Linking ATP and allosteric sites to achieve superadditive binding with bivalent EGFR kinase inhibitors

Bivalent molecules consisting of groups connected through bridging linkers often exhibit strong target binding and unique biological effects. However, developing bivalent inhibitors with the desired activity is challenging due to the dual motif architecture of these molecules and the variability that can be introduced through differing linker structures and geometries. We report a set of alternatively linked bivalent EGFR inhibitors that simultaneously occupy the ATP substrate and allosteric pockets. Crystal structures show that initial and redesigned linkers bridging a trisubstituted imidazole ATP-site inhibitor and dibenzodiazepinone allosteric-site inhibitor proved successful in spanning these sites. The reengineered linker yielded a compound that exhibited significantly higher potency (~60 pM) against the drug-resistant EGFR L858R/T790M and L858R/T790M/C797S, which was superadditive as compared with the parent molecules. The enhanced potency is attributed to factors stemming from the linker connection to the allosteric-site group and informs strategies to engineer linkers in bivalent agent design.


Introduction
Molecules that simultaneously bind to distinct sites within biological targets are increasingly sought after in drug development.This binding strategy is often accomplished through bivalent (or heterobifunctional) compounds, which comprise dual functional motifs connected by a covalent linker. 1,28][29][30] Despite the simple premise, diverse efforts spanning several decades have shown that superadditivity is scarce and the majority of cases fall short of achieving the expected improvement in target a nity. 20,21,28Additionally, case studies have offered suggestions for ideal binding properties of optimally linked compound, 20,21,28,30 but little is known regarding a general structural-based strategy for swiftly maximizing fragment linkers.
5][36] The kinase domain of the epidermal growth factor receptor (EGFR) is an established drug target in non-small cell lung cancer (NSCLC) where oncogenic mutations often predict clinical responsiveness to treatment with certain TKIs. 37Indeed, clinically-effective TKIs are often selective for EGFR activating mutations L858R (LR) and exon19del, 38 as well as drug resistant T790M (TM) gatekeeper and C797S (CS) mutants.
Promising pre-clinical results have been seen where combinations of ATP and allosteric inhibitors show synergistic tumor regression in vivo in addition to delayed acquired drug resistance. 39,40Importantly, the EGFR allosteric inhibitor binding site is adjacent to the ATP pocket and cobinding of structurallycompatible inhibitors within both sites 41,42 has been shown to enable structural changes consistent with biochemical experiments where these two inhibitor types exhibit cooperative binding. 42The unique effects enabled by combinations of ATP and allosteric inhibitors, as well as the structural proximity of their binding sites, have led to the recent development of ATP-allosteric bivalent molecules that simultaneously occupy these sites. 43,44 this study, we synthesized a novel series of bivalent EGFR kinase inhibitors that simultaneously bind the ATP and allosteric sites and differ with respect to the structure of the site-bridging linker.Strikingly, we nd that distinctly linked compounds exhibit considerable differences in potency where one linker exhibits superadditivity, and another is virtually inactive.Structural characterization indicates that the linker structure induces conformational differences and intermolecular interactions that provide a unique side-by-side comparison of functionally divergent linking strategies.Cocrystal structures and molecular dynamics simulations of these bivalent inhibitors enable the dissection of the speci c properties of the linker that afford strong binding, which informs novel design strategies.

Results
Due to the structural proximity of the allosteric and ATP (orthosteric) sites within the EGFR kinase, we sought to explore alternatively linked bivalent compounds that span these pockets.Starting motifs were selected and derived from established ATP-competitive inhibitors based on trisubstituted imidazole molecules (5-7), [45][46][47][48][49] and the mutant-selective allosteric 5,10-dihydro-11H-dibenzo[b,e] [1,4]diazepin-11one inhibitors 8 and 9 (herein denoted as "benzo" for simplicity). 50We synthesized a set of bivalent ATPallosteric inhibitors bridged by an N-linked methylene (1) and C-linked amide (2-4) (Scheme 1).To combine fragments for the N-linked derivative 1 we utilized a cross-coupling focused reaction route.The motif of the allosteric site was thereby assembled by slight adjustments of previously described conditions for derivatives of the allosteric inhibitor 8. 50 The subsequent Miyaura borylation afforded the corresponding boronic acid pinacol ester, which was Suzuki coupled with the imidazole core of the orthosteric scaffold.Bromination and Suzuki coupling with the hinge-binding motif yielded 1 after acidic deprotection (Scheme 2).For the synthesis of the allosteric motif of the C-linked series we applied a Buchwald-Hartwig amination for coupling of methyl anthranilate with 3-bromo methyl anthranilate.The product was re uxed in acetic acid to obtain the methyl-dibenzodiazepine-9-carboxylate by means of an intramolecular aminolysis.Saponi cation of the remaining ester and amide coupling of thereof resulting carboxylic acid with amines of corresponding orthosteric motifs yielded C-linked derivatives 2, 3 and 4 after deprotection (Scheme 3, Supplementary Scheme S2).
