Quantitative interaction analysis permits molecular insights into functional NOX4 NADPH oxidase heterodimer assembly

Protein–protein interactions critically regulate many biological systems, but quantifying functional assembly of multipass membrane complexes in their native context is still challenging. Here, we combined modeling-assisted protein modification and information from human disease variants with a minimal-size fusion tag, split-luciferase–based approach to probe assembly of the NADPH oxidase 4 (NOX4)-p22phox enzyme, an integral membrane complex with unresolved structure, which is required for electron transfer and generation of reactive oxygen species (ROS). Integrated analyses of heterodimerization, trafficking, and catalytic activity identified determinants for the NOX4-p22phox interaction, such as heme incorporation into NOX4 and hot spot residues in transmembrane domains 1 and 4 in p22phox. Moreover, their effect on NOX4 maturation and ROS generation was analyzed. We propose that this reversible and quantitative protein–protein interaction technique with its small split-fragment approach will provide a protein engineering and discovery tool not only for NOX research, but also for other intricate membrane protein complexes, and may thereby facilitate new drug discovery strategies for managing NOX-associated diseases.

Protein-protein interactions (PPIs) 4 are vital for the regulation of biological systems. Current methods in unperturbed cellular environments rely mostly on antibodies, fluorescent tags, or fragment complementation. Pitfalls are often antibody specificity and sensitivity, tag-induced changes in protein localization and function, or the absence of dynamic quantification. Understanding and interfering with dimer formation of multipass receptors, transporters, or enzyme complexes containing numerous interhelical contacts may be feasible by reversible split-luciferase comple-mentation assays (1), but it is often impeded by the sensitivity of the functional conformation to tag-induced perturbations. Newly developed approaches such as reversible protein-fragment complementation coupled with stability-optimized bright luminescence may permit detailed analysis of assembly and disassembly of these complexes in living cells (2).
Assembly of multipass integral membrane enzymes or receptors by homo-or heterodimerization is governed by the folding and packing of their transmembrane segments, thereby enabling PPIs within the membrane, which directly participate in stabilization of the complex and signal transduction across lipid bilayers. Reactive oxygen species producing NADPH oxidases form membrane-bound complexes composed of 1-5 minimally required components. The ability to transfer electrons from NADPH across the membrane to molecular oxygen resides in several distinct features of the catalytic NOX domain but requires also in all NADPH oxidases (except for NOX5; NOX1-4 and DUOX1-2) heterodimerization with specific integral membrane proteins (p22 phox , DUOXA1-2). Heterodimerization is required for NOX/DUOX stabilization and exit from the endoplasmic reticulum and may direct intracellular trafficking (3). In vivo mechanistic analysis of heterodimerization is particularly difficult for NADPH oxidases, as the complex structure is unresolved, and enzyme activity is easily impaired by amino acid changes, insertions, or additions. If interference with the ability to heterodimerize or to disrupt the formed complex could be achieved, it might lead to a new class of NOX inhibitors.
The NADPH oxidase NOX4, a primarily transcriptionally regulated, constitutively active enzyme that generates H 2 O 2 (4), is a promising drug target for fibrotic and metabolic disease (5). Heterodimer formation of NOX4 (6 transmembrane ␣-helices (TMHs)) with p22 phox (4 TMHs) in the endoplasmic reticulum is thought to occur after incorporation of two bis-histidinecoordinated low-potential hemes, followed by dimer trafficking to the plasma membrane or, in some cell types, to intracellular membrane compartments. To date, NOX mutagenesis/chimera studies or rare human loss-of-function variants (CYBA (p22 phox ), CYBB (NOX2), and NOX1) provide the only information about particular residues or structural elements of the NOX1-4 heterodimers required for NOX activity. NOX4-dependent H 2 O 2 generation is contingent on heterodimerization, NADPH/FAD binding, electron transfer across the membrane via heme to molecular oxygen, and intramolecular conversion of superoxide to H 2 O 2 by an unidentified structure-induced mechanism. Reactive oxygen species generation and NADPH/ oxygen consumption are the only quantifiable readouts for the catalytic activity of all NOX/DUOX enzymes, impeding molecular insight into unique contacts at PPI interfaces that could be exploited for selective NOX isoform targeting. Here, we describe dynamic quantification of NOX4-p22 phox assembly in a cellular and functional context, leading to discovery of hot spot residues and interaction surfaces required for heterodimerization.

