Conformation-locking antibodies for the discovery and characterization of KRAS inhibitors

Small molecules that stabilize inactive protein conformations are an underutilized strategy for drugging dynamic or otherwise intractable proteins. To facilitate the discovery and characterization of such inhibitors, we created a screening platform to identify conformation-locking antibodies for molecular probes (CLAMPs) that distinguish and induce rare protein conformational states. Applying the approach to KRAS, we discovered CLAMPs that recognize the open conformation of KRASG12C stabilized by covalent inhibitors. One CLAMP enables the visualization of KRASG12C covalent modification in vivo and can be used to investigate response heterogeneity to KRASG12C inhibitors in patient tumors. A second CLAMP enhances the affinity of weak ligands binding to the KRASG12C switch II region (SWII) by stabilizing a specific conformation of KRASG12C, thereby enabling the discovery of such ligands that could serve as leads for the development of drugs in a high-throughput screen. We show that combining the complementary properties of antibodies and small molecules facilitates the study and drugging of dynamic proteins. Antibodies that lock KRAS mutants in inactive conformations facilitate drug discovery.

with Cys12 to inhibit KRAS G12C protein 29 . Thus, the viability of strategies targeting this pocket in other KRAS mutants lacking this critical mutant cysteine residue remains to be determined. In addition, the consequence of covalent inactivation of KRAS G12C has not been investigated due to the lack of reagents capable of visualizing endogenous KRAS G12C covalent modification in tumor cells. Such reagents could be powerful tools to investigate the dynamics and homogeneity of KRAS G12C covalent modification following inhibitor treatment in individual cell populations, which may shed light on how tumor cells acquire resistance to KRAS G12C inhibitors.
Here we describe an antibody platform to identify conformation-locking antibodies for molecular probes, which we refer to as CLAMP. We applied this platform to covalently modified KRAS G12C and discovered two classes of CLAMPs. One such CLAMP (1A5) enables detection of covalently modified KRAS G12C within individual cancer cells and tumors, providing a direct biomarker for KRAS G12C inhibitor-mediated target engagement, and can be coupled with RAS pathway markers to assess pathway inhibition and subsequent rebound at the cellular level. Other CLAMPs also exhibit affinity for KRAS G12C -GDP alone, and one CLAMP (2H11) stabilizes an open conformation of the SWII pocket in the absence of ligand. This feature substantially improves the affinity of multiple noncovalent inhibitors for KRAS G12C and KRAS WT , and enabled the discovery of ~100 new SWII pocket ligands via a high-throughput screen. We propose that KRAS CLAMPs are an important biology and drug discovery tool for investigation of KRAS G12C covalent modification in vivo and will provide a unique platform for identifying chemical matter targeting the KRAS SWII pocket.

Discovery and characterization of KRAS SWII pocket CLAMPs.
We developed a method to identify KRAS CLAMPs using an in vitro selection strategy with synthetic antibody libraries. We applied this method to KRAS, which exhibits conformational heterogeneity in its switch I (SWI) and SWII regions, to identify antibodies that would stabilize an open SWII pocket conformation compatible with that induced by KRAS G12C covalent inhibitors (Fig. 1a). To identify KRAS CLAMPs, we leveraged distinct conformations of KRAS G12C : unmodified KRAS G12C -GDP, KRAS G12C -GDP with a covalent inhibitor and KRAS G12C -GMPPCP (nonhydrolysable GTP mimetic) (Fig.  1b). Four rounds of biopanning were performed in which the synthetic phage libraries were incubated in solution with biotinylated KRAS G12C -GDP covalently modified with GNE-1952, a tool G12C inhibitor 30 (Extended Data Fig. 1a). To drive selections towards the unique conformation of covalently inhibited KRAS G12C -GDP, selections were done in the presence of excess of KRAS G12C -GDP and KRAS G12C -GMPPCP in solution. To enable the discovery of antibodies capable of stabilizing the open SWII pocket conformation in the absence of inhibitor, we also performed three rounds of selection with KRAS G12C -GDP + GNE-1952 followed by a fourth round with KRAS G12C -GDP. After phage enzyme-linked immunosorbent assay (ELISA) screens to confirm specificity, we generated IgGs for eleven unique clones and characterized their binding specificity by ELISA (Fig. 1c) and surface plasmon resonance (SPR) (Fig. 1d, Extended Data Fig. 1b and Supplementary Table 1).
The 11 antibodies were assayed for their ability to recognize covalently modified KRAS G12C with additional KRAS G12C covalent molecules ARS-853 and ARS-1620 (refs. 22,24 ) (Extended Data Fig. 1a) to filter out CLAMPs recognizing only GNE-1952-bound KRAS G12C -GDP. All antibodies showed affinity to KRAS G12C -GDP covalently modified by GNE-1952, ranging from ~1 to 139 nM (Supplementary Table 1). However, one group of antibodies (five clones) was specific for the GNE-1952-bound KRAS G12C -GDP conformation while a second group (six clones) broadly recognized modified KRAS G12C -GDP (GNE-1952-, ARS-853-and ARS-1620-bound KRAS G12C -GDP conformations) and were referred to as KRAS CLAMPs since they appeared to be conformation rather than chemotype specific. Epitope mapping analysis revealed that these two groups bound distinct but partially overlapping epitopes on GNE-1952-bound KRAS G12C -GDP (Extended Data Fig. 1c).