We rst sought to understand the degree to which these alternatively linked motifs in uence the ability to inhibit recombinant EGFR kinase activity.We carried out biochemical IC 50 value determination using HTRF-based activity assays with puri ed EGFR kinase domains (Table 1, Fig. 1).Strikingly, the N-linked 1 was observed to be limitedly potent against WT and mutant EGFR with IC 50 values ≥ 1 µM while the Clinked inhibitors 2-3 show substantially lower IC 50 values of 1.2-1.5 nM for LR and 51-64 pM for LRTM and LRTMCS (Table 1, Fig. 1).The C797-targeting irreversible C-linked analogue 4 was slightly less potent as a reversible inhibitor of LRTMCS, and additional time-dependent activity measurements showed that this molecule was most effective against LR (Table 2).To put these biochemical IC 50 values into proper context, we next compared them to structurally related ATP-and allosteric-site analogues 5-9.The ATPsite imidazole motifs 5,6 and the original allosteric inhibitors 8,9 inhibit LRTM and LRTMCS at IC 50 values ≥ 6-10 µM and ~ 39-59 nM, respectively, indicating that the C-linked bivalent molecules are 10 3to-10 6 -fold more potent over the parent motifs.The matched covalent analogue 4 inhibits LRTMCS reversibly with an IC 50 value 100-fold better than the orthosteric-only 7, showcasing the additional reversible binding gained by interactions in the allosteric pocket in this covalent scaffold.Seeing as how the allosteric motif and linker in 2-4 is different from the N-linked 8,9, we synthesized matching C-linked compounds of the benzo scaffolds and assayed them against LRTM (Supplementary Schemes S1 & S3 and Supplementary Figure S1).To our surprise, 10, which is the closest structural analogue to 2-4, is virtually inactive and related aminothiazole-containing analogues 11 and 12 are slightly more active but with IC 50 values ≥ 10 µM (Supplementary Figure S1).The relative inactivity of the matched benzo analogues 10-12 (≥ 10 µM) and ATP site analogues 5 and 6 (≥ 6 µM) demonstrates that the linker in the C-linked bivalent inhibitors 2 and 3 (51-59 pM) enables the markedly improvement in potency.Analysis of these IC 50 values allows for the estimation of the higher-limit of the linking coe cients, which are consistent with superadditivity (E < 1), as done previously (Supplementary Table S1). 28,30While the overall values are estimated to be 0.5-1.0M − 1 , they represent upper limits due to the C-linked allosteric 10 exhibiting virtually no activity against LRTM.The estimate of a lower-limit linking coe cient less than 1 is in line with 2 and 3 exhibiting superadditivity and con rms the C-linked amide as one of only a few examples of linked bivalent compounds exhibiting this behavior. 20,21,28Corresponding calculations for lower limits of N-linked bivalent 1 indicate that this compound has distinctively higher linking coe cients (> 1.9 x 10 7 M − 1 ) in line with the IC 50 activity measurements and show 10 7 -fold differences compared to C-linked amides.To characterize the binding modes of these bivalent inhibitors inspired by overlapping features in cocrystal structures (Fig. 2A), we determined X-ray cocrystal structures through soaking EGFR(T790M/V948R) crystals with the compounds, which reliably crystallize EGFR in the inactive (αC-helix "out") conformation (Fig. 2B,, Supplementary Table S2).A 2.1 Å-resolution cocrystal structure of 1 shows the imidazole and benzo groups bound within the ATP and allosteric sites, respectively, with the benzo moiety adopting an "outward" conformation (Fig. 2C, Supplementary Figure S2A-C).Analogously, a 2.2 Å-resolution cocrystal structure of 2 indicates that this compound is bound identically at the ATP site as 1, but with an opposite "inward" conformation within the allosteric pocket (Fig. 2D-E, Supplementary Figure S2D-F).Additional intermolecular interactions are observed for 2 such as H-bonding with T854 and D855 enabled by the C-linked amide, which are not possible in the N-linked methylene 1 (Fig. 2C-D).The side chain of K745, the catalytic lysine, exhibits a "swing" toward the benzo ketone in the case of 2 binding opening a position on the imidazole, which now binds a solvent water (Fig. 2D).The conformation of the allosteric benzo moiety in uences the position of the A-loop for the cocrystal structure of 1 and 2 (Fig. 2F-G, Supplementary Figure S3).Despite the parent allosteric inhibitors 8 and 9 being best matched in terms of the N-linked 1, the binding conformation of the C-linked 2 corresponds most