The effects of signaling proteins are often determined by the subcellular location and kinetics of their conformational changes. To fully understand cell physiology, we must therefore characterize these conformational changes in living cells and animals. This has been accomplished using fluorescent protein biosensors 1,2 , but because biosensors must interact with native cell components they are all burdened by some level of cell perturbation. Perturbation can be reduced by enhancing the sensitivity of the biosensor (to reduce biosensor concentration) and by using biosensor designs that minimize unwanted interactions with native proteins. Cell perturbation by biosensors is becoming an important limitation as new studies strive to image multiple biosensors simultaneously, or as researchers reach for low abundance proteins 3 .
Diverse and creative biosensor designs have provided access to many different target proteins 2 , but these designs perturb cell physiology to different degrees and via different mechanisms. One prominent group of biosensors is based on "affinity reagents" (AR) that bind selectively to a specific conformation of the target protein [3][4][5][6][7][8][9][10][11][12][13][14][15] . AR have been derived from naturally occurring downstream proteins, antibody fragments, or libraries of engineered scaffolds 5,8,16,17 .
To visualize conformational change, fluorophores can be positioned on the AR and the target protein to generate FRET when the AR binds to the 'activated' protein conformation. In other iterations, binding reunites fragments of luciferase or fluorescent proteins, or enables interaction with a fluorogenic substrate 14,18,19 . For FRET, the AR-target interaction can be intramolecular, with fluorophores usually positioned somewhere along a chain connecting the AR to the target 9,10,15,[20][21][22][23][24] , or intermolecular, with fluorophores on the separate AR and target 3,7,25 . Dual chain and single chain biosensors each possess unique advantages, but the approach reported here is most useful for dual chain biosensors (see discussion), and here we will exemplify the approach by developing dual chain biosensors for GTPases.
Importantly, the AR of a dual chain biosensor can interact with both endogenous and fluorescent versions of the target protein. Because the endogenous protein bears no fluorophore, the interaction perturbs normal physiology without yielding any FRET. In addition, the fluorescent version of the target protein can interact with nonfluorescent endogenous ligands, again perturbing cell physiology without producing any valuable information. To ameliorate these problems, we modified dual chain biosensors by introducing complementary point mutations in the interacting surfaces of the fluorescent target protein and the AR (Figure 1). The goal was to reduce the interactions of biosensor components with endogenous proteins, while maintaining their interactions with one another (Figure 1). We refer to these modified binding surfaces as orthogonal interfaces, because of their reduced cross-reactivity with endogenous proteins.
We focused on improving biosensors for the GTPases Rac1, Cdc42 and Rap1, because we have extensive experience quantifying cell perturbation by these biosensors 3,25 . Several laboratories have produced FRET biosensors for these GTPases 7,10,15,21,24 , and almost all could potentially benefit from orthogonal interface design. The ARs for these biosensors are fragments of naturally occurring downstream effectors that have evolved tight, specific binding to the activated GTPase conformation 6,[26][27][28] . Unfortunately, this means that the fluorescent AR also interacts with endogenous, nonfluorescent GTPase, and that the exogenous fluorescent GTPase binds to nonfluorescent, endogenous effector proteins (Figure 1). This generates dominant negative effects (from fluorescent AR binding up GTPase) and overexpression artefacts (from exogenous fluorescent GTPase).
In this study we generated the orthogonal biosensor interfaces by introducing either "knobinto-hole" mutations or charge swap mutations, strategies used previously to assemble bispecific antibodies [29][30][31] and to rewire GTPase signaling networks using an orthogonal GTPase/GEF pair 32 .
We employed multi-state design protocols in the molecular modeling software Rosetta, using energy calculations to find mutations that maintain the desired interactions while destabilizing unwanted interactions 33 . For the GTPases Rac1, Cdc42 and Rap1, "knob-into-hole" mutations generated AR and GTPases that bind well to each other and have weakened interactions with endogenous proteins. For the most successful targets, Rac1 and Cdc42, we showed that the new orthogonal biosensors reported activation dynamics like those reported previously. They were activated by upstream guanine exchange factors, and there was a substantial reduction in dose-dependent perturbation of cell morphodynamics, which is regulated by these GTPases.
Finally, a computational model was used to examine the relationship between desirable FRET interactions and off-target interactions, as a function of affinity between biosensor components and endogenous proteins. This indicated that the orthogonal interfaces could produce normal concentrations of GTPase-effector complex (i.e. downstream effects) over a broader range of biosensor concentrations than native interfaces.
To quantitatively probe how orthogonal biosensors with redesigned interfaces would impact GTPase signaling and biosensor sensitivity, we produced a simple computational model based on mass-action kinetics. The model incorporated endogenous GTPases and effectors, together with fluorescent exogenous GTPase and fluorescent affinity reagents derived from effectors, and therefore competing with endogenous effectors. We evaluated the concentration of effective cellular signaling complexes (endogenous or fluorescent GTPase bound to endogenous effector), inactive complexes (endogenous or fluorescent GTPase bound to the AR, so unable to signal downstream), and complexes composed of fluorescent GTPase and fluorescent AR (contributing to FRET and reporting on GTPase activity)(Figures 1A and 1B). We assumed that any complexes which included endogenous effector were functional signaling complexes, for endogenous or fluorescent GTPase. We also assumed that inactive GTPase would not bind to either endogenous effectors or biosensor AR. Given that it would be nearly impossible to model each potential effector independently, we assumed a single generic effector whose affinity was one parameter of the model. This enabled us to vary the fraction of GTPase that is active, and the concentration of biosensor components, to examine effects on the concentration of GTPase-effector complexes.
For wild type biosensors, the model showed a narrow window of biosensor concentrations that did not alter the number of functional complexes by greater than 5% (Figure 1C). If we used a theoretical orthogonal GTPase-AR interface that produces 10-fold reduction in the affinity of cross partners (endogenous GTPase with affinity reagent, and fluorescent GTPase with endogenous effector), while keeping the orthogonal affinity the same as wild type, the window where less than 5% deviation occurred was considerably expanded (Figure 1C), and perturbation was reduced across all concentration ranges. The effect was greatest for lower levels of GTPase activation, which most closely resemble previously reported native GTPase behaviors 34 .
