Validation of an allosteric binding site of Src kinase identified by unbiased ligand binding simulations

Abstract

Allostery plays a primary role in regulating protein activity, making it an important mechanism in human disease and drug discovery. Identifying allosteric regulatory sites to explore their biological significance and therapeutic potential is invaluable to drug discovery; however, identification remains a challenge. Allosteric sites are often “cryptic” without clear geometric or chemical features. Since allosteric regulatory sites are often less conserved in protein kinases than the orthosteric ATP binding site, allosteric ligands are commonly more specific than ATP competitive inhibitors. We present a generalizable computational protocol to predict allosteric ligand binding sites based on unbiased ligand binding simulation trajectories. We demonstrate the feasibility of this protocol by revisiting our previously published ligand binding simulations using the first identified viral proto-oncogene, Src kinase, as a model system. The binding paths for kinase inhibitor PP1 uncovered three metastable intermediate states before binding the high-affinity ATP-binding pocket, revealing two previously known allosteric sites and one novel site. Herein, we validate the novel site using a combination of virtual screening and experimental assays to identify a v-type allosteric small-molecule inhibitor that targets this novel site with specificity for Src over closely related kinases. This study provides a proof-of-concept for employing unbiased ligand binding simulations to identify cryptic allosteric binding sites and is widely applicable to other protein-ligand systems.

Introduction

Allostery has been studied for over a half-century leading up to our current understanding as an intrinsic property of dynamic proteins for propagating signals from one site to distal functional sites within a given protein [1–4]. Allosteric regulation is critical to signal transduction by eukaryotic protein kinases, which represent one of the largest superfamilies of proteins with over 500 known family members [5–8]. Protein kinases share a structurally conserved catalytic kinase domain that allosterically integrates stimuli from regulatory domains, post-translational modifications within the kinase domain, and interactions with substrates and effector proteins [5]. Dysregulation of kinase activity by genetic translocation and mutation drives many diseases, including various forms of cancer, which has led to extensive drug discovery efforts targeting protein kinases [9–12]. However, most FDA-approved small-molecule kinase inhibitors are ATP-competitive and target the highly conserved ATP-binding site of the homologous catalytic kinase domain [13]. The high sequence and structural conservation of the orthosteric ATP-binding site poses a challenge for developing highly specific kinase inhibitors [14]. An alternative strategy for kinase inhibition involves targeting allosteric sites outside of the conserved orthosteric ATP-site to achieve the necessary selectivity profiles for safe and effective inhibition. Therefore, identifying allosteric sites is vital for understanding allosteric regulatory mechanisms in kinases that can be utilized for targeted inhibition.

The growth of kinases as successful therapeutic targets has led to >5,500 published kinase domain structures representing roughly 300 kinases [15]. The wealth of structural information has advanced the field of structure-based drug design by providing mechanistic insights into kinase activation states and ligand binding modes. However, these high-resolution structures exemplify static snapshots that do not reflect the dynamic conformational ensemble of the catalytic kinase domain [16]. Kinases undergo large-scale conformational changes during the ligand binding process, and intermediate kinase conformations are not represented in crystal structures of protein-inhibitor complexes. Therefore, studying ligand binding pathways and the encompassed intermediate protein conformations and ligand interactions can elucidate transient cryptic binding pockets that may be allosterically relevant [17–19].

Long-timescale unbiased MD simulations of beta-blockers binding to the β₂-adrenergic receptor (β₂AR) revealed several metastable hidden binding sites in the extracellular vestibule of the protein [17]. A hidden allosteric site in the extracellular vestibule was later confirmed through an x-ray crystal structure of the human M2 muscarinic acetylcholine receptor with positive binding of an allosteric modulator [20]. These two studies suggest that ligand binding pathways that reveal metastable, intermediate binding sites can be employed as a predictive tool for allosteric binding site identification.

To further determine whether ligand binding pathways can be utilized to identify cryptic allosteric sites in protein kinase domains, we revisited previously published long-timescale unbiased ligand binding simulations that revealed the process of ligand binding to Src KD using small-molecule kinase inhibitors PP1 and dasatinib [18]. The binding simulations demonstrated that PP1 and dasatinib bind to Src kinase through a four-step process: (i) adsorption onto the protein surface, (ii) diffusion across the protein surface, (iii) residence in low-affinity, intermediate binding sites, and (iv) binding to the highest affinity site, which was often accompanied by conformational rearrangement of the binding site. These simulations reproduced the experimentally determined binding conformation in detail, including the orientation of the ordered water molecules that mediate binding interactions.

In step iii of the ligand binding process, PP1 and dasatinib bind to low affinity secondary binding sites. Two of these sites have been previously identified as allosteric sites (PIF and MYR pockets), whereas the third is a novel site in the C-lobe, which we term the G-loop site (Figure 1a). PP1 (Figure 1b) binding outside of the ATP binding site is consistent with a previous report of PP1 as a non-ATP-competitive ligand of Src [21]. Interestingly, the G-loop site is located at the base of the allosteric regulatory network in Src kinase [22], signifying its potential for allosteric modulation of kinase activity. This study set out to determine whether the G-loop site is indeed an allosteric regulatory site that can be targeted with small molecules.

Figure 1: Targeting a proposed allosteric site.

(a) Density map of kinase inhibitor PP1 on Src kinase during unbiased ligand binding simulations with main intermediate binding sites notated. (b) Chemical structure of ATP-competitive inhibitor PP1 employed in simulations. (c) Simulation snapshot of PP1 occupying the G-loop site (bottom), which is not fully formed in the X-ray structure employed as the starting model for the ligand binding simulations (top) (PDB-entry 1Y57) [24]. (d) Allosteric catalytic spine in Src kinase domain [22] interfaces with the predicted G-loop binding site (purple). Allosteric catalytic and regulatory spine residues are highlighted in yellow and blue, respectively. Key regulatory structural elements are color-coded (P-loop = blue, helix-αC = orange, activation loop = magenta). G-loop binding pocket is stabilized when Src kinase is in the active conformation (Lys295-Glu310 salt bridge is formed, DFG Asp404-in) (small box). PP1 bound to predicted G-loop site with visible binding site residues notated (large box). Panels (a) and (c) are adopted from [18].

We observed by unbiased MD simulation that PP1 binds to the G-loop site and does not dissociate for the course of the simulation (SI Movie 1). The G-loop binding pocket was occluded in the starting conformation of the simulation and only formed transiently with ligand-induced stabilization, explaining why this site has been unobserved through x-ray crystal structures of Src (Figure 1c). Here, we validated and characterized the predicted G-loop site by virtually screening for ligands that preferentially bind and stabilize this site in Src. First, we performed a computational docking screen against the G-loop binding pocket. We used a representative conformation of the kinase extracted from the MD simulation in which PP1 remains bound to the G-loop site. Second, we tested 49 top-scoring ligands for their ability to modulate Src kinase activity. Of those compounds, three activated the kinase by more than 2.5 standard deviations, and four compounds inhibited the kinase by more than 2.5 standard deviations. We focused our characterization on the small-molecule compound 1C, which acts as a partial and non-competitive inhibitor of Src kinase activity. Compound 1C inhibits Src in biochemical and cellular assays and is specific for Src over the closely related Hck and Abl kinases which are sequence divergent in the allosteric G-loop binding site. We validated the binding of 1C and introduced binding site mutations that abrogate allosteric inhibition.