closely to the allosteric inhibitor, 8 (Fig. 2H).Despite this difference in binding mode, the length of the linker is comparable between 1 and 2 (Supplementary Figure S4).To gain a more complete understanding of the inhibitors binding and provide the insights in the activity differences, we have performed 20 µs long molecular dynamic (MD) simulations (10 replicas x 1 µs per compound) based on our cocrystal structures of 1 and 2 (Supplementary Table S3, Supplementary Figure S5-S6).We nd excellent correspondence of the ligand interaction patterns between the simulations and experimental structures including some minor variations with respect to water-mediated H-bonds with ligands not evident from the cocrystal structures (Fig. 2I-J).Additionally, computer-aided docking has provided a pose for the covalent 4 similar to what is observed for 2 with the expected orientation for covalent bond formation with C797 and consistent with our earlier structural and functional studies (Supplementary Figure S7). 44,49These cocrystal structures and MD simulations indicate these structurally related compounds with different linker structures exhibit alternative inhibitor binding modes within the allosteric site, side chain orientations, and intermolecular interactions.
The wide range of potency observed for the C-linked 2 and the N-linked 1 motivated us to more completely understand the structural basis that enables this difference in activity.Appreciating that mobility can contribute to binding, we assessed compound rigidity from crystallographic B-factors of the bound ligands 1 and 2 (Fig. 3A-B, Supplementary Figure S8). 51This is made possible due to several commonalities shared between these cocrystal structures, including as they originate from the same protein, unit cells, and atomic resolutions (2.1 Å for 1 and 2.2 Å for 2).Generally, the ATP-binding imidazole in both compounds are comparably rigid with B-factors below the structure average while a notable increase is observed for the allosteric moiety in 2 and to a much lesser extent in 1 (Fig. 3A-B).To gain deeper insight, we performed generalized Born and surface area solvation (MM-GBSA) calculations on MD trajectories using our cocrystal structures (Fig. 1C-D).These calculations provide Free energies of binding where 2 exhibits greater a nity than 1 (ΔΔG = 9.5 kcal/mol), consistent with the difference in IC 50 values (Table 1), which is enabled by enhanced van der Waals and H-bonding interactions (Supplementary Table S4-5, Supplementary Figure S9).Additionally, MM-GBSA ligand energy calculations indicate a ~ 3.4-fold lower energy for 2 compared to 1, implying that 2 possesses a greater degree of structural complementarity within the kinase binding sites (Supplementary Table S4, Supplementary Figure S10).Superior binding of 2 is also aided by the "inward" benzo binding mode as this conformation is capable of full displacement of energetically unfavorable water molecules and the "outward" conformation of 1 does not allow complete displacement (Fig. 3C, Supplementary Figure S11).Further energetic analysis indicates that the potential energy of 2 pertaining to conformation within the binding mode of 2 is ~ 4-fold more favorable than the corresponding conformation of 1 (Fig. 3D-E).The overall pictures obtained by the crystallographic B-factors and MD simulations indicate that the superior potency of the C-linked compounds is due to a variety of factors that all stem from the structure of the linker allowing for improved mobility and pocket complementarity within the allosteric site.How the structure of the linker impacts the binding mode of these compounds is best visualized in terms of torsion angles observed in the MD simulations (Fig. 3F-G, Supplementary Figure S12).The most unique rotatable bond in 2 is the C-C bond that connects the linker amide to the benzo via the back pocket phenyl ring and allows for enhanced mobility of the group within the allosteric pocket (orange arrow Fig. 3G).This rotatable bond is the key structural element that allows for tight binding of this compound to EGFR since the other rotatable bonds in the linkers of 1 and 2 are comparably rigid and anchored to the relatively static ATP-site imidazole.These experimental and theoretical studies reveal the molecular factors that enable effective bivalent inhibitor binding, which can all be attributed to the "inward" conformation within the allosteric pocket enabled by the enhanced linker-enabled mobility of the C-linked scaffold.