The six KRAS CLAMPs could be further divided into two classes: the class I CLAMP (1A5) had high specificity for covalently bound KRAS G12C -GDP with >100-fold improved affinity for this target compared to KRAS G12C -GDP, whereas the class II CLAMPs (1E5, 2H11, 2A3, 3A12 and 4G12) recognized KRAS G12C -GDP independently of inhibitor presence, as measured by ELISA and SPR (Fig.  1c,d and Supplementary Table 1). To further establish antibody specificity in cells, we performed an immunoprecipitation experiment with 1A5 and 2H11 on lysates from HCC1171 KRAS G12C mutant cells treated with ARS-1620 or DMSO control. We found that both class I and II CLAMPs specifically immunoprecipitated covalently bound KRAS G12C -GDP but not KRAS G12C -GDP alone (Fig. 1e), suggesting that the affinity of the CLAMPs was insufficient to bind to KRAS G12C -GDP or KRAS G12C -GTP in cells. To further explore the potential of these antibodies, we focused on representative class I (1A5) and class II (2H11) antibodies.

Tracking KRAS G12C covalent modification in cells.
Since the 1A5 CLAMP showed the highest specificity for covalently modified KRAS G12C -GDP compared to KRAS G12C -GDP, we tested whether the 1A5 CLAMP could specifically detect inhibitor-bound KRAS G12C in cells. Immunofluorescence (IF) staining with the 1A5 CLAMP resulted in detection of covalently modified KRAS G12C in HCC1171 KRAS G12C -mutant cells treated with a variety of KRAS G12C covalent inhibitors, including GNE-1952, ARS-853, ARS-1620 and AMG 510 (Fig. 2a,b), thus confirming the ability of 1A5 to recognize a common three-dimensional epitope induced by various KRAS G12C covalent inhibitors. In contrast, there was no detectable 1A5 CLAMP staining in HCC1171 KRAS G12C -mutant cells treated with DMSO, or in HCT116 KRAS G13D -mutant cells treated with GNE-1952 (Extended Data Fig. 2a). The intensity of 1A5 CLAMP staining in ARS-1620-treated HCC1171 cells appeared to be both dose and time dependent (Fig. 2b,c) and was consistent with results from immunoblotting for covalently inhibited KRAS G12C , as measured both by an upward electrophoretic mobility shift of KRAS G12C protein band migration and inhibition of RAS pathway markers pERK and pMEK in a bulk population of cells (Extended Data Fig.  2b, lanes 1 and 2). Studies by IF revealed that ARS-1620 covalent modification of KRAS G12C in cells occurred in a fairly synchronous fashion at both the plasma membrane and in punctate compartments of cells ( Fig. 2a-c). Since antibodies specific for RAS-GDP do not exist, staining with the 1A5 CLAMP probably provides information on the localization of KRAS G12C -GDP in cells at, or shortly after, covalent modification. Moreover, the uniformity and kinetics of 1A5 CLAMP staining indicate that covalent modification of KRAS G12C may occur independent of cell cycle stage. Since covalent modification of KRAS G12C -GDP is dependent on intrinsic GTP hydrolysis 21,22 , these findings additionally suggest that intrinsic KRAS G12C -GTP hydrolysis is also not cell cycle dependent (Fig.  2b,c). The 1A5 CLAMP was also used to visualize covalent modification of KRAS G12C -GDP by ARS-1620 in a number of KRAS G12C lines expressing very low levels of KRAS G12C protein (Fig. 2d), highlighting the sensitivity of 1A5 CLAMP in detection of covalently modified KRAS G12C -GDP.
We next evaluated the utility of combining the 1A5 CLAMP and pS6 antibody to simultaneously measure KRAS G12C covalent modification and RAS signaling in individual cells using flow cytometry. Similar to the IF studies, the 1A5 CLAMP specifically detected covalent modification of KRAS G12C -GDP in a dose-dependent manner (Fig. 2e). This staining correlated with an observed dose-dependent decrease in pS6 levels in most of the population (Fig. 2e). Notably, a small population (~1%) of 1A5 + cells remained high for pS6, sug-gesting that this population of cells maintains RAS signaling despite covalent modification of KRAS G12C .