We also assessed whether the orthogonal interface could improve biosensor signal-to-noise by increasing the proportion of biosensor components that were bound to each other rather than to endogenous cell components. For normal biosensors, expression levels must be chosen to produce sufficient FRET despite nonproductive interactions. Orthogonal interfaces could increase the proportion of expressed biosensor components that actually produce FRET, enabling lower expression levels and/or higher signal/noise. The model predicted that the orthogonal biosensors produced increased numbers of FRET complexes relative to similar concentrations of conventional biosensors (Figure 1D), although the effect was strongly dependent on the level of GTPase activation. This model can be used by others to see how the KD of their biosensor interactions could impact performance, so we produced an interactive Graphical User Interface (GUI) (See Methods and Supplemental Figure 1).
We set out to test experimentally whether we could produce functional orthogonal interfaces, and whether they would reduce cell perturbation. Mutations were introduced in thoroughly characterized 'wild type' biosensors from our laboratory 3,7,25 . Protein design simulations were performed for the GTPase Rac1 partnered with the CRIB domain from PAK1, and the GTPase Rap1 with its effector RalGDS. In each case, the goal was to identify a small number of mutations that would preserve a favorable interaction between the mutant pair but destabilize binding of biosensor components to wild type sequences. Multiple sets of simulations were performed for each complex, each simulation focusing on a different region of the interface (see Methods). For Rac1 and PAK1, two separate knob-into-hole designs were predicted to be favorable. In the first design a bump was created with the mutation P21F in PAK1, and the matching hole was made with the Rac1 mutation Y23A (Figure 2A, B). In the second design a hole was introduced in PAK1, F24G, and a bump, T25F, on Rac1. These regions of the Rac1/PAK1 interface are conserved in the interaction of Cdc42 with PAK1, so we also carried the Y23A design into testing Cdc42 partnered with PAK1 (P21F).
The best-scoring orthogonal sequences for Rap1 and RalGDS were based on charge swap mutations rather than knob-into-hole mutations. Like many GTPase effectors, RalGDS binds to Rap1 by strand pairing with the switch 1 region of Rap1 (Figure 2C, D). Sidechain-sidechain interactions at the interface occur either "above" or "below" the paired -strands. Above the strand pair (the side closer to the GTP binding site), D33 and D38 from Rap1 interact with K52 and K32 from RalGDS respectively. Below the strand pair, E37 from Rap1 interacts with R20 from RalGDS.
The multi-state design simulations selected a variety of alternate charge swap mutations for these salt bridge pairs, including D33K from Rap1 partnered with K52E from RalGDS (Figure 2C, D), E37R partnered with R20E, and D38K partnered with K32E. In addition to the charge swap designs, we also tested a knob-into-hole design, where I36H (Rap1) was predicted to interact favorably with S28W (RalGDS). Both the charge swap and knob-into-hole designs were carried into the lab for experimental testing.
For an initial screen, we compared the amount of FRET generated by orthogonal biosensor designs, compared to wild type biosensors (Supplemental table 1). From this screen, we had success with multiple knob-into-hole designs: Rac1 designs Y23A/P21F and T25F/F24G, the Cdc42 design Y23A/P21F, and the Rap1 design I36H/S28W. Of the charge swap designs, Rap1:D33K with RalGDS K52E showed some promise. However, this pair had reduced interaction between the orthogonal partners. Further screening of alternative residues at position 52 led us to K52Q, which showed >95% wildtype FRET (Supplemental table 1). This mutation and the above knob-into-hole designs were carried forward.
To function as biosensors, the new molecules had to accurately respond to upstream regulators. Although the mutations were selected to avoid known binding interfaces for upstream interactions 35 , we experimentally tested for the ability of upstream guanine nucleotide exchange factors (GEFs) to activate the new biosensors. Wildtype and orthogonal biosensors were expressed in HEK-293t cells and titrated with specific GEFs (Dbl, Vav2, Trio, and Tiam for Rac1; Dbl, Vav2, Intersectin, and Asef for Cdc42; and EPAC, RapGEF1 and CalDAG for Rap1). The corrected FRET/Donor ratio was plotted against GEF transfected for wildtype and orthogonal biosensors for each GEF. Although there was cell to cell variation in the amount of GEF expressed, these experiments were averaged across large fields of cells, and gave consistent ratios of average GEF fluorescence relative to DNA transfected 36 . The orthogonal Rac1 and Cdc42 biosensors incorporating the Y23A/P21F mutation were activated to wildtype levels (Figure 3A and 3B). However, Rac1 T25F/F24G and both orthogonal Rap1 biosensors showed reduced sensitivity to upstream regulators (Supplemental figure 2A-C). Rap1 I36H/S28W appeared quite insensitive to GEF activation (Supplemental figure 2B), while Rap1 D33K/K52Q (Supplemental figure 2A) and Rac1 T25F/F24G (Supplemental figure 2C) both required higher levels of GEF to reach maximal activation. Rap1 D33K/K52Q, for example, required 3-fold more RapGEF1 that did wild type Rap, suggesting that the D33K mutation reduced the GEF affinity of orthogonal Rap1 (Supplemental figure 2A). These experiments highlight the necessity of testing new orthogonal designs with a panel of regulators to ensure that the biosensors are responding correctly. In live cell imaging experiments, we also saw that these biosensors (Rac1 T25F and Rap1 D33K) reported a different distribution of GTPase activation than wildtype biosensors (data not shown).
We therefore restricted our further studies to the Cdc42 and Rac1 (Y23A) biosensors.
To more precisely determine the orthogonal mutations' effects on binding affinities, we purified wildtype and orthogonal versions of the GTPases and affinity reagents. The binding affinities were then determined for different combinations of GTP and AR using Biolayer interferometry (BLI).
