Protein Engineering-Enabled Cryo-EM Investigation of Small GTPases

Abstract

Small GTPases play important roles in cellular signaling. Due to their small sizes (~21 kDa), structural studies of small GTPases have been predominantly performed using x-ray crystallography in which crystal lattice contacts made it challenging to define unperturbed conformations of the key switch regions. Here, we develop a protein-engineering strategy that enables cryo-EM analysis of small soluble proteins and applied to RAS. We fused the C-terminal α5 helix of the RAS globular domain to a small protein BRIL by forming a continuous helix, which leaves most RAS surfaces exposed to the solvent and unperturbed, followed by the complex formation with an anti-BRIL Fab. This engineered complex with an increased molecular weight, termed “RAS-lollipop”, enabled single-particle cryo-EM of RAS. Using this approach, we determined the cryo-EM structure of NRAS, whose structural studies using crystallography have been the least successful among the RAS isoforms. We revealed the conformations of the switch region and α5 helix that differ from those observed in published crystal structures, and also defined the binding site of an NRAS-specific monobody. We uncovered an unexpected surfactant-like property of this monobody, which reduces orientation biases of particles on cryo-EM grids. Together, this work establishes a platform for visualizing small GTPases and potentially other small proteins with minimal perturbation of their surfaces.

Graphical Abstract

Introduction

Small GTPases are essential molecular switches that regulate a wide range of cellular processes, including proliferation, differentiation, and survival^1,2. Small GTPases have a common architecture consisting of a globular (G) domain and a C-terminal tail that is post-translationally lipidated for membrane localization. These proteins function by cycling between the inactive GDP-bound state and the active GTP-bound state, a transition mediated by nucleotide exchange and GTP hydrolysis. This switching is accompanied by conformational changes in two flexible surface loops, known as switch I (SW1) and switch II (SW2).

Structural studies of small GTPases, predominantly performed using x-ray crystallography, have been foundational to understanding their molecular mechanisms of their functions and pathological mutations. However, a major limitation in crystallographic studies is that the switch regions are frequently engaged in extensive crystal contacts, leading to artificially constrained conformations and obscuring biologically relevant states. Because the switch regions are largely surface exposed, it is difficult to form well-packed crystals without involving the switch regions in lattice formation.

Cryo-electron microscopy (cryo-EM) provides an opportunity to visualize proteins in solution without crystal packing. However, the small molecular weight of small GTPases makes it difficult for single-particle cryo-EM to determine their structure. Two recent protein-engineering approaches increased the effective particle size of RAS, a major oncogenic small GTPase family^3,4, thereby enabling cryo-EM investigation. KRAS was captured by a designed protein cage displaying a KRAS-binding DARPin by the Yeats group⁵, and a disulfide-rigidified anti-KRAS Fab from the Sudhamsu lab⁶, respectively. Both strategies achieved high resolution KRAS structures, but the binders directly interact with the switch regions, likely constraining their conformational landscape away from the unperturbed state.

In this work, we developed a protein engineering strategy that increase the effective particle size without directly interacting with the switch regions. We use BRIL (thermostabilized apocytochrome b562), a four-helix bundle widely used to aid crystallization and structural determination of small membrane proteins, particularly GPCRs^7–9. In these studies, the C-terminal helix of a protein of interest is connected to the N-terminal helix of BRIL in such a way that the two helices form a continuous helix. More recently, a synthetic anti-BRIL Fab with a rigidified elbow was developed by the Kossiakoff laboratory and, when combined with BRIL, increases the effective molecular weight and asymmetry of small membrane proteins to facilitate cryo-EM analysis^10,11. We envisioned that the BRIL-Fab scaffolding could be applied to small, soluble proteins such as RAS, to expand its utility.

RESULTS

Design of HRAS-lollipops

Based on the observation that the G domain of small GTPases terminates with a long α-helix (α5), we envisioned that this helix can be seamlessly fused to the first helix of BRIL to create a rigid connection. Binding of the anti-BRIL Fab (BAG2)^10,12 to this fusion protein would result in the formation of a complex (MW ~80 kDa) that can overcome the size limitation of single-particle cryo-EM. In addition, the switch regions should remain unperturbed and solvent-exposed (Fig. 1a). We named this strategy “RAS-lollipop” based on the overall shape (a hand holding a lollipop).

Figure 1. RAS-lollipop strategy.

(a) Schematic of the RAS-lollipop design. RAS is fused at its C-terminal α5 helix to BRIL. An anti-BRIL Fab, BAG2, is added to increase the particle size (the light chain shown in light blue, and the heavy chain in blue). The poly-alanine linker is shown in red. SW1 is shown in magenta and SW2 is shown in green. The RAS-BRIL fusion model is generated using AlphaFold2.

(b) The elution profile from a size-exclusion chromatography column of HRAS-BRIL with BAG2. The elution positions of molecular weight standards are indicated with arrows.

(c) Cryo-EM 2D class averages of HRAS-lollipops showing severe preferred orientation.

(d) Cryo-EM 2D class averages after the addition of monobody NS1. Multiple particle views and clear secondary-structure features were observed.