We also surveyed a set of commercially available antibodies to KRAS to determine their conformational specificity. We identified two antibodies (EP1125Y and Ras10) with comparable affinity to KRAS G12C , with or without covalent modification, by immunoprecipitation and ELISA (Fig. 3a,b), indicating that these antibodies are not specific to the conformation induced by covalent modification. In contrast, the iDab antibody, which was reported to be highly specific for HRAS-GTP, showed little to no binding to covalently modified KRAS G12C -GDP but bound to both KRAS G12C -GDP and KRAS G12C -GMPPCP by ELISA (Fig. 3b), and in immunoprecipitation studies (Fig. 3a) 3 . Since the iDab antibody binds an epitope that spans both the SWI and SWII regions, the SWII conformation induced by covalent modification of KRAS G12C -GDP probably prevents iDab binding.
Given the complementary specificities of the 1A5 CLAMP and iDab, we wondered whether 1A5 CLAMP and iDab antibodies could be used in combination to costain and visualize both unmodified and covalently modified KRAS G12C within the same cell. We conducted IF experiments with both 1A5 CLAMP and iDab antibodies on HCC1171 KRAS G12C cells treated with a dose titration of ARS-1620. Similar to previous IF experiments with 1A5, we detected an increase in 1A5 staining that correlated with high concentrations of ARS-1620 treatment. Furthermore, increased 1A5 staining coincided with decreased staining with the iDab antibody, confirming that these two antibodies stain different KRAS G12C conformations (Fig. 3c). Costaining with both antibodies allowed us also to monitor the resynthesis of KRAS G12C . Treatment of KRAS G12C cells with ARS-1620 or ARS-853 for 16 h resulted in almost complete covalent modification of KRAS G12C based on the appearance of 1A5 and disappearance of iDab staining, respectively, by IF, and was consistent with an electromobility shift of KRAS based on immunoblot analysis ( Fig. 3d and Extended Data Fig. 2b). Following drug washout, the appearance of unmodified KRAS G12C was apparent as early as 6 h and was increased at 24 h as evident by immunoblot analysis (Extended Data Figure 2b). Similar to immunoblot analysis, 24-h drug washout studies also showed a significant decrease in 1A5 staining and an increase in iDab antibody staining by IF (Fig. 3d).
Tumor cell adaptation to ARS-1620 has been attributed to new synthesis of KRAS G12C protein not bound to drug and subsequent reactivation of pERK, a marker of RAS pathway activity 31 . Accordingly we investigated whether the 1A5 CLAMP staining signal decreased over time due to new synthesis of KRAS G12C that remains unmodified, and whether that decrease correlated with the appearance of a pERK signal. We first treated NCI-H358 KRAS G12C mutant cells with 1 μM ARS-1620 for various time points (4, 24, 48, 72 and 96 h) to assess target occupancy, pERK inhibition and potential rebound. In cells treated with 100 nM and 1 μM ARS-1620, we observed an increase of alkylated KRAS G12C as evident by the enhanced intensity of the 1A5 CLAMP signal within the first 48 h, followed by a drop between 72 and 96 h, suggesting accumulation of unmodified nascent KRAS G12C . Correspondingly, pERK intensity demonstrated a rapid decrease as early as 4 h followed by a sharp rebound after 72 h, suggesting that the rebound of pERK is probably due to insufficient target occupancy by ARS-1620 (Fig. 3e). We then treated cells with 300 nM MRTX849 (a more potent KRAS G12C covalent inhibitor) at the clinically relevant dose. Unlike ARS-1620, treatment with MRTX849 resulted in a long-lasting increase in 1A5 CLAMP signal intensity that correlated with sustained suppression      of pERK signaling 96 h post treatment (Fig. 3e). This study demonstrates that potent KRAS G12C covalent inhibitors overcome pERK rebound in KRAS G12C mutant tumors, and serves as proof of concept for using CLAMPs to study long-term drug dynamics and acquired resistance to therapy. We next sought to determine whether the 1A5 CLAMP could detect covalently modified KRAS G12C in human tumor xenograft experiments. Detection of covalently inhibited KRAS G12C with the 1A5 CLAMP by immunohistochemistry (IHC) using formalin-fixed, paraffin-embedded tissue was not possible, probably due to the harsh formalin treatment that destroys the conformational epitope recognized by 1A5 CLAMP. However, ARS-1620 covalently modified KRAS G12C was readily detectable by IHC in tumor samples prepared as unfixed fresh frozen tissues (Fig. 4a). In addition, we observed a subtle trend towards stronger staining in samples treated with 200 mg kg -1 ARS-1620 compared to those treated with 50 mg kg -1 (Fig. 4a). Tumors expressing lower amounts of KRAS G12C covalently modified by ARS-1620 were also detected by the 1A5 CLAMP (Fig. 4b). Similar to in vitro cell experiments, the 1A5 CLAMP detected KRAS G12C that was covalently modified by ARS-1620 in fluorescent activated cell sorter (FACS) experiments with ex vivo tumor samples, and could be combined with the pS6 (Fig. 4c). These results thus show that the 1A5 CLAMP enables measurement of direct target engagement of KRAS G12C inhibitors in KRAS G12C mutant tumor samples and enables single-cell analysis with markers of RAS pathway activation.