Development of a NOX4-p22 phox PPI assay
Understanding NOX4 heterodimerization is hindered by lack of structural information. We and others have attached tags at the enzyme's N or C termini. NOX4 activity tolerates some short (V5) (6) or large (GFP) (7) N-terminal tags, albeit only certain sequences permit maturation and can be visualized. The partner protein p22 phox can be modified at the C terminus without loss of expression or its ability to heterodimerize with NOX2 (8). Ambasta et al. (9) used C-terminal NOX4-yellow fluorescent protein and p22 phox -cyan fluorescent protein fusions for FRET studies. We and others postulated that the first intracellular loop (B-loop) and the penultimate C-terminal region of NOX4 form an interface required for enzymatic activity (10 -12), a scenario that would exclude NOX4 C-terminal additions. Indeed, a C-terminal GFP fusion markedly reduced NOX4 expression and H 2 O 2 generation ( Fig.  1) independently of linker length or linker composition (not shown). Thus, quantitative and functional PPI analysis of NOX4-p22 phox and other NOX isoform complexes will require reducing the size of fusion tags as recently accomplished by the reversible split-luciferase complementation assay NanoBiT (SB-11aa, LB-157aa) (2) and strategically placing these fusions without disrupting any of the structural features required for NOX4 function.
SB incorporation into the second cytosolic loop (D-loop) of NOX4 (SB3 and SB4) was not tolerated (Fig. S1B), supporting D-loop involvement in heterodimerization (6). Replacement of penultimate C-terminal NOX4 sequences with SB abolished H 2 O 2 production (Fig. S1B), whereas insertion of SB between FAD 2 and NADPH 1 (SB5) permitted enzyme trafficking and catalytic activity. This indicates that the SB5-containing com-

NOX4 complex assembly
plex was properly formed but did not provide the appropriate distance for SB/LB interaction, as SB5 is localized on the opposite side in the folded model (Fig. 2F, SB5 surface red). This observation reveals the critical importance of a multipronged approach when designing and validating interaction pairs. Comparison of NOX4-SB5 with the adjacent NOX4-SB6 identified a novel and unique region in the NOX4 C terminus that can be structurally modified without altering enzyme activity. Docking of the NanoLuc structure (PDB entry 5IBO) to the SB1-NOX4/p22-LB2 model indicates an approximate distance of 40 Å between the NOX4 N terminus and the p22 phox C terminus, necessitating long linkers to accommodate the shorter 25-Å distance between the SB and LB termini. The model suggests that C-terminal fusions, such as the combination NOX4-LB1/p22-SB1, will form a complex. This interaction was confirmed, albeit, as expected from the earlier GFP studies (Fig. 1), with markedly reduced catalytic activity, although protein expression and cell-surface localization were comparable with catalytically active SB1-NOX4/p22-LB2 (Fig. 2, G-J). In contrast to N-terminal GFP-NOX4 (7), introduction of LB at the NOX4 N terminus (LB2-NOX4/p22-SB1) resulted in significantly reduced PPI and H 2 O 2 generation (Fig. 2, K and L), indicating that SB versus LB orientation and linker length as well as the choice of fusion protein are critical aspects for NanoBiT association and NOX4 activity.