(e) Low-resolution 3D reconstruction obtained from 854 micrographs of HRAS-3A-BRIL/BAG2/NS1. The AlphaFold-predicted HRAS-3A-BRIL model and BAG2 structure (PDB:6CBV) were fitted into the map. The NS1/HRAS crystal structure (PDB: 5E95) was aligned to the HRAS G-domain of the predicted HRAS-3A-BRIL model.

To test the feasibility of the RAS-lollipop approach, we first applied it to HRAS. We began with a construct comprising residues 1–166 of HRAS, fused to BRIL through a short poly-alanine extension of three to four residues between α5 of HRAS and helix 1 of BRIL (Table S1), following the linker designs of the Kossiakoff and Traaseth groups^10,11. Because of our ongoing studies on oncogenic mutations at Q61, we used HRAS(Q61L) here, which hereafter we refer to as HRAS for brevity. We then assembled HRAS-BRIL with the Fab BAG2 to generate a ~90 kDa complex (Fig. 1b).

In the initial cryo-EM screening, all of the HRAS-BRIL/Fab complexes tested exhibited severe preferred orientation. Particles consistently gave a single dominant view, and no 3D reconstruction could be obtained (Fig. 1c). After our unsuccessful surfactant screening, we introduced an additional binding partner, NS1, a monobody that binds HRAS and KRAS in the α4/α5 region without affecting the switch regions¹³. NS1 is agnostic about oncogenic mutations, consistent with its bind mode. The inclusion of excess NS1 in the sample substantially improved particle distribution, producing diverse 2D classes with clear density for HRAS (Fig. 1d).

From a small dataset of 854 micrographs, we obtained a low-resolution (~7 Å) reconstruction of the HRAS-3A-BRIL/Fab with NS1. The AlphaFold2-predicted model fit well into the density (Fig. 1e). Alignment of the available HRAS-NS1 crystal structure to the HRAS portion of the model placed NS1 appropriately into the additional density (Fig. 1d). Because a high-resolution crystal structure of HRAS in complex with NS1 is already available, we did not pursue this particular sample further. These results demonstrate the feasibility of the RAS-lollipop scaffold for cryo-EM analysis.

Design and functional validation of the NRAS-lollipops

In the RAS family, KRAS and HRAS have been extensively studied with hundreds of crystal structures, but the G domain of NRAS remains relatively underexplored, with only 15 structures deposited in the Protein Data Bank, and no available solution NMR structures. To extend the RAS-lollipop approach to NRAS, we made several designs and assessed their particle behavior using cryo-EM. All studies were performed using GDP-bound NRAS(Q61K) mutation. In cells, the GTP-bound state of RAS is frequently stabilized by interactions with effector proteins that directly engage the switch regions. We therefore initially focused on the GDP-bound state to characterize the conformational state that is more likely to be present in cells, although in retrospect we acknowledge that NRAS(WT) would have been a better choice. Ongoing studies include NRAS(WT) in the GTP-bound state and NRAS Q61 mutants in complex with mutant-selective monobodies, and will be reported elsewhere. For brevity, we refer to NRAS(Q61K) simply as NRAS hereafter.

Before initiating structural studies, we verified that the BRIL fusion to NRAS did not block effector binding or exhibit reduce protein stability (Fig. S1). Similar to HRAS-BRIL/Fab, NRAS-BRIL/Fab particles exhibited severe preferred orientation, yielding predominantly a single projection and preventing reconstruction (Fig. 2a). Based on our experience with the HRAS lollipops and NS1 monobody, we tested the addition of an NRAS-binding monobody, Mb24. Our previously published study¹⁴ established that Mb24 is a pan-NRAS inhibitor and agnostic about oncogenic mutations, and suggested that Mb24 likely binds an area distal to the switch regions. We proposed that its epitope might overlaps with that of NS1, but the exact binding site was unknown¹⁴. The inclusion of excess Mb24 in the sample resolved the orientation issue and we observed diverse 2D class averages (Fig. 2b).

Figure 2. Optimization and functional validation of the NRAS-lollipop.

(a) 2D class averages of the NRAS-169-BRIL/Fab complex showing severe preferred orientation.

(b) 2D class averages of the NRAS-169-BRIL/Fab complex in the presence of excess Mb24. Multiple particle views were observed.

(c) 2D averages of NRAS-169-BRIL/Fab/Mb24 from a greater number of micrographs.

(d) 2D averages of NRAS-168-BRIL/Fab/Mb24, with the NRAS/Mb24 region is circled.

(e) 2D averages of NRAS-167-BRIL/Fab/Mb24.

(f) Schematics of the functional assay validating that NRAS-BRIL retains biological activity. H358 cells (KRAS(G12C)-dependent) transfected with NRAS(WT)-BRIL or NRAS(Q61L)-BRIL were treated with sotorasib (a G12C specific inhibitor). Only cells expressing oncogenic RAS that is insensitive to sotorasib proliferate.

(g) Growth curves of cells transfected with NRAS(WT) with and without BRIL fusion. The right panel shows data with sotorasib added 20 h after transfection.

(h) Growth curves of cells transfected with NRAS(Q61L) with and without BRIL fusion, shown in the same manner as in g. The data shown in (g) and (h) are representative of two biological replicates.