We next sought to gauge the biological activity of our C-linked scaffolds as novel ATP-allosteric bivalent inhibitors.The Michael acceptor-containing C-linked analogue 4, designed to target C797 as done previously, 44,[46][47][48] was found most effective at suppressing LRTM phosphorylation (pY1068) as well as downstream pERK and pAKT in the human NSCLC cell line H1975 below 1 µM concentration dosed for 6 hours (Fig. 4A).Additional studies with H3255 (LR), H3255GR (LRTM) and HCC827 (exon19 delE746-A750) cells exhibit analogous suppression of EGFR pY1068 phosphorylation, slightly better potency in H3255 and H3255GR, cells indicating that 4 broadly targets diverse EGFR mutations (Fig. 4B-C, Supplementary Figure S13).The reversible binding 2-3 are also effective in H1975 cells, however to a lesser extent than 4, while the N-linked 1 exhibits limited ability to suppress phosphorylation (Supplementary Figure S13).We have also determined antiproliferation effects in human cancer cell lines H1975 and HCC827 treated with 4, which show that this novel EGFR inhibitor is active at concentrations ~ 100-500 nM overall ~ 60-fold less potent than AZD9291 in both cell lines (Supplementary Table S6, Supplementary Figure S14).While activity in H1975 and H3255 cells is expected based on biochemical data (Fig. 1), 4 is unexpectedly effective against HCC827 cells that harbor the prominent EGFR exon19 delE746-A750 mutation as allosteric pocket binding compounds are typically ineffective against this mutation (Fig. 4C). 52Additional antiproliferative activity experiments in Ba/F3 cells are consistent with above observations, and the ~ 220 nM EC 50 value for LR is markedly improved compared to our earlier covalent bivalent ATP-allosteric inhibitors indicating that the benzo-derived scaffold exhibits improved cellular activity (Supplementary Table S7). 44Furthermore, we con rm that 4 is selective across the kinome exhibiting a selectivity score of S(35) = 0.084 (Supplementary Table S8) and metabolically stable in liver microsome assays (Supplementary Figure S15).These experiments indicate that the C-linked bivalent inhibitor 4 is capable of targeting cellular EGFR in biological context of several prevalent oncogenic activating mutants and motivates further efforts to optimize the potency and medicinal chemistry properties for translational studies.

Discussion
][55] However, their complex dual-motif chemical structure is challenging to optimize, especially with respect to the structure of the linking group as subtle changes to structure can have considerable effects on biological function. 22,56,5722]30 For these reasons we rationalize that studies of alternatively linked bivalent molecules, which exhibit a broad range of potency, would reveal structural insights that can be used to improve the processes in optimizing bivalent compounds.The bivalent ATP-allosteric EGFR inhibitors reported here provide a structural basis for linker potency and offer previously unconsidered strategies for linker design.
The molecules in this study are informative since they differ only in terms of linker structure and exhibit signi cantly different biochemical potencies (> 10 6 -fold in LRTM and LRTMCS, Figure 1).The origins of this sizable potency range can be understood on the basis of how these compounds bind to the EGFR kinase domain (Figure 2).The "inward" allosteric benzo conformation of the C-linked scaffold 2 is found to best match the parent allosteric fragment 8 (Figure 2G) and MD simulations indicate that stronger binding of 2 is due, in part, to structural complementarity with the allosteric site.Matching the binding modes of parent fragments is a well-appreciated objective in FBDD linking, 28 which is consistent with the structures and biochemical potencies of 1 versus 2. Interestingly, the binding mode "inward" conformation of 2 most closely resembles the allosteric-only 8 despite possessing the N-linked structure of 1, and demonstrates how minute alterations in linker structure can manifest as major conformational differences in target binding sites. 50We further characterize the linker structure to be the key factor in enabling benzo mobility provided by variations crystallographic B-factors of 1 and 2 in cocrystal structures (Figures 3A-D, Supplementary Figure S8), MM-GBSA ligand energy calculations (Supplementary Table S4, Supplementary Figure S10) and rotatable bond torsion pro les (Supplementary Figure S12), which attributes the enhanced binding of 2 to a single unique C-C bond within the linker (orange in Figure 3G, Supplementary Figure S12).This potency-enabling aspect of the structure of 2 is only possible by shifting the linker connection (N-to C-linked) on the benzo motif (Table 1, Figure 1 & 4).Several previous studies have highlighted the importance of enabling proper exibility for effective linkers, 30,58 but to our knowledge our compounds are distinctive with respect to the range of potency (>10 6 -fold) and binding energy despite the relatively subtle differences in linker structure (Figure 1, ΔG Supplementary Table S4).
