Deep learning guided design of dynamic proteins

Abstract

Deep learning has advanced the design of static protein structures, but the controlled conformational changes that are hallmarks of natural signaling proteins have remained inaccessible to de novo design. Here, we describe a general deep learning-guided approach for de novo design of dynamic changes between intra-domain geometries of proteins, similar to switch mechanisms prevalent in nature, with atom-level precision. We solve 4 structures validating the designed conformations, demonstrate modulation of the conformational landscape by orthosteric ligands and allosteric mutations, and show that physics-based simulations are in remarkable agreement with deep-learning predictions and experimental data. Our approach demonstrates that new modes of motion can now be realized through de novo design and provides a framework for constructing biology-inspired, tunable and controllable protein signaling behavior de novo.

Many natural proteins can adopt multiple conformational states, and the controlled interconversion between these states often underlies biological functions and their regulation (1, 2). Binding partners, environmental conditions, post-translational modifications, allosteric effectors, and mutations can modulate protein function by changing the populations of active and inactive conformations (3, 4). Notably, many of the conformational changes required for function and regulation do not involve large-scale structural rearrangements; rather, critical motions often consist of localized rotation, tilting, and sliding of secondary structure elements within a domain. These types of motion are hallmarks of central biological processes, such as the regulation of kinases (5) and signaling via G-protein coupled receptors (GPCRs) (6) (Fig. S1).

Despite the functional importance of local conformational changes in natural proteins, predictive design of similar dynamics de novo has remained elusive (7). One key challenge is the difficulty in parameterizing sufficiently accurate physics-based energy functions to design sequences that can adopt multiple conformations (multi-state design, (8)), since the energy gap between user-specified folded states, off-target folded states, and disordered states is typically small especially for intra-domain conformational changes. Here, recent advances in deep learning that have enabled robust design of sequences conditioned on structure (9, 10) could provide more sensitive assessment of how changes to sequence impact conformational equilibria. A second challenge is to generate alternative conformational states de novo (not borrowed from existing proteins) that are designable, i.e. there must exist an amino acid sequence that can adopt these alternative conformations. To date, there is no generalizable method to address this problem (11). Consequently, there are only a few examples of de novo designed dynamic proteins and none on the local scale prevalent in natural regulators. Previous pioneering studies include the design of fold switches (12, 13), side chain dynamics (14), and controllable changes in coiled-coil assemblies (15). Most commonly, however, design of protein switches has involved domain-level rearrangements or hinge-like motions of rigid bodies, where conformational changes are driven by intermolecular protein-protein interactions and the multi-state design problem is simplified as either most atomic interactions within the rigid bodies remain the same (16) or one of the states becomes disordered (17). As a result, most classes of conformational switch mechanisms, effectors, and their combinations have been inaccessible and hence completely unexplored by de novo design.

Design approach for dynamic proteins with tunable two-state equilibria

Inspired by the scale and modes of conformational changes prevalent in natural signaling proteins, we sought to develop a generalizable approach to design non-native controllable conformational switching in proteins de novo, providing a basis for designing tunable and complex signaling behavior beyond what exists in nature (7). Specifically, we aimed to design sequences with multiple energetic minima in structure space, each corresponding to a well-defined conformational state (Fig. 1A). We focused on conformational changes involving intra-domain reorientation of secondary structural elements, mimicking a dominant mechanism in naturally occurring proteins (in contrast to the easier problem of movement of otherwise static sub-domains about a hinge). As in biological regulation, we aimed to design mechanisms for modulating the conformational equilibrium by inputs such as orthosteric ligands (binding within the region of conformational change) and allosteric perturbations (acting at distal sites coupled to the active site (18)).

Fig. 1. Generalizable approach for the deep learning guided design of dynamic proteins.

(A) Schematic of design goal to engineer dynamic proteins in a two-state equilibrium that can be controlled by orthosteric ligands and allosteric perturbations. (B) Main stages of the approach: (1) De novo generation of alternative states that differ in their geometry (light blue region) using systematic conformational sampling (Fig. S2), followed by in silico and experimental validation of single state designs. (2) Deep learning (DL) guided sequence/structure search to focus sampling during multi-state design (MSD) at key positions for determining state preference and their neighbors. (3) Combination of physics-based molecular dynamics (MD) simulations, mutual information analysis (MutInf), and deep learning models to determine state-specific residue interaction networks and identify mutations capable of modulating the conformational landscape, followed by experimental validation. (C-D) Application to generate two designable states with distinct conformations coupled to ligand binding (Ca²⁺). (C) (Top row, light blue) Binding-competent state 1 structure with Ca²⁺ binding site (inset) (PDB ID: 1SMG) shown in two orientations. (Bottom row, teal) De novo generated alternative (binding-incompetent) state 2 model shown in two orientations. To couple Ca²⁺ binding to the designed conformational change, the Ca²⁺ binding site is significantly reshaped in state 2 to disfavor Ca²⁺ binding (inset). (D) Overlay of the NMR structure of a state 2 single-state design (teal) with its AF2 prediction (grey) shows excellent agreement (Cα RMSD = 0.98Å, excluding loops).

Our design approach (Fig. 1B) involves three general stages: The first stage identifies alternative structural states through (i) generation of a library of de novo candidate states that differ in their local geometries using systematic physics-based conformational sampling (Fig. S2, Methods) (19), and (ii) assessment of the designability of these states. To assess single-state designability of the generated backbones, we designed sequences for each state, evaluated them in silico (Methods), and characterized them experimentally. The second stage performs a deep learning guided search in sequence and structure space that (i) restricts the search space by identifying the minimal set of residues required to define each state, and (ii) designs sequences at these positions that are simultaneously compatible with pairs of states (multi-state design). The third stage seeks to identify perturbations that modulate the designed protein conformational landscape by (i) integrating physics-based simulations and deep learning predictions to determine state-specific interaction networks and (ii) predicting mutations leading to tunable conformational switching.

For the proof-of-concept application described here, we required the conformational landscape to be responsive to a ligand input, where the ligand-binding site (orthosteric site) changes conformations between states and thereby couples ligand binding to conformational switching. For simplicity, we used an engineered Ca²⁺ binding protein (derived from the N-terminal domain of troponin C) as our ligand-binding-competent starting state (“state 1”) (Fig. 1C, top row). The wild-type protein consists of two EF hand motifs (sites I and II), which both bind Ca²⁺ in the low micromolar range. We instead used a variant (PDB ID: 1SMG) with an E41A point mutation in site I, which weakens the affinity of site I to the millimolar range while retaining moderate micromolar affinity in site II (20). Additionally, the E41A mutant does not undergo conformational change upon Ca²⁺ binding (Fig. S3). By using an existing binding state as a starting point and designing a conformational change encompassing the binding site, we demonstrate here how our protocol can be generally applied to engineering controllable conformational changes into natural proteins; alternatively, one could generate a (static) binding-competent state de novo (21).

Generation and experimental validation of de novo design alternative states

To generate a structurally defined alternative state (“state 2”) (Fig. 1C, bottom row), we sampled de novo orientations of a contiguous protein segment including loop III, helix C, and Ca²⁺ binding site II (which we refer to as the “reshaped region”) (Fig. S2) using the loop-helix-loop unit combinatorial sampling algorithm (LUCS), previously shown to be capable of generating static proteins that differ in the local geometry of user-defined protein segments (Methods) (19). We only kept design models with the same length as our input but allowed the secondary structure in the reshaped region to vary (i.e., some regions may be part of a loop in the input but adopt a helical structure in the output or vice versa, a common transition in natural proteins). We then used Rosetta to design sequences optimal for each of the output models (“single-state designs”) and filtered these designs computationally (Methods). This procedure generated a library of approximately 1×10³ diverse conformations with an average Cα RMSD of 7.1Å in the reshaped region, which is comparable in scale to functional conformational changes in natural signaling proteins (Fig. S1). From this library, we selected 11 designs (each corresponding to a unique backbone) for experimental testing.