The allosteric G-loop binding site provides insights into the allosteric regulation of Src kinase, the development of allosteric inhibitors, and the overall drug binding process. Here we propose a potentially generalizable workflow involving a combination of unbiased ligand binding MD simulations, virtual screening, and experimental validation that can be applied to any allosterically regulated protein with a known structure. Since proteins are highly dynamic, we propose that most may be targetable with small-molecule allosteric ligands. Although it is more challenging to develop high-affinity allosteric inhibitors than orthosteric ones, allosteric inhibitors may provide adequate handles for other pharmacological strategies such as targeted degraders (PROTACs), which tend to require lower affinity to the target to be efficacious [23].

Results

Unbiased MD simulations predict a novel allosteric site in Src Kinase

Based on the conformational stability of the PP1-bound pocket, we selected a representative conformation of the G-loop binding pocket stabilized by PP1 from our previously published unbiased ligand binding simulations [18]. The MD-derived protein structure of Src kinase domain was in the active conformation: (i) helix- αC is rotated inwards, facilitating salt bridge formation between Lys295-Glu310, and (ii) Asp404 of the DFG-motif is facing into the ATP binding site (chicken c-Src numbering) (Figure 1d). Compared to the starting conformation of the binding simulation, based on PDB-entry 1Y57 [24], residues 479–481 moved to reveal a new pocket in the C-lobe of the kinase, which we designated as the G-loop site (Figure 1c). The opening of the G-loop site was primarily attributed to the displacement of Met481, with the movement of the sidechain sulfur atom by 4.87 Å. The G-loop site is formed by hydrophobic residues Trp446, Ile450, Val461, Tyr463, Met466, Met481, Trp499, five prolines (Pro462, Pro464, Pro482, Pro484, Pro485), and polar residues Thr453, Thr457, Gln474, and Cys483 (Figure 1d). The pocket has a volume of 400 ų and an fpocket [25] druggability score of 0.89 on a scale ranging from 0 to 1. Scores larger than 0.5 are deemed druggable, indicating that the unbiased ligand binding simulations identified a promising pocket for ligand development.

Virtual docking screen yields potential binders of the G-loop site

We performed a computational docking screen to identify ligands that bind and stabilize the predicted G-loop site in Src. Our goal was to find potential binders of the G-loop site that show improved binding affinity compared to PP1, which served as the reference compound in our docking screen. Using DOCK6.6, we virtually screened 230,000 compounds from the ZINC library of commercially available compounds (Figure 2a, Materials and Methods) [26]. Of the 230,000 molecules docked, 209,151 unique molecules were successfully processed to completion. The failed compounds were either duplicates in the ZINC library or did not successfully dock to the predicted allosteric site.

Figure 2. Docking screen identifies potential binders of the G-loop site that alter kinase activity.

(a) Overview of ligand docking strategy. We docked 230,000 compounds from the ZINC library and ranked them by DOCK Cartesian energy score (DCEsum). We selected the top 100,000 compounds and clustered them according to chemical similarity. Within clusters, we ranked compounds for DCE_sum-dependent qualities (DCE_sum and Lig_eff) or DCE_sum-independent footprint scores (FPS). The total score considers both DCE_sum and FPS_sum. The highest-ranking cluster heads yielded a final selection of compounds that were further narrowed for biochemical characterization. Histograms of (b) DCE_sum, (c) FPS_sum, (d) Total score, and (e) molecular weight from the top 100,000 (green) and the top 123 compounds (purple) for the allosteric site on Src kinase. (f) Effects of the top 49 ligands from the computational docking screen on Src kinase activity at 100 μM relative to DMSO control. Error bars represent ± SEM of triplicate experiments.

After docking, the compounds were ranked by DOCK Cartesian score (DCE_sum), the sum of the Van der Waals and electrostatic energies. The top 100,000 compounds were then subjected to post-processing (see Materials and Methods). The top 100,000 compounds had DCE_sum scores ranging from −70.18 kcal/mol to −30 kcal/mol, and the final selection had scores ranging from −70.18 to −31.97 kcal/mol (Figure 2b). The final selection showed enrichment for molecules with DCE_sum scores less than −40 kcal/mol, which improved compared to reference compound PP1 (−35 kcal/mol) (Figure 2b). DCE_sum scores in the final selection were distributed bimodally around −60 and −35 kcal/mol, which is likely due to the final selection of DCE_sum-dependent clusterheads (DCE_sum, Total Score, Lig_eff) and Footprint dependent scores (Figure 2b). The hits selected based on DCE_sum scores populated the lower energy peak, whereas compounds selected for footprint similarity did not necessarily lead to DCE_sum score enrichment.

We included the Ligand Efficiency (Lig_eff) metric, defined as the DCE_sum score divided by the molecular weight of the ligand, as a selection criterion to include smaller molecules that may be useful scaffolds for chemical editing [27]. The molecules with the 20 highest Lig_eff scores were added to the final selection. For the 100,000 ligands, Lig_eff scores ranged from 0.06 to 0.30 kcal/Da, while the final selection ranged from 0.08 to 0.30 kcal/Da. In the final selection, we observed enrichment for compounds with scores higher than 0.25 kcal/Da, an increase in comparison to reference compound PP1 (0.15 kcal/Da).

We employed footprint-based scoring to score compounds by binding mode similarity in comparison to the reference binder PP1 [28]. Footprint-based scoring has been shown to improve the ability of DOCK to identify the correct pose of known binding ligands [28]. The footprint plot shape for each compound was compared to the PP1 reference plot using Euclidian similarity. Compounds received a unitless footprint score with 0 indicating a binding footprint identical to PP1 and higher values indicating footprint differences. For the 100,000 ligands, the Van der Waals component of the Footprint score ranged from 0.98 to 16.43, while the final selection ranged from 1.01 to 7.04. Furthermore, the electrostatic Footprint scores ranged from 0.30 to 11.20 for the top 100,000 ligands and 0.30 to 10.68 for the final selection. The combined van der Waals and electrostatic footprint score (FPS_sum) ranged from 1.87 to 22.19 for the top 100,000 and from 1.87 to 17.09 for the final selection (Figure 2c).

The Total DOCK score is the sum of DCE_sum and FPS_sum. The Total scores for the top 100,000 ligands ranged from −55.86 to −11.76 kcal/mol, while the final selection ranged from −55.86 to −27.05 kcal/mol (Figure 2d). The final selection showed a bimodal distribution centered at −50 and −30 with enrichment in the range of less than −45 kcal/mol compared to the top 100,000, a group that scored better than PP1 (−30 kcal/mol) (Figure 2d). Notably, the bimodal distribution suggests that while compounds recapitulate the binding modalities of PP1 by footprint score alone, we could potentially identify more effective and specific G-loop site inhibitors with lower DOCK scores and alternative binding modalities.

The final selection can be compared to the top 100,000 using various molecular descriptors calculated by the MOE software [29]. This evaluation showed that the molecular weight distributions (Figure 2e) and rotatable bonds were very similar between the two groups, indicating no enrichment for either criterion. Of the docked compounds, 49 were selected for experimentation (SI Table 1).

