, identifying potential therapeutic options in human diseases where the hyperactivity of the GSK3 kinases has been implicated, such as diabetes, cardiovascular, and central nervous system disorders.


Study design

This study sought to design a set of isochemogenic and paralog-selective GSK3 inhibitors and validate the use of our GSK3α inhibitor as a targeted therapy in AML. Paralog-selective GSK3 inhibitors (BRD0705, BRD3731, and BRD0320) were designed. Selectivity and on-target effects were validated using in vitro and in cellulo assays (kinome assays evaluating selectivity against >300 kinases; BRET luminescence assays; β-catenin accumulation through protein quantification, assessment of phosphorylation modifications, nuclear translocation evaluated by both transcriptional reporter assays and immunofluorescence, Y279-GSK3α–selective phosphorylation impairment, GSK3α KO single clones, and selectivity pocket mutant forms of GSK3). Effects of BRD0705 in AML were studied in vitro in six AML cell lines and five primary patient AML blast samples and in vivo in three mouse models, including human orthotopic xenograft and syngeneic mouse models with different genetic backgrounds. Sample size was chosen in light of the fact that these in vivo models were historically highly penetrant, aggressive, and consistent. Blinded observers visually inspected mice for obvious signs of illness, such as loss of appetite, hunched posture, and lethargy. Mice were randomly assigned to each treatment group. The number of experimental replicates is specified in each figure legend.

Plasmids and cell infection

The β-catenin reporter 7xTCF-luc-mCherry was a gift from W. C. Hahn (Dana-Farber Cancer Institute). LentiCRISPR v2 was a gift from F. Zhang (Addgene plasmid #52961). We generated lentiCRISPR v2–NT [single-guide RNA (sgRNA): GTAGCGAACGTGTCCGGCGT] and lentiCRISPR v2–GSK3α#12 (sgRNA: TACACGTGAGGCAAGGGTTG) vectors following the GeCKO Lentiviral CRISPR toolbox cloning protocol (59). The GSK3α pWZL overexpression vector was generated as previously described (34). The pMMP LUC-NEO vector was a gift from A. Kung (Columbia University).

For virus production, 12 μg of the above plasmids and 6 μg of pCMV-GAG/POL and pCMV-VSVG (for retroviral infection) or 6 μg of pCMV8.9 or psPAX2 and pCMV-VSVG (for lentiviral infection) packaging vectors were transfected into the HEK 293T packaging cell line using X-tremeGENE 9 (Roche), and the resulting viral supernatants were harvested as previously described (34). The lentiviruses and retroviruses were then concentrated using PEG-it Virus Precipitation Solution or Retro-Concentin (SBI System Biosciences), respectively. Cells were infected using RetroNectin (Clontech) following the manufacturer’s recommendations for retroviral vectors and as previously described for lentiviral plasmids (34).

RNA sequencing

Genome expression profiling. U937 cells were treated in triplicate with either vehicle or 10 μM BRD0705, BRD3731, or BRD0320 for 24 hours. Total RNA was extracted with an RNeasy kit (Qiagen) and profiled by RNA-seq (HiSeq, Illumina) at the BioMicroCenter of the Massachusetts Institute of Technology (MIT). The total number of reads for individual samples ranged from 21 to 24 million base pairs. Quality control tests for the unmapped reads were performed using the FASTQC software ( The reads were aligned to the GRCh37/hg19 human genes by using Bowtie v2 (60). The average percentage of uniquely mapped reads in the aligned data was 84.90%, with an SD of 2.7%.

Fragments per kilobase million (FPKM) scores for genes were computed on the basis of the RSEM R software package (61). Expression data were estimated as log2(FPKM). The significance of the differential expression between the control and treatment phenotypes was estimated on the basis of the EBSeq method implemented in the EBSeq R library (62) by using the significance cutoff P ≤ 0.05 for the posterior probability. The aligned RNA-seq data for this experiment have been deposited in the Gene Expression Omnibus (GEO) database (GSE109987).