We compared wildtype GTPase:wildtype AR, wildtype GTPase:orthogonal AR, orthogonal GTPase:wildtype AR, and the fully orthogonal GTPase:AR pair (Figure 3C). For Cdc42, the wildtype GTPase bound to the wildtype AR with an affinity of 65 nM, and this dropped to 2.68 M when paired with the orthogonal AR. The orthogonal GTPase bound to wildtype AR with a low affinity of 1.3 M but a much higher affinity of 385 nM to the orthogonal AR (Figure 3C). Although the orthogonal:orthogonal affinity was five-fold lower than the wildtype:wildtype affinity, the orthogonal:orthogonal affinity was still much higher than the wild type:orthogonal interactions, particularly those involving the orthogonal AR, where the orthogonal GTPase had almost 10-fold higher affinity than the wildtype GTPase. Similar data was obtained with Rac1 (Figure 3C); The wildtype interaction was strongest, followed by the orthogonal interactions, and off-target interactions were lowest. Given that the Rac1 did not have as high an affinity for PAK1 as did Cdc42, we were concerned that the orthogonal interaction might not produce enough signal for imaging. Therefore, in an attempt to bring orthogonal Rac1 binding to wild type levels, other residues were tested at the originally selected positions (i.e., P21 => F, L, Y, M, or W, and Y23 => G, V, S, or A) but these produced no improvement in FRET over the Y23A/P21F pair (Supplemental Table 1).
Finally, we incorporated the experimentally derived binding affinities in the computational model described above. This predicted that both the Rac1 and Cdc42 orthogonal designs would permit a larger range of biosensor concentrations with minimal perturbation (Figure 3D). The Cdc42 biosensor was predicted to be essentially unperturbing across all concentrations examined, while the Rac1 sensor was only perturbing at the highest concentrations.
We previously used Rac1 and Cdc42 biosensors to investigate GTPase activation in migrating mouse embryonic fibroblasts (MEFs). We therefore compared the GTPase activation reported by orthogonal and wildtype biosensors in these cells. We addressed the fact that varying ratios of GTPase to AR expression could affect the extent to which GTPase activation generated FRET.
A fixed GTPase:AR ratio was generated using a single construct to express the two biosensor chains. Tandem viral skip sequences were used between the genes to ensure that there was complete separation. Fluorescence activated cell sorting (FACS) showed that the optimized skip sequences efficiently expressed two separate chains (Supplemental figure 3A). In these experiments we did not overexpress biosensor components as we did in the experiments of Figure 3, where the cells were essentially used as 'cuvettes' to examine the interactions of overexpressed components. At the lower expression levels, the GTPase:AR ratios produced by the skip sequences generated good signal/noise for the Cdc42 biosensor, but produced weaker activation readouts for Rac1, presumably because of the lower affinity of the AR for Rac1. By using two AR genes for each GTPase gene in the Rac1 expression construct (with skip sequences between the three genes) a clear readout of Rac1 activation was obtained. The increased AR levels were able to generate FRET levels similar to those seen with the wild type biosensor (Supplementary Figure 3B).
When we expressed the Cdc42 and Rac1 orthogonal biosensors in MEFs we saw activation distributions qualitatively similar to those of wildtype biosensors (Figure 4A) 7 . For more precise comparison we quantified the correlation of Rac1 and Cdc42 activity with edge velocity (Figure 4B), as done previously 7 . Briefly, a series of discrete windows were drawn around the edge of the cell. Biosensor activity and cell edge velocity were calculated for each window, and these were tracked together as the cell underwent protrusion and retraction. This allowed us to map cell edge velocity relative to nearby biosensor activity using correlation analysis. The correlations of wild type and orthogonal biosensors were similar. However, there were slight shifts in the timing of maximum correlation. This could be due to experimental variation, lower interaction between biosensor and endogenous effectors, or reduced cell perturbation by the orthogonal biosensor.
With the orthogonal biosensor, the correlation between edge velocity activity extended further into the cell, perhaps due to higher signal/noise ratio for the orthogonal biosensor.
Using our previously developed quantitative assays of motility/morphological dynamics 7 , we asked whether the orthogonal interfaces reduced cell perturbation (Figure 4C). At levels of expression where wild type biosensor was shown to produce normal protrusion behavior 3,25 , there was little difference between wild type and orthogonal sensors. However, when we looked at cell populations with higher wild type biosensor expression levels, where we could detect a significant perturbation of morphological dynamics (Figure 4D), both the Rac1 and Cdc42 biosensors decreased edge velocity, and the Rac1 biosensor also reduced the frequency of protrusion. We examined the morphological dynamics of cells with orthogonal biosensors at these higher expression levels. Unlike the wild type biosensors, we saw no significant perturbation. In these studies, biosensor cells were compared to control cells expressing only a membrane-associated fluorophore for edge tracking (Figure 4 D, E). These data showed that the orthogonal biosensors are less perturbing than wild type biosensors.
In this paper we showed that altering the interface between components of dual chain FRET biosensors could reduce cellular perturbation and improve signal/noise. Molecular modeling of protein interfaces predicted mutations that could minimally perturb FRET-producing interactions between biosensor components, while reducing unwanted interactions between the biosensor and endogenous proteins. Because the orthogonal interfaces increased the proportion of biosensor molecules involved in productive interactions, less biosensor was needed to achieve acceptable signal, and concentration-dependent perturbation of cell morphodynamics was substantially reduced. Importantly, the mutations did not perturb the ability of the biosensors to accurately reflect activation by upstream guanine exchange factors. Qualitative and quantitative measures of activation indicated that the orthogonal biosensors accurately reported GTPase dynamics. Our computational model indicated that orthogonal biosensors should produce normal cell behavior over a wider range of concentrations than wild type biosensors. Some published GTPase biosensors are based on a single chain, with the AR connected to the GTPase, rather than the two chain design of the Rac1 and Cdc42 biosensors developed here.