Compounds from virtual screen alter Src kinase activity

Our virtual screen predicted binders of Src kinase independent of their effect on kinase activity; therefore, we screened the effects of the 49 top-scoring leads on Src kinase enzymatic activity in vitro. We tested the effects of these compounds in triplicate at 100 μM with Src wild-type kinase domain (Src wt KD). At 100 μM, four compounds inhibited Src kinase activity, and three compounds activated the kinase by more than 2.5 standard deviations (SD) (Figure 2f). We further validated the binding of these seven leads to Src wt KD by STD-NMR. We were not able to detect binding by STD-NMR for the activator compounds 2E, 4A, and 5D, and inhibitor compound 6F. Their dissociation rate constants may be outside of the detection limits for STD-NMR (k_off = 10 – 10⁵ s⁻¹, K_D = μM – mM). In addition, it is possible that these compounds bind to multiple sites (specific or non-specific) and result in an averaged and undetectable STD-NMR signal. We performed ¹H-¹⁵N HSQC chemical shift perturbation (CSP) experiments with compounds 2E, 4A and 5D and indeed found that the chemical shift perturbations were distributed across the kinase domain without a clear confirmation of the G-loop binding site (SI Figure 1). We observed positive binding via STD-NMR for three compounds: 1C, 1E, and 6C. Compound 1E is mitoxantrone, an FDA-approved type II DNA topoisomerase inhibitor widely used in cancer therapy [30, 31]. It has been shown that mitoxantrone can act as an ATP- and peptide-competitive kinase inhibitor [32, 33]. Given our focus on allosteric inhibitors, we excluded 1E from further investigation. We prioritized compound 1C for further characterization of the remaining two inhibitors due to its chemical similarity to PP1 and enhanced solubility.

Compound 1C is a commercially available racemic mixture of tert-butyl (3R, S)-3-(4-hydroxyphenyl) piperazine-1-carboxylate. Our docking results with the 3R stereoisomer revealed the hydroxy-phenyl group buried in the allosteric pocket, forming hydrophobic interactions with the side chains of residues Ile450, Thr453, Tyr463, and Met481 (Figure 3a). In addition, the piperazine core formed hydrophobic interactions with Thr457 and Pro482, and the tert-butyl group formed hydrophobic interactions with Met466 and Pro464.

Figure 3: Experimental validation of computational docking reveals compound 1C is a non-competitive inhibitor of Src.

(a) Simplified structure of 1C docked to the MD-derived Src structure with labeled binding site residues on helices-αF and -αG, and adjacent loops (b) Percent inhibition of kinase activity by 1C revealed inhibition of Src wt KD with an EC₅₀ of 61.4 ± 34.4 μM. (c) Kinase activity assay in the presence of increasing concentrations of ATP (0–800 μM) and 0 μM 1C (blue, circles), 50 μM 1C (purple, squares), and 100 μM 1C (green, triangles) (left). Src wt KD K_M for ATP and reaction V_max plotted against 1C concentration (right). (d) Kinase activity assay in the presence of increasing concentrations of substrate peptide (0–800 μM) and 0 μM 1C (blue, circles), 50 μM 1C (purple, squares), and 100 μM 1C (green, triangles) (left). Src wt KD K_M for substrate peptide and reaction V_max plotted against 1C concentration (right). All error bars represent ± SEM of triplicate experiments.

Compound 1C is a non-competitive Src inhibitor

Next, we employed a continuous coupled-kinase assay to evaluate the dose-response relationship between 1C and Src wt KD [34]. We found that 1C partially inhibited Src in a dose-dependent manner by approximately 40% at 150 μM and exhibited a half maximum effect (EC₅₀) at a concentration of 61.4 ± 34.4 μM (Figure 3b). We used NanoBRET to measure target engagement between 1C and Src in live cells [35–38]. Briefly, full-length Src kinase fused to a C-terminal NanoLuc Luciferase (Src-nLuc) was expressed in human embryonic kidney (HEK293T) cells and incubated with a cell-permeable fluorescent tracer that binds to the ATP-binding site. Target engagement between the ATP-competitive tracer and Src-nLuc results in the emission of a Bioluminescent Resonance Energy Transfer (BRET) signal. The addition of a test compound causes dose-dependent displacement of the tracer and decrease of the BRET signal. This technique has been successfully employed to quantify compound engagement for Abl kinase with ATP-competitive kinase inhibitors, as well as allosteric inhibitors that cause tracer displacement through altered conformational dynamics of the ATP-site [35, 38]. 1C displayed binding to full-length Src-nLuc with a cellular affinity (IC₅₀) of 1.31 ± 0.152 mM at equilibrium (SI Figure 2). To establish that 1C allosterically inhibits Src kinase activity and does not directly compete with ATP or substrate peptide, we determined its inhibitory mechanism of action. We investigated the effects of varying 1C concentrations (0 μM, 50 μM, and 100 μM) on K_M and V_max for ATP and substrate peptide (Rs1 peptide sequence: AEEEIYGEFAKKK) using the kinase assay previously described [34]. We found that 1C did not increase the apparent K_M for ATP or substrate peptide, suggesting that 1C does not inhibit the kinase by competing with either ligand (Figure 3c and d). As expected for an inhibitory compound, we observed a reduction in reaction V_max with increasing 1C concentrations. From these data, we can conclude that 1C behaves as a V-type non-competitive inhibitor of Src.

Compound 1C is specific to Src kinase

Next, we assessed whether Src kinase inhibition by 1C is due to non-specific binding or general interference with the biochemical and cellular assays. We used the Src family kinase Hck and the closely related Abl kinase as controls due to their sequence divergence at the predicted G-loop site (Figure 4a). We employed differential scanning fluorimetry (DSF) to test whether 1C alters the thermostability of these kinases as a read-out for ligand-induced protein aggregation, a leading mechanism for artifactual protein inhibition [39]. We found that 1C significantly increased the melting temperature for Src by 1.6 ± 0.09 °C but did not alter the melting temperatures for Abl or Hck (Figure 4b). These results indicate that inhibition of Src by 1C is due to ligand binding rather than artifactual.

Figure 4. 1C is selective for Src kinase.

(a) Sequence alignment of G-loop binding site residues for Src, Abl, and Hck. (b) Differential scanning fluorimetry thermal denaturation temperatures for Src, Abl, and Hck wt kinase domains in the absence and presence of 1C. Error bars represent ± SEM of triplicate experiments. (c) ¹H NMR-STD Amplification factors for 1C binding to Src, Abl, and Hck wt kinase domains. Amplification factors were calculated from the average of five hydrogen peaks originating from 1C. Data represent the average amplification factors ± SD. (d) Percent inhibition of kinase activity by 1C for Src, Abl, and Hck wt kinase domains. (e) Percent inhibition of kinase activity by 1C for Src wt, Src T453W, and Src M481K kinase domains. (d,e) Percent inhibition was calculated for each inhibitor concentration in comparison to kinase activity reactions containing no inhibitor. Error bars represent ± SEM of triplicate experiments. 0.1234 (ns), 0.0002 (***), <0.0001 (****).