The NOX4-p22 phox PPI may serve as a high-throughput screening assay
High-throughput screening for small molecule modulators targeting PPI interfaces can identify lead compounds with increased selectivity for disrupting signaling circuits (13). Existing NOX inhibitors are pan-flavoenzyme inhibitors or radical scavengers or NOX inhibitors lacking isoform selectivity. We reported that complex formation between NOX4 and p22 phox differs from other NOX isoforms (7), opening the door for selective PPI inhibitors. In the case of constitutively active NOX4, enzyme activity could alter PPI readout by oxidation (overexpression output 30 -100 nmol of H 2 O 2 /h/10 6 cells). However, the SB1-NOX4/p22-LB2 luminescence signal was not affected by released or exogenously supplemented H 2 O 2 (Fig. 3). After miniaturization to 1536-well format, selected compounds currently used as NOX inhibitors were tested in the NOX4 PPI assay. As expected, diphenyleneiodonium (DPI), which targets flavin via irreversible phenylation (14), or GKT compounds, which can inhibit NOX1-4 (15), but also p22 phoxindependent NOX5 and DUOX enzymes (16, 17), did not alter PPI (Fig. S3). Some compounds reduced at high concentrations the luminescence signal of the NOX4/p22 phox PPI but decreased then also the signal generated by an unrelated PPI pair or by the cell-free NanoLuc substrate control. Thus, cer- Figure 2. Development of a PPI assay for the NOX4-p22 phox complex combined with functional analysis. A, topological scheme of the NOX4/p22 phox complex with selected interaction pairs. B and G, determination of PPI by NanoLuc luminescence (RLU percentage of control). C and H, catalytic activity (H 2 O 2 release). D and I, protein expression (immunoblotting). E and J, cell-surface localization using anti-NOX4 antibody to an extracellular domain. F, NOX4-p22 phox homology model in surface presentation depicting location of NanoBiT assembly in a cartoon representation (SB1-NOX4/p22-LB2; red arrow indicates SB5 location). K and L, PPI (luminescence) and H 2 O 2 generation of SB versus LB N-terminally tagged NOX4 matched with full-length versus truncated p22 phox . Error bars, S.E. Analysis for G and H was an unpaired, two-tailed Student's t test; B, C, K, and L, one-way ANOVA with post hoc Tukey's multiple-comparison test. n.s., not significant; **, p Յ 0.01; ***, p Յ 0.001; ****, p Յ 0.001.

NOX4 complex assembly
tain compounds used as nonspecific NOX inhibitors (at up to 50 M) caused a higher false positive hit rate in the NanoBiT assay than reported previously for a LOPAC 1280 library (10 M) (18), emphasizing the need for robust counterscreens.