Connecting the AR and GTPase is valuable because it simplifies expression and image analysis 37 . The single chain design may confer many of the advantages of the orthogonal mutations described here, because the AR is more likely to interact with the target protein than with separate endogenous GTPases. However, in single chain biosensors, the intramolecular interaction increases the effective affinity of the AR for the GTPase. This has been shown to substantially alter the inactivation rate of the GTPase 14 . A difficulty with dual chain designs is the variations in relative expression levels of the two separate chains. Here we showed that an optimized viral skip sequence could be used to produce ratios of the two chains with tight tolerance. Dual chain designs are easier to engineer. Fluorophore orientation and linkers are optimized to produce maximal FRET upon AR-target binding. This is simpler than incorporating two fluorophores in a linked AR -target molecule such that there is a sufficient difference between the FRET of two conformations, and the activated conformation induces high FRET. It is also easier to maintain interactions with endogenous ligands when the GTPase does not have to carry a chain bearing an affinity reagent interacting with the GTPase.
The interface mutations reduced perturbation more effectively for Cdc42 than for Rac1. This may have been due to differences in their effector binding sites. Cdc42 effectors almost all use a canonical CRIB domain to bind to the same site on Cdc42, so that our mutations are broadly effective. In contrast, Rac1 has more diverse binding interactions. Crystal structures of Rac bound to effectors, such as p67phox and PRK1, have demonstrated binding away from the sites of our orthogonal mutations 38,39 . These additional effector interaction sites could be targets for additional orthogonal mutations.
There are now myriad biosensor designs, which enable study of diverse protein structures. Some designs respond to endogenous, unaltered target proteins while others, like the designs addressed here, require modification of the target. Where modification is required, orthogonal interfaces can be useful to reduce off target binding. Techniques to engineer protein interfaces will continue to improve 40,41 . Advances in understanding allosteric networks will help target sites that are affected by active site conformational changes but are far from ligand binding sites [42][43][44] .