Given the notable potency enhancement between the Nand C-linked scaffolds, we further considered how these compounds may inform strategies for more e cient linker optimization.30,59 If we consider the evolution of the Nto C-linked scaffold, the structural basis for the large potency enhancement is related to how the linker is connected with the benzo allosteric motif.The shift in the linker "point of connection" from 1 to 2 introduces several attributes that are not necessarily possible in varying what is considered to be more conventional linker structural properties, such as length or alternative functional groups.To our knowledge, linker optimization of this sort has not been discussed previously, however, it is worth noting the parallels with pioneering work by Fesik and co-workers and their work pertaining to Bxl-x L inhibitors.60,61 Speci cally their linked lead compound, connected through a central trans-ole n, was shown to exhibit improved binding by a shift to a linear acylsulfonamide linker structure that varies with respect to different point of connection, and ultimately lead to the FDA-approved drug venetoclax. 61,62In terms of general optimization strategies, we propose that once linked, early-phase exploration of points of connection to the linked motifs are likely to result in large changes to drug potency (Figure 5).While understandably cumbersome in terms of synthesis, we rationalize that arriving at the optimal linker structure at the onset of a project is the most e cient scenario and as such would streamline lead optimization strategy.Additionally, this strategy may be especially helpful in cases where structural information regarding the linked molecule and target is challenging to obtain.
The iterative nature of drug optimization makes developing new medicines reliant on brute force structural explorations and serendipity, 20,59,63 but results from our work offer new strategies to more effectively optimize highly sensitive regions of bivalent inhibitors.The dissection of the structural basis of our alternatively linked molecules informs a new perspective on linker design, and also demonstrates that insights can be afforded through studies of linker-dependent effects on three-dimensional binding and how they translate to function.We rationalize that multi-site model systems, such as EGFR and others, represent constructive frameworks for evaluating drug design strategies amenable to structure determination with the ultimate goal to streamline structure-based drug optimization.Additionally, this work presents new examples of bivalent scaffolds unique within the diverse repertoire of EGFR tyrosine kinase inhibitors, which we show here are active in human NSCLC cell lines across the most prevalent oncogenic activating mutations (L858R, T790M, and exon19del).Future work will be directed toward expanding evaluations of bivalent inhibitor applicability broadly across the human kinome and improve our understanding of structure-based drug design.

Protein expression and puri cation
The EGFR kinase domain (residues 696-1022) was cloned into pTriEx with an N-terminal 6xHisglutathione S-transferase (GST) fusion tag followed by a TEV protease cleavage site.EGFR WT, L858R, L858R/T790M, L858R/T790M/C797S was expressed after baculoviral infection in SF9 cells and EGFR(T790M/V948R) was expressed in SF21 cells.Brie y, cells were pelleted and resuspended in lysis buffer composed of 50 mM Tris pH 8.0, 500 mM NaCl, 1 mM tris(2-carboxyethyl) phosphine (TCEP), and 5% glycerol.Cells were lysed via sonication prior to ultracentrifugation at >200,000 g for 1 h.wavelengths.4 was dosed between 0 and 10 µM in 24-point curves with 1.5-fold dilutions.Fluorescence, determined with identical reactions but lacking puri ed enzyme or crude cell lysate was subtracted from the total uorescence signal for each time point, with both determined in duplicate, to obtain corrected relative uorescence units (RFU).Corrected RFU values then were plotted vs. time and the reaction velocity for the rst ~40 min (initial reaction rates) was determined from the slope using GraphPad Prism (La Jolla, CA) with units of RFU/min.

HTRF Assays
Biochemical assays for EGFR domains were carried out using a homogeneous time-resolved uorescence (HTRF) KinEASE-TK (Cisbio) assay, as described previously. 65Assays were optimized for ATP concentration of 100 µM with enzyme concentrations WT EGFR 10 nM, L858R 0.1 nM, L858R/T790M at 0.02 nM and L858R/T790M/C797S at 0.02 nM.Inhibitor compounds in DMSO were dispensed directly in 384-well plates with the D300 digital dispenser (Hewlett Packard) followed immediately by the addition of aqueous buffered solutions using the Multidrop Combi Reagent Dispenser (Thermo Fischer).Compound IC 50 values were determined by 11-point inhibition curves (from 1.0 to 0.00130 μM) in triplicate.The data was graphically displayed using GraphPad Prism version 7.0, (GraphPad software).The curves were tted using a non-linear regression model with a sigmoidal dose response.