Structural analysis of the CLAMP-KRAS G12C complex.
We hypothesized that the 2H11 class II CLAMP might stabilize the open conformation of the SWII pocket in KRAS G12C -GDP, based on its ability to recognize both the unbound and covalently modified conformations of KRAS G12C . We therefore determined the crystal structure of KRAS G12C -GDP in complex with 2H11 fragment antigen binding (Fab) at 2.2-Å resolution (Fig. 5a). We found that, rather than binding to the KRAS G12C SWII pocket, the 2H11 Fab contacts the outer surface of the SWII region to stabilize an open conformation of the pocket in an allosteric manner (Fig. 5b). The 2H11 Fab complementarity-determining regions (CDRs) H1 and H3 contribute to the majority of direct contacts with KRAS G12C (Fig.  5c). The long, 13-residue CDR H3 loop directly engages the SWII region by inserting Trp99 into a small hydrophobic pocket, known as the DCAI pocket 32,33 , and is surrounded by KRAS G12C residues Lys5, Leu6, Val7, Ser39, Asp54, Leu56, Tyr71, Thr74 and Gly75 (Fig. 5b). Notably, 2H11 exploits this site with a chemically similar tryptophan side chain (Extended Data Fig. 3a). CDR H1 contacts KRAS near the C-terminal end of the SWI region and packs against a portion of the RAS-binding domain binding site. Unlike iDab 3 , 2H11 makes minimal contact with the residues in the N-terminal half of SWI (Extended Data Fig. 3c), consistent with its observed ability to bind to both KRAS-GDP and GTP states. CDR L2 and H2 also contribute several van der Waals contacts with KRAS. CDR L2 contacts the N-terminal tip of the SWII helix, providing additional stabilization to the SWII loop but without overly restricting the SWII conformation. As shown in Fig. 5d, the most flexible part of SWII, Gln60-Ala66, is completely free from direct contact with 2H11, maintaining flexibility in the pocket. The SWII pocket lies in a stable, open conformation that resembles the shape of previously published ligand-bound KRAS structures, suggesting easy accessibility for ligand binding 28 . We then determined the crystal structure of 2H11 in complex with KRAS G12C covalently modified by GNE-1952, and compared it to 2H11 in complex with unbound KRAS G12C (Fig. 5d) to test the above hypothesis. While GNE-1952 caused a shift in the main and side chain conformation of SWII residues (SWII Cα atoms: root The sidechain of His95, which forms a hydrogen bond with the quinazoline nitrogen of the inhibitor, is the only residue outside SWII that changed conformation following GNE-1952 binding. This interaction appears to be common for quinazoline scaffold compounds 22 , and we show that the His95 swing is not affected by the presence of 2H11 Fab. Comparison of two inhibitor complex structures, with and without 2H11 Fab (Extended Data Fig. 3c), revealed a highly similar ligand-binding mode ( Supplementary  Fig. 3b) and atomic interactions although, in the absence of Fab, the SWII helix tilted further inward to the pocket and the ligand displayed a concerted shift (SWII Cα atom RMSD = 0.87 Å). This shift is probably influenced by crystal packing interactions, and indicates that SWII maintains some level of flexibility following ligand binding.
Since 2H11 employs an allosteric site away from G12C, we wondered whether 2H11 could also recognize the GDP-bound states and potentially stabilize the open conformation of the SWII region in other KRAS mutants. We evaluated binding of 2H11 to a panel of KRAS mutants by ELISA (Fig. 5e). Notably, 2H11 exhibited strong binding to KRAS G12V -GDP, KRAS G12R -GDP and KRAS Q61H -GDP but much weaker binding to KRAS G13D -GDP and KRAS WT -GDP (Fig. 5e). The similar degree of binding to G12V, G12R and Q61H  KRAS WT was immobilized on a SPR chip, and small molecules were injected at concentrations ranging from 1 to 50 μM in the absence or presence of 2H11. b, Fragment screening of a library containing 3,060 fragments in the presence of 2H11 increased the primary hit rate by >twofold (from 1.5 to 3.3% in the presence of 2H11 CLAMP). The plot shows the affinity distribution of the confirmed hits for KRAS G12C fragment screens with (orange) and without 2H11 CLAMP (blue). Comparison between the screens demonstrates the ability of 2H11 to increase the overall number of hits, especially in the higher-affinity (K D < 500 μM) groups. c, Fragment screening against KRAS G12C with 2H11 CLAMP uncovered SWII ligands previously below the hit criteria threshold. GNE-2897 and GNE-9764 exhibit a shift in potency between KRAS G12C alone (green) and KRAS G12C coimmobilized with 2H11 Fab (blue). d, Crystal structure of GNE-2897 bound in the SWII region of KRAS G12C + 2H11 Fab. KRAS residues surrounding the binding site are depicted in ribbon diagram and surface, and the SWII region is colored green. Key residues are labeled. Compounds are depicted as a stick model with carbon atoms colored gold. Dotted lines indicate hydrogen bond interactions.
suggests that these mutants may adopt a shared SWI/SWII conformation that is distinct from G13D, G12C and WT.