To rapidly screen the de novo state 2 sequences experimentally, we displayed each design fused to a C-terminal c-Myc tag on the surface of yeast (Methods) and used surface display levels (which are known to be correlated with stability) as a proxy for designability. Although we only tested a few single-state designs in this study, one could screen thousands of designs with yeast display to identity many more designable conformational states. 10 out of 11 designs (all except #615) showed high surface display levels (Fig. S4). We decided to further characterize design #6306 as it had a very different conformation in the reshaped region compared to state 1 (Fig. 1C, bottom row), involving both rotation and translation of the reshaped helix C. Moreover, the Ca²⁺ binding loop was substantially restructured and partially helical, resulting in an unfavorable conformation for binding. We solved the nuclear magnetic resonance (NMR) structure of #6306 and saw excellent agreement (Cα RMSD = 0.98Å, excluding loops) between the experimentally solved structure (teal) and the AlphaFold2 (AF2) (22) model of the design (grey) (Fig. 1D, Fig. S5). In contrast to state 1 (1SMG), design #6306 did not bind Ca²⁺ in site II at Ca²⁺ concentrations up to 1mM (Fig. S6), as expected due to significant restructuring of the binding site. These results confirm that the backbone of design #6306 is both designable and unfavorable for ligand binding, making it suitable as our binding-incompetent conformation (state 2) for two-state design.

Multi-state design of dynamic proteins

We next aimed to identify sequences that were simultaneously compatible with the conformations of both state 1 and state 2. Rather than allowing all residues in the reshaped region and their contacting residue positions to be designable, we sought to generate multi-state designs able to populate both states to varying degrees despite having high sequence similarity. To do so, we used deep learning-based structure predictions (AF2) to shrink the searchable sequence space and focus sampling at key positions for determining state preference. We reasoned that this approach would allow us to better interpret sequence differences predicted to shift the preference for one state versus the other, identify potential sites for allosteric mutations, and facilitate experimental characterization. To achieve this, we used AF2 to identify mutations to design #6306 (predicted to adopt state 2) that would increase sequence similarity to 1SMG (state 1) without affecting the predicted structure (Methods) (Fig. 1B). We then used the resulting sequence with the highest sequence identity to the state 1 sequence (1SMG), but still predicted to fold into state 2, as input for multi-state design (position-tied ProteinMPNN) (Methods, Fig. S7A-B) (9). Sites where mutations caused significant structural perturbations were typically positions where there were large changes to solvent accessible sidechain surface area, hydrogen bonding networks, or steric packing between states (Table S1). The final set of multi-state designable residues included these positions and their neighbors, decreasing the set of designable residues from 37 to 25 (Table S2). The structures of putative switch design sequences were then predicted with AF2 to assess their compatibility with both states.

Our design approach identified a family of sequences that had AF2 structure predictions that were either in state 1, state 2, or a combination of both including structural intermediates (Fig. 2A). The designs differed from the original state 1 sequence (1SMG) by n=18 mutations and from the high sequence identity single-state state 2 design by n=15 mutations, but strikingly differed from each other at only one residue position, 89, which was located outside the reshaped region and distal from the Ca²⁺ binding site. Position 89 was hence predicted to act as an allosteric site where perturbations, in this case mutations, were predicted to change the populations of states in the reshaped region, including the conformation of the distal Ca²⁺ binding site. In particular, smaller hydrogen bond donors and acceptors at position 89 were biased toward state 2 by forming a hydrogen bond with the backbone of loop III, bringing it closer to the central helix D. Conversely, bulky and/or hydrophobic amino acids pushed loop III outward into a conformation more consistent with state 1. The AF2 confidence metric (pLDDT) of the reshaped region distal to site 89 for this sequence family varied considerably depending on the amino acid identity at residues 89 (Fig. S7C).

Fig. 2. Two-state equilibrium in fast exchange shifted by allosteric mutations.

(A) The protocol in Fig. 1B predicted a family of sequences differing only at position 89 (X), where the amino acid identity at position X determined whether the 5 AF2 predicted models were entirely in state 2 (left), mixed (middle), or entirely in state 1 (right). Depicted AF2 models (grey cartoons, with reshaped region colored by state 1 (light blue) or state 2 (pink) conformations are for the underlined amino acid at position X (shown as sticks). Small polar residues favored state 2 by hydrogen-bonding with the backbone of loop III (bottom, left), while bulky and/or hydrophobic residues favored state 1 by pushing loop III away from the central helix (bottom, right). (B) 2D ¹H,¹⁵N-HSQC spectra of S89, N89, and I89, with several well-resolved peaks (shaded ovals) showing chemical shift changes consistent with a two-state equilibrium in fast exchange between state 1 preferred (I89), state 2 preferred (S89) and intermediate (N89). Inset shows ¹Hn chemical shift changes between I89 and S89 colored on the AF2 model of I89, consistent with the designed conformational change in the reshaped region (the reshaped region is circled by dashed line in panels B-E). (C) Agreement between the NMR structure of S89 (red) and its AF2 prediction (grey) (Cα RMSD = 1.31Å excluding loops). Inset shows the hydrogen bond formed between S89 and loop III in the NMR structure consistent with AF2 predictions in (A). (D) NMR models for I89 were consistent with sampling both state 1 (blue, left) (Cα RMSD = 1.67Å excluding loops) and state 2 (pink, right) (Cα RMSD = 1.31Å excluding loops), with residue ILE 69 buried in proposal #1 (state 1) and solvent-exposed in proposal #2 (state 2). (E) ¹⁵N on-resonance R_1ρ relaxation rates for design I89 plotted per residue (top) and visualized on the AF2 structure of design I89 (bottom) indicate low microsecond exchange in the regions predicted to undergo significant conformational changes (median predicted change in Cβ-Cβ distances between states shown on the x axis colored by magnitude). Residue numbering of all designs includes the N-terminal thrombin cleavage site scar (grey in (A)).

Allosteric modulation of the conformational landscape by mutation

To test the prediction that our designs will adopt two different defined conformations whose relative populations are dependent on the amino acid identity at position 89, we selected five designs that covered a range of AF2 predicted behaviors: state 2 preferred (S89, N89), state 1 preferred (I89, K89), and mixed (R89). We recorded 2D ¹H,¹⁵N-HSQC NMR spectra for each of these designs (Fig. 2B, Fig. S8) and focused structural characterization on three representative designs (S89, N89, I89; analysis of R89, predicted to adopt multiple conformational states, was complicated by evidence of peak broadening, Fig. S8). Remarkably, the 2D ¹H,¹⁵N-HSQC NMR spectra were drastically different even though the designs differed only by a single point mutation: 67 out of the 92 backbone amide peaks in the spectrum of I89 (state 1 preferred) had significantly different chemical shifts (ΔδHn > 0.03ppm or ΔδN > 0.4ppm) compared to that of S89 (state 2 preferred), suggesting different chemical environments of these residues (Fig. 2B, Table S3). Moreover, for well-separated peaks, the chemical shifts of S89 and I89 were at the two ends of a range, with N89 being intermediate. This finding was overall consistent with a simple model in which the designed proteins are in an equilibrium between two states in fast exchange on the NMR timescale where the observed chemical shifts reflect the population average of the two states (though we cannot exclude the possibility of additional exchange processes from these data). To analyze this behavior further, we assigned the backbone ¹Hn and ¹⁵N chemical shifts of S89 and I89 and plotted the differences on the AF2 model of I89 (Fig. 2B). We found that the changes to chemical shifts were not merely localized near the mutated position 89, but also at more distal residues throughout the reshaped region, including Ca²⁺ binding site II and its neighboring residues, consistent with a change in the ensemble averaged conformation of the reshaped region.