Comparative marker analysis. The 15 samples available in the data were separated into four groups: one corresponding to vehicle (six replicates), the second to BRD0705 (three replicates), the third to BRD3731 (three replicates), and the fourth to BRD0320 treatment (three replicates). The Comparative Marker Selection module from GenePattern v3.9.6 (63) was used to identify individual genes that were differentially expressed between treated and vehicle conditions. The log2(FPKM) expression data were analyzed by applying a two-sided signal-to-noise ratio (SNR) test followed by 1000 permutations of phenotype labels. The settings for the SNR parameters were as follows: log-transformed data, yes; complete, no; balanced, no; smooth P values, yes.

Molecular signatures for genes down-regulated or up-regulated by each of the BRD0705, BRD3731, and BRD0320 compounds versus vehicle were defined on the basis of the cutoffs SNR permutation P ≤ 0.05, Benjamini-Hochberg FDR ≤ 0.05, and absolute fold change [FC, log2(FPKM)] ≥ 1.5. The BRD0705 molecular signature consisted of 193 genes down-regulated and 55 genes up-regulated between treated samples and vehicle (table S9), the BRD3731 molecular signature consisted of 203 genes down-regulated and 187 genes up-regulated between treated samples and vehicle (table S10), whereas the BRD0320 molecular signature consisted of 975 genes down-regulated and 556 genes up-regulated between treated samples and vehicle (table S11).

GSEA analysis. The GSEA (v2.1.0) software (64, 65) was used to identify functional associations of the molecular phenotypes induced by BRD0705, BRD3731, and BRD0320 with a compendium of gene signatures including the MSigDB v5.0 (65) collections c2 of 4725 curated gene sets and c6 of 189 oncogenic signatures, the DMAP collection of hematopoietic lineage differentiation signatures (66), and the GSK3 inhibitor CHIR99021 gene set signature [GSE54056 (67)]. Gene sets with fewer than 15 genes or more than 500 genes were excluded from the analysis. Gene sets with an FDR ≤ 0.25 and a nominal P ≤ 0.05 were considered significant hits. The results were visualized on GSEA plots and heatmaps for selected gene signatures and with the Enrichment Map software, which organizes the significant gene sets into a network called an “enrichment map.” In the enrichment map, the nodes correspond to gene sets, and the edges reflect significant overlap between the nodes according to a two-tailed Fisher’s exact test. The size of the nodes is correlated with the number of genes in the gene set. The hubs correspond to collections of gene sets with a unifying class label according to gene ontology (GO) biologic processes.

Statistical analysis

Statistical analysis was performed using Microsoft Excel, PRISM 5.03 (GraphPad), or the indicated software for more dedicated analysis. Data were analyzed using a nonparametric Mann-Whitney test (with the assumption of no Gaussian distribution of the group) unless otherwise specified, and the threshold of significance (α) was always set at 0.05.


Materials and Methods

Fig. S1. GSK3 conservation across species.

Fig. S2. Compounds’ electron density maps.

Fig. S3. X-ray crystal structure of GSK3 from U. maydis.

Fig. S4. Schematic of backend interactions.

Fig. S5. Hinge backend interaction measurements and energy plots.

Fig. S6. Synthesis of pyrazolodihydroquinolinones.

Fig. S7. Spectra for compounds in Fig. 2B.

Fig. S8. X-ray crystal structures of hGSK3B bound to BRD3731 and BRD0705 and hGSK3B (D133E) bound to BRD0705.

Fig. S9. Live cell target engagement analysis for GSK3α and GSK3β using NanoBRET.

Fig. S10. Heterogeneous effects of BRD3731 on differentiation in AML cell lines.

Fig. S11. Differentiation and impaired stemness and mitochondria transcriptional programs triggered by BRD0705.

Fig. S12. Heterogeneous effects of BRD3731 on colony formation in AML cell lines.

Fig. S13. Pharmacokinetic, efficacy, and tolerability properties of BRD0705.

Fig. S14. Glycogen accumulation induced by BRD0705 in AML.

Table S1. X-ray data reduction statistics for x-ray co-crystal structures of hGSK3β with BRD0209, BRD3731, and BRD0705.

Table S2. Crystallographic refinement statistics for x-ray co-crystal structures of hGSK3β with BRD0209, BRD3731, and BRD0705.

Table S3. Kinome selectivity for BRD0320 (Carna Biosciences).

Table S4. Kinome selectivity for BRD5648 (Carna Biosciences).