Since our DSF results revealed selective thermal stabilization of Src over Abl and Hck, we employed STD NMR to further examine the binding selectivity of 1C. To compare the binding of 1C between Src, Abl, and Hck, we determined the average amplification factors for five protons originating from 1C. Consistent with our DSF data, we found that 1C binding to Abl and Hck was significantly reduced relative to Src (Figure 4c). The defective binding of 1C to Abl and Hck was further exemplified by the lack of their inhibition in the continuous coupled-kinase assay [34]. We found that 1C does not inhibit Abl or Hck kinase activity in vitro (Figure 4d). In addition, using the NanoBRET assay to measure intracellular target engagement (previously described), we found that 1C does not inhibit full-length Abl (Abl1 bound to an N-terminal NanoLuc Luciferase) with a projected cellular affinity (IC₅₀) greater than 100 mM (SI Figure 2). From these results, we can conclude that 1C is not a pan-assay interference compound (PAIN) but a specific Src kinase inhibitor.

Compound 1C binds to the G-loop site of Src kinase

Next, we set out to confirm the binding site of 1C on Src kinase domain. Since 1C exhibited selectivity for Src over Abl and Hck, we compared the sequence differences between these three kinases at the predicted G-loop site (Figure 4a). The most notable sequence difference is at Src Thr453, which is substituted with a tryptophan and methionine in Abl and Hck, respectively. Residue Thr453 is part of helix-αF, which is involved in allosteric modulation through interactions with Src’s catalytic regulatory spine [40] and allosteric network [22] (Figure 1d). Modeling mutation T453W into the MD-derived structure resulted in occlusion of the G-loop binding pocket. Therefore, we predicted that Src T453W would be uninhibited by 1C due to defective binding.

Src T453W exhibits more than a two-fold higher K_M for ATP relative to Src wt KD and increased thermal stability by approximately 3°C (SI Figure 3a and b). We have shown previously that the allosteric network is disrupted by mutation D454A, resulting in a >2-fold increase in K_M for ATP [22]. Therefore, it is not surprising that a stabilizing mutation in helix-αF in direct contact with the allosteric network would affect ATP binding. As seen for Abl and Hck, we found that 1C does not inhibit Src T453W kinase activity (Figure 4e).

We mutated a second G-loop binding site residue that was predicted to interact with 1C. Met481 is located within the loop region between helices-αG and -αH and was one of the primary residues found to move and expose the G-loop binding pocket in simulations. Our docking results showed hydrophobic interactions between 1C and Met481; therefore, we mutated Met481 to lysine to test whether a larger, charged residue would disrupt 1C binding and impair inhibition. As seen with G-loop site mutant T453W, M481K is also not inhibited by 1C (Figure 4e). Collectively, by introducing two mutations that reside on opposite interfaces of the G-loop binding pocket, we observed a loss of 1C inhibition.

Binding of 1C induces long-range conformational and dynamic changes in Src.

To understand how 1C binding to Src KD elicits its allosteric inhibitory effects, we employed ¹H-¹⁵N TROSY-HSQC to investigate conformational changes in the protein. Previously, we have shown by NMR that dasatinib binding to Src wt KD induces chemical shift perturbations (CSPs) across the entire kinase domain, including residues located on helix-αG [41]. This indicates allosteric coupling between the ATP-binding site and helix-αG, which has been previously established both experimentally and computationally [22, 41–43]. Since dasatinib is a type-I inhibitor that binds to Src with nanomolar affinity, we studied the effects of 1C in the presence of dasatinib to stabilize the kinase with occupation of the ATP-site.

Here, we identified long-range CSPs and resonance intensity perturbations (IPs) induced by 1C binding to Src·dasatinib. CSPs greater than 1 SD from the mean were defined as significant. IPs were calculated as the ratio of normalized absolute signal intensities for the two inhibitor-bound states (I_{dasatinib·1C} / I_dasatinib). The changes in intensity could be due to altered exchange with the solvent due to a static conformational change (e.g., a residue becomes more solvent exposed upon 1C binding), or due to the altered exchange between multiple conformations (e.g., faster interconversion between states upon 1C binding). We cannot distinguish between the two, but since the computational model of 1C bound to Src kinase does not indicate large structural changes of the protein, it is possible that the conformational dynamics of the protein change. Therefore, IP ratios greater than one can indicate increased amide proton exchange with the solvent or faster conformational interconversion, whereas IP ratios less than one can indicate decreased exchange or slower conformational interconversion. IP ratios greater than 1 SD from the mean ratio > 1 and ratios greater than 1 SD from the mean ratio < 1 were defined as significant.

The binding of 1C to Src•dasatinib causes CSPs and IPs in the G-loop binding site that propagate to distal regulatory elements in the N-lobe of the kinase (Figure 5a–d). We observed significant CSPs (>3 SD from the mean) for G-loop binding site residues Ile450 and Thr457 on helix-αF (Figure 5a and c). In addition, Glu486, located between helices-αG and -αH, also exhibited a significant CSP (>3 SD from the mean) (Figure 5a and c). Glu486 is located directly adjacent to the predicted G-loop binding site residues Pro482, Pro484, and Pro485, for which CSPs were unobservable due to lack of amide protons in proline residues. In addition, we observed that IP ratios decreased significantly for residues Ile450, Thr457, and Glu486, which can be expected for residues involved in ligand binding as they are hindered from the solvent (Figure 5b and d). Backbone amides for G-loop residues 479–481 all exhibited increased IP ratios, with binding site residues Tyr479 and Met481 exhibiting significant IPs (Figure 5b and d). This observation is consistent with our simulation data, showing that outward displacement of G-loop residues 479–481 forms the G-loop site. The CSPs and IPs for these predicted binding pocket residues are consistent with a binding event occurring in this region.

Figure 5. Binding of 1C to Src KD•dasatinib induces CSPs and dynamic changes in G-loop site and N-lobe regulatory sites.

¹H-¹⁵N NMR CSPs of Src•dasatinib induced by 1C binding were analyzed by (a) histogram and (c) CSP (Δδ(ppm)) structure mapping to the Src•dasatinib structure (PDB 3QLG). The yellow, orange, and red histogram dashed lines represent CSP magnitudes corresponding to 1, 2, and 3 standard deviations from the mean. Backbone amide resonance intensity ratios for Src•dasatinib bound to 1C relative to Src•dasatinib were analyzed by (b) histogram and (d) structure mapping to the Src•dasatinib structure (PDB 3QLG). The black dashed histogram line indicates the overall mean intensity ratio. Orange and red dashed histogram lines represent 1SD and 2SD greater than the mean intensity ratio >1. Green and blue dashed histogram lines indicate the 1SD and 2SD less than the mean intensity ratio <1.

The hinge region that connects the N- and C-terminal lobes (residues 339–345) and regulates inter-lobe motions also displayed significant CSPs and IPs. For example, Tyr340, which is directly adjacent to hinge residue Met341 that hydrogen bonds to dasatinib, displayed among the largest CSPs (> 3 SD) and IP ratio decreases upon binding 1C (Figure 5a and c). The significant CSPs and IPs within the hinge region support the notion that 1C binding allosterically propagates conformational changes from the kinase C-lobe to the N-lobe. The proximity of hinge-residues to dasatinib may also reflect alterations in the inhibitor binding site.