CLAMPs enable discovery of weak SWII ligands. The CLAMP-mediated stabilization of the open conformation of the SWII pocket in KRAS G12C -GDP would then be predicted to improve the noncovalent affinity for ligands that bind in the SWII pocket. We developed a robust SPR assay to detect small-molecule ligands binding to the KRAS G12C -GDP SWII pocket using a referencing strategy and cocapture method (Methods), and measured the affinity of various SWII pocket covalent ligands (GNE-1952, ARS-853, ARS-1620 and a noncovalent version of GNE-1952 that lacked the reactive acrylamide moiety; Extended Data Fig. 1a) to KRAS G12C -GDP in the presence and absence of 2H11 Fab (referred to as 2H11 below). Additionally, we included KRAS WT -GDP to determine whether 2H11 could stabilize the SWII pocket in other KRAS variants. We found that inhibitor affinity was greatly enhanced in the presence of 2H11 (Fig. 6a and Extended Data Fig. 4a). We next sought to leverage this finding for the discovery of new SWII ligands.
Due to the dynamic nature of the SWII pocket, all previous efforts to discover compounds that bind this region have required the use of electrophilic functional groups that can form covalent bonds with Cys12 (refs. 21,26 ). However, most compounds in screening libraries typically do not carry chemically reactive groups and thus such libraries would be of limited use for screening against KRAS G12C . Given that 2H11 can stabilize the SWII pocket and substantially enhance the binding of weak ligands, we reasoned that this CLAMP would enable fragment-based lead discovery for the SWII pocket of KRAS G12C -GDP. We screened a fragment library of 3,060 compounds in the presence of 2H11 and identified 3.3% of these with K D < 1,500 μM ( Fig. 6b and Supplementary Table 2). A third of these compounds showed K D < 500 μM in the presence of 2H11 (Fig. 6b and Supplementary Table 2). Furthermore, fragment screening in the presence of 2H11 resulted in >twofold increase in overall hit rate and a 13% increase in the number of fragments with K D < 500 μM, highlighting the ability of the 2H11 CLAMP to enable weak ligand discovery ( Fig. 6b and Supplementary Table 2). We then determined the affinity of two ligands, GNE-2897 and GNE-9764, for KRAS G12C in the presence or absence of 2H11. The binding affinity of GNE-2897 and GNE-9764 to KRAS G12C in the presence of 2H11 ranged between K D = 600 and 1,200 μM (Fig. 6c). In contrast, neither ligand showed any significant affinity to KRAS G12C in the absence of 2H11 (Fig. 6c). We next solved the crystal structure of GNE-2897 bound to KRAS G12C -GDP-2H11 ( Fig. 6d and Supplementary Fig. 4b). We found that GNE-2897 binds in the back of the SWII pocket, the site commonly occupied by covalent inhibitors 8 . Notably, GNE-2897 shares similarity with an elaborated potent KRAS G12C -GDP covalent inhibitor recently published 34 , demonstrating that the 2H11 CLAMP can be used to identify new, diverse fragments that can be further optimized into new leads.

Discussion
While RAS proteins are major oncogenic drivers of many cancers, the lack of stable pockets has hindered drug development. The recent discovery of KRAS G12C covalent inhibitors revealed a cryptic pocket within the SWII region and sets a precedent for clinical activity for future ligands of mutant KRAS. Despite this success, whether other KRAS mutants can be targeted via the SWII pocket remains to be determined.
The discovery of KRAS CLAMPs has revealed new insights into KRAS covalent modification and can be used to visualize and track inhibitor-bound KRAS G12C in both cells and in vivo tumor models. Although the covalent modification of KRAS G12C appeared to be homogenous in preclinical studies, modification of KRAS G12C and subsequent tumor response in patient tumors is probably more heterogeneous. Implementation of a 1A5 CLAMP-based flow cytometry or IHC assay, along with markers of KRAS pathway activity such as pERK in clinical trials, should enable rapid detection of covalent modification to confirm target engagement and may inform potential mechanisms of acquired resistance.
Biophysical and structural studies showed that the 2H11 CLAMP can induce a conformation with an open SWII pocket in KRAS G12C -GDP, can improve the affinity of SWII ligands and may enable future efforts to target this pocket more broadly across KRAS mutants. While this strategy may be successful for other mutant proteins with sufficient intrinsic GTP hydrolysis, other KRAS mutants (for example, KRAS Q61H ) appear to be constitutively bound to GTP, making it unclear whether targeting of these mutants is possible 20 . Recent work using a non-natural cysteine suggests that it is possible to identify compounds that engage the SWII pocket in both GDP-and GTP-bound states 35 . Future efforts to identify KRAS CLAMPs that also bind and stabilize the GTP state to promote a stable open SWII pocket, irrespective of the nucleotide-bound state, may be possible and provide a foundation for small-molecule screens using a CLAMP-KRAS mutant complex to expand the targeting of the SWII pocket to all RAS mutants.