To assess whether our designs indeed adopted the two specific designed conformational states, we solved the structures of S89 and I89 by NMR (Fig. 2C-D). The structure of S89 was in excellent agreement with its AF2 prediction (state 2, Cα RMSD = 1.31Å, excluding loops), including at the sidechain level (Fig. 2C, Fig. S9). In contrast, for I89, automated structure determination using ARTINA (Methods) did not converge to a single proposed conformation in the reshaped region (Fig. 2D, Fig. S9). Instead, the top two structures proposed by ARTINA resembled the designed state 1 and state 2 conformations, respectively, based on assigned distance restraints consistent with both states 1 and 2 (Fig. S10). The residue predicted to undergo the largest conformational change in our computational models (isoleucine 69) was buried in proposal #1 but solvent-exposed in proposal #2, hallmarks of the state 1 and 2 conformations, respectively. Taken together, our structures confirm that the designs adopt the two intended conformations but with different populations (Table S4). Although S89 primarily adopts state 2 as predicted, I89 samples both state 1 and state 2, and a single-state structure cannot fully account for all ensemble-averaged distance restraints simultaneously.

To probe the suggested dynamics in design I89 more directly, we first collected a series of 2D ¹H,¹⁵N-HSQC spectra from 5°C to 35°C at 5°C intervals. We observed temperature-dependent changes in peak intensity localized to residues in the reshaped region and their neighbors, consistent with changes in chemical environment due to dynamics in the reshaped region (Fig. S11). Moreover, peak intensities were higher at 35°C compared to 5°C, as expected for a system being in fast exchange at higher temperatures and slowing to fast-intermediate exchange with decreasing temperature, manifesting as line broadening. Additionally, we measured R_1ρ values for design I89 with on-resonance rotating frame relaxation dispersion at an effective field strength of 3kHz and observed higher R_1ρ values for residues in the reshaped region and their neighbors (Fig. 2E). This behavior is consistent with chemical exchange on the micro-to-millisecond timescale accelerating measured relaxation rates (Fig. 2E, Table S5), as well as our estimate of an upper bound on the exchange time from chemical shift data τ_ex << 10ms (Table S6). Finally, further NMR dynamics experiments and analyses including R_1ρ relaxation dispersion measurements allowed us to tighten the upper bound for exchange in the reshaped region to 42μs (Methods, Fig. S12-13).

Orthosteric modulation of the conformational landscape by ligand binding

To probe whether Ca²⁺ binding also modulates the state populations by preferentially stabilizing state 1, we recorded 2D ¹H,¹⁵N-HSQC spectra with and without Ca²⁺ for each point mutant (Fig. S14). As expected from the designed conformational change, we found that Ca²⁺ addition caused significant chemical shift perturbations (ΔδHn > 0.03ppm or ΔδN > 0.4ppm) throughout the reshaped region and its neighboring residues, affecting approximately 30 additional peaks when compared to the single-state binding-incompetent design #6306 (Fig. 3A, Table S7). Given that our NMR structural ensemble for S89 is in agreement with our binding-incompetent structure (state 2) while I89 has a substantial binding-competent (state 1) population (Fig. 2C-D), the direction of chemical shift changes with Ca²⁺ was indeed consistent with a shift in the equilibrium toward state 1 (Fig. 3B, Table S6). We next measured the Ca²⁺ binding affinity by monitoring Ca²⁺ concentration-dependent chemical shift changes of residues in the reshaped Ca²⁺ site and indeed found increasing affinity with an increasing estimated population of the binding-competent state 1, with an overall ~10-fold difference in Ca²⁺ binding affinity between designs I89 and S89 (Fig. 3C). These results confirm that mutations at the distal position 89 - more than 15Å away from the closest Ca²⁺ binding residue – allosterically modulate the Ca²⁺ binding site conformational equilibrium and therefore binding affinity, consistent with previous approaches using mutagenesis to probe allosteric coupling (23, 24).

Fig. 3. Modulation of the conformational landscape by ligand binding.

(A) Changes in I89 ¹Hn chemical shifts upon adding 10eq of Ca²⁺ visualized on the AF2 model, showing significant chemical shift perturbations distal to the Ca²⁺ binding sites, particularly in the reshaped region and in neighboring residues. (B) 2D ¹H,¹⁵N-HSQC spectra of S89, N89, and I89 with 10eq of Ca²⁺ show ligand-dependent chemical shift changes for residues in the reshaped region consistent with Ca²⁺ shifting the population distribution toward state 1 (arrows). (C) Ca²⁺ titration curves for I89 (dark blue), N89 (medium blue), and S89 (light blue) for residues V71 (circles) and D78 (triangles) in the reshaped Ca²⁺ binding site, which demonstrate increasing affinity as the estimated binding-competent state population increases. The estimated K_d values are 1.6±0.2mM (V71) and 3.9±0.2mM (D78) for I89, 20±2mM (V71) and 11.6±0.4 (D78) for N89, and 22±2mM (V71) and 14.6±0.4 (D78) for S89. (D) The Ca²⁺ bound NMR structure of I89 is in excellent agreement with its AF2 prediction (Cα RMSD = 1.34Å, excluding loops), and the observed binding site backbone is consistent with the known EF hand binding motif with modeled-in Ca²⁺ (right).

Finally, we solved the NMR structure of I89 in the presence of Ca²⁺. The holo structure was in excellent agreement with our computational state 1 model (Cα RMSD = 1.34Å, excluding loops) and the backbone conformation of binding site II was consistent with an EF hand binding motif (Fig. 3D). Though holo I89 had many more distance restraints consistent with state 1 and less consistent with state 2 compared to apo I89, we still observed several nuclear Overhauser effect-derived distance restraints consistent with state 2 even in the presence of excess Ca²⁺, suggesting residual dynamics (Fig. S15). Taken together, these results suggest that the tested family of sequences adopted the two designed conformational states of the reshaped region in solution, where the populations of these states (Table S6) can be modulated by both allosteric mutations (Fig. 2) and Ca²⁺ binding (Fig. 3).

Integration of physics-based simulations to tune conformational equilibria

To further probe the atom-level interactions underlying the designed conformational switching, we ran molecular dynamics (MD) simulations for designs I89 and S89 with and without Ca²⁺ (Fig. 4A, Fig. S16, Fig. S17). We observed reversible transitions between states 1 and 2 for design I89 in the absence of Ca²⁺ in excellent agreement with our design predictions and experimental data (Fig. 4A, Fig. S18, Movie S1). When fitting a Markov state model (MSM) on our aggregate 36μs Ca²⁺-free I89 simulation data clustered by Cα RMSD of the reshaped helix to each state, we estimated a lower bound of ~3μs for our overall timescale of exchange (Fig. S19, Fig. S20). Combined with the upper bound from the R_1ρ data of ~42μs (Fig. S13), these results indicate that our reshaped region undergoes low microsecond exchange, comparable to motions of similar scale in natural proteins (25). In contrast, we did not observe conformational switching when simulating design I89 with Ca²⁺, and Ca²⁺ remained bound to site II during the simulation (Fig. 4A, lower plot), in agreement with the experimental data showing that Ca²⁺ stabilizes state 1. Design S89 remained in the state 2 conformation throughout the entire course of a 2μs simulation in the Ca²⁺-free condition (Fig. S16). Though we did not observe a transition to state 1 in the presence of Ca²⁺, we did see larger fluctuations in the reshaped region for which the RMSD to state 1 decreased compared to trajectories without Ca²⁺ (Fig. S16, Fig. S17). The lack of state transitions at this timescale may be due to insufficient sampling, though this behavior is consistent with the NMR structure of S89 (Fig. 2C) and the Ca²⁺ binding data (Fig. 3C). The MD results provide compelling support of exchange between the designed states on a low microsecond timescale for design I89, show that Ca²⁺ preferentially stabilizes state 1 in I89, and are consistent with allosteric modulation of the designed switch at position 89.

Fig. 4. Physics-based simulations reveal molecular mechanisms underlying switch behavior.