Significant CSPs in the N-lobe were predominantly observed for residues within or directly adjacent to structural regulatory elements, including the P-loop and helix-αC. The glycine-rich P-loop is a flexible and dynamic structure whose conformation has been linked to ATP-competitive drug selectivity and drug resistance [44, 45]. We observed significant CSPs for P-loop residues Leu273 and Cys277 (>2SD from the mean), which reflect conformational changes in the P-loop that differ from the dasatinib-bound structure (Figure 5a and c). Cys277 also exhibited a significant IP ratio decrease (Figure 5b and c). Interestingly, Cys277 is critical for redox-dependent kinase activation and has been targeted by several covalent Src inhibitors, highlighting its environmental sensitivity [46–48].

Helix-αC is one of the three major structural elements that contribute to the transition between active and inactive conformations. Outward displacement of helix-αC breaks the catalytically important salt bridge formed between conserved residues Glu310 and Lys295 [49]. Notably, Glu310 in helix-αC and the catalytic lysine Lys295 both exhibited IP ratios that increased significantly (≥ 2SD > mean ratio > 1) (Figure 5b and d). Given that this salt bridge is critical for kinase activation, it is possible that the increased resonance intensity for these residues reflects faster amide proton exchange with the solvent or multi-conformational interconversion that result in the weakening of their interactions. In addition, helix-αC residues Leu308, Met314, Leu317, and residues within the flexible loop and beta-sheet following it (His319, Lys321, Gln324, Leu325, Tyr326) all exhibited significant CSPs (Figure 5a and c). Residues located at the beginning of helix-αC (residues 308–311) displayed IP ratios that increased significantly, whereas residues at the end of helix-αC (residues 312–317) exhibited the opposite (Figure 5b and d). Together, the CSPs and IPs observed for helix-αC may explain the decrease in kinase activity upon binding to 1C.

Although the overall magnitude of CSPs observed is small (< 0.1 ppm), this could be attributed to an unbound fraction of Src resulting in weaker CSPs. Alternatively, the model of 1C bound suggests small conformational changes to Src compared to ATP-competitive inhibitors such as dasatinib. Previously, we found CSPs induced by dasatinib were deemed significant if greater than 0.17 ppm (> 1 SD from the mean) [41], whereas CSPs induced by 1C were deemed significant if greater than 0.012 ppm (> 1 SD from the mean). This difference highlights the greater effect that dasatinib has on Src’s conformational change compared to 1C. Since dasatinib binding to the ATP-site results in long-range protein stabilization, this is also expected to restrict changes in the protein upon binding to 1C and can contribute to smaller CSPs.

Compound 1C is a conformation-selective Src inhibitor

Given that Src was in the active conformation when the G-loop site was revealed in simulations, we sought to determine whether 1C binding to the G-loop site is conformation-selective. We used two conformation-selective inhibitors to evaluate 1C binding by DSF and STD-NMR: (1) dasatinib which binds to the “DFG-Asp-in/αC-in” active conformation [50] and (2) DasDFGOII, based on the chemical scaffold of dasatinib, which binds to the “DFG-Asp-out, αC-in” inactive conformation [51]. Although these inhibitors are both categorized as “αC-in,” DasDFGOII differs from dasatinib by forming hydrogen bonds with Glu310 and Asp404, a common feature of Type II inhibitors [51]. In a previous study, we compared the effects of dasatinib and DasDFGOII on Src KD backbone conformation and dynamics [41]. We found that CPSs that span from the N-terminus to helix-αD (residues 251–350) were similar in identity and magnitude for dasatinib and DasDFGOII. However, differential CSP patterns arose in the C-lobe, specifically in helices-αEF, αF, and αG [41]. Given that the binding of dasatinib and DasDFGOII induce different conformational changes for C-lobe elements associated with the G-loop site, we evaluated 1C binding by STD-NMR as a read-out for G-loop site formation.

We performed DSF thermal melt analysis to determine whether 1C can further stabilize the Src•dasatinib and Src•DasDFGOII complexes. 1C was added in 100-fold excess to compensate for its weak binding affinity compared to the nanomolar binding affinities for the ATP-competitive inhibitors dasatinib and DasDFGOII [51]. We found that 1C significantly stabilizes the Src•dasatinib complex by 1.0 ± 0.14 °C (Figure 6a). In contrast, we found that 1C significantly destabilizes the Src•DasDFGOII complex by 1.9 ± 0.41 °C (Figure 6a). Jointly, these data illustrate that 1C selectively binds and stabilizes the active conformation of the kinase.

Figure 6. 1C exhibits conformation selectivity for Src in the active conformation.

(a) Differential scanning fluorimetry thermal denaturation analysis to determine the change in melting temperature for Src wt KD in the presence and absence of 1C and ATP competitive inhibitors dasatinib and DasDFGOII. (b) Chemical structure of compound 1C with annotated protons used for amplification factor quantification. (c) Amplification factors for 1C protons in the presence of Src, Src•dasatinib, and Src•DASDFGOII. (d) Normalized percent inhibition for 1C dose-response at 100 μM for Src wt KD and Src wt 3D. 0.1234 (ns), 0.0332 (*), 0.0021 (**).

To extend this analysis, we evaluated 1C binding by STD-NMR and found that the relative affinity for 1C improved when Src is bound to dasatinib and in the active conformation, resulting in >2-fold increase in the average amplification factor for all peaks measured (Figure 6b and c). In contrast, the relative for 1C binding to Src•DasDFGOII was significantly impaired, rendering only 2 out of 7 measurable proton peaks in the STD spectrum, where 4 peaks were not present, and 1 was at the noise level (Figure 6c). Together, the presence of 1C proton peaks H_g,h and H_j and lack of benzyl region peaks may indicate 1C binding outside of the G-loop site in the presence of DasDFGOII. Together, these data suggest that dasatinib and 1C bind cooperatively, whereby 1C increased the thermostability of the Src•dasatinib complex and dasatinib increased the relative affinity for 1C compared to apo Src (Figure 6a and c). Conversely, our results reflect that DasDFGOII induces changes that do not favor G-loop site formation. Anti-cooperative binding between DasDFGOII and 1C was displayed by a significant decrease in the relative affinity for 1C in the presence of DasDFGOII (Figure 6c). Destabilization of the Src•DasDFGOII complex in the presence of 1C further supports anti-cooperative binding (Figure 6a). These observations are consistent with our simulation data that revealed the G-loop binding pocket when the kinase was in the active conformation.

We next determined the extent of 1C inhibition in an alternative Src construct containing its regulatory SH3 and SH2 domains (Src wt 3D). The isolated kinase domain is constitutively active in solution, whereas SH3-SH2 regulatory domains stabilize the autoinhibited inactive kinase conformation [52]. We found that 1C inhibits Src wt KD by approximately 5% more than Src 3D at 100 μM (Figure 6d). This result has two implications: first, it supports our STD-NMR results showing that 1C binds preferentially to the active kinase conformation, and second, the lower amplitude of inhibition for Src 3D can be explained by Src 3D already being partially autoinhibited as expected.

G-loop site ligand selectively modulates binding kinetics of ATP-site ligands.