Cellular Antiproliferative Activity Assays H1975 and HCC827 cells were obtained from the lab of Dr. Pasi Jänne (Dana-Farber Cancer Institute, 2022) cultured at 37°C in RPMI 1640 media (Corning, 1004-CV) supplemented with 10% fetal bovine serum (Tissue Culture Biologicals, 35-010-CV) and 1% penicillin and streptomycin (P/S, Corning, 30-002-CI) and seeded overnight in a 96-well plate at a density of 60000 cells/mL.Cells were dosed with inhibitors to a nal DMSO 1% in triplicate for 37°C for 72 hours.Cellular inhibition of growth was assessed by MTT viability assay according to the manufactures protocol (OZ Biosciences).Parental Ba/F3 cells was a generous gift from the laboratory of Dr. David Weinstock (in 2014), Dr. Pasi Jänne (2020) both of the Dana-Farber Cancer Institute and was used to generate the wildtype EGFR, L858R, and L858R/T790M EGFR mutant Ba/F3 cells as performed previously. 52 66Brie y, all Ba/F3 cells were cultured in RPMI1640 media with 10% fetal bovine serum and 1% penicillin and streptomycin.All cell lines were tested negative for Mycoplasma using Mycoplasma Plus PCR Primer Set (Agilent) and were passaged and/or used for no longer than 4 weeks for all experiments.Assay reagents were purchased from MilliporeSigma (Cat# R7017-5G).Ba/F3 cells were plated and treated with increasing concentrations of inhibitors in triplicate for 72 hours.Compounds were dispensed using the Tecan D300e Digital Dispenser.Cellular growth or the inhibition of growth was assessed by resazurin viability assay to a nal 1% DMSO.All experiments were repeated at least 3 times and values were reported as an average with standard deviation.
Western blotting H1975, H3255, H3255GR and HCC827 lung adenocarcinoma cells were cultured in RPMI 1640 media (Corning, 1004-CV) supplemented with 10% FBS (Tissue Culture Biologicals, 35-010-CV) and 1% penicillin-streptomycin (P/S, Corning, 30-002-CI).H1975 and HCC-827 cells were seeded in 6-well plates with 400,000 cells per well and incubated overnight.H3255 and H3255GR cells were seeded in 12-well plates with 200,000 cells per well and were grown to con uency after 48 hours.Cells were treated the next day after replacing fresh media for 6 hours as indicated in the gure legends.Culture medium was removed, cells washed with PBS, and lysed with lysis buffer containing 5M NaCl, 1M TRIS pH 8.0, 10% SDS, 10% Triton X-100 and a tablet of protease and phosphatase inhibitor.Protein lysate concentration was analyzed using Pierce BCA kit (ThermoFisher, 23225 Modeling and structure for MD simulations Molecular modeling was conducted using Maestro (Schrödinger Release 2023-1, Schrödinger LLC, New York, NY, 2021) and the OPLS4 force eld. 69The crystal structures 1 (PDB ID 8FV3) and 2 (PDB ID 8FV4) were utilized for modeling the complexes of compound 1 and compound 2, respectively.Prior to modeling, the protein structures were prepared using the Protein Preparation Wizard 70 (Maestro 2021.4,Schrödinger LLC, New York, NY, USA) with default settings, which involved adding hydrogen atoms and correcting any missing side chains.In the compound 1 complex, the disordered residues in the activation loop required were rebuilt as following: The A859-A871 region was rebuilt based on the chain B of the 8FV4 crystal structure using the chimera homology modeling approach, followed by further minimization of the region using the OPLS4 force eld within a selected interval and a 3Å region around the selected residue interval.Additionally, the E872-E874 residues were added using Maestro's cross-link proteins tool, followed by region minimization in the OPLS4 force eld.For the compound 2 complex, the L862-A871 region of the activation loop was rebuilt based on the chain B of the 8FV4 crystal structure using chimera homology modeling.Similarly, the E872-K875 residues and S784 residue were added using Maestro's cross-link proteins tool, followed by region minimization in the OPLS4 force eld within a selected interval and a 3Å region around the selected residue interval.The initial validation of the individual models was assessed using the Ramachandran plot.