Integrated analysis of NOX4 and CYBA mutations identifies PPI hot spot residues
Naturally occurring or induced missense mutations constitute an ideal tool for PPI validation and the mapping of heterodimer interfaces. Introduction of mutations that affect NOX4 cytosolic domains and are primarily connected to NOX catalytic activity, such as T94D (NOX4-specific Bloop) (10), P437H (conserved NADPH 1 -binding site), or R523D (conserved GRP sequence), did not alter PPI (Fig. 4, A  and B). Heterodimerization was also not affected in SB1-NOX4 H222Q/C226V , a double mutation in the second extracellular loop (E-loop) (Fig. 4B), yet this mutant failed to translocate in the majority of cells (Fig. 4, E and F). In HEK293 cells, which retain the majority of NOX4 on intracellular membranes and cannot be used for trafficking analysis, single H222Q and C226V mutations generated intracellular superoxide instead of H 2 O 2 (19). We show here that replacing certain E-loop residues located close to TMH5 impairs trafficking of the assembled heterodimer, which could be due to reduced binding to ER chaperones, such as calnexin (20). Mutation of histidines serving as heme ligands in TMH3 and TMH5 (H119L and H119L/H194L) rendered NOX4 catalytically inactive and reduced PPI significantly (Fig. 4, A-D). In accord, treatment of SB1-NOX4/p22-LB2expressing cells with the heme synthesis inhibitor succinyl acetone reduced luminescence and H 2 O 2 generation (not shown). These results indicate that NOX4 follows the NOX2 paradigm, where heme incorporation is a prerequisite for heterodimerization (21). Nonstabilized NOX4 His-119 and His-119/194 mutants failed to translocate to the plasma membrane (Fig. 4, E and F).
Missense variants in CYBA (p22 phox ) cause the immunodeficiency disorder chronic granulomatous disease due to defective assembly of the multimeric NOX2 complex. In A22 0 chronic granulomatous disease patients, CYBA variants lead to loss of both p22 phox and NOX2 protein, as heterodimerization is required for the stability of both proteins in neutrophils (22). These rare CYBA variants may affect all NOX family members that rely on p22 phox as a dimerization partner (NOX1-4). Because the CYBA p.Gly24Arg variant in TMH1 is rather common (http://structure.bmc.lu.se/idbase/CYBAbase) (48), 5 it was introduced into p22-LB2 and co-expressed with SB1-NOX4. H 2 O 2 production and heterodimerization were significantly reduced, and ϳ50% of the SB1-NOX4/p22-LB2 G24R complex was trapped in the ER in epithelial cells (Fig. 5, A-E). Mutant p22 phox mice (Cyba nmf333 , p22 phox Y121H) display a Nox2 and Nox3 loss-of-function phenotype (23), whereas NOX4 activity is not affected by this mutation (7), suggesting increased structural flexibility of the NOX4/p22 phox complex and/or differences in the interaction surface with p22 phox . SB-LB tagging slightly reduced luminescence and H 2 O 2 production of SB1-NOX4/p22-LB Y121H (20 -25%), but the majority of NOX4 translocated to the cell surface (Fig. 5, A-E). Inducing a ␤ turn in TMH4 by replacing Tyr-121 with proline (p22 phox Y121P) is predicted to disrupt the complex. In accord, heterodimerization, catalytic activity, and cell-surface expression of SB1-NOX4/p22-LB2 Y121P were markedly reduced. A close-up of the NOX4-p22 phox interface shows that p22 phox forms an extensive network of hydrophobic interactions with NOX4 via Leu-Leu and Leu-Val contacts. The clustering of numerous hydrophobic contacts at the heterodimer interface will create a leucine zipper-like structure. Owing to the position of Tyr-121 in the last p22 phox TMH, a proline substitution will alter the helix structure by breaking hydrophobic contacts and the hydrogen bond between NOX4 Ile-182 and p22 phox Thr-64 (Fig. 5F). Arg-90 in p22 phox forms a hydrogen bond with the p22 phox Gly-24 backbone residue and NOX4 Thr-163. Thus, mutation of Gly-24 to arginine will destabilize both the p22 phox fold due to repulsive charge and the hydrophobic interaction network between p22 phox and NOX4 (Fig. 5F). Hence, the p22 phox TMH4 penultimate amino acid residues discriminate between NOX isoforms and affect selectively heterodimerization with p22 phox , whereas changes in the p22 phox Gly-24 region (TMH1) influence heterodimerization with NOX4, NOX2, and probably NOX1/3.

Discussion
NADPH oxidases constitute a prime drug target for the treatment of diseases associated with chronic superoxide/H 2 O 2 overproduction, such as fibrotic, cardiovascular, or metabolic diseases (16). Almost all currently available small molecule inhibitors are not NOX isoform-selective and will interfere with redox signaling in a nonspecific manner. Inhibitors of the NOX electron transport chain are usually nonspecific due to high identity or homology of the involved domains and amino acid residues, although a NOX2-selective, NADPH-competitive compound was described recently (24). Therapeutic disruption of the NOX membrane protein complex could lead to isoform selective compounds but is difficult to achieve, as the structure of the complex remains unresolved. A partial structure of a NOX5 homolog expressed in the cyanobacterium Cylindrospermum stagnale has been reported (25), but NOX5-type oxidases differ substantially from NOX1-4 and DUOX1-2 enzymes, as heterodimerization is not required. Using knowledge accumulated by NOX protein engineering, disease-associated missense mutations, and NOX complex modeling, we developed a NOX4/p22 phox PPI assay that distinguished between catalytic versus structural alterations, identified two hot spots in p22 phox (TMH1 and TMH4), and revealed NOX4 loop involvement (loops D and E) in complex formation and/or trafficking. Hence, this PPI assay will not only support a deeper understanding of the catalytic center of the NOX4 enzyme, but can also assist in functional analysis of human diseaseassociated variants. Development of this assay expands our repertoire for assessing protein-protein interactions, as until now immunoprecipitation of NOX or p22 phox from detergent-containing extracts was the only means for analyzing complex formation. Immunoprecipitation depends on the antibody selectivity and affinity, is highly variable, and is not quantitative. 5 Please note that the JBC is not responsible for the long-term archiving and maintenance of this site or any other third party hosted site.