With additional particle views now available, we assessed how the length of the NRAS-BRIL fusion helix influenced the visibility of the NRAS G-domain. If the helix is too long, it can introduce flexibility and reduce the visibility of the G-domain, as particles are primarily aligned using the Fab/BRIL portion, which contributes most of the mass and signal. If the helix is too short, there is a risk of steric clash between NRAS and the Fab, which might distort the structure and/or destabilize the complex. We started with NRAS-169-BRIL, which fuses NRAS 1–169 directly to BRIL. The fusion helix of this construct has the same length as the HRAS-3A-BRIL, as our HRAS portion terminates at position 166 (Table S1). The 2D class gallery showed clear densities for the Fab and BRIL, but is missing the densities for NRAS or Mb24, likely due to high flexibility of the fusion helix (Fig. 2c). We then tested NRAS-168-BRIL and NRAS-167-BRIL, with the NRAS α5 helix shortened by one or two residues. The shortened constructs show clear and strong NRAS/Mb24 density in some 2D averages (Fig. 2d and e, circled in red). NRAS-168-BRIL provided more diverse orientations and was selected for subsequent experiments.

Having identified an effective design for NRAS lollipop, we also assessed whether NRAS-BRIL fusion retained the biological activity of NRAS using a cellular assay (Fig. 2f). We utilized the H358 lung adenocarcinoma cell line whose proliferation is dependent on the activity of an oncogenic RAS mutant, KRAS(G12C). Inhibition of endogenous KRAS(G12C) with an inhibitor, sotorasib, leads to growth arrest and cell death, which could be rescued with the introduction of an oncogenic RAS that is resistant to sotorasib. This assay is conceptually similar to the widely used BaF cell-based assay for transformation activity¹⁵. We transfected H358 cell lines with different GFP-tagged NRAS-BRIL constructs, where the BRIL is inserted between the G-domain and the C-terminal tail. The tail is essential for proper localization. The GFP signal of NRAS-BRIL localized on the cell membrane similar to NRAS without BRIL, showing the BRIL fusion does not interfere with the membrane association (Fig. S2). Expression of wild-type NRAS (NRAS(WT)) or NRAS(WT)–BRIL failed to rescue growth under sotorasib treatment (Fig. 2g). These results suggesting that the BRIL fusion does not abolish normal regulatory control of NRAS. By contrast, expression of oncogenic NRAS(Q61L) or NRAS(Q61L)-BRIL markedly increased cell viability compared with the NRAS WT controls (Fig. 2h), although NRAS(Q61L)-BRIL showing a lower viability than NRAS(Q61L). We used NRAS(Q61L) mutant in this assay as a stringent functional test of NRAS signaling, as Q61L remains sensitive to GAP-mediated hydrolysis leading to the GDP-loaded state and its nucleotide exchange is stimulated by SOS^16,17. These results demonstrate that NRAS-BRIL retain the intrinsic signaling function and effector engagement in cells.

Cryo-EM structure of NRAS determined with the RAS-lollipop approach

Using the NRAS-168-BRIL construct in complex with BAG2 and Mb24, we collected a dataset of 12,324 micrographs. Final 3D refinement yielded a reconstruction at ~3.28 Å global resolution (Table S2, Fig. S3 & S4). The BAG2 Fab and NRAS-168-BRIL components were well defined in the map and has the highest resolution (Fig. 3a and S4). Local resolution analysis indicated that the Fab and BRIL regions reached <3 Å, while the NRAS G-domain resolved to ~3.3–4.0 Å (Fig. S3). Despite local discontinuities and side chain density missing in some regions, the overall map quality allowed us to build an atomic model excluding Mb24 (Fig. 3a).

Figure 3. Cryo-EM structure of GDP-bound NRAS-168-BRIL.

(a) Cryo-EM density map and ribbon model of the NRAS-168-BRIL/Fab/Mb24 complex at 3.28 Å global resolution.

(b) AlphaFold2-predicted model fitted into the cryo-EM map using the BRIL region using BRIL region. The model is colored according to pLDDT confidence scores.

(c) Comparison of the experimentally determined (left) and the AlphaFold2-predicated (right) fusion helix from residues 151 to 188. Sidechains are omitted for clearance. Residues forming the π-like turn are colored in purple. AlphaFold2-predicted model is colored based on pLDDT confidence score.

(d) Zoomed-in view of the π-like turn spanning residues 161–166 of NRAS. Dashed lines indicate hydrogen bonds involving the backbone atoms.

(e) The comparison of the α4 to α5 region of NRAS-168-BRIL (left) and canonical NRAS (right, PDB: 3CON). There is a three-amino acid register shift on the α5 helix and representative residues are labeled.

(f) Model-map fit of the NRAS-168-BRIL fusion helix. The right panel is the attempted model built using the canonical α5 register colored in green. The threshold level of the map is set to 0.155.

(g) The conformation of the switch regions of the NRAS•GDP structure determined using the RAS-lollipop approach. SW1 and SW2 regions are colored in yellow. Inset: a comparison with SW regions in available NRAS•GDP crystal structures (PDB entries: 3CON, 6ZIO, and 6ZIR).