Molecular Dynamics Simulations
Desmond MD engine 71 was used for the MD simulations with OPLS4 69 force eld.The system was solvated in an orthorhombic box (minimum distance of 10 Å to the edges from the protein).A temperature of 300 K was used for membrane patch pre-equilibration.The water was described with the TIP3P 72 model.The nal systems comprised ~44 k atoms.All simulations were run in the NpT ensemble (T = 310 K, Nosé-Hoover method; p = 1.01325 bar, Martyna-Tobias-Klein method) with default Desmond settings.Reversible reference system propagator algorithms (RESPA) integrator with 2 fs, 2 fs, and 6 fs timesteps were used for bonded, near and far, respectively.Short-range coulombic interactions were calculated using 1 fs time steps and 9.0 Å cut-off value, whereas long-range coulombic interactions were estimated using the Smooth Particle Mesh Ewald method, which is a su ciently good approximation to treat long-range interactions on large timescales. 73The system was relaxed using the default Desmond protocol before the production simulation.
A total of 20 simulation replicas were prepared individually, each simulated with a random seed for a duration of 1μs for compound 1 and compound 2.This resulted in a cumulative simulation time of 20μs (10 replicas x 1μs x 2 compounds).All production simulations were conducted using consistent settings as previously described.The simulation interaction diagram provided by the Maestro package (Schrödinger, LLC, New York, NY) served as the foundation for the analysis of the simulations.
Protein-ligand interactions, as well as hydrophobic interaction frequency, RMSD, and torsional conformations of rotatable bonds were analyzed by the Simulation Interaction Analysis tool (scripts: event_analysis.py;analyze_simulation.py)(Schrödinger LLC).The default settings were used in the de nition of the interactions, where the following parameters were applied: H-bonds, a distance of 2.5 Å between the donor and acceptor with ≥120 and ≥90° for donor and acceptor angles, respectively; πcation interactions, a 4.5 Å distance between the positively charged and aromatic group; π-π interactions, stacking of two aromatic groups face-to-face or face-to-edge; water bridges, a distance of 2.8 Å between the donor and acceptor with ≥110 and ≥90° for donor and acceptor angles, respectively.

MM-GBSA
The molecular mechanics energies with generalized Born and surface area continuum solvation (MM-GBSA) were calculated with Prime Thermal MM/GBSA. 74,75Each 2nd frame of MD was used for MM-GBSA calculations (5010 complexes proceeded for an individual complex x 2 ligands).MM-GBSA calculations report energies for the ligand, receptor, and complex structures as well as energy differences relating to strain and binding and are broken down into contributions from various terms in the energy expression.

WaterMap
WaterMap 76,77 simulations were using Maestro, and the system was solvated in TIP3P water box extending at least 10 Å beyond the truncated protein in all directions.5 ns MD simulation was performed, following a standard relaxation protocol, and the water molecule trajectories were then clustered into distinct hydration sites.Entropy and enthalpy values for each hydration site were calculated using inhomogeneous solvation theory.

Figures
Figure 1 The linker structure bridging the ATP and allosteric site enables signi cantly altered biochemical activities of bivalent inhibitors.HTRF-based biochemical activity dose-response curves for the reversible

Figure 4 The
Figure 4

Table 1
Biochemical EGFR IC 50 values (nanomolar) against WT and mutant EGFR kinase domains.
Imidazole pH 8.0 was added to the supernatant for a nal concentration of 40 mM and owed through a column containing Ni-NTA agarose beads.The resin was washed with lysis buffer supplemented with 40 mM imidazole and eluted with lysis buffer containing 200 mM imidazole.Eluted EGFR kinase domain was dialyzed overnight in the presence of 5% (w/w) TEV protease against dialysis buffer containing 50 mM Tris pH 8.0, 500 mM NaCl, 1 mM TCEP, and 5% glycerol.The cleaved protein was passed through Ni-NTA resin to remove the 6xHis-GST fusion protein and TEV prior to size exclusion chromatography on a prepgrade Superdex S200 column in 50 mM Tris pH 8.0, 500 mM NaCl, 1 mM TCEP, and 5% glycerol.Fractions containing EGFR kinase of ≥95% purity as assessed by Coomassie-stained SDS-PAGE were concentrated to approximately 4 mg/mL as determined by Bradford assay or absorbance.Diffraction data was processed and merged in Xia2 using