(A) Ca²⁺-free 2μs molecular dynamics (MD) trajectory of I89 (teal) showing the designed transition between a binding-competent state 1 (light blue) and binding-incompetent state 2 conformation (pink) (middle panels). The protein remains in state 1 in the presence of Ca²⁺ (bottom) and coordinates Ca²⁺ consistent with an EF hand (inset). C⍺ RMSD is measured for the reshaped helix compared to our computational models for states 1 (black line) and 2 (blue line) after aligning on the regions of the non-reshaped backbone with regular secondary structure. (B) Mutual informational analysis (heatmap) revealing a correlated network of residues (light blue) connecting the Ca²⁺ binding loop (green) with the allosteric mutation site (dark blue), shown in two orientations (top panels, reshaped helix circled by a dashed line). Middle panels show interaction details of a key step in which I69 becomes solvent exposed in state 2, enabling the formation of distinct contacts by Y43. (C) State-specific interactions for state 1 and state 2 (colored) observed during a 2μs Ca²⁺-free MD simulation for I89. (D-F) In silico and experimental validation of state-specific interaction networks using mutations. (D) Frame2Seq (F2S) predictions for Y64F and K68E; score refers to the negative log-likelihood difference between the mutated and original sequence, where the negative change in score predicts the mutations disfavor state 2. (E) AF2 predictions showing higher pLDDT (greater confidence) for state 1 for mutants on the bottom compared to the original sequence on the top. (F) 2D ¹H,¹⁵N-HSQC NMR data showing chemical shift changes consistent with an increased state 1 population in the mutants (arrows denoting direction of shifts from primarily state 2 (S89) towards a larger population of state 1 (I89)).

Because of the remarkable agreement between the design predictions (Fig. 2A), MD simulations (Fig. 4A), and NMR data (Fig. 2B-E; Fig. 3), we next asked whether the MD results could (i) explain the mechanism of allosteric modulation and (ii) make testable prospective predictions to further validate this mechanism. We first performed a mutual information analysis of side chain torsional dynamics (26) in the apo I89 MD trajectories where a state transition was observed. We observed a network of hydrophobic core residues coupling torsional motions of the Ca²⁺ binding site II (residues 70–76) to distal residues in loop III (residues 50–58) and helix D (residues 80–94) (Fig. 4B). Allosteric residue 89 directly faces loop III from central helix D. Combined with our experimental results on allosteric modulation (Fig. 2, Fig. 3C), our MD results therefore suggest a mechanism where the amino acid identity at residue 89 causes differences in sterics (I89) or hydrogen bonding (S89) interactions between helix D and loop III (Fig. 2A), which in turn couple through the identified correlated hydrophobic network to allosterically affect the conformation of the distal Ca²⁺ binding site (Fig. 4B). Further analysis of the mutual information data revealed two unique interaction networks that are specific for state 1 or state 2, respectively (Fig. 4C). State 1 appeared to be stabilized by hydrophobic interactions involving burial of I69, π-π stacking between Y64 and Y88, and less steric bulk near I89. Conversely, in state 2 I69 is pointed outwards to the surface, allowing for Y43 to form new contacts in a network of hydrogen-bonding and electrostatic interactions mediated by Y43, Y64, K68, and E81. Taken together, the mutual information analysis of the MD trajectories suggests two extensive state-specific interaction networks that link the reshaped region and the allosteric site 89.

To test these state-specific interactions, we predicted mutations of the identified networks that would favor state 1 and assessed them by (i) scoring with a structure-conditioned masked language model (Frame2seq) (10), (ii) predicting the mutant structures with AF2, and (iii) collecting 2D ¹H,¹⁵N-HSQC spectra. We reasoned that a Y64F mutation should be disruptive in state 2, as it is unable to hydrogen bond with Y43 and E81, but neutral in state 1, where phenylalanine can still form a π-π stacking interaction with Y88. Likewise, a K68E mutation should be disruptive in state 2, as it cannot form a stabilizing electrostatic interaction with E81, but neutral in state 1 where it is solvent-exposed. In both cases, Frame2seq predicted a higher likelihood for the mutated amino acid compared to the original amino acid given state 1 and a lower likelihood given state 2, AF2 predicted these mutants to adopt state 1 with higher confidence compared to the original I89 sequence, and the 2D ¹H,¹⁵N-HSQC spectra were consistent with the expected shift in population toward state 1 (Fig. 4D-F). In summary, our results explain the observed allosteric modulation of the conformational landscape by mutations at the lynchpin position 89 and validate predictions to further tune the switch equilibrium.

Discussion

Our results demonstrate a general approach to design proteins with two distinct conformational states specified by the designer whose interconversion can be modulated both by ligand concentration (orthosterically) and by mutations to distal sites (allosterically). The approach generalizes beyond prior de novo designed switches based primarily on domain replacement or hinge-like motions (11) where most of the atomic interactions between the rigid bodies remain constant (16, 17). In contrast, our designs interchange between distinct sets of atomic interactions and demonstrate that new modes of motions – inspired by those present in regulator superfamilies such as kinases and GPCRs - can now be realized through de novo design, greatly expanding accessible functional space.

One notable observation is the strong correspondence between the deep-learning based predictions, experimental data, and physics-based simulations. This agreement provided testable hypotheses on the mechanisms underlying the bistability of the switch and allowed us to modulate the conformational equilibrium at the atom-level. We attribute these findings at least in part to specific features of our approach, in particular the search through sequence and structure space that narrowed designable positions to those predicted to be strong determinants of the engineered conformational change (Fig. 1B). The speed and inference offered by deep learning-based protein sequence design and structure prediction enabled this approach that, ultimately, designed distinct residue networks stabilizing the two structural states. This level of insight into a designed system is essential to advance de novo design of allosteric regulation.

The synergy between deep learning and physics-based simulations - demonstrated here for the de novo design of dynamic proteins - may be useful for developing future design methods that would enable predictive control over conformational landscapes and the timescale of exchange (we note that our current design method does not explicitly consider transition state barriers). Emerging approaches to these problems include training models on simulation and/or experimental data to accurately and rapidly predict conformational ensembles in silico, which remains a major challenge (27, 28), and using machine learning to parameterize new force-fields that may allow for good estimates of energy landscapes without the resource demand of all-atom simulations (29, 30). Future development of deep learning integrated with unique dynamics data offered by experiment and simulation could allow for one-shot conformational ensemble prediction (31, 32). Such models could ultimately be applied to design entire user-defined conformational landscapes. It is encouraging that our approach can already successfully design motion involving intricate changes to the inter-residue interaction network between states, despite the individual models for sequence and structure prediction never having been explicitly trained for such a task.

Functional conformational changes in natural proteins are often complex, typically not well-understood, and hence difficult to study and manipulate using engineering approaches. In contrast, our design method both generates two-state conformational changes similar to those in nature, as well as successful predictions how to modulate the populations of these states. The proof-of-concept application here is an exciting example of designing a de novo alternative state into a naturally occurring protein. This result suggests that our approach may be useful for controlling natural functions through tunable conformational switching, which would open many new avenues to study and manipulate the roles of natural switches in the cellular context. Moreover, our approach should be useful to generate entirely bio-orthogonal protein switches. The approach and concepts described here hence have broad relevance to both modulate and deconstruct the requirements for natural signaling and construct complex switch-like signaling systems entirely from the ground up.

Materials and Methods

Residue numbering

Please note that all designs reported in this manuscript (including deposited structures/computational models, experimental data, and molecular dynamics simulation analyses) are indexed to include a four residue N-terminal thrombin cleavage scar (grey in Fig. 2A in the main manuscript). We adopted this residue numbering convention throughout (unless we are directly referring to the 1SMG PDB structure) to be consistent with the design constructs characterized structurally. We note that the scar was not explicitly modeled during generation of alternative de novo states, single-state design of these states, or multi-state design, which used a scar-less sequence with numbering starting at one (see Readme note for deposited scripts).