Our CLAMP platform has led to the discovery of antibodies that recognize conformations existing within the dynamic KRAS switch regions. The success of this strategy relies on the presence of a covalent ligand, highlighting the importance of locking conformationally dynamic proteins in different states to enable the discovery of unique conformation-specific antibodies. Additionally, we provide experimental evidence that CLAMPs can also induce and lock a conformation in the absence of ligand, and thus enable small-molecule discovery efforts against transient pockets within conformationally dynamic proteins. More generally, our work has broad implications for antibody-assisted small-molecule drug discovery against dynamic biological targets.

Online content
Any methods, additional references, Nature Research reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/ s41587-021-01126-9.

Methods
Phage selection. Phage selection was performed using existing synthetic Fab phage display libraries 36,37 . The pooled library was cycled through three to four rounds of binding in solution to biotinylated KRAS G12C -GDP + GNE-1952 (concentrations started at 500 nM and were gradually decreased to 10 nM in the final round). The solution was captured on neutravidin beads (Promega), blocked with 5 μM biotin, washed three times for 30 s each in PBS + 0.5% bovine serum albumin (BSA) + 0.1% Tween 20 (PBSBT) and eluted with 100 mM HCl. The eluted phage was neutralized with 1 M Tris-HCl pH 8.0 before overnight amplification in Escherichia coli XL1-blue (Stratagene) with the addition of M13KO7 helper phage (New England Biolabs). To enrich for binders specific to covalently modified KRAS G12C , selections were performed in the presence of an excess of either soluble KRAS G12C -GDP or KRAS G12C -GMPPCP at 1 μM. After selection, individual colonies were picked and grown overnight at 30 °C in 96-well, deep-well plates in 2xYT medium supplemented with carbenicillin and helper phage. Phage supernatants were used in phage ELISAs against KRAS G12C -GDP + GNE-1952, KRAS G12C -GDP and KRAS G12C -GMPPCP to identify target-specific clones. Antibody ELISA against mutant KRAS-GDP proteins. KRAS-GDP proteins were directly coated in triplicate at 10 μg ml -1 on Maxisorb plates (Thermo Scientific) in PBS and incubated overnight at 4 °C. Plates were blocked for 2 h at 25 °C using 4% BSA. Serial dilutions of 1A5 and 2H11 antibodies, starting at 10 μg ml -1 , were added for 1-2 h at 25 °C with shaking. Plates were developed and read as described above.
Epitope binning. Epitope-binning experiments were performed in HBS-P+ (GE Healthcare) running buffer at 25 °C on an array-based imager (IBIS MX96) as described previously 39 . Briefly, 10 μg ml -1 anti-KRAS antibody was amine coupled to the surface in 10 mM sodium acetate pH 4.5 and the surface was quenched with 1 M ethanolamine. Epitope-binning experiments were done by initially flowing 2 μM KRAS G12C -GDP + GNE-1952 over immobilized antibodies and over 10 μg ml -1 of each of the anti-KRAS antibodies in solution. Enough time was allowed for association of the antigen before addition of antibody. Before the addition of the next antibody in solution, the surface was regenerated with 10 mM glycine pH 2.5.
Cell lines and tissue culture. Cell lines were obtained, characterized and controlled for quality as described 40 . All cell lines were cultured in RPMI with 10% fetal bovine serum supplemented with 2 mM L-glutamine.
Immunofluorescence and high-content imaging. Between 0 and 40,000 cells per well were seeded into Poly-L-lysine-coated 96-well plates (Cell Carrier Ultra, PerkinElmer) and treated the following day with KRAS G12C inhibitors at the indicated concentrations and durations. Cells were then washed twice with cold PBS, fixed with 3% paraformaldehyde (PFA) for 20 min at 25 °C, washed for 10 min with PBS and then PFA was quenched with 50 mM NH 4 Cl for 10 min at 25 °C. Cells were washed twice with PBS for 5 min and permeabilized with 1× Perm/ Wash Buffer (BD, Fisher Scientific) for 20 min at 25 °C. Cells were then incubated with primary antibody (1A5 or iDab or both) diluted in Perm/Wash buffer at the indicated concentration for 2 h at 25 °C. Cells were then washed three times with Perm/Wash buffer for 10 min each and then incubated with conjugated fluorescence secondary antibody (1:500; Alexa488 anti-human (catalog no. 709-546-149) and Alexa647 anti-rabbit (catalog no. 711-606-152) or anti-rat (catalog no. 712-606-150) at 1:500, from Jackson ImmunoResearch Laboratories) at 25 °C for 60 min. Next, 100 ml of 300 nM DAPI was added to each well for 15 min followed by two washes with Perm/Wash buffer and one with 1× PBS before imaging. Imaging was done on the Opera Phenix HCS machine (PerkinElmer) using the ×40 water immersion lens and confocal mode for enhanced membrane scanning ability. Four or five fields were acquired; analysis and quantification were conducted on Harmony (PerkinElmer) software and plotted using GraphPad Prism.