We hope that the approach described here can provide access to currently inaccessible molecules, including low abundance proteins. It can enable multiplexed imaging with reduced toxicity and provide GTPase biosensors that can be used together with optogenetic and chemogenetic approaches. 𝑑[𝐺 𝑛 * ] 𝑑𝑡 = -𝑘 𝑜𝑓𝑓 𝑛 [𝐺 𝑛 * ] + 𝑘 𝑜𝑛 𝑛 [𝐺 𝑛 ] -𝑘 + 𝑛𝑛 * [𝐺 𝑛 * ][𝐸] + 𝑘 - 𝑛𝑛 * [𝐺 𝑛 * 𝐸] -𝑘 + 𝑛𝑓 * [𝐺 𝑛 * ][𝐴] + 𝑘 - 𝑛𝑓 * [𝐺 𝑛 * 𝐴] 𝑑[𝐺 𝑛 ] 𝑑𝑡 = -𝑘 𝑜𝑛 𝑛 [𝐺 𝑛 ] + 𝑘 𝑜𝑓𝑓 𝑛 [𝐺 𝑛 * ] -𝑘 + 𝑛𝑛 [𝐺 𝑛 ][𝐸] + 𝑘 - 𝑛𝑛 [𝐺 𝑛 𝐸] -𝑘 + 𝑛𝑓 [𝐺 𝑛 ][𝐴] + 𝑘 - 𝑛𝑓 [𝐺 𝑛 𝐴] 𝑑[𝐺 𝑓 * ] 𝑑𝑡 = -𝑘 𝑜𝑓𝑓 𝑓 [𝐺 𝑓 * ] + 𝑘 𝑜𝑛 𝑓 [𝐺 𝑓 ] -𝑘 + 𝑓𝑛 * [𝐺 𝑓 * ][𝐸] + 𝑘 - 𝑓𝑛 * [𝐺 𝑓 * 𝐸] -𝑘 + 𝑓𝑓 * [𝐺 𝑓 * ][𝐴] + 𝑘 - 𝑓𝑓 * [𝐺 𝑓 * 𝐴] 𝑑[𝐺 𝑓 ] 𝑑𝑡 = -𝑘 𝑜𝑛 𝑓 [𝐺 𝑓 ] + 𝑘 𝑜𝑓𝑓 𝑓 [𝐺 𝑓 * ] -𝑘 + 𝑓𝑛 [𝐺 𝑓 ][𝐸] + 𝑘 - 𝑓𝑛 [𝐺 𝑓 𝐸] -𝑘 + 𝑓𝑓 [𝐺 𝑓 ][𝐴] + 𝑘 - 𝑓𝑓 [𝐺 𝑓 𝐴] 𝑑[𝐸] 𝑑𝑡 = -𝑘 + 𝑛𝑛 * [𝐺 𝑛 * ][𝐸] + 𝑘 - 𝑛𝑛 * [𝐺 𝑛 * 𝐸] -𝑘 + 𝑛𝑛 [𝐺 𝑛 ][𝐸] + 𝑘 - 𝑛𝑛 [𝐺 𝑛 𝐸] -𝑘 + 𝑓𝑛 * [𝐺 𝑓 * ][𝐸] + 𝑘 - 𝑓𝑛 * [𝐺 𝑓 * 𝐸] -𝑘 + 𝑓𝑛 [𝐺 𝑓 ][𝐸] + 𝑘 - 𝑓𝑛 [𝐺 𝑓 𝐸] 𝑑[𝐴] 𝑑𝑡 = -𝑘 + 𝑛𝑓 * [𝐺 𝑛 * ][𝐴] + 𝑘 - 𝑛𝑓 * [𝐺 𝑛 * 𝐴] -𝑘 + 𝑛𝑓 [𝐺 𝑛 ][𝐴] + 𝑘 - 𝑛𝑓 [𝐺 𝑛 𝐴] -𝑘 + 𝑓𝑓 * [𝐺 𝑓 * ][𝐴] + 𝑘 - 𝑓𝑓 * [𝐺 𝑓 * 𝐴] -𝑘 + 𝑓𝑓 [𝐺 𝑓 ][𝐴] + 𝑘 - 𝑓𝑓 [𝐺 𝑓 𝐴] 𝑑[𝐺 𝑛 * 𝐸] 𝑑𝑡 = 𝑘 + 𝑛𝑛 * [𝐺 𝑛 * ][𝐸] -𝑘 - 𝑛𝑛 * [𝐺 𝑛 * 𝐸] 𝑑[𝐺 𝑛 𝐸] 𝑑𝑡 = 𝑘 + 𝑛𝑛 [𝐺 𝑛 ][𝐸] -𝑘 - 𝑛𝑛 [𝐺 𝑛 𝐸] 𝑑[𝐺 𝑛 * 𝐴] 𝑑𝑡 = 𝑘 + 𝑛𝑓 * [𝐺 𝑛 * ][𝐴] -𝑘 - 𝑛𝑓 * [𝐺 𝑛 * 𝐴] 𝑑[𝐺 𝑛 𝐴] 𝑑𝑡 = 𝑘 + 𝑛𝑓 [𝐺 𝑛 ][𝐴] -𝑘 - 𝑛𝑓 [𝐺 𝑛 𝐴] 𝑑[𝐺 𝑓 * 𝐸] 𝑑𝑡 = 𝑘 + 𝑓𝑛 * [𝐺 𝑓 * ][𝐸] -𝑘 - 𝑓𝑛 * [𝐺 𝑓 * 𝐸] 𝑑[𝐺 𝑓 𝐸] 𝑑𝑡 = 𝑘 + 𝑓𝑛 [𝐺 𝑓 ][𝐸] -𝑘 - 𝑓𝑛 [𝐺 𝑓 𝐸] 𝑑[𝐺 𝑓 * 𝐴] 𝑑𝑡 = 𝑘 + 𝑓𝑓 * [𝐺 𝑓 * ][𝐴] -𝑘 - 𝑓𝑓 * [𝐺 𝑓 * 𝐴] 𝑑[𝐺 𝑓 𝐴] 𝑑𝑡 = 𝑘 + 𝑓𝑓 [𝐺 𝑓 ][𝐴] -𝑘 - 𝑓𝑓 [𝐺 𝑓 𝐴] Here, 𝑘 + and 𝑘 -are the rates of formation and splitting of the corresponding complexes, and 𝑘 𝑜𝑛 and 𝑘 𝑜𝑓𝑓 are the rates of the activation and deactivation of the corresponding endogenous and tagged GTPases. In the steady state, the model equations can be written in terms of 𝐾 𝑑 values and activity fractions as [𝐺 𝑛 * ][𝐸] [𝐺 𝑛 * 𝐸] = 𝑘 - 𝑛𝑛 * 𝑘 + 𝑛𝑛 * = 𝐾 𝑑 𝑛𝑛 * , [𝐺 𝑛 ][𝐸] [𝐺 𝑛 𝐸] = 𝑘 - 𝑛𝑛 𝑘 + 𝑛𝑛 = 𝐾 𝑑 𝑛𝑛 , [𝐺 𝑛 * ][𝐴] [𝐺 𝑛 * 𝐴] = 𝑘 - 𝑛𝑓 * 𝑘 + 𝑛𝑓 * = 𝐾 𝑑 𝑛𝑓 * , [𝐺 𝑛 ][𝐴] [𝐺 𝑛 𝐴] = 𝑘 - 𝑛𝑓 𝑘 + 𝑛𝑓 = 𝐾 𝑑 𝑛𝑓 , [𝐺 𝑓 * ][𝐸] [𝐺 𝑓 * 𝐸] = 𝑘 - 𝑓𝑛 * 𝑘 + 𝑓𝑛 * = 𝐾 𝑑 𝑓𝑛 * , [𝐺 𝑓 ][𝐸] [𝐺 𝑓 𝐸] = 𝑘 - 𝑓𝑛 𝑘 + 𝑓𝑛 = 𝐾 𝑑 𝑓𝑛 , [𝐺 𝑓 * ][𝐴] [𝐺 𝑓 * 𝐴] = 𝑘 - 𝑓𝑓 * 𝑘 + 𝑓𝑓 * = 𝐾 𝑑 𝑓𝑓 * , [𝐺 𝑓 ][𝐴] [𝐺 𝑓 𝐴] = 𝑘 - 𝑓𝑓 𝑘 + 𝑓𝑓 = 𝐾 𝑑 𝑓𝑓 , [𝐺 𝑛 * ] [𝐺 𝑛 ] = 𝑘 𝑜𝑛 𝑛 𝑘 𝑜𝑓𝑓 𝑛 = 𝐾 𝑎𝑐𝑡 𝑛 , [𝐺 𝑓 * ] [𝐺 𝑓 ] = 𝑘 𝑜𝑛 𝑓 𝑘 𝑜𝑓𝑓 𝑓 = 𝐾 𝑎𝑐𝑡 𝑓 .
In our model, we assume that the total number of the proteins in all forms is conserved on the time scale of the observation:
Therefore, the steady-state solutions of the differential equations are determined by fourteen user-defined parameters: eight 𝐾 𝑑 values (𝐾 𝑑 𝑛𝑛 * , 𝐾 𝑑 𝑛𝑛 , 𝐾 𝑑 𝑛𝑓 * , 𝐾 𝑑 𝑛𝑓 , 𝐾 𝑑 𝑓𝑛 * , 𝐾 𝑑 𝑓𝑛 , 𝐾 𝑑 𝑓𝑓 * , 𝐾 𝑑 𝑓𝑓 ), two activation constants (𝐾 𝑎𝑐𝑡 𝑛 , 𝐾 𝑎𝑐𝑡 𝑓 ), and four total concentrations (𝐺 𝑛 Total, 𝐸 Total, 𝐺 𝑓 Total, 𝐴 Total). Notice that in the GUI, the activation constantans are entered as the percentage of active molecules ( 100 * [𝐺 𝑛 * ] [𝐺 𝑛 * ]+[𝐺 𝑛 ] = 100 * 𝐾 𝑎𝑐𝑡 𝑛 1+𝐾 𝑎𝑐𝑡 𝑛 and 100 * [𝐺 𝑓 * ] [𝐺 𝑓 * ]+[𝐺 𝑓 ] = 100 * 𝐾 𝑎𝑐𝑡 𝑓 1+𝐾 𝑎𝑐𝑡 𝑓
).