Table S5. Kinome selectivity for BRD0705 (Carna Biosciences).

Table S6. Kinome selectivity for BRD3731 (Carna Biosciences).

Table S7. Final diffraction statistics for the mutant hGSK3β (D133E)/BRD0705 complex crystal used in structure determination.

Table S8. Final refinement statistics for the mutant hGSK3β (D133E)/BRD0705 complex structure.

Table S9. Lists of genes differentially expressed between control and BRD0705-treated U937 cells.

Table S10. Lists of genes differentially expressed between control and BRD3731-treated U937 cells.

Table S11. Lists of genes differentially expressed between control and BRD0320-treated U937 cells.

Table S12. Top enriched gene sets from the functional groups predicted in BRD0705 versus DMSO enrichment map.

Table S13. Cytogenetics and molecular and clinical characteristics of patient samples used in this study.


Acknowledgments: We thank R. Suto and A. White (Xtal BioStructures) for helpful discussions. NB-4 and MV4-11 cells were provided by R. Levine (Memorial Sloan Kettering Cancer Center), and MOLM13 was provided by B. Ebert (Dana-Farber Cancer Institute). The β-catenin reporter 7xTCF-luc-mCherry was provided by W. C. Hahn (Dana-Farber Cancer Institute). LentiCRISPR v2 was provided by F. Zhang (Addgene plasmid #52961). The pMMP LUC-NEO vector was provided by A. Kung (Columbia University). Funding: This work was supported by the Stanley Medical Research Institute and grants from the National Cancer Institute (R01 CA140292 and R35CA210030) (K.S.) and the Children’s Leukemia Research Association (K.S.). K.S. is a Leukemia and Lymphoma Society Scholar. L.B. is an MD-PhD candidate of “Ecole de l’INSERM Liliane Bettencourt” and a recipient of Philippe Foundation fellowship. A.P. is a recipient of the ATIP-AVENIR and the European Research Council research programs. Author contributions: F.F.W., L.B., A.J.C., M.W., J.R.S., L.R., A.P., T.K., O.H., A.K., E.S., Y.-L.Z., J.Q.P., M.T.H., K.S., and E.B.H. designed experiments. F.F.W., L.B., A.J.C., M.W., J.R.S., J.P.G., L.R., A.P., A.C., M.B., T.K., X.S., and Y.-L.Z. performed experiments. F.F.W., L.B., A.J.C., M.W., J.R.S., L.R., A.P., G.A., O.H., E.S., Y.-L.Z., J.Q.P., M.T.H., K.S., and E.B.H. analyzed and interpreted the data. A.J.C. and T.K. performed MD experiments and analyzed data. F.F.W., M.W., J.R.S., and E.B.H. designed and/or synthesized compounds. G.A. performed statistical analysis, biostatistics, and computational analysis of the RNA-seq. I.G., D.J.D., and R.M.S. provided patient samples. The manuscript was written by F.F.W. and L.B. with input from all authors. Competing interests: The following patent applications were filed (US-2014-0107141) or granted (U.S. Patent 9,096,594). F.F.W., J.Q.P., E.B.H., M.W., and K.S. are inventors on patent application WO2014059383 A1 held/submitted by the Broad Institute, Dana-Farber Cancer Institute, and General Hospital Corporation that covers the small molecules described in this article and their use in disease. F.F.W., E.B.H., and E.S. are inventors on patent application US 20160375006 A1 held/submitted by the Broad Institute and MIT that covers the use of the small molecules described in this article in disease. F.F.W. served as a paid consultant for company that licensed the molecules described in this article, until end of December 2017. Data and materials availability: The authors declare that data supporting the findings of this study are available within the paper and its supplementary information document. Structure factors and coordinates have been deposited in the PDB under the following accession codes: 5KPK (hGSK3β bound to BRD0209), 5KPM (hGSK3B bound to BRD3731), 5KPL (hGSK3β bound to BRD0705), and 5T31 (hGSK3β mutant bound to BRD0705). BRD0705, BRD3731, BRD0320, and BRD5648 can be obtained through a materials transfer agreement (MTA) from the Broad Institute of Harvard and MIT. The aligned RNA-seq data have been deposited in the GEO database (GSE109987).

Stay Connected to Science Translational Medicine

Navigate This Article