We have shown above that conformation-selective ATP-competitive inhibitors affect the binding of 1C. Next, we speculated that this communication between the ATP-binding site and the G-loop site should work reversibly. Therefore, we aimed to determine how the G-loop site inhibitor 1C affects the binding kinetics of conformation-selective inhibitors binding to the ATP binding site. We determined the binding kinetics for type-I and -II inhibitors dasatinib and DasDFGOII to Src wt KD in the presence and absence of 1C using our previously described stopped-flow assay [53]. The binding kinetics of the type-II inhibitor DasDFGOII resembles that of other type-II inhibitors: at low inhibitor concentrations, the observed rate constant increases linearly with inhibitor concentration. However, at DasDFGOII concentrations above 6.25 μM, a conformational change in the Asp-Phe-Gly motif (DFG-flip) of Src becomes rate-limiting, and the observed rate constant of inhibitor binding gradually levels out [54]. The conformational transition from DFG_in to DFG_out becomes rate-limiting due to the thermodynamic penalty associated with adopting this conformation in Src kinase [53]. Our previously published kinetic model for this process allows us to derive the rate constant for DasDFGOII binding (k_on) and the rate constant for the DFG-flip (k₊₁*) [54].

1C decreased the DasDFGOII forward binding rate constant (k_on) to Src wt KD by approximately 38 % (0.245 ± 0.005 s⁻¹ μM, and 0.151 ± 0.007 s⁻¹ μM) (Figure 7a). We found that the rate constant for the DFG-flip (k₊₁*) is approximately 2.4-fold slower in the presence of 1C compared to without 1C (0.99 ± 0.06 s⁻¹ and 2.4 ± 0.18 s⁻¹, respectively). As a control, we tested the effect of 1C on the binding kinetics of the type I inhibitor dasatinib, which are not rate limited by the DFG-flip. When compared to Src wt KD without 1C, we observed that 1C does not affect the binding rate constant (k_on) for dasatinib (8.90 ± 0.619 s⁻¹ μM and 8.93 ± 0.204 s⁻¹ μM, respectively) (Figure 7b). Therefore, differences in DasDFGOII binding kinetics are indeed attributed to changes in the rate of the DFG-flip in the presence of 1C.

Figure 7. The G-loop site allosterically modulates ATP-competitive binding kinetics.

Kinetics of (a) type-II inhibitor DasDFGOII and (b) type I inhibitor dasatinib binding to Src (blue, circles), Src + 1C (purple, squares), and Src T453W (green, triangles). Experiments were done in triplicate and the average observed rate constants (k_obs) ± SEM were plotted against increasing inhibitor concentrations and fit to a straight line in GraphPad Prism (version 9) to derive k_on (left). Association rates derived from the slopes of straight lines fitted in (a) and (b) (right). 0.1234 (ns), 0.0002 (***).

Next, we wanted to determine if the T453W substitution on helix-αF impacts the binding rates for ATP-competitive inhibitors dasatinib and DasDFGOII. In comparison to Src wt, Src T453W exhibited a slower forward DasDFGOII binding rate constant by almost 41% (0.245 ± 0.005 s⁻¹ μM and 0.145 ± 0.005 s⁻¹ μM, respectively), a similar extent to the effect of 1C (Figure 7a). We found that Src T453W exhibited a slower DFG-flip rate constant (k₊₁*) relative to Src wt (1.6 ± 0.28 s⁻¹ μM and 2.4 ± 0.18 s⁻¹, respectively). The higher thermal stability of Src T453W relative Src wt KD•1C (ΔT_m of 3 °C and 1.6 °C, respectively) likely compensates for the thermodynamic penalty of the DFG-flip and may account for the subtle difference in flip rate. In comparison to Src wt KD, Src T453W did not affect dasatinib binding kinetics (8.90 ± 0.619 s⁻¹ μM and 8.53 ± 0.373 s⁻¹ μM, respectively) (Figure 7b). Cumulatively, we observed that 1C and the helix-αF mutation T453W both affect ATP-competitive ligand binding kinetics and result in slower type-II binding rates. These kinetics data suggest that occupying the G-loop site either by mutation or a ligand reduces the DFG-flip rate and thereby shifts the conformation equilibrium towards the active state, causing slower type-II binding rates.

Discussion

Allosteric regulatory sites present alternative means for targeted kinase inhibition and circumventing challenges associated with ATP-competitive inhibitors, such as off-target inhibition and resistance. However, allosteric sites that can be targeted by allosteric modulators are often challenging to identify or predict in the absence of stabilizing ligands. In the present study, we reexamined our previously published unbiased ligand MD simulations [18] to predict a novel cryptic allosteric binding site in Src kinase, which we term the G-loop site. Here, we have reported observations from in-depth analyses of the predicted G-loop binding pocket, virtual ligand screening against the MD-derived structure of the predicted site, and biochemical and biophysical assays characterizing a potential lead compound. We demonstrated the feasibility of this integrative workflow by identifying compound 1C as a small-molecule ligand that binds to the novel G-loop site by virtual ligand docking, site-directed mutagenesis, and NMR. We showed that 1C is a V-type non-competitive inhibitor that partially inhibits Src kinase in biochemical and cellular assays and exhibits high specificity for Src over the closely related Abl and Hck kinases. While compound 1C only partially inhibits Src activity by roughly 40% at 150 μM, this is consistent with findings for the myristate-pocket allosteric inhibitor of Abl kinase GNF-2, which partially reduces the activity of SH3-SH2 containing Abl constructs by up to 60–80%, albeit at much lower inhibitor concentration [54]. Further medicinal chemistry derivatization and ADME improvements of GNF-2 yielded GNF-5. Not only does GNF-5 exhibit excellent specificity for Abl, but when combined with ATP-competitive inhibitor imatinib, dual inhibition can suppress and overcome resistance mutations [55]. Future medicinal chemistry efforts beyond the scope of this study could analogously improve the inhibitory capacity of 1C to create higher-affinity G-loop site allosteric inhibitors.

The concentration-dependent reduction in reaction V_max without alterations in K_M (Figure 3c and d) led us to conclude that 1C inhibits Src without competing for ATP or substrate peptide binding. A potential inhibitory mechanism was revealed by our ¹H-¹⁵N TROSY NMR data, whereby significant increases in amide signal intensities for Lys295 and Glu310 were observed (Figure 5b and d). Since salt bridge formation between Lys295 and Glu310 is critical for kinase activity, the increased signal intensities may reflect a weakening of this salt bridge through increased proton exchange rates between backbone amides and the solvent or fast interconversion between multiple conformational states. Additionally, STD-NMR revealed that the relative affinity for 1C in the presence of dasatinib was enhanced, whereas defective binding of 1C was observed in the presence of DasDFGOII. The conformational selectivity of 1C was further demonstrated by slower type-II binding kinetics observed for DasDFGOII binding to Src wt KD bound to 1C. We attribute this to slower DFG-flip rates observed in comparison to Src wt KD. The similar effects observed for Src T453W suggest that occupation of the G-loop site by a ligand or mutation causes a shift in the conformational equilibrium toward the active state by increasing the thermodynamic penalty associated with adopting the DFG_out conformation. Modulation of the thermodynamic penalty of the DFG_out conformation by the G-loop site is supported by several patient-derived imatinib resistance mutations located in the G-loop binding site region in the closely related Abl kinase [56]. Abl•imatinib resistance mutations have been found at c-Src residue positions 473, 476, 482, and 461, all of which are in or directly adjacent to the G-loop site, which may reflect evolutionary divergence at this site. In simulations, Abl mutation E473V (c-Src numbering) was found to increase the thermodynamic penalty of the DFG-flip by 6 kcal/mol [57].