Generating alternative backbone conformations

The starting Ca²⁺ binding-competent structure (PDB ID: 1SMG) was obtained from the Protein Data Bank. Conformer 1 of the NMR ensemble was relaxed in Rosetta with all heavy atoms in the backbone and sidechains restrained.

relax.linuxgccrelease -database $ROSETTA/main/database -ex1 -ex2 -use_input_sc -flip_HNQ -no_optH false -relax:constrain_relax_to_start_coords -relax:coord_constrain_sidechains -relax:ramp_constraints false -s 1SMG.pdb

The loop-helix-loop unit (“reshaped region”, residues 49–72) involving the higher affinity Ca²⁺ binding loop (residues 68–72 and 77) was reshaped with loop-helix-loop unit combinatorial sampling (19) using loop libraries of length 2–5 residues. The choice of loop lengths to sample were based on an analysis of loops commonly occurring in protein structures (for a detailed description, see (19)). Note that these residue ranges follow 1SMG residue numbering, i.e. these residues correspond to residues 53–76 for the reshaped region and 70–76 and 81 for Ca²⁺ binding site II in our designs and presented data.

Only models having the same length as the input were written out. Since the input structure was primarily helical, alanine was used as the clash check residue rather than valine. Any outputs that had a reshaped helix orientation compared to the input (Cα RMSD < 3Å) were filtered out to generate alternative conformations that were experimentally distinguishable. Any outputs that had a reshaped helix Cα RMSD > 10Å compared to the input were also discarded, as these structures often lacked a well-packed hydrophobic core. This left n=921 candidate alternative apo backbone conformations for further analysis.

Single-state sequence design

A residue i was defined as pointing toward the reshaped region if the cosine of the angle between the vector pointing from Cα_i to Cα_j and the vector pointing from Cα_i to Cβ_i was greater than 0.5, where j is any residue in the reshaped region. Residues in the reshaped region as well as all residues within 10Å (Cα-Cα distance) from the reshaped region and pointing toward the reshaped region were designable, i.e. allowed to change amino acid identity and rotamer conformation. Residues within 8Å (Cα-Cα distance) and pointing toward designable residues were repackable, i.e. constant amino acid identity but allowed to change rotamer conformation. Residues in the reshaped region that were highly conserved in EF hand motifs were held fixed during design (i.e. D66, D68, G69, S70, G71, T72, E77, following PDB: 1SMG numbering convention). Allowable amino acids at designable residue positions were determined by the extent of residue burial using the Rosetta LayerDesign task operation (33). Cysteine and histidine were disallowed at all positions to avoid disulfide bond formation and pH sensitivity issues. Rosetta FastDesign and RotamerTrial protocols using extra rotamers for χ₁ and χ₂ (enabled by the ex1 and ex2 options in the ExtraRotamersGeneric task operation) were applied during design. Relevant scripts and command line prompts can be found at https://github.com/Kortemme-Lab/local_protein_sequence_design. 20 designs per backbone were generated.

In silico validation of single state designs

9-mer and 3-mer fragments for structure prediction were generated using the make_fragments.pl script distributed with Rosetta (34).

make_fragments.pl -verbose -id design_id -frag_sizes 3,9 -n_frags 200 -n_candidates 1000 sequence.fasta

A biased ab initio structure prediction simulation was run as a preliminary screen for the lowest scoring design (using the ref2015 Rosetta energy function (35)) of each backbone. This was done by using only the top three 9-mer/3-mer fragments closest in Cα RMSD to the desired backbone and generating 30 decoys. If low-RMSD decoys did not have low Rosetta energies or the lowest energy decoy was >3Å in Cα RMSD to the desired backbone, the design was removed from further analysis, as it would be unlikely to fold into the desired structure in an unbiased simulation. Full Rosetta ab initio structure prediction simulations (20,000 decoys, 200 fragments at each sliding window) were run for the remaining designs.

AbinitioRelax.linuxgccrelease -abinitio:relax -use_filters true -abinitio::increase_cycles 10 -abinitio::rg_reweight 0.5 -abinitio::rsd_wt_helix 0.5 -abinitio::rsd_wt_loop 0.5 -relax::fast -in:file:fasta sequence.fasta -in:file:frag3 <fragments_3mer_file> -in:file:frag9 <fragments_9mer_file> -psipred_ss2 <ss2_file_from_frag_generation> -nstruct <num_output> -out:sf <score_file_output> -out:file:silent <silent_file_output>

We selected 11 designs for experimental characterization (Fig. S4, yeast display), which had good agreement between the design model in the reshaped region and either the lowest energy Rosetta ab initio decoy structure or the top-ranking (by pLDDT) AF2 structure (Cα RMSD < 2Å). All designs selected had high RMSD to the original holo state in the reshaped region (Cα RMSD > 4.5Å).

Yeast display screen for stable single state designs

Gene fragments of selected designs flanked by homology regions (5’: ggagggtcggcttcgcatatg, 3’: ctcgagggtggaggttccgaacaacagcttatttctgaagaggacttgta) were ordered from Integrated DNA Technologies (Table S8). 10ng of synthetic DNA was amplified in a 25μLreaction using Q5 Hot Start High-Fidelity DNA Polymerase (NEB #M0494) for 30 cycles by PCR. A strong single band at the expected size was observed for all fragments. The reaction product was cleaned using Zymo Research DNA Clean & Concentrator kits for transformation of EBY100 yeast using the protocol described in (36). We used a modified version of the pETcon vector known as pETcon v3 (RRID: Addgene_41522) that had been altered to (i) remove a long single-nucleotide stretch near the cloning region, (ii) add an N-terminal 6x-His tag to normalize intact protein level to surface display level, and (iii) include constitutively fluorescent protein Venus as the dropout sequence to aid in colony selection. Single colonies were picked, inoculated into 5mL of SD-CAA medium, and incubated at 30°C overnight for each design. The starter culture was diluted 1:10 into SG-CAA to induce protein production. A volume corresponding to 5×10⁵ cells was added to a 96 well plate and spun at 3000g for 20min to pellet the cells. Cells were washed and resuspended in PBS then incubated with anti-c-Myc Mouse mAb (Alexa Fluor 647 conjugate) (Cell Signaling Technology #2233) and 6x-His Tag Antibody Alexa Fluor 488 conjugate (Fisher Scientific #MA1–135-A488) shaking at room temperature for 30min. Excess unbound antibody was washed away with chilled PBS + 1% BSA and analyzed using a Beckman Coulter Cytoflex flow cytometer. Events were gated by forward scattering area and back scattering area to collect the live cell population then by forward scattering width and forward scattering height to select individual cells for further analysis. Cells were then gated by fluorescence intensity where the threshold separating displaying from non-displayed cells was set such that 1% of an uninduced, unstained control would pass the gate. The fraction of anti-c-Myc positive cells from single cell events was used to estimate the level of expressed intact protein level on the surface of yeast as a proxy for thermostability (37).

Protein expression and purification for experimental screening

Plasmids (pET-28a(+)) encoding the selected designed proteins were ordered from Twist Bioscience. These constructs included an N-terminal 6x-His tag and thrombin cleavage site (MGSSHHHHHHGLVPRGSHM). Escherichia coli BL21(DE3) cells were transformed with these plasmids. Colonies were inoculated into 5mL LB medium and cultured at 37°C overnight. Starter cultures were diluted 1:100 into 1L of fresh LB medium and cultured at 37°C until the OD600 reached 0.6–0.8. Then IPTG was added to a final concentration of 300μM to induce protein expression at 37°C overnight. Cell cultures were centrifuged at 8000g for 10min to spin down the cells. Cell pellets were then lysed by resuspending in 4mL/g pellet of B-PER (Thermofisher #78243) with a dissolved cOmplete protease inhibitor cocktail EDTA-free tablet (Roche #COEDTAF-RO) and incubating at room temperature for 15min. Cell lysate was centrifuged at 18000g for 30min to separate the soluble and insoluble fractions. The soluble fraction was mixed with an equal volume of equilibration buffer (50mM Tris pH 7.5, 300mM NaCl, 10mM imidazole) and 1mL HisPur Ni-NTA resin slurry volume (Thermo Scientific #88222) to pull down the His-tagged proteins. Ni-NTA resin beads were washed 3 times with wash buffer (50mM Tris pH 7.5, 300mM NaCl, 25mM imidazole). For screening assays (e.g. size exclusion chromatography, dynamic light scattering, circular dichroism spectroscopy), the protein was then eluted 3 times with 0.5mL elution buffer (50mM Tris pH 7.5, 300mM NaCl, 250mM imidazole). Purity and amount of designed protein was estimated by Coomassie-stained SDS-PAGE throughout purification (BIO-RAD #456–1095).