aimless and dials.The structure was determined by molecular replacement with the program PHASER using the inactive kinase domain EGFR(T790M/V948R) kinase from our previous work excluding the LN3844 ligand (PDB 6WXN).Repeated rounds of manual re tting and crystallographic re nement were performed using COOT and Phenix.The inhibitor was modeled into the closely tting positive F o − F c electron density and then included in following re nement cycles.Statistics for diffraction data processing and structure re nement are shown in TableS1.Due to a mixture of difference map density with contributions from both AMP-PNP and 2 in in Chain C we elected to leave this chain without bound ligands.kinase domain enzyme concentrations of WT EGFR at 2.0 nM, LR at 1.0 nM, and LRTM at 2.0 nMin nal solutions of 52 mM HEPES pH 7.5, 1 mM ATP, 0.5 mM TCEP, 0.011% Brij-35, 0.25% glycerol, 0.1 mg/ml BSA, 0.52 mM EGTA, 10 mM MgCl 2 , 15 μM Sox-based substrate (AQT0734).BSA was not included in this experiment to prevent interference with irreversible inhibitor characterization via off-target binding.All reactions were run for 240 minutes at 30 °C.Time-dependent uorescence from the Soxbased substrate was monitored in PerkinElmer ProxiPlate-384 Plus, white shallow well microplates (Cat.#6008280) Biotek Synergy Neo 2 microplate reader with excitation (360 nm) and emission (485 nm) Crystallization and structure determination EGFR(T790M/V948R) pre-incubated with 1 mM AMP-PNP and 10 mM MgCl 2 on ice was prepared by hanging-drop vapor diffusion over a reservoir solution containing 0.1 M Bis-Tris (pH = 5.5), 25% PEG-3350, and 5 mM TCEP (buffer A for crystals soaked with compound 1) or 0.1 M Bis-Tris (pH = 5.7), 30% PEG-3350 TCEP (buffer B for crystals soaked with compound 2).Drops containing crystals in buffer A and B were exchanged with solutions of both buffers containing ~1.0 mM 1 or 2 were exchanged three times for an hour and then left to soak overnight.Crystals were ash frozen after rapid immersion in a cryoprotectant solution with buffer A or BA containing 25% ethylene glycol.X-ray diffraction data of T790M/V948R-compound 2 crystals was collected at 100K at the Advanced Light Source a part of the Northeastern Collaborative Access Team (NE-CAT) on Beamline 24-ID-C.While data on T790M/V948Rcompound 1 crystals were collected at 100K at the National Synchrotron Light Source II 17-ID-2 64 .Time-dependent Kinase Inhibition AssaysBiochemical assays were performed with commercially available EGFR WT, cytoplasmic domain (669-1210), GST-tagged, Carna (Cat#/Lot#: 08-115/21CBS-0127H), EGFR [L858R], cytoplasmic domain (668end), GST-tagged, SignalChem (Cat#/Lot#: E10-122BG/D2411-4), EGFR [T790M/L858R] cytoplasmic domain (669-1210), GST-tagged, Carna (Cat#/Lot#: 08-510/12CBS-0765M).Reactions were performed with MS analysis.All compound incubations were conducted at least in triplicates.Additionally, a negative control containing BSA (20 mg/mL) instead of liver microsomes and a positive control using verapamil instead of compound were performed.A limit of 1% organic solvent during incubation was not exceeded.Sample separation and detection were performed on an Alliance 2695 Separations Module HPLC system (Waters Corporation, Milford, MA, USA) equipped with a Phenomenex Kinetex 2.6 μm XB-C18 100 Å 50 x 3 mm column (Phenomenex Inc., Torrance, CA, USA) coupled to an Alliance 2996 Photodiode Array Detector and a MICROMASS QUATTRO micro API mass spectrometer (both Waters Corporation, Milford, MA, USA) using electrospray ionization in positive mode.Mobile phase A: 90% water, 10% acetonitrile and additionally 0.1% formic acid (v/v), mobile phase B: 100% acetonitrile with additionally 0.1% formic acid (v/v).The gradient was set to: 0-2.5 min 0% B, 2.5-10 min from 0 to 40% B, 10-12 min 40% B, 12.01-15 min from 40 to 0% B at a ow rate of 0.7 mL/min.Samples were maintained at 10 °C, the column temperature was set to 20 °C with an injection volume of 5 μL.Spray, cone, extractor, and RF lens voltages were at 4 kV, 30 V, 8 V and 2 V, respectively.The source and desolvation temperatures were set to 120 °C and 350 °C, respectively, and the desolvation gas ow was set to 750 L/h.Data analysis was conducted using MassLynx 4.1 software (Waters Corporation, Milford, MA, USA). ) and Maestro 12.8.117.The receptor grid was generated from the EGFR(T790M/V948R) kinase domain from Chain D of PDB ID 8FV4 (compound 2), and ligands were prepared with LigPrep.The best binding poses were ranked on the basis of the lowest docking and glide score values.67,68