The allosteric sites revealed in simulations (PIF, MYR, and novel G-loop) were predicted using two different inhibitors, PP1 and dasatinib, indicating that the prediction is robust and is not highly sensitive to the chemical probe used in the simulation screening. The use of small molecules or chemical fragments as probes that screen the dynamic surface of proteins for emerging binding pockets resembles an induced-fit process, where initial binding of the probe allows for further structural rearrangements in the emerging pocket to accommodate the ligand. Based on these results, studying the ligand binding process by unbiased ligand binding simulations with a panel of small molecule probes could be used to screen for novel allosteric regulatory sites on proteins.

Materials and Methods

Binding site mapping and virtual docking screen

A representative snapshot from the unbiased MD simulation where PP1 occupied the G-loop binding site and did not dissociate for the length of the simulation was extracted as a PDB file and employed for the docking screen (SI Movie 1). The binding site was evaluated using the fpocket [25] software suite by physicochemical characteristics and druggability score. The ligand and solvent molecules were removed from this structure, and UCSF Chimera [58] was used to add hydrogen atoms and charges to the kinase using the ff99sb force field [59]. The G-loop binding site was prepared for DOCK using a published three-step method [60–62] using the suggested parameter values from the DOCK 6.6 manual [63].

The Aldrich CPR and Sigma Aldrich (Building Blocks) subsets from ZINC 12 were downloaded (http://zinc.docking.org) for the screen. These ~240,000 ligands were filtered to obtain those with a charge between −2 and +2, less than 15 rotatable bonds, and molecular weight greater than 180 g/mol, leaving 230,000 ligands. The filtered set of ligands was then divided into groups of 10,000, yielding 24 subsets of ligands docked to Src using the DOCK6.6 suggested parameters for flexible ligand docking [63, 64]. The single lowest grid energy pose was retained. The docked output was minimized in Cartesian space to prepare for footprint scoring using DOCK6.6 and a simplex coefficient of 5.0 was employed. The minimized results were re-scored using a Footprint descriptor score in DOCK6.6, as previously described [28]. Here, the standard Euclidean similarity metric was compared to the reference, which was PP1 bound to the G-loop site.

The top 100,000 molecules (ranked by DOCK energy score) were then clustered using the program MOE [65]. Descriptors such as the Lipinski rules were also calculated. The top-scoring ligand in each cluster (clusterhead) was taken onto the next step to increase diversity. The clusterheads were then ranked by 6 metrics: (i) DOCK Cartesian energy (DCE_Sum), (ii) Ligand efficiency (Lig_eff). (iii) FPS_vdw, (iv) FPS_es, (v) FPS_sum, and (vi) Total score (sum of DCE_sum and FPS_sum) (Figure 2a). The top 50 compounds obtained by each method were chosen. After three-dimensional inspection and accounting for overlap between the lists, 49 compounds were designated candidates for experimental screening.

Protein expression and purification

The kinase domain constructs of human c-Abl (residues 229–511), chicken c-Src (251–533), murine Lck (residues 227–509), and human Hck (166–445) were cloned into a modified pET-28a vector (Novagen) containing a tobacco etch virus (TEV) protease cleavable N-terminal hexahistidine tag. Proteins were co-expressed with full-length YopH phosphatase from Yersinia pseudotuberculosis [66, 67] and GroEL/Trigger factor chaperone, as previously described [53, 68]. Cultures were grown at 37°C to an OD_600nm between 0.4 and 0.8 and cooled for 1 hour with shaking at 16°C prior to induction with 1mM IPTG for 16 hours at 16°C. Proteins were purified as described previously for c-Abl and c-Src [68]. Proteins were eluted in 20 mM Tris (pH 8.0), 100 mM NaCl, 5% glycerol, 1 mM DTT, concentrated using 10K MWCO concentrator tubes (Corning), snap-frozen in liquid nitrogen, and then stored at −80°C until use.

Isotopically labeled ¹⁵N Src KD GG253 was employed for ¹H-¹⁵N TROSY-HSQC NMR experiments, which differs from Src wt KD only in the N-terminal tail through deletion of residues 249–253. Isotopically labeled ¹⁵N Src wt KD was employed for ¹H-¹⁵N HSQC NMR experiments. The same plasmid and expression system as non-labeled Src wt protein was employed, but instead, bacterial cells were grown in M9 minimal media containing 1 g/L ¹⁵NH₄Cl as a nitrogen source and 4 g/L glucose as a carbon source. The purified ¹⁵N Src KD GG253 protein was concentrated to 200–350 μM and buffer exchanged into NMR buffer (50 mM MES, pH 6.4, 300 mM NaCl, 10 mM MgCl₂, 1 mM TCEP) at 10 °C. Protein was snap-frozen in liquid nitrogen and stored at −80°C.

Mutations in the chicken c-Src kinase domain were introduced by Quikchange site-directed mutagenesis (Agilent) and verified by DNA sequencing.

Kinase assays

Kinase activity was measured using a continuous spectrophotometric kinase assay [34] with Rs1 substrate peptide (Rs1 sequence: AEEEIYGEFAKKK) [69]. Reactions (75 uL) contained kinase buffer (100 mM Tris (pH 7.5), 10mM MgCl2), 0.5 mM ATP, 0.5 mM substrate peptide, 1 mM phosphoenolpyruvate, 0.6 mg/mL NADH, 75 U/mL pyruvate kinase, and 105 U/mL lactate dehydrogenase. The decrease in absorption at 340 nm was measured between 30 and 45 minutes at 30°C in a Synergy Neo2 multimode microplate reader (BioTek), as described previously for imatinib [53]. Background kinase activity was determined and subtracted by measuring kinase activity without substrate peptide. For the initial screening assay, 300 μM peptide and 200 μM ATP were employed. Screened compounds were considered hits if they changed the initial velocity by more than 2.5 SD of the DMSO control. For IC₅₀ experiments, compound titrations were prepared in 100% DMSO and 3.75 μL was added to each reaction, yielding a 5% final DMSO concentration. For K_M experiments, Src wt KD was incubated with 0 μM, 50 μM, or 100 μM compound 1C, yielding a final 5% DMSO concentration for all samples. ATP or substrate peptide titrations ranged from 0 μM to 800 μM (final concentration). The initial velocities were plotted and fit to the Michaelis-Menten equation in GraphPad Prism (version 9). Concentrations of Src wt KD used for these assays was 33 nM.

Saturation transfer difference NMR

Saturation Transfer Difference NMR (STD-NMR) was performed as described in [70]. Proteins were buffer exchanged into 20 mM Tris pH 8.0, 250 mM NaCl, 10 mM MgCl₂. Samples contained 40 μM kinase, 500 μM compound, and 100 μM ATP-competitive inhibitor (when present). A Bruker 850 MHz instrument was employed, and time was set at 2.5 s for screening with the on-resonance frequency set at 0.57 ppm. Amplification factors (A_STD) were quantified for 1C protons using equation 1.1.

$$A_{STD}=\frac{I_{STD}}{I_0}\times \frac{\left[L\right]}{\left[P\right]}$$ (Eq. 1.1)

Where I is the amplitude of the ¹H peak, [L] is ligand concentration, and [P] is kinase concentration. For screening and kinase selectivity experiments, A_STD for 1C was calculated for 1C protons H_a1, H_a2, H_i1, H_i2, and H_j. For conformation selectivity experiments, A_STD for 1C was calculated for 1C protons H_a1, H_a2, H_c,d, H_g,h, H_i1, H_i2, and H_j.