Circular dichroism spectroscopy

Circular dichroism (CD) data were collected on a Jasco J-710 spectrometer. Purified designs were diluted to a final salt concentration of 10mM KCl (pH 6.7). The concentrations of diluted samples were approximately 2μM determined by a Bradford assay. CD spectra were measured using a 1mm cuvette at 25°C. Melting curves were recorded at 208nm from 25°C to 95°C using a rate of 1°C/min (Fig. S21).

Analytical size exclusion chromatography

Protein samples purified by His-tag pull down were analyzed using a Superdex 75 10/300 GL (Cytiva #29148721) (switch designs I89, N89, R89, and S89) or the Superdex 200 10/300 GL (Cytiva #28990944) (single-state designs #6306 and wildtype 1SMG) size exclusion column with 100mM KCl (pH 6.7) (Fig. S21, Fig. S22).

Dynamic light scattering

Size measurements of protein samples purified by His-tag pull down using dynamic light scattering were measured on a Zetasizer Nano S90 (Fig. S22). Samples were filtered (0.22μm) prior to measurement to remove large particles. Number distribution of hydrodynamic diameter was plotted to provide an estimate of particle size.

Protein expression and purification for NMR structure calculation

[U-¹³C, U-¹⁵N]-labeled proteins were expressed by inoculating E. coli colonies into 5mL M9 minimal medium that included 4g/L [U-¹³C]-glucose (99%) and 0.5g/L [U-¹⁵N]-NH₄Cl (99%), which were grown at 37°C overnight. Starter cultures were then diluted 1:50 into 500mL fresh [U-¹³C, U-¹⁵N]-labeled M9 minimal medium and grown at 37°C until the OD600 reached 0.6–0.8. Then IPTG was added to a final concentration of 300μM to induce protein expression at 37°C overnight. The expressed proteins were purified by following the Ni-NTA resin pull down protocol described in the protein purification section with the following modifications: Rather than eluting the His-tagged protein from the beads, the beads were washed with and resuspended in thrombin cleavage buffer (20mM Tris-HCl, 150mM NaCl, pH 8.4). 4U of biotinylated thrombin (Novagen #696022) per 1mg of protein was added to the beads, and the immobilized His-tagged protein was cleaved from the beads while rocking at room temperature for 20h. To improve purity of the cleaved protein, the beads were pelleted, and the supernatant was added to fresh Ni-NTA resin. The beads were pelleted once more, and the supernatant was mixed with 32μL of streptavidin agarose slurry (Novagen #696022) per U of thrombin for 30min at room temperature. The sample was then filtered (0.45μm) and incubated with 100μM EDTA for 5min at room temperature to remove residual amounts of Ca²⁺. The sample was then buffer exchanged into 100mM KCl pH 6.7 by dialysis overnight at 4°C and concentrated to approximately 180μL for NMR experiments. Ca²⁺ was then added (if applicable) to the desired concentration.

Structure determination by NMR

5% D₂O was added to samples. The final protein concentrations determined by Bradford assay were approximately 400μM. NMR spectra were all measured at 298.1K. Two dimensional (2D) ¹H,¹⁵N-HSQC (pulse program: fhsqcf3gpph), 2D ¹H,¹³C-HSQC (pulse program: hsqcetgpsisp2), 36ms 3D HCCH-TOCSY (pulse program hcchdigp3d) and 120ms 3D simultaneous ¹³C/¹⁵N-NOESY-HSQC (pulse program: noesyhsqcgpsismsp3d) spectra were measured using a Bruker NEO 800 MHz spectrometer with a 5mm TCI H&F-C/N-D CryoProbe. 3D CACB(CO)NH (pulse program: hncocacbgpwg3d), 3D CACBNH (pulse program: hncacbgpwg3d), 3D Hcc(co)NH (pulse program: hccconhgpwg3d2), 3D Cc(co)NH (pulse program: hccconhgpwg3d3), 2D HBCBCGCDHDGP (pulse program: hbcbcgcdhdgp), and 2D HBCBCGCDCEHEGP (pulse program: hbcbcgcdcehegp) were collected on a Bruker Avance 600 MHz spectrometer with an Inverse 5mm H-C/N-D cryoprobe. Spectra were processed in TopSpin 3.6.3. For the Ca²⁺-bound structure, we also collected the following experiments on the Bruker Avance 600 MHz spectrometer: 2D TROSY (trosyargpphwg) and 3D ¹H-¹³C NOESY-TROSY (noesytrosyargpphwg) to aid in the assignment of aromatic nOes. Automated peak picking, resonance assignment, and structure calculations were performed by ARTINA (38), and dihedral angle restraints were generated by TALOS (39). Manual inspection of ARTINA chemical shift assignments was done in CCPNMR v. 3.1.0 and corrections were made as necessary (40). The manually curated chemical shift list was then included as input for a subsequent structure calculation run in ARTINA. The output candidate structure (an ensemble of n=20 conformers) with the lowest CYANA target function value was refined in XPLOR-NIH-3.7 with the refine.py script included in the distribution (eginput/gb1_rdc/refine.py) (41). The 20 lowest scoring structures out of 100 were then refined in explicit water with the wrefine.py script (eginput/gb1_rdc/wrefine.py). The ensemble of the refined structures was validated using the PDB validation server (42). For the apo I89 structures (i.e. proposals #1 and #2 from ARTINA), we re-calculated the structures in CYANA after excluding a small subset of distance restraints that could be unambiguously assigned to the other state prior to refinement. For the Ca²⁺ bound I89 structure, the structure was first calculated with no Ca²⁺ restraints using ARTINA as described above. The Ca²⁺ restraints were then included during refinement in XPLOR-NIH-3.7, which were set to 2.8Å between the Ca²⁺ ion and the oxygens involved in coordinating Ca²⁺ (43). The structural statistics from the PDB validation server can be found in Table S9.

Two-state sequence design

Initially, we attempted to use Rosetta-based methods for multi-state design. We first tried a restrained convergence algorithm, which allowed each state to explore sequence space independently and encouraged convergence to a single sequence by incrementally increasing an energy bonus if the same amino acid was sampled at corresponding positions in both states (44). This method did not appear to work well for states varying in conformation, as it was unlikely for positions that differed greatly in solvent accessible surface area or secondary structure to converge on the same amino acid, even with an energy bonus. We also attempted to use a genetic algorithm that allowed users to input a custom fitness function as input (45). However, this method performed Monte Carlo optimization of the Rosetta energy function over discrete rotamer space and was consequently computationally expensive compared to rotamer-free design methods. Due to what appeared to be low in silico success rates, we decided to use ProteinMPNN, a deep learning-based design method capable of quickly generating tens of thousands of multi-state designs (9). This method has since been shown to successfully generate proteins where domains can be hinged open and closed upon binding a peptide effector (16).

In general, a PDB file containing both the single-state alternative state 2 conformation and the natural Ca²⁺ binding protein state 1 conformation separated by a distance much larger than the approximate diameter of either state was generated by PyMOL. State 1 was conformer 1 of the PDB ID: 1SMG NMR ensemble relaxed using Rosetta with restraints on heavy atom positions. Ca²⁺ was removed from the structure, as this implementation of ProteinMPNN did not support ligand atoms.