NOX4 complex assembly
By permitting quantitative monitoring of conformational dimerization, the PPI assay constitutes also a novel tool for multiplexed high-throughput screening to identify candidate NOX inhibitors targeting either complex assembly or complex disruption or inducing allosteric modulation. The minimal steric burden inherent in this reversible, sensitive PPI system conveys a distinct advantage for intricate complex assemblies with multiple sensitive structural features. Our preliminary test of this NOX4 PPI assay indicates that current compounds used to inhibit NADPH oxidases do not target NOX4-p22 phox het-

NOX4 complex assembly
erodimerization. The utility of this PPI for NOX4 inhibitor discovery can be improved by inducible stable incorporation of the PPI pair into NOX4/p22 phox -deficient cells. Establishing similar assays for other NOX isoforms may reveal additional determinants for assembly, improve prediction of human disease variant effects, and further support development of NOX isoform-selective PPI drug candidates. Requiring only a minimal-size fusion tag on one of the protein interaction partners was critical for maintaining a catalytically active, properly assembled oxidase enzyme. Although not tolerated by NOX4, insertion of the small split fragment into loop or surface regions of multipass membrane proteins will provide a distinct advantage for real-time association studies in cells. This approach will probably be enabling for structure-function studies and drug discovery platforms of a wide variety of membrane protein complexes of unknown structure.   (1-172)-GS linker-LB). GS (sequence GSS-GGGGSGGGGSSG and multiple variations, 4 -30 aa) and AA linkers (5 aa) were used; when compared, the linker length did not substantially alter the performance of the SB-LB interaction pairs. Point mutants SB1-NOX4 (T94D, P437H, R523D, H222Q/C226V, H119L, and H119L/H194L) and p22-LB2 (G24R, Y121H, and Y121P) were generated by PCR site-directed mutagenesis and verified by sequencing. Mutations were also introduced into pcDNA3.1-NOX4 WT, pcDNA3.1-V5-V5-NOX4, or pcDNA3.1-p22 phox WT. NOX4 was also cloned into pEGFP-C1 (28-aa linker) and pEGFP-N1 (8-aa linker).

Cell culture and transfection
NOX/DUOX-deficient COS7 cells stably expressing p22 phox (26) were maintained in Dulbecco's modified Eagle's medium with 10% fetal bovine serum. Most experiments were performed with NOX/DUOX/p22 phox -deficient NCI-H661 cells maintained in RPMI 1640 medium with 10% fetal bovine serum. Transient transfections of H661 and COS-p22 phox cells were performed using Dreamfect Gold (Oz Biosciences) or Lipofectamine 3000 (Invitrogen) according to the manufacturers' instructions. Briefly, cells were plated at a density of 5 ϫ 10 4 cells/well in a 12-well plate and transfected with a total of 375 ng of DNA (2:1 ratio of NOX4/p22 phox DNA) per well for 48 h. Medium was changed the day after transfection.

Measurement of H 2 O 2 production
H 2 O 2 release was measured by monitoring homovanillic acid (HVA) fluorescence as described previously (4). Briefly, cells were washed in PBS with Ca 2ϩ and Mg 2ϩ and incubated at 37°C for 1 h in HVA solution (100 mM HVA, 4 units/ml horseradish peroxidase in PBS with Ca 2ϩ and Mg 2ϩ ). The reaction was terminated by the addition of Stop buffer (0.1 M glycine, 0.1 M NaOH, pH 12, and 25 mM EDTA in PBS). Fluorescence was read on a Biotek Synergy HT plate reader (320-nm excitation, 420-nm emission). Fluorescence readings were converted into nmol of H 2 O 2 based on a H 2 O 2 standard curve and normalized to cell lysate protein content.