For washout experiments, cells were plated as described above, treated with KRAS G12C inhibitor and incubated with compound for 18-24 h. One plate was imaged after 24 h as control and the others were washed twice with cold 1× PBS. Plates were then incubated for either 24 or 48 h with 150 ml of complete compound-free medium and stained and imaged as described above.
Immunoprecipitation studies. HCC1171 cells were treated with DMSO or 5 μM ARS-1620 for 24 h. Cells were collected and lysed in IP lysis buffer (Thermo Fisher Scientific, no. 87787) with the addition of 300 mM NaCl, protease inhibitors (Roche, no. 11836170001) and phosphatase inhibitors (Roche, no. 4906845001). Antibody (1 μg) was mixed with 200 µg of lysate in 100 μl of IP buffer and incubated on ice for 1 h. Next, 15 μl of protein A/G MAG beads in 200 μl of IP lysis buffer was added to each antibody/lysate mix with rotation at 4 °C overnight. Beads were washed three times with IP lysis buffer then suspended in 1× LDS loading buffer and boiled at 95 °C for 10 min. Lystates were subjected to immunoblot analysis and probed with KRAS antibody no. 12063-1-AP (1:1,000 dilution).
In vivo tumor studies. All procedures were approved and conformed to the guidelines and principles set by the Institutional Animal Care and Use Committee of Genentech, and were carried out in a facility accredited by the Association for the Assessment and Accreditation of Laboratory Animal Care. Female C.B-17 SCID (inbred) mice aged 16-17 weeks and weighing 24-27 g were obtained from Charles River Laboratories. The mice were housed at Genentech in standard rodent micro-isolator cages and were acclimated to study conditions for at least 3 days before tumor cell implantation. Mice were inoculated with 5 million NCI-H358 non-small cell lung carcinoma cells (suspended in a 1:1 mixture of Hank's balanced salt solution containing Matrigel at a 1:1 ratio) in both the left and right flank subcutaneously. Tumors were monitored until they reached a mean volume of 400-600 mm 3 . Mice were given a single dose of either 0 (vehicle, 100% Labrasol), 50 or 200 mg kg -1 ARS-1620 orally by gavage in a volume of 100 μl. Plasma and tumor samples were collected at 8 or 24 h post dosing.
Immunohistochemistry studies. Frozen optical coherence tomography-embedded tumor samples were sectioned to 5-μm thickness using a cryostat, and acetone-fixed at 25 °C. Before IHC staining, sections were air-dried overnight at 25 °C. Endogenous peroxidase activity was quenched with glucose oxidase solution for 60 min at 37 °C. Sections were loaded onto a BOND III autostainer, incubated with the 1A5 CLAMP and diluted to 3 μg ml -1 in 3% BSA/PBS for 60 min at 37 °C. Sections were then incubated in rabbit anti-Human IgG (Jackson ImmunoResearch, no. 309-005-082) and diluted to 5 μg ml -1 in 3% BSA/PBS for 2H11 CLAMP Fab in vitro biotinylation. The 2H11 Fab was buffer exchanged from the storage buffer into PBS using Zeba Spin columns (7K MWCO) from Thermo Fisher Scientific. The 2H11 Fab was mixed in a 1:2 molar ratio with EZ-Link NHS-PEG4-Biotin (Thermo Fisher Scientific) for 2 h at 25 °C. To stop the reaction the mixture was desalted 2× through Zeba Spin columns (7K MWCO) into PBS. Biotinylation was confirmed by pulldown using streptavidin Phytips (Phynexus) and analyzed using Bioanalyzer (Agilent). This reagent was then termed 2H11 CB Fab.

SPR surface preparation for 2H11 CLAMP comparison.