The state 2 conformation was generated with the following design task in mind: beyond individual bistable designs, we wanted to design a family of sequences that had significantly different state population distributions in the absence of ligand despite having high sequence identity. We reasoned it would be more likely to design such a set of sequences if we first restricted the allowable sequence space during design. Initially, the set of potential designable residues for two-state design included the reshaped region and its neighbors in both states as defined for single-state design (n=37 residues). To exclude positions that did not change in environment significantly between states or were otherwise not critical in determining the preferred structure, we individually mutated each designable position in the state 2 design to the corresponding state 1 amino acid and predicted the structural impact of each “reversion” using AF2 (though one could also use an in silico deep mutational scanning approach to evaluate other potential amino acids at each position). If the predicted structure was still consistent with the state 2 conformation (Cα RMSD < 1.5Å), this mutation was considered to be in the “tolerated” sequence space of state 1. We then input a state 2 sequence containing all “tolerated” amino acid reversions into AlphaFold2 and used the best-ranked structure by pLDDT as the state 2 state during two-state design. Of note, this prediction was still highly similar to the original single-state design in structure (Cα RMSD = 1.15Å in the reshaped helix (residues 59–69) and 1.21Å over the entire backbone) and had an average pLDDT of 84.6 in the reshaped helix. The final set of designable residues included: (1) all positions that were not reverted to the corresponding state 1 amino acid through this process and (2) their neighbors (defined as in single-state design) (n=25 residues) (Table S2).

All corresponding positions between states were tied together across chains during design with equal weighting of each state. During preliminary rounds of design, we observed the last turn of the reshaped helix (residues 68–72) fraying at the C-terminal region in some of the AF2 structure predictions. To generate designs more closely matching our desired target state, we performed a sequence analysis of the designs closest in RMSD to our input structure for state 2 and found that there was a strong preference for positively charged residues and those capable of hydrogen bonding at position 62 (i.e., K, R, Q), which could then interact with D66. In subsequent rounds of two-state design, we biased this position towards those amino acid types. This kind of iterative in silico refinement of the design process to increase sequence / structure compatibility is common and should be generalizable across input structures. ProteinMPNN was used to generate 10⁴ sequences, which were input without MSA generation into ColabFold for structure prediction (46). Designs selected for experimental characterization can be found in Table S10.

NMR experiments to characterize dynamics

[U-¹⁵N]-labeled protein was prepared as described above for structure calculation except unlabeled glucose was used in place of [U-¹³C]-glucose (99%) in culture media. All experiments were done on the Bruker Avance NEO 800 MHz spectrometer with the cryoprobe specifications mentioned previously. The concentration of samples ranged from 100–200μM in 100mM KCl, pH 6.7, 5% v/v D₂O. An initial screen for Ca²⁺ modulation of conformational dynamics was done by collecting 2D ¹H,¹⁵N-HSQC spectra of all two-state designs with and without 10eq Ca²⁺ (pulse program: fhsqcf3gpph). To further probe which residues may be dynamic in the absence of Ca²⁺, a 2D ¹H,¹⁵N-HSQC temperature series from 5°C to 35°C in 5°C increments was collected for design I89. To characterize motions occurring on a micro-to-millisecond timescale, we acquired ¹⁵N rotating frame (R_1ρ) relaxation dispersion experiments using spin-lock field strengths (ω_SL) of 1.0kHz, 1.25kHz, 2.0kHz, 2.5kHz, and 3.0kHz (pulse program: hsqctretf3gpsitc3d) at T=288.1K and 298.1K with delay times of 2, 4, 6, 8, 10, 20, 30, 50, 70, 100ms, where the 30ms experiment was repeated to assess reproducibility and estimate error (47, 48). R_1ρ rates were calculated using the Bruker Dynamics Center 2.8.4. We additionally measured longitudinal relaxation (R₁) rates in the laboratory frame at both temperatures using delay times of 20, 40, 80, 200, 500, 900, and 1400ms (pulse program: hsqct1etf3gpsitc3d) (49) and then converted the R_1ρ rates to observed transverse relaxation (R₂) rates, accounting for differences in the site-specific R₁ rates and tilt angles. Heterogeneity in the temperature dependence of R₂ suggested there may be multiple exchange processes occurring on the micro-to-millisecond timescale (Fig. S12A). To further probe this possibility, we acquired a two-point ¹⁵N CPMG experiment (pulse program: hsqctrexetf3gpsi3d) at ν_CPMG = 2000Hz and 25Hz at T=298.1K (Fig. S12B, upper panel) (50). These data again indicated there were at least two exchange processes occurring at distinct timescales: a slower process detectable by ¹⁵N CPMG in helix B/site I and a faster process detectable by ¹⁵N R_1ρ at high ω_SL in the reshaped region (see Fig. S12 for details). The slower process may have originated as a feature of our single-state design for state 2, as we observed lower peak intensity in that region consistent with line broadening due to exchange (Fig S12B, lower panel).

As detailed in Fig. S13, we fit the ¹⁵N R_1ρ relaxation dispersion data globally across all residues to a simple model of the sum of two independent exchange processes occurring at separate timescales in the fast-exchange limit (51) as follows:

$${\mathrm{R}}_2={\mathrm{\Phi }}_{\text{ex},\text{fast}}{\tau }_{\text{ex},\text{fast}}/\left(1+{\tau }_{\text{ex},\text{fast}}{\omega }_{\text{eff}}\right)+{\mathrm{\Phi }}_{\text{ex},\text{slow}}{\tau }_{\text{ex},\text{slow}}/\left(1+{\tau }_{\text{ex},\text{slow}}{\omega }_{\text{eff}}\right)+{\mathrm{R}}_{\mathrm{2,0}}$$ (Eq. 1)

where ${\mathrm{\Phi }}_{\text{ex,i}}$ is the product of the state populations and chemical shift difference for a given spin (i.e. p_A,ip_B,iΔω_i²) and R_2,0 is the intrinsic transverse relaxation rate in the absence of exchange. The two exchange timescale (${\tau }_{\text{ex}}$) values were each sampled on a 2D grid with values ranging from 42–199μs (1.25-fold lower and higher, respectively, than the lowest and highest measured $1/{\omega }_{\text{eff}}$ values). Given each pair of fixed ${\tau }_{\text{ex}}$ values, the parameters ${\mathrm{\Phi }}_{\text{ex},\text{fast}}$, ${\mathrm{\Phi }}_{\text{ex},\text{slow}}$, and R_2,0 were determined using linear least squares fitting for all residues. ${\mathrm{\Phi }}_{\text{ex}}$ values were constrained to be greater than or equal to zero by removing the respective term from the fit if it produced a negative non-physical value.

Approximation of state populations using chemical shifts

To approximate the chemical shifts of pure state 1 or state 2, we used the chemical shifts of switch design I89 in the presence of Ca²⁺ and design S89 without Ca²⁺, respectively, and assumed a 5% minor state population for each. Peaks that were well-resolved in the 2D ¹H,¹⁵N-HSQC spectrum and corresponded to residues facing or in the reshaped region but distal from position 89 and Ca²⁺ site II were selected for analysis. To estimate state populations, we assume that each peak position was a population-weighted average of the two endpoints.