The GUI allows graphing the results in two ways (Supplemental Figure 1). In the Graph Mode "Steady State", the GUI displays the concentrations of the complexes as bar plots and allows one to change 𝐺 𝑓 Total and 𝐴 Total using two horizontal sliders. In the Graph Mode "Surface", the GUI displays 3D plots for any use-selected component or complex as a function of 𝐺 𝑓 Total and 𝐴 Total. The user can adjust the range and grid step for these variables.
The GUI enables exploring the model solutions in four regimes. The regime denoted as "None"
not imply any assumptions about the component interactions, so that 𝐾 𝑑 values for all eight complexes [𝐺 𝑛 * 𝐸], [𝐺 𝑛 * 𝐴], [𝐺 𝑓 * 𝐸], [𝐺 𝑓 * 𝐴], [𝐺 𝑛 𝐸], [𝐺 𝑛 𝐴], [𝐺 𝑓 𝐸], [𝐺 𝑓 𝐴] can be modified by the user. In the general dual-chain regime "DC general", the formation of complexes with the inactive forms of the endogenous and tagged GTPases is prohibited ([𝐺 𝑛 𝐸]=[𝐺 𝑛 𝐴]=[𝐺 𝑓 𝐸]=[𝐺 𝑓 𝐴] = 0) by presetting the corresponding 𝐾 𝑑 s to very large values (10 6 ). In the dual-chain orthogonal regime "DC Orth", the GUI additionally pre-sets 𝐾 𝑑 values so that (𝐾 𝑑 𝑛𝑓 * , 𝐾 𝑑 𝑓𝑛 * ) ≫ (𝐾 𝑑 𝑛𝑛 * , 𝐾 𝑑 𝑓𝑓 * ), which ensures reduced cross-interactions between active endogenous GTPase 𝐺 𝑛 * and the affinity reagent 𝐴, and between active fluorescently-tagged GTPase 𝐺 𝑓 * and the endogenous effector 𝐸. Finally, the single-chain regime "SC" is modeled by pre-setting 𝐾 𝑑 values so that the interactions are dominated by [𝐺 𝑛 * 𝐸], [𝐺 𝑓 * 𝐴], and [𝐺 𝑓 𝐴] due to the physical proximity of fluorescently tagged
GTPase and the affinity reagent 𝐴, in the biosensor design.
Finally, using "Plot Optimization Surfaces" in the "Tools" of the GUI menu, the user can display the deviation of the total amount of the effector 𝐸 in complexes [𝐺 𝑛 * 𝐸] + [𝐺 𝑓 * 𝐸] (i.e., perturbed by the presence of the biosensor) from the unperturbed concertation of the GTPase-effector complex [𝐺 𝑛 * 𝐸] when 𝐺 𝑓 Total = 𝐴 Total = 0. The curve in the (𝐺 𝑓 Total, 𝐴 Total)-plain corresponding to the minimal deviation is also calculated as a polygonal fit of a user-selected degree through a number of points along the selected axis 𝐺 𝑓 Total or 𝐴 Total, as specified in the "Fit Parameters" area.
Software is available at https://github.com/tsygankov-lab/orthogonal_biosensors
The Rac1 interaction with the CRIB domain from PAK1 was modeled by threading the Rac1 sequence onto the crystal structure of Rac3 bound to the CRIB domain (PDB code: 2QME). The Rap1/RalGDS complex was modeled by threading the Rap1 sequence onto the structure of Ras bound to RalGDS (PDB code: 1LFD 45 ). Ras and Rap1 have identical sequences in the switch 1 region, which is the binding interface for RalGDS. Multistate design simulations in Rosetta were performed as described previously 29 . In each simulation, a small cluster of residues (2-6) at the interface were allowed to mutate. During the multistate simulation, the sequences under consideration were threaded onto models of the mutant/mutant pair, the mutant/wildtype pair and the wildtype/mutant pair as well as the wildtype and mutant proteins as monomers. Binding energies were calculated by subtracting the energy of the monomers from the respective complexes. The multistate algorithm uses a genetic algorithm to "evolve" sequences to optimize a specified fitness function 33 . We used a fitness function with terms for the target (mutant/mutant) binding interaction, off target interactions (mutant/wildtype), overall complex stability (total energy of the complex) and stability of the monomers. Top scoring sequences favored the mutant/mutant interaction over interactions with wildtype partners while maintaining the total energy of the complex and the unbound monomers.
The GTPase FLARE.dc1g biosensors are based on previously published biosensor designs 3,7,27 . In brief, YPet was fused upstream to residues 60 -145 of human PAK1 for Rac1 and Cdc42 biosensors, and to residues 743-830 of human RalGDS for the Rap1B biosensor. Turquoise2 was fused to the N-terminus of the full-length GTPases. The two biosensor chains were expressed as one open reading frame with two consecutive 2A viral peptide sequences, from Porcine teschovirus-1 (P2A) and Thosea asignavirus (T2A), directly inserted between them, with no additional residues, leading to generation of two separate biosensor chains. The tandem viral sequences caused complete separation of the two biosensor components as previously described 27 while retaining consistent molar ratios (Supplemental figure 4). In the Rac1 biosensors where two AR were expressed from a single gene, the same pair of viral skip sequences was used between the two AR. Orthogonal mutations were introduced using Q5mediated site directed mutagenesis.
Image acquisition HEK-293t cells were plated in 96-well plates with flat µ-clear plastic bottoms (Greiner bio-one) coated with poly-l-lysine (Sigma). Cells were transfected in triplicate and imaged after 24 h. For affinity-reagent binding experiments, cells were transfected with the full-length GTPase fused to Turquoise2, with mCherry2 co-translated using the 2A viral sequences described in "Biosensor design" above. Cells were co-transfected with serial dilutions of the YPet tagged affinity reagent.