A series S SA (streptavidin) chip was inserted into a Biacore 8K (Cytiva). The instrument was primed into running buffer (50 mM HEPES pH 7.5, 150 mM NaCl, 0.2% (w/v) PEG3350, 0.1% CM-dextran (w/v), 0.1 mM TCEP, 10 mM MgCl 2 , 100 nM GDP and 2% (v/v) DMSO). This SPR chip and buffer were used for all subsequent KRAS G12C experiments described in this paper. KRAS G12C was captured on FC2-4 to 1,900 RU then data were collected, all referenced to FC1. 2H11 CLAMP IgG, Fab and CB Fab were injected once at 200 nM for 180 s, following the immobilization of the KRAS G12C , to saturation and allowed to stabilize for 1,000 s. SPR fragment screen with KRAS G12C and 2H11 CLAMP Fab. The fragment screen was performed using Biacore 8K+ (Cytiva). Alkylated KRAS G12C was captured to yield 2,000 RU on flow channel 1 (FC1) to serve as the reference. KRAS G12C was captured on FC2 within 5% of the reference channel capture level, and data were collected in FC 2-1 mode. For the screen with the 2H11 CB Fab, this was injected at 200 nM for 600 s over both channels. Subsequently, all channels were blocked by injection of 180 nM amine-PEG-biotin (Thermo Fisher). The fragment library of 3,060 compounds was injected as a single injection at 500 μM with an association/dissociation time of 20 s at a flow rate of 60 μl min -1 . A control sample was injected every 24 injections to monitor surface activity over the course of the screen, and to normalize response between the different injection channels. Blanks (running buffer) were injected every 12 injections. All screen data were analyzed using Biacore Insight Software (3.0.12, Cytiva). All data were blank subtracted and then normalized for control response 17 . Primary hits were determined by calculating the mean and standard deviation of the baseline points (excluding the top and bottom 5% of responses); the primary hit cutoff was determined to be three standard deviations above the baseline mean. The overall hit rate was determined to be the product of the primary and confirmation hit rates.

SPR compound and fragment affinity determination with and without 2H11
CLAMP. Affinity determination of fragments and compounds was done using either the Biacore 8k+ or S200. Fragments were tested for dose-response against KRAS G12C at 15.6-500 μM, and compounds were tested at 3.12-100 μM with a 20-s association\dissociation time at a flow rate of 65 μl min -1 . All sensorgrams were analyzed with Biacore Insight (3.0.12, Cytiva) or S200 Evaluation software (1.1, Cytiva) using a 1:1 affinity model. Diffraction data collection and structure determination. The diffraction data of KRAS G12C /GNE-1952, KRAS G12C /2H11 and KRAS G12C /GNE-1952/2H11 crystals were collected using monochromatic X-rays at either the Stanford Synchrotron Radiation Lightsource (SSRL) beamline 12-2 or Advanced Light Source (ALS) beamline 5.0.2 using a PILATUS3 6M detector. The rotation method was applied to a single crystal for each complete dataset. The crystals were kept at cryogenic temperature throughout the data-collection process. Data reduction was performed using the program XDS 41 and the CCP4 program suite 42 . Data reduction statistics are shown in Supplementary Table 3.

Crystallization of KRAS
The diffraction data of KRAS G12C /2H11/GNE-2897 crystals were collected using monochromatic X-rays at ALS beamline 5.0.2 using a PILATUS3 6M detector. The rotation method was applied to a single crystal for the complete dataset. Crystals were kept at cryogenic temperature throughout the data-collection process. Data reduction was done using the program XDS 41 and the CCP4 program suite 42 . Diffraction patterns were clearly anisotropic. We performed anisotropic scaling using program STARANISO 43 . Data reduction statistics are shown in Supplementary Table 3.
The structures of KRAS G12C /2H11 and KRAS G12C /GNE-1952/2H11 were phased by molecular replacement (MR) using the program Phaser 42 . A previously published crystal structure of KRAS G12D (PDB: 4DSU) and a Fab structure (PDB: 3R1G) were used as MR search models. Manual rebuilding was performed with the graphics program COOT 44 . Structures were further refined iteratively with the programs REFMAC5 (ref. 45 ) and PHENIX 46 using maximum-likelihood target functions, anisotropic individual B-factor refinement and TLS refinement to obtain the final statistics shown in Supplementary Table 3.
The structures of KRAS G12C /GNE-1952 were phased by MR using Phaser 42 . A previously published crystal structure of KRAS G12D (PDB: 4DSU) was used as the MR search models. Varying density corresponding to GNE-1952 binding to the SWII pocket was present in the initial Fo-Fc map. The ligand structure was built into the electron density using COOT 44 . Subsequent manual rebuilding of the entire protein structure was also performed in COOT. The structures were further refined iteratively with REFMAC5 (ref. 45 ) and PHENIX 46 using maximum-likelihood target functions, anisotropic individual B-factor refinement and TLS refinement to obtain the final statistics shown in Supplementary Table 3.
The structure of KRAS G12C /GNE-2897 was phased by MR using Phaser 42 . The crystal structure of KRAS G12C /2H11 was used as the MR search models. Varying density corresponding to GNE-2897 binding to the SWII pocket was present in the initial Fo-Fc map (Extended Data Fig. 4b). The ligand structure was built into the electron density using COOT 44 . Subsequent manual rebuilding of the entire protein structure was also performed in COOT. The structure was further refined iteratively with REFMAC5 (ref. 45 ) and PHENIX 46 using maximum-likelihood target functions, anisotropic individual B-factor refinement and TLS refinement to obtain the final statistics shown in Supplementary Table 3.