Two-dimensional NMR

¹H-¹⁵N TROSY-HSQC NMR samples contained 200 μM Src GG253 in 50 mM MES (pH 6.4), 300 mM NaCl, 1 mM TCEP. Dasatinib was added in 2.5-fold excess, and 1C was added in 5-fold excess (final DMSO-d₆- < 5%). Deuterium oxide was added to NMR samples at a final concentration of 10% for signal locking. All ¹H-¹⁵N TROSY-HSQC spectra were obtained at 25 °C on a Bruker Avance III HD spectrometer operating at a ¹H frequency of 850 MHz equipped with a Triple Resonance (TCI) 13C-enhanced 5 mm cryogenic probe using the Bruker implemented pulse sequence trosyf3gpphsi19.2 [71, 72]. All spectra were acquired with the following parameters: Scans= 64, Size of FID (F2:H/F1:15N) = 2048/128, Spectral Width (F2:H/F1:15N) = 15.9791 ppm/ 36 ppm. The spectra were processed using the software Bruker topspin3.6.2. The backbone assignments for Src wt KD + Dasatinib were transferred from published Src wt + Dasatinib assignments in BMRB [73]. The chemical shift differences and intensity differences were analyzed using the software NMRFam Sparky [74]. The significance of the chemical shift difference and intensity difference was determined using their respective standard deviation values and graphed by GraphPad Prism version 8.0.2.

¹H-¹⁵N HSQC NMR samples contained 200 μM Src wt KD in 50 mM MES (pH 6.4), 250 mM NaCl, and 10 mM DTT. Dasatinib was added in 1.5-fold excess and compounds (2E, 4A, and 5D) were added in 2-fold excess (final DMSO-d₆ = 2.5%). The ¹H-¹⁵N HSQC spectra were obtained at 25 °C using a Bruker 700 MHz NMR instrument equipped with an Inverse Triple Resonance (TXI) 5 mm cryogenic probe. The Bruker implemented pulse sequence hsqcfpf3gpphwg was used. All spectra were acquired with the following parameters: Scans = 32, Size of FID (F2:H/F1:15N) = 2048/128, Spectral Width (F2:H/F1:15N) = 15.9408 ppm/ 36 ppm. The spectra were processed using the software Bruker topspin. The chemical shift differences were analyzed using the CCPN software suite.

Differential scanning fluorimetry thermal shift assay

Protein thermal stability and protein-ligand interactions were determined experimentally by a differential scanning fluorimetry thermal shift assay. In this assay, 2 μM kinase was incubated with DMSO or drug at 10X the expected K_D for 1 hour at 4°C. After incubation, 5X SYPRO Orange dye was added to each sample, and plates were heated for 90 minutes in 1.0-degree increments from 5°C to 95°C in a StepOne Real-Time PCR Machine (Thermo Fisher Scientific). Melting curves were normalized to obtain relative fluorescence values between 0 and 1. Normalized fluorescent melt curves were fit to a Boltzmann sigmoidal curve in GraphPad Prism version 9 to obtain V₅₀ (Y=Bottom+(Top-Bottom)/(1+exp((V50-X)/Slope))). Statistical significance was determined using a One-Way ANOVA with correction for multiple comparisons using Dunnett’s T3 multiple comparisons test in GraphPad version 9 (0.1234 (ns), 0.0332 (*), 0.0021 (**), 0.0002 (***), <0.0001 (****).

Stopped-flow kinetics

The binding kinetics of ATP-competitive inhibitors dasatinib and DasDFGOII were monitored by a stopped-flow spectrofluorimetric ligand-binding assay at 10° C, as described previously [53]. A stopped-flow system was employed (SX20 series, Applied Photophysics Ltd.) to monitor changes in intrinsic tryptophan fluorescence of the protein over time. Protein was mixed with buffer containing increasing concentrations at a 1:1 ratio, comprising a total volume of 200 μL. Pseudo-first order conditions were obtained by using at least a 10-fold molar excess of inhibitor. For each inhibitor concentration, there were at least three replicate measurements made, and replicate traces were averaged prior to fitting. The decay in intrinsic protein tryptophan fluorescence upon drug binding was fit to a single exponential equation with a sloping baseline using Pro-Data Viewer (Applied Photophysics Ltd.) to derive the observed rate constant (k_obs) [54]. Each experiment was repeated in triplicate, and the average k_obs values ± SEM were plotted against drug concentration in GraphPad Prism version 9. The k_obs values ± SEM were fit to a straight line to derive association rates based upon the relationship: k_obs = k_on [drug] + k_off. Statistical significance was determined using a One-Way ANOVA with correction for multiple comparisons using Dunnett’s T3 multiple comparisons test in GraphPad version 9 (0.1234 (ns), 0.0332 (*), 0.0021 (**), 0.0002 (***), <0.0001 (****).

NanoBRET drug-target engagement assays

Bioluminescence resonance energy transfer (BRET) can be used as a proximity-based measure of drug binding to kinase targets in live HEK293T cells [37]. BRET is observed between a NanoLuciferase (nLuc) tag on the full-length protein kinase and a tracer molecule (a BODIPY fluorophore attached to an ATP-competitive inhibitor scaffold). A test compound of interest binds to the protein kinase, displaces the tracer, and reduces BRET in a dose-dependent manner. The full-length Src and Abl NanoBRET constructs were a kind gift from the Promega Corporation. Standard NanoBRET manufacturer’s protocol (Promega) was followed, with slight modifications as listed below. Briefly, Src and Abl open-reading frames were cloned in frame with a C-terminal nLuc and N-terminal nLuc, respectively and used to transfect HEK293T cells at a density of 2×10⁵ cells/mL for twenty hours in white 384-well tissue culture plates (Corning #3570). Src and Abl-transfected cells were then incubated with BRET Kinase Tracer K4 (Promega) at their Tracer IC₅₀s of 80 nM and 40 nM respectively and serially diluted 1C in a mixture of 50% OptiMEM media without phenol red and 50% DMSO. The system was allowed to equilibrate for two hours at 37°C and 5% CO₂. BRET was measured upon adding NanoGlo Substrate and Extracellular Inhibitor (Promega) in a Biotek Synergy Neo2 plate reader at luminescence wavelengths of 450 nm and 610 nm. The BRET ratios (610 nm/450 nm) were graphed using GraphPad Prism version 9, and curves were fit to a four-parameter equation (Y=Bottom + (Top-Bottom)/(1+(IC50/X)^HillSlope)).

Abbreviations

FDA

Food and Drug Administration

PTK

protein tyrosine kinase

TROSY

transverse relaxation-optimized spectroscopy

HSQC

heteronuclear single quantum correlation

CSP

chemical shift perturbation

IP

intensity perturbation

STD-NMR

saturation transfer difference – nuclear magnetic resonance

KD

kinase domain

Src 3D

Src kinase containing the Src homology 3, 2 and 1 domains

MD

molecular dynamics

SD

standard deviation

SEM

standard error of the mean

DSF

differential scanning fluorimetry

ADME

Adsorption distribution metabolism excretion