Dilutions were carried out in 96-well plates and transfection complexes were formed using Plus
Images were analyzed using MATLAB (Mathworks). Briefly, 4 fields were imaged for each well and the intensity was summed for each channel. These intensities were then background subtracted using values from wells that were mock transfected, and ratios were calculate using these background-subtracted values. For binding affinity experiments R= (FRET -α(Donor)β(Acceptor))/Cherry where R is the Ratio, FRET is the total FRET intensity as measured, α is the bleed-through of the donor into the FRET signal, β is the bleed-through of acceptor into the FRET signal, and Donor, Acceptor, and Cherry are the donor and acceptor intensities as measured through direct excitation. Curves from different days were normalized such that values for the wildtype curve spanned from 0 to 1. Then 9 transfections across 3 days were plotted on a single chart using Graphpad prism, plotting R against Acceptor values to produce binding affinities using the fit Specific binding with Hill slope. For regulator titration experiments R= (FRET -α(Donor)
-β(Acceptor))/Donor. Curves from different days were normalized such that values for the wildtype curve spanned from 0 to 1. Nine transfections across 3 days were then plotted on a single chart using Graphpad prism.
Plasmids encoding biosensor components were transformed into E. coli BL-21 (DE3) cells (New England Biolabs) and 5mL starter cultures were grown overnight at 37C in Lysogeny Broth (LB) supplemented with 50μg/mL kanamycin. The cultures were diluted 1:100 in Terrific BrothTB media (RPI) supplemented with kanamycin and grown at 37 °C to an OD600 of ~0.6. The temperature was subsequently lowered to 16 °C and the cultures were grown to OD600 of ~0.8
and induced with isopropyl β-D-1-thiogalactopyranoside (IPTG) to a final concentration of 0.5mM.
The cultures were shaken for 16-18 hours at 200 rpm. Cell pellets were collected by centrifugation at 14,500xg and lysed in B-PER Reagent (ThermoFisher Scientific) according to the manufacturer's protocol. The lysates were centrifuged at 14,000 xg for 30 minutes at 4C. The supernatant was incubated with Ni-NTA beads (Qiagen) equilibrated with native binding buffer (20mM NaH2PO4, 500M NaCl, 2% glycerol, 10mM imidazole, pH=7.5) for an hour at 4C. The beads were settled by centrifugation, and the supernatant was removed by pipetting. The beads were washed with wash buffer (50mM NaH2PO4, 500M NaCl, 2% glycerol, 30mM imidazole, pH=7.5) after which the protein was eluted with elution buffer (20mM NaH2PO4, 500M NaCl, 250mM imidazole, pH=7.5). Protein expression and purity were confirmed by SDS-PAGE analysis, and quantified spectrophotochemically at 280nm on a Nanodrop 2000 (ThermoFisher Scientific). The eluted protein was buffer exchanged into and concentrated on 3kDa MW spin MW spin columns (ThermoFisher Scientific).
Binding by Biolayer Interferometry.
To determine the dissociation constants (KDs) between GTPases and affinity reagents, binding was measured by Biolayer Interferometry (BLI) on an Octet R4 instrument (Sartorius) using AR2G biosensors. The target affinity reagents were coupled to the biosensors according to the manufacturer's guidelines. The biosensors were then dipped into 5 concentrations of GTPases in HBS-EP+ buffer. The association phase was carried out for 300 seconds and the dissociation for 600 seconds. Regeneration of the binding surface was done in 10mM glycine-HC, pH=2 for 5 x 5 seconds with a 5 second baseline stabilization. A steady state model was used to fit for KDs. Data analyses were carried using Octet Analysis Studio software, version 13.0.
Cos7 cells were plated in 6 well-plates and were transfected with biosensor plasmids using Fugene6 (Promega). After 20-24hrs, cells were replated onto coverslips coated with fibronectin (5µg/ml 37C overnight, Sigma) and allowed to attach in DMEM/10% FBS for 2hrs. The media was then replaced with Hams/F12 with 5% FBS, 10mM HEPES, 100 µm Trolox, and 0.5mM Ascorbate and cells were imaged in an attofluor chamber (Thermo Fisher) at 37C. Cells were imaged using a 40X, 1.3 NA Si objective on an Olympus IX-81 inverted microscope and using Metamorph screen acquisition software (Molecular Devices) and mercury arc lamp illumination. Filters used were Ex: 434/17, Em 482/35 with a dichroic mirror with reflectance bands at 462/523 nm (all Semrock) Images were obtained on a Flash4 sCMOS camera (Hamamatsu). The camera dark current was determined by obtaining images for each camera without excitation, and the dark current was subtracted from all images. Images were corrected for shading due to uneven illumination by taking images of a uniform dye solution under conditions used for each wavelength, normalizing this image to an average intensity of 1 to produce a reference image for each wavelength, and then dividing the images corrected for dark current by the shading correction reference image. Images were segmented into binary masks separating cell and noncell regions using the segmentation package "MovThresh" 46 , which is based on the Otsu algorithm 47 . These binary masks were then used to calculate cell edge velocity and protrusion retraction frequencies.
After the corrections described above, the cell images were divided into rectangular windows along the entire cell edge (see Cell Windowing Analysis below). Sampling windows located at the cell edge were used to derive a time course of edge velocity as the windows tracked the morphological changes of the cell over time. Cell edge velocities and frequencies were calculated from the tracking of these windows. Statistical comparisons of cell edge velocities were tested using Dunnetts multiple comparisons test. Software for the statistical analysis of edge motion is available at https://github.com/DanuserLab/Protrusion_Retraction_Database.
MEFs stably expressing biosensors under control of the tet-off system were produced using retroviral transduction. Biosensor expression was induced 48 hr prior to imaging through trypsinization and culturing without doxycycline. On the day of imaging, cells were replated onto coverslips coated with fibronectin (5µg/ml 37C overnight) using trypsin/EDTA (Corning) and allowed to attach in DMEM /10%FBS. After 2 hrs the media was replaced with Hams/F12 with 5% FBS, 10mM HEPES, 100 µm Trolox, 0.5mM Ascorbate, and cells were allowed to equilibrate.
After a further 30mins, cells were imaged in a closed chamber at 37C. Cells were imaged using an Olympus 40X, 1.