Luciferase assay for protein-protein interactions
The Nano-Glo live cell assay kit (Promega) was used as follows. Cells were seeded at 7.5 ϫ 10 3 cells/well in a white, clearbottom 96-well plate 24 h before transfection (10 ng each of NOX4 and p22 phox , combined with 480 ng of pBluescript DNA). After 48-h transfection, medium was removed and replaced with 100 l of Opti-MEM medium for 1 h at 37°C. The Nano-Glo reagent was prepared as per the manufacturer's instructions and added to each well immediately before the luminescence reading was taken. Luminescence was measured at 10 -20-s intervals for 15 min on a Berthold Centro 960 LB and reported as relative light units (RLU). For quantitative comparison of SB-LB interactions, the peak values at the 2-3min time point were used.

Fluorescence microscopy
H661 cells were seeded at 2.5 ϫ 10 4 cells/well on poly-Llysine-coated glass coverslips and transfected as described. Cells were fixed in 3% paraformaldehyde, followed by permeabilization with 0.1% Triton X-100 and blocking in PBS containing 1.5% BSA and 3% goat serum. Cells were stained with antibodies directed to NOX4, calnexin (MAB3126, Chemicon), or GM130 (610822, BD Transduction Laboratories) antibodies. Secondary antibodies were goat anti-rabbit Alexa Fluor 488 and goat anti-mouse Alexa Fluor 647 (Thermo Fisher). Nuclear stain was obtained using DAPI (Sigma). Images were taken on an Olympus Fluoview FV1000 confocal microscope (Olympus) using a ϫ60 oil objective. Images were processed using ImageJ (version 1.48) and Photoshop CC.

Protein modeling, docking, and molecular dynamics simulation
A three-dimensional model of the SB1-NOX4/p22-LB2 complex was constructed in multiple steps. Ab initio NOX4 and p22 phox secondary structure prediction was performed using the ROSETTA-based Robetta webserver (28). The model of the transmembrane domains of NOX4 was generated using the homology modeling program Modeler 9v14 (29). To identify the most suitable templates, a BLAST search of PDB was performed with the NOX4(1-305) sequence and cyanobacterium Cylindrospermum stagnale Nox5 (PDB 5O0T), and four cytochrome c oxidoreductase complexes with the highest homology served as initial templates. The NOX4(1-305) sequence was aligned to the templates using Clustal (http://www.ebi.ac.uk/ Tools/msa/clustalo/) 5 (49) and alignment was adjusted manually, taking into account the secondary structure prediction and previously published data (6, 7, 10, 30 -34). Modeling was per-NOX4 complex assembly formed with the default parameters using the "allHmodel" protocol to include hydrogen atoms, the "HETATM" protocol to include heme with restricting heme-binding histidine residue side-chain distance, and a thorough molecular dynamics MD optimization and refinement protocol. The secondary structure of each individual loop was further optimized using the "loopmodel" protocol (35). As per the procedure described above, the p22 phox (1-121) transmembrane domains were generated by combining previously available structural information and secondary structure prediction (7,32,36,37). The transmembrane domain of NOX4 was docked to its cytosolic C-terminal NADPH/FAD binding domain (PDB entries 3A1F and 5O0T (38)) using HADDOCK (39,40). Docking was performed with most of the parameters set to default using the web server version of HADDOCK with a Guru interface including heme, FAD, and NADPH in the docking run. Semiflexible segments and active residues were manually defined (transmembrane, FAD, or NADPH domain each represented as one segment), whereas passive residues were defined automatically. HADDOCK proposed only one probable cluster of similar models, and the model with the lowest Van der Waals, electrostatic, and desolvation energy was chosen. The chosen model also had the shortest FAD-heme distance. The transmembrane p22 phox domains (4-TMH model for best fit) were then docked to NOX4 using HADDOCK, as described above. Using Modeler, the p22 phox C terminus, SB1, and LB2 (PDB entry 5IBO) were added, and its folding was simulated using a NAMD 2.11 (http://www.ks.uiuc.edu/Research/namd/) 5 (50) plug-in in the VMD version 1.9.2.27 program (http://www.ks.uiuc.edu/ Research/vmd/) 5 (51) for 10 ms. Intermediate and final structures were evaluated in PyMOL.