The RAS-BRIL fusion helix was unambiguously resolved, and, interestingly, revealed an overall fold markedly different from the AlphaFold2 prediction. When the predicted model was fit into the cryo-EM density using BRIL, the NRAS G-domain did not fit the density (Fig. 3b). In the AlphaFold2 model, the α5 helix of NRAS and BRIL helix 1 fusion form a continuous, ideal α-helix. In contrast, our experimental map showed substantial structural rearrangement from an ideal α-helix in the connecting segment between BRIL and NRAS. We observed a π-like turn spanning R161-Y166 and unwinding of the helix between residues 167–166 (Fig. 3c). In this conformation, R161 forms a hydrogen bond with the i+5 residue, Y166, rather than the canonical, i+4 partner, Q165 (Fig. 3d).

We also observed a register shift of the α5 segment, residues 151–160 by three residues relative to the canonical NRAS structures observed in published crystal structures (Fig. 3e). These features are all well supported by the density (Fig. 3f left). A model based on the canonical α5 placement does not fully account for the density in this region. Several side chains are not well accommodated, and, when BRIL helix 1 is maintained in a physically reasonable conformation, the canonical placement leaves a gap in the density at the fusion junction (Fig. 3f right). In contrast, the refined model provides a continuous connection between NRAS α5 and BRIL helix 1 and yields a more consistent interpretation of the map (Fig. 3f left).

The deviation of NRAS α5 from the canonical conformation likely reflects intrinsic plasticity of this region of NRAS. An analysis of available NRAS-GDP crystal structures supports this notion: the B-factors are variable at the first few residues of α5 (151–155), remain low and uniform from residues 156–160, begin to rise from residue 161, and finish high at the C-terminal end from 164–168 (Fig. S5). This trend of B-factors suggests local mobility towards the end of α5. We speculate that structural constraints imposed by the BRIL fusion shifted the energy balance between the canonical and observed conformations.

In the NRAS portion, most of both switch regions were resolved with no additional density nearby, suggesting that the switch regions are not involved in additional protein-protein interactions, as expected (Figs. 3a and S4d). SW1 adopted a conformation distinct from available NRAS crystal structures (RMSD ~3.5–5.7 Å), while SW2 adopted a short helix spanning R68-T74, and the position of SW2 helix was close to the state 2 conformation but still distinct from crystal structures (RMSD 2.0–6.1 Å) (Fig. 3g). These differences likely reflect inherent flexibility that allows these regions to take on variable conformation under spatial restrictions imposed by crystal packing (Fig. S6). The remainder of the NRAS G-domain aligned closely with previously determined RAS family structures, confirming conservation of the small GTPase fold, except for the α5 region as described above. Together, these results establish the first cryo-EM structure of NRAS with SW regions unperturbed by crystal packing and reveal conformational features that are not evident in the previously reported structures. The switch regions are intrinsically flexible, although Mb24 does not directly engage these regions, its binding may indirectly influence their conformational landscape. The observed conformations therefore likely represent one of multiple accessible states. Further classification with larger particle numbers may help to resolve additional conformational variability.

Mapping the Mb24 epitope on NRAS

We next sought to identify the epitope of Mb24. Based on our earlier studies, we initially suggested that Mb24 might bind near the α4-α5 interface, similar to the known binding site of the monobody NS1^13,14. When we examined our cryo-EM reconstruction at lower density thresholds, we observed an additional lobe of density adjacent to the α3-α4 region that was not accounted for by the NRAS-BRIL/Fab model (Fig. 4a). This extra density appeared consistently across multiple half-maps and classification subsets, indicating that it was reproducible and likely contributed by Mb24. Nevertheless, the Mb24 density remained weaker than the NRAS density, with a reported overall resolution of 5.86 Å and local resolution between 5–7 Å (Fig. S7a and b), suggesting partial occupancy or conformational heterogeneity. The polypeptide chain could not be traced well enough to allow de novo model building.

Figure 4. Identification and validation of the Mb24 binding site on NRAS.

(a) Low-threshold cryo-EM density map of NRAS-168-BRIL/BAG2/Mb24. AlphaFold-predicted Mb24 model is fitted into the extra density colored in cyan.

(b) Close up of NRAS-Mb24 interaction site. NRAS specific residues are colored in red. Mb24 density is shown in cyan.

(c) Location of the inferred Mb24-binding site relative to the known NS1-binding site on the α4-α5 interface. NRAS model and AlphaFold-predicted model of Mb24 are fitted to cryo-EM density. NS1-HRAS crystal structure (PDB:5E95) was aligned using the RAS G-domain.

(d) BLI responses at the end of the association phase (300 s) for wild-type NRAS and mutants in which NRAS specific residues in α3 were mutated to their respective KRAS counterparts. Control indicates a sensor with no protein immobilized. Representative data from n = 3.

(e) BLI sensorgrams of NRAS-K135R, pre-mixed NRAS-K135R/Mb24, or Mb24 alone binding to immobilized NS1 (top), and NRAS-K135R, pre-mixed NRAS-K135R/NS1, or NS1 alone binding to immobilized Mb24 (bottom). Each experiment was performed once, but the two experiments, top and bottom, complement each other.

To interpret this feature toward epitope mapping, we generated an AlphaFold2-predicted structure of Mb24 and fitted the predicted model into the map. The predicted model showed high confidence for the conserved β-sandwich scaffold, whereas the confidence for the loop regions was lower as expected (Fig. S6). The best-fitted orientation positioned the CD loop and D strand around the NRAS α3 helix, with the FG loop extend to the α3–α4 interface (Fig. 4b). This spatial arrangement is consistent with our observation that Mb24 acts as a “side binder” in which positions in the CD and FG loops and the D strand were modified to confer target binding¹⁸. In this pose, several NRAS-specific residues on α3 are located close to the Mb24 density and could contribute to the interaction (Fig. 4b). Among them, A91 has most contact with the density and appears critical. In HRAS and KRAS, residue 91 is a glutamic acid rather than an alanine whose larger size and negative charge of glutamic acid would likely destabilize this interface. Because the local resolution of Mb24 is limited, we sought to experimentally validate this model using targeted mutagenesis and biolayer interferometry (BLI). NRAS-specific residues on α3 were individually substituted with their KRAS counterparts. A91E abolished Mb24 binding, but S87T, N94H, and L95H retained robust interaction (Fig. 4d). Importantly, all the mutants, including A91E, maintained binding to RAF-RBD (Fig. S8), confirming that their global fold and effector interface were intact. Thus, position 91 is the main determinant of Mb24 specificity to NRAS. Together, the biochemical and mutagenesis data provide strong support for the proposed Mb24 binding model.

Interestingly, this deduced epitope is away from the known NS1 epitope at the α4-α5 interface (Fig. 4c). In our previous study, we tentatively inferred that Mb24 and NS1 bind to overlapping epitopes from a competitive binding assay¹⁴. To further validate the Mb24 epitope, we re-examined the binding of Mb24 and NS1 in a more rigorous manner. R135 is a determinant of NS1 specificity to HRAS/KRAS, where NRAS has a lysine. NS1 binds an NRAS variant harboring the K135R mutation¹⁴. K135R does not affect Mb24 binding to NRAS. We performed a competition binding experiment using BLI. Instead of simply testing binding of NS1 and Mb24 in a sequential manner as in our previous study, NRAS was pre-mixed with a high concentration of Mb24, and this mixture was then applied to biosensor tips immobilized with NS1. Robust binding was still observed in this experiment as well as a complementary experiment in the reciprocal orientation, indicating that Mb24 and NS1 can bind simultaneously to NRAS, although their order of association seems to affect binding kinetics observable using BLI (Fig. 4e).

Collectively, these results indicate that Mb24 recognizes a NRAS-specific surface centered on A91 of the α3 helix, providing a molecular basis for its isoform selectivity. Moreover, this binding mode of Mb24, together with its ability to inhibit NRAS-mediated signaling, suggests that NRAS can be inhibited by targeting the α3 helix probably through the disruption of its nanoclustering¹⁴.

Mb24 acts as an additive that reduces orientation bias

While testing the RAS-lollipop on HRAS, we serendipitously found that the NRAS-specific monobody Mb24 alleviated the preferred-orientation problem even though it does not bind HRAS¹⁴. As shown above, cryo-EM data collected on the HRAS-3A-lollipop/Fab showed severe orientation bias, with particles adopting almost exclusively one view and preventing 3D reconstruction (Fig. 1b). Unexpectedly, when a high concentration (~15 μM) of Mb24 was included during grid preparation under the same conditions, the resulting micrographs showed improved particle distributions. 2D class averages revealed additional orientations with clear secondary-structure features, and the HRAS G-domain became visible in several classes (Fig. 5a, circled in red).

Figure 5. Mb24 reduces orientation bias without directly binding to the particles.

(a) 2D class averages of the HRAS-166–3A-BRIL/BAG2 sample prepared in the presence of high-concentration Mb24 (~15 μM) showing multiple orientations of the particles.

(b) A proposed mechanism by which Mb24 alleviates preferred orientation by coating the air-water interface.

(c) 2D classes averages of thyroglobulin sample prepared in the presence of different additives. Orientations that are underrepresented in the no-additive condition but enriched upon addition of additives are outlined in red.

S.K. is a co-founder, receives consulting fees and hold equity in Aethon Therapeutics; is a co-founder and holds equity in Revalia Bio; has received research funding from Aethon Therapeutics, Argenx BVBA, Black Diamond Therapeutics, and Puretech Health; has received consulting fees from Eisai, all outside of the current work. A.K. is the spouse of S.K. and shares the conflicts listed above. The remaining authors declare no conflict of interest.

This observation suggests that Mb24 have a surfactant-like property. Because the target-binding surface of Mb24 includes one β-sheet characteristic of “side binders”, the Mb24 molecule is more hydrophobic on one surface of the β-sandwich and more hydrophilic on the other. We speculate that this biased surface hydrophobicity may allow it to accumulate at the air-water interface and shield protein particles (Fig. 5b). Proteins with similar amphipathic properties have been reported to alleviate preferred-orientation problems in cryo-EM^19,20. This finding broadens the applicability of the RAS-lollipop strategy, because Mb24 eliminates the need to include a monobody that binds the protein of interest to resolve the preferred orientation problem.

To assess whether this effect extends beyond RAS lollipops, we tested Mb24 with thyroglobulin (Tg), an independent system with known preferred orientation. In single-particle cryo-EM analysis, Mb24 improved angular sampling (cFAR = 0.21, SCF = 0.91) relative to the no-additive condition (cFAR = 0.09, SCF = 0.72), although to a lesser extent than CHAPS (cFAR = 0.33, SCF = 0.88), while a second monobody, NS1, showed less effect (cFAR = 0.18, SCF = 0.89) (Fig. 5c). These results suggest that Mb24 can modulate particle orientation without binding to the system of interest, and that this effect may extend to other monobody scaffolds.

Discussion

In this study, we used protein engineering to extend the established BRIL-fusion strategy to water-soluble proteins, small GTPase RAS. Using the RAS-lollipop approach, we obtained a 3.3 Å structure of GDP-bound NRAS, resolving both switch regions in the absence of crystal lattice contacts or a binder involving the region. The requirement for our strategy is that the protein of interest terminates with a helix to which BRIL can be fused via a contiguous helix. Clearly, it can be applied to other small GTPases that share the molecular architecture with RAS. We envision that it is generalizable for studying other small soluble proteins that possess terminal helices. Potential candidates beyond small GTPases include E3 ubiquitin ligase adaptors^21,22 and protein phosphatases²³. These proteins are slightly larger than RAS and their terminal helices often project away from the active sites. Determining the structures of these proteins using the lollipop approach could reveal transient conformations and additional binding pockets for drug discovery.

The lollipop design addresses an important limitation of conventional binder-assisted structural biology, in which the fiducial or crystallization chaperone raised to directly bind the target protein often binds to a functionally important surface of the target protein²⁴. Functional surfaces are inherently primed for molecular recognition, whereas other regions of the protein surface are typically evolved to avoid strong interactions. As a result, antibodies, monobodies, nanobodies, and other binders frequently bind functional surfaces used for natural ligands, with buried surface areas often exceeding 800 Å^{2 25,26}. Although such a binder is a powerful tool for identifying a druggable site and a druggable conformation of the target protein, it imposes strong constraints on the epitope conformation. By contrast, the RAS-lollipop approach can keep functional surfaces free and accessible to their native partners, as demonstrated in our cell-based studies (Fig. 2f–h), as well as synthetic binders (Fig. 1e and 4a–c).

The success of the RAS-lollipop strategy depends on the engineering of the connecting helix that rigidly holds the RAS-BRIL fusion. Our investigation revealed that even a minor increase in linker length can significantly impact the rigidity of the connecting helix and hence the rigidity of the overall complex. Specifically, the insertion of four alanine residues in the HRAS-BRIL fusion introduced sufficient flexibility to blur the signal of the HRAS G-domain in 2D averages (Fig. 1c) but we were able to rigidify the helix by attaching the NS1 monobody that binds α5, a part of the linker helix (Fig. 1d–e). When extending this platform to other proteins, the maximum allowable linker length will likely depend on the rigidity of the target protein’s terminal helix and how tightly it packs against its core.

Currently, the PDB contains cryo-EM structures for ≥35 different proteins solved using the BRIL/Fab system. Most of these studies applied the scaffold to GPCRs, typically fusing two helices of BRIL to two transmembrane (TM) helices of the GPCR to create two anchor points that rigidify the linkage^27–29. Other successful cases involved fusing BRIL to loop regions³⁰, again using two anchor points. In cases where high resolution was achieved with a single-helix fusion¹¹, it is possible that the detergent micelle that is not modeled in the final structure enhanced the rigidity of the linker helix and/or served as the second anchor point for rigidifying the fusion protein. We envision that a second contact can be introduced between RAS and BRIL, for example by using computational design, which may enable structure determination of the RAS portion without attaching a monobody.

From a data processing perspective, we found that the strong signal from the larger BRIL-Fab scaffold (~62 kDa) tended to dominate particle alignment. Achieving a high-resolution reconstruction required a precise masking strategy for local refinement to focus alignment on the NRAS portion. At approximately 20 kDa, the NRAS G-domain may be near the lower size limit at which sufficient signal is available for reliable particle alignment and refinement using current single-particle cryo-EM methodologies. This challenge also stems from the flexibility of RAS-BRIL fusion proteins. Design improvement to nearly completely rigidify the RAS-BRIL junction should eliminate the need for masking.

A recurring challenge we encountered with the BRIL/Fab scaffold was the severe preferred orientation of the particles. Although we tested multiple linker lengths that alter the relative orientation of the RAS G-domain to the BRIL/Fab scaffold, preferred orientation persisted. These data suggest that this problem is likely caused by the large, flat surfaces of the BRIL/Fab scaffold, which can dominate the interactions with the grid and air-water interface. Because the RAS portion make only minor contribution to the overall shape and surfaces, changing the relative position of RAS within the particle cannot strongly affect the preferred orientation of the entire particle. Our serendipitous discovery of the surfactant-like property of Mb24 enabled us to overcome this challenge (Fig. 2, 5). The mechanism underlying this property needs to be further characterized in future studies. We speculate that, as a small and stable protein that is easy to produce, further engineering of Mb24 may enable the development of off-the-shelf protein additives to mitigate preferred orientation for diverse cryo-EM projects.

During the preparation of this manuscript, a related approach for enabling cryo-EM studies of KRAS by extending α5 using a coiled-coil fusion was reported³¹. Both approaches establish promising starting points for generalizable frameworks that expand the accessible size range of cryo-EM, enabling studies of small GTPases and other small soluble proteins while minimally perturbing their functional surfaces. Such structures will in turn advance mechanistic studies and drug discovery. Furthermore, we envision that, the lollipop design could be combined with other protein engineering approaches to enhance its impact. While our current design uses a 62 kDa BRIL/Fab scaffold, adding the Fab-binding nanobody in the Megabody form³² could provide a larger (>100 kDa) fiducial marker, with the potential of altering the surface interaction between the particle and the grid to avoid preferred orientation issues and help solve structures of even smaller targets.

Methods

Protein Expression and purification

All proteins (HRAS-Q61L-A3-BRIL, HRAS-Q61L-A4-BRIL, NRAS-Q61K-168-BRIL, NRAS-Q61K-169-BRIL, monobodies, RAF-1 RBD (residues 51–131)) were produced with N-terminal tags consisting of His₆, Avi-tag and a TEV protease recognition site using the pHBT vector³³. pHBT-Mb(NRAS_24) has been deposited to Addgene (ID 255052). The proteins were produced in E. coli BL21(DE3). To produce biotinylated proteins, E. coli BL21(DE3) harboring the pBirA plasmid was used as a host and grown in the presence of 50 μM biotin. Expressed proteins were purified using Ni-Sepharose columns (GE Healthcare) via gravitational flow, followed by dialysis in Tris-buffered saline (TBS, 50 mM Tris-Cl pH 7.5 containing 150 mM NaCl) for all non-RAS proteins. TBS containing 5 mM MgCl2 was used for RAS proteins. Samples were further purified using a Superdex 75 size exclusion column on an AKTA Pure Systems (GE Healthcare). BAG2 Fab was produced using an expression vector containing a synthetic gene (modified from a gift from Nate Traaseth¹¹) in the 55244 strain (ATCC) and purified using a protein G affinity column, as described previously³⁴.

Nucleotide exchange of RAS

Purified RAS proteins used in binding experiment were prepared by diluting stock protein (typically 20–250 uM) 25 times with 20 mM Tris-Cl buffer pH 7.5 containing 5 mM EDTA, and 1 mM nucleotide (GDP or GTPγS). Samples were incubated a 30°C for 30 mins. MgCl₂ was then added to a final concentration of 20 mM and the solution was further incubated on ice for at least 5 minutes prior to use.

Biolayer interferometry analysis

BLI experiments were performed on an Octet Red96 instrument (Sartorius). Biotinylated proteins were immobilized on streptavidin biosensor tips. Analytes were diluted in Analyte buffer (50 mM Tris, 150 mM NaCl, 0.005% Tween 20, 5 mM MgCl2, 0.5 % BSA, 1 uM Biotin, pH 7.5). BLI signals were analyzed using Octet Data analysis Software (Sartorius).

Cell Culture

NCI-H358 was validated externally via IDEXX and maintained in RPMI-1640 high glucose with sodium pyruvate, L-glutamine (Thermo), supplemented with 10% FBS (Gemini Bio-products) and antibiotics-antimycotics (Gibco). The absence of mycoplasma contamination was confirmed periodically using a PCR-based mycoplasm testing kit (LiLIF).

H358 cells (50000 cells per well) were seeded in a 96 well plate (Corning, #353072) for 1 day prior to transfection. On the day of transfection, media were replaced with antibiotic-free complete media (RPMI-1640 supplemented with 10% FBS). The cells were transfected with vectors encoding eGFP-fused RAS and RAS-BRIL using lipofectamine 3000 (Thermo Fisher Scientific) according to the manufacturer’s protocol. The cell culture plate was maintained at 37 °C and 5% CO₂. and monitored utilizing an Incucyte S3 (Sartorius). Data were processed and analyzed using Incucyte S3 Software.

Cryo-EM Sample Preparation

For each sample, purified RAS-BRIL and BAG2 were mixed with 1.5:1 molar ratio at room temperature for 30 minutes, and the RAS-BRIL/BAG2 complex was purified using a Superdex75 10/300 column (Cytiva). The complex was concentrated to ~3 μM. When indicated, monobodies were added to a final concentration of ~15 μM and incubated RT 30 minutes before grid freezing. A sample (3 μl) was applied on in-house prepared gold foil grids with 1.2/1.3 300 mesh. Right before use, the grids were glow-discharged for 30 s in the PELCO easiGlow Glow Discharge Cleaning System (Ted Pella, Inc.). All the grids were frozen in liquid ethane using a Vitrobot Mark IV (Thermo Fisher Scientific) using the following settings: no waiting time, blot time 2.5 s, blot force 5, chamber temperature 4 °C, and chamber humidity 100%.

For thyroglobulin experiment, 10 mg of bovine thyroglobulin (Sigma-Aldrich catalog No. T9145) was dissolved in 600 μl PBS and purified using a Superose 6 increase column (Cytiva). The purified fraction was concentrated to 7 mg/mL. A sample (3 μl) was applied on in-house prepared gold foil grids with 1.2/1.3 300 mesh. 0.2%(w/v) CHAPS or 15 μM of different monobodies were added right before grid freezing.

Cryo-EM Data Collection and Processing

In the screening sessions, the cryo-EM data were collected on a Talos Arctica with a K3 camera (Gatan) at the NYU Cryo-EM Laboratory using Leginon 3.6^35,36 at a magnification of 45,000x, corresponding to a super-resolution pixel size of 0.4335 Å (for a nominal pixel size of 0.867 Å after binning by 2). Movies were collected at a total dose of ~50 e⁻/Ų. Intermediate frames were recorded every 0.04 seconds for a total of 40 frames per micrograph. Images were collected at a nominal defocus range of 1.6 to 2.7 μm. Ice thickness was determined as described^36,37.

For the NRAS-168-BRIL/Fab/Mb24 complex, cryo-EM data were collected at the NYU Cryo-EM Laboratory using Leginon 3.6^35,36. Data was acquired on a Titan Krios (Thermo Fisher Scientific) with a K3 camera (Gatan) at a magnification of 105,000×, corresponding to a super-resolution pixel size of 0.4125 Å (or a nominal pixel size of 0.825 Å after binning by 2). Zero-loss images were obtained using an energy filter slit width of 20 eV. Movies were collected at an accumulated dose of 47.22 e⁻/Ų. Intermediate frames were recorded every 0.04 s for a total of 50 frames per micrograph. All images were collected at a nominal defocus range of 0.5 to 2.5 μm. Ice thickness was determined as described^36,37. On-the-fly processing was performed using MotionCor2 v 1.5³⁸ and CTFFIND4 v 4.1.13³⁹ under control of Appion⁴⁰.

Data were further processed by importing the micrographs in cryo-SPARC⁴¹. First, a small subset of 834 micrographs were used. Particles were picked using Blob Picker on all micrographs with a contrast transfer functions (CTF) fit of better than 6.0 Å, followed by particle extraction (300 pixels box size binned to 150 pixels), and multiple subsequent steps of 2D classification and selection for particle cleaning. The 2D-cleaned particles were used for Topaz training⁴². The trained picking model was used to pick the full dataset of 12,324 micrographs. Topaz picked particles were duplicate-removed and cleaned using 3 rounds of 2D-classification. Then the particles were used to perform a heterogeneous Ab initio reconstruction to generate three 3D maps (classes). The maps obtained and the associated particles were processed further by performing multiple steps of heterogeneous refinement using decoy 3D volumes to both refine the maps and further clean the particle subsets. 3D classifications were also performed to clean up the particles. After these passages, the remaining particles were re-extracted without binning (300 pixels box size) and used to perform nonuniform refinement⁴³. A total of 64,660 particles was used, giving a map with an overall resolution of 3.50 Å. Global and Local CTF refinements were performed. Then, reference-based motion correction was performed on the final set of particles. A final step of local refinement was performed, using a mask that encloses map densities corresponding to NRAS and Mb24, resulting in the final map with an overall resolution of 3.28 Å. Focused 3D classification and 3D variability analysis of the Mb24 region was performed, but did not reveal distinct conformational classes or well-defined variability. The mask exploited for local refinement processes was realized using UCSF Chimera software⁴⁴. An AlphaFold predicted atomic model of the NRAS G-domain was used as a starting model and was subjected to real-space refinement in PHENIX 1.18.2⁴⁵, including simulated annealing, to reduce model bias. The model was refined against the experimental cryo-EM density map, and regions of discrepancy were manually rebuilt in Coot⁴⁶ based on the observed density. The final structures was validated using PHENIX and PDB deposition tools, obtaining the statistics summarized in Table S2. Cryo-EM maps and structures were visualized and imaged using UCSF Chimera X⁴⁷ and Pymol (Schrodinger, LLC.).

Thermal denaturation assay

Protein thermal stability was assessed using a SYPRO Orange-based thermal denaturation assay. Purified HRAS or HRAS-BRIL fusion protein was diluted to 5 μM in TBS. SYPRO Orange dye (Life Technologies S-6650) was added from the manufacturer’s stock solution to a final of 1:250 dilution. Samples were prepared in a total volume of 25 μL in triplicated in 384-well PCR plates, sealed with optical film.

Fluorescence was monitored during a temperature ramp from 25 °C to 95 °C at a rate of 0.5 °C/min using a CFX384 thermocycler. SYPRO Orange fluorescence was recorded using the “FRET” settings. Melting temperatures (Tm) were determined from the peak of the first derivative curve of fluorescence with respect to temperature. Data were analyzed using Prism9 and Excel, and reported Tm values represent the mean of 3 independent measurements.

Highlights.

• Overcomes the size limitation of cryoelectron microscopy applied to small proteins

• Protein engineering to stably link two proteins via terminal helices

• Enables cryo-EM studies of RAS small GTPases and their switch regions

• Epitope mapping of a NRAS-specific monobody inhibitor

• Monobody additive minimizes preferred orientation in cryo-EM samples

Data Availability

The cryo-EM map and model have been deposited to the Electron Microscopy Database (EMD74594) and the Protein Data Bank (9ZR7).