3 NA Si objective on an Olympus IX-81 inverted microscope and using Metamorph screen acquisition software (Molecular Devices) and mercury arc lamp illumination. Excitation filters used were -CFP/FRET 434/17 and YFP 510/10, using a dichroic mirror with reflectance bands at 462/523 nm (all Semrock). For emission, a TuCam dual camera adaptor (Andor) was fitted with a 509 nm long pass imaging flat dichroic mirror and the filters used were CFP 482/35 and YFP/FRET: 549/50 (all Semrock). Images were obtained using Flash4 sCMOS cameras (Hamamatsu).
Biosensor activation levels were measured in living cells by determining the ratio of FRET to donor emission on a pixel-by-pixel basis. Donor and FRET images were aligned using fluorescent beads as fiduciaries to produce a transformation matrix using the Matlab function "cp2tform"
(Matlab, The Mathworks Inc.). This was then applied to the Donor image using the Matlab function "imtransform". The camera dark current was determined by obtaining images for each camera without excitation, and the dark current was subtracted from all images. Images were corrected for shading due to uneven illumination by taking images of a uniform dye solution under conditions used for each wavelength, normalizing this image to an average intensity of 1 to produce a reference image for each wavelength, and then dividing the images corrected for dark current by the shading correction reference image. Background fluorescence was removed by subtracting, at each frame, the intensity of a region containing no cells or debris. Images were segmented into binary masks separating cell and non-cell regions using the segmentation package "MovThresh" 46 , which is based on the Otsu algorithm 47 . The Donor channel was used for segmentation, as it had the highest signal to noise, particularly at the cell edge. The masks were then applied to all channels, setting non-cell regions to zero intensity. The images were corrected for bleed-through and ratios were obtained using the following equation (using data from control cells expressing donor or acceptor alone to obtain the bleed-through coefficients α and β): R= (FRET -α(Donor) -β(Acceptor))/donor where R is the Ratio, FRET is the total FRET intensity as measured, α is the bleed-through of the donor into the FRET signal, β is the bleed-through of acceptor into the FRET signal, and Donor and Acceptor are the donor and acceptor intensities as measured through direct excitation. These ratio images were then corrected for photobleaching.
For constitutive migration the whole cell average was fitted to a double exponential curve, and this curve was used to normalize. Pseudocolor scales were produced without considering the lowest and highest 5% of ratio values, to eliminate spurious pixels and normalizing so the lowest value was 1.
After the corrections described above, the cell images of the ratiometric biosensor activity were compartmentalized into layers of rectangular windows along the entire cell edge as previously described 3 . To construct the sampling windows at a constant distance from the cell edge we This in silico compartmentalization of the cell allowed us to represent the biosensor activity in a cell-shape invariant space. For each frame of the movies, the biosensor activity was averaged within the area of each sampling window, resulting in a set of matrices representing the biosensor activity of a layer of windows with a fixed distance from the cell edge. Software to perform these image data transformations is available at https://github.com/DanuserLab/Windowing-Protrusion.
Pearson's correlation coefficient 𝜌(𝑎
1 (𝑡), 𝑎 2 (𝑡)) 𝜏 between two activity time courses 𝑎 1 (𝑡) and 𝑎 2 (𝑡) was computed as a function of the time lag 𝜏 using the Matlab function (xcov). This function implements the mean corrected and normalized correlation functions as follows: 𝜌(𝑎 1 (𝑡), 𝑎 2 (𝑡)) 𝜏 = 𝐶𝑜𝑣(𝑎 1 (𝑡) -𝑎 1 , 𝑎 2 (𝑡 -𝜏) -𝑎 2 ) √𝑉𝑎𝑟(𝑎 1 (𝑡) -𝑎 1 √𝑉𝑎𝑟(𝑎 2 (𝑡) -𝑎 2
The operators Cov(.) and Var(.) denote the covariance and the variance of the mean corrected time courses, respectively. The variables 𝑎 1 and 𝑎 2 denote the mean value of the respective time course. Correlation functions were first calculated for pairs of time courses per window between edge velocity and biosensor intensity. For each correlation curve we computed the level, which is exceeded by no more than 5% of correlations between two random time series. This level (~0.1 for all our data) depended on the duration of the videos and the number of sampling windows.
Regions of the correlation curve that extend above this confidence interval are therefore considered significant with a confidence of 95%. In addition, through bootstrap analysis of the variation among all correlation curves, we computed confidence intervals for the mean crosscorrelation. This reflected the number and consistency of sampling windows across experiments. Fluorescent AR binds to endogenous GTPase, soaking up AR without producing FRET, and generating dominant negative effects. In the orthogonal biosensor this interaction is reduced. Bottom: Endogenous effectors can bind to fluorescent GTPases, increasing effector activation without generating FRET. This interaction is reduced for orthogonal GTPases. (C) Heatmaps show how the formation of GTPase-effector signaling complexes is affected by the concentrations of fluorescent GTPase and AR (concentrations relative to endogenous GTPase and effector). Each column shows this for a different level of GTPase activation. For wild type biosensors, increasing perturbation (warm colors, see scale) is predicted as GTPase concentration increases, particularly for low levels of GTPase activation. In contrast, orthogonal interfaces with a 10-fold reduction in off-target affinities show perturbations lower than 10% across all GTPase concentrations, and concentration ranges that induce perturbation <5% (black regions) extend across more GTPase and effector concentrations. (D) Graph shows the increased incorporation of fluorescent biosensor components into productive, FRET-producing complexes, for orthogonal biosensors that have a 10-fold reduction in affinity for endogenous proteins. The percent increase in FRET-producing complexes, for orthogonal relative to wild type biosensors, is shown on the y axis. complex formation (concentration is relative to endogenous concentration) using the experimentally determined affinities from (C) to calculate the number of signaling complexes as in Figure 1C. biosensor cells compared to control cells not expressing biosensor. (E) Relative expression levels of cells in D. For D and E, individual cells are plotted, central horizontal line is mean +/-95% C.I. (n= cells, Control n=38, Wt Cdc42, n=7; Orthogonal Cdc42, n=10; Wt Rac1 n=12; Orthogonal Rac1, n=13 * p=0.0148, ** p=0.0034, **** p<0.0001).