Testing of compounds in a miniaturized plate format
The transfection protocol described below is the result of optimization of multiple parameters, including cell number to transfect and response to linearity, DNA versus Lipofectamine ratio, duration of protein expression, assay tolerance to DMSO, and statistical quality (Z value).
Transient transfection of SB1-NOX4/p22-LB2 pairs-NCI-H661 cells were maintained using RPMI 1640 medium (Thermo Fisher Scientific) with 10% fetal bovine serum (Bioconcept, Allschwil, Switzerland). All cell cultures were maintained inside a humidified incubator at 37°C and 5% CO 2 . Transfection using Lipofectamine 3000 (Thermo Fisher Scientific) was performed according to the manufacturer's protocol with a DNA/Lipofectamine ratio of 1:2.22 (w/v). A transfection enhancer, the 3000 enhancer reagent (1:2 DNA/reagent (w/v)), was used along with the Lipofectamine 3000 transfection reagent for all transfections. Transfection reagents were diluted in Opti-MEM TM (Thermo Fisher Scientific) and added to the cell suspension before cell distribution to a 1536-w plate with black walls and transparent bottom, tissue culture-treated (792091-191, Greiner bio-one, St. Gallen, Switzerland). Typically 3 ng of plasmid DNA for 500 cells per 4 l were transferred to each well of the 1536-well plate (with a 1:1 ratio of SB1-NOX4/p22-LB2 (w/w)). The transfection protocol for KRAS-LB/cRAF G12C-SB is identical to the one used for SB1-NOX4/p22-LB2.
Nano-Glo assay-Detection of NanoLuc Luciferase was performed following the manufacturer's protocol (Live Cell Method, Promega, Madison, WI). Compounds were tested by using a "disruptive mode" protocol. In these experiments, compounds are evaluated for their ability to disrupt an already formed complex following expression of NOX4 and p22 phox pairs. To this end, cells were washed with HBSS containing 20 mM Hepes (pH 7.4) 20 -24 h after cell transfection and incubated in 2 l of HBSS containing 20 mM Hepes (pH 7.4). 40 nl of compounds in serial concentrations (Novartis Compound archive as 10 mM solutions in 100% DMSO or from powders diluted in 90% DMSO/water (v/v) diluted in 100% DMSO) were then added using an acoustic dispenser (Echo550, Labcyte Inc., Sunnyvale, CA) and incubated for another 1 h at 37°C in a humidified atmosphere with 5% CO 2 . The assay is very tolerant to DMSO, as this solvent could be used up to a 2.7% (v/v) final concentration without any negative impact on the signal amplitude. After incubation, plates were equilibrated at room temperature for 5-10 min. NanoLuc activity was detected by adding 2 l of furimazine (NanoLuciferase substrate, a coelenterazine derivative) diluted in the buffer provided by the manufacturer or in PBS (final dilution of 1:100 in the well). Luminescence was measured for 10 s on an FDSS700 kinetic reader with 1-s integration times. Data are expressed as average luminescence counts over 10 s of four replicate wells.

Statistical analysis
All experiments were performed at least three times in triplicate (the error bars indicate mean Ϯ S.E., n Ն 3). Graphs for NanoBiT interaction assays are representative examples of at least three independent experiments. Bar graphs for NanoBiT interaction assays represent the RLU values obtained at the 2-3-min time point and include at least three separate experiments. For comparison of cell-based assays with transiently expressed mutants, the internal SB1-NOX4/p22-LB2 control was set to 100% for quantification. When indicated, a one-way ANOVA with post hoc Tukey's multiple-comparison test was performed. Statistical significance is indicated on the bar graphs: n.s., not significant; *, p Յ 0.05; **, p Յ 0.01; ***, p Յ 0.001; ****, p Յ 0.001, comparing samples as indicated.