Determining Ca²⁺ binding affinity

The expression and purification of [U-¹⁵N]-labeled protein were as described previously to yield ~1mM protein. Stock solutions of 10mM, 100mM, and 1M CaCl2 in 100mM KCl, 95% H₂O/5% D₂O were prepared and adjusted to pH 6.7 with HCl. For each titration point, a volume of stock solution was added such that the [Ca²⁺]/[L] ratio approximately doubled, ranging from [Ca²⁺] = 0.025–51.2mM. The titration was monitored by collecting a series of 2D ¹H,¹⁵N-HSQC spectra and tracking the chemical shift of residues in Ca²⁺ binding site II that were well-resolved at all titration points (V71, D78). We fit the data to the following equation to calculate the apparent K_d for each residue:

$$\mathrm{\Delta }{\delta }_{\mathrm{o}\mathrm{b}\mathrm{s}}=\mathrm{\Delta }{\delta }_{\mathrm{m}\mathrm{a}\mathrm{x}}\{\left(\mathrm{n}[\mathrm{P}]+[\mathrm{L}]+{\mathrm{K}}_{\mathrm{d}}\right)-{[{\left(\mathrm{n}[\mathrm{P}]+[\mathrm{L}]+{\mathrm{K}}_{\mathrm{d}}\right)}^2-4\mathrm{n}[\mathrm{P}][\mathrm{L}]]}^{1/2}\}/2\mathrm{n}[\mathrm{P}]$$ (Eq. 2)

where $\mathrm{\Delta }{\delta }_{\text{obs}}$ is the change in the observed shift from the Ca²⁺-free state, $\mathrm{\Delta }{\delta }_{\text{max}}$ is the maximum shift change on saturation (treated as a free variable during fitting), [P] and [L] are the total protein and ligand concentrations, respectively, and n is the number of sites with approximately (within an order of magnitude) similar ${\mathrm{K}}_{\mathrm{d}}$ values. To account for binding to site I, which is of comparable affinity, we set n equal to 2, as described previously (52).

Molecular dynamics simulations

Simulations of designs were performed using GROMACS 2022.5 (53). The initial structures used for the triplicate 2μs simulations of I89 and S89 in the presence/absence of Ca²⁺ were the AMBER-relaxed AF2 predictions for the design with the N-terminal thrombin cleavage site scar included to be consistent with experimentally characterized designs. For simulations with Ca²⁺, we initiated the trajectory with the Ca²⁺ ion in site II. To more extensively sample the conformational space of design I89, we collected 7 additional 1μs long trajectories starting from the AMBER-relaxed AF2 prediction. Furthermore, we sampled design I89 initiated from different starting conformations using snapshots from previously collected trajectories that had (1) low Cα RMSD of the reshaped helix to state 2 to initiate 10 additional 1μs trajectories (three of which were extended to 2μs) and (2) intermediate Cα RMSD of the reshaped helix to both states 1 and 2 to initiate 10 additional 1μs trajectories from an intermediate conformation (frames designated by triangles in Fig. S18). The proteins were parameterized using the a99SB-disp force field and water molecules were parameterized using the a99SB-disp water model (54). The systems were solvated and neutralized as in Table S11. Energy minimization of the system was performed with the steepest descent minimization algorithm to a tolerance of 1000.0 kJ mol⁻¹ nm^-1. Equilibration was performed in the NVT ensemble for 1000ps at 300K using the Berendsen thermostat. Systems were then equilibrated in the NPT ensemble for 100ps at a target pressure of 1 bar at 300K maintained by the Berendsen thermostat with position restraints on all heavy atoms. Bond lengths and angles of protein atoms were constrained with the LINCS algorithm and water constraints were applied using the SETTLE algorithm. The PME algorithm was used for electrostatics with a grid spacing of 1.8nm. Van der Waals forces were calculated with the Verlet cut-off scheme using a 1.2nm cut-off distance. Snapshots were saved every 80ps. Initial structures were AlphaFold2 structure predictions of two-state designs (including the residues GSHM at the N-terminus corresponding to a scar from the thrombin cleavage site) relaxed in AMBER. The Cα RMSD of the reshaped region (residues 53–76) were calculated over the course of a 1μs simulation using GROMACS 2022.5 command line tools.

Mutual information analysis

To quantify correlations between residue dihedral angles during the apo I89 simulation, the first and third replicates (in which a state 1 to state 2 transition was observed) were split into 200ns blocks, discarding the first 100ns. For each block, dihedral angle data was extracted with:

gmx chi -f md.xtc -s md_1_us.tpr -phi -psi -omega -rama -all -maxchi 4 -HChi -b 5000

We then calculated the mutual information matrix using the MutInf method (23):

dihedral_mutent -x <base dir> -d / -o 6 -n 7 -w 30 -p 0 -c “yes” -a “yes” I89.reslist > I89_mutinf.out

where the *.reslist file was a list of all residues. The file containing mutual information between pairs of residues, with zero diagonal (* bootstrap_avg_mutinf_res_sum_0diag.txt) was visualized a heatmap.

Markov state modeling

For each simulation frame, we first aligned the Cα atoms of the non-reshaped region with well-ordered secondary structure (residues 7–17, 20–33, 40–53, 79–90) to our computational models for state 1 or state 2 and calculated the Cα RMSD of the reshaped helix (residues 59–69) to each reference state. We then used mini-batch k-means clustering on these two RMSD values, assessing models fit to 2, 5, and 10 clusters using all aggregate simulation data from I89 simulations in the absence of Ca²⁺ (Fig. S18). Markov state models (MSMs) were then constructed for each number of total states using a range of lag times. For each model, mean first passage times (MFPTs) were computed between the clusters most representative of our designed state 1 and 2. MFPT values appeared to converge as the number of states increased where the absolute difference between MFPTs computed from a 5-state versus 10-state MSM model was less than 10μs for transitions from state 2 to state 1 and less than 0.5μs from state 1 to state 2 (Fig. S20A). As both cluster sizes provided similar results, we chose 5 states for further analysis. We did not sample lag times greater than 300ns as that would result in discarding over one-third of our simulation data. The implied timescales for the four slowest modes of the 5-state MSM were then plotted (Fig. S20B), and a lag time of 200ns was chosen for further analysis as the timescales begin to plateau at this point, suggesting that this MSM is a faithful representation of the underlying dynamics (55). Clusters, MSMs, and MFPTs were calculated using MSMBuilder v3.8 (56). MFPTs from the MSM were converted to implied radian units (i.e. divided by a factor of 2π) for direct comparison to NMR experimental data in the main text. The complete list of MFPTs (as output by MSMBuilder without implied radian units) and stationary probabilities can be found in Table S12. Alignment, RMSD calculations, and each cluster medoid structure file were performed and generated using MDAnalysis v2.8 (57).

Frame2seq scoring of point mutants

To predict the effects of single point mutations on the conformational equilibrium of switch design I89, we computed sequence likelihoods given structure using Frame2seq (10). Frame2seq models learn to approximate P(sequence|structure) via a masked language modeling objective. Pseudo log-likelihood (PLL) has been explored for scoring sequences (58), which we adapted for a structure-conditioned ranking task. For each conformational state, we separately output Frame2seq model PLL by providing the structure and the sequence as input with a mask introduced at the mutated position. We then subtracted the single point mutant negative PLL from the wildtype negative PLL (the reference point) to compute the score as follows:

$$\text{score}=\mathrm{l}\mathrm{o}\mathrm{g}(\mathrm{p}({\mathrm{x}}_{\mathrm{i}}={{\mathrm{x}}_{\mathrm{i}}}^{\mathrm{w}\mathrm{t}}\mid {{\mathrm{x}}_{-\mathrm{i}}}^{\mathrm{w}\mathrm{t}},\mathrm{Y})-\mathrm{l}\mathrm{o}\mathrm{g}(\mathrm{p}({\mathrm{x}}_{\mathrm{i}}={{\mathrm{x}}_{\mathrm{i}}}^{\mathrm{m}\mathrm{t}}\mid {{\mathrm{x}}_{-\mathrm{i}}}^{\mathrm{m}\mathrm{t}},\mathrm{Y})$$ (Eq. 3)

where $\mathit{i}$ is the mutated position, ${\mathit{x}}^{\mathit{w}\mathit{t}}$ is the wildtype sequence, ${\mathit{x}}^{\mathit{m}\mathit{t}}$ is the mutant sequence, ${\mathit{x}}_{-\mathit{i}}$ is the sequence $\mathit{x}$ with a mask introduced at position $\mathit{i}$, and $\mathit{Y}$ is the structure.

Supplementary Material

Figures S1 to S22

Tables S1 to S12

Movie S1

Data and materials availability:

Expression plasmids for single-state and switch designs have been deposited to Addgene with accession codes 231958, 231959, 231960, 231961, 231962, 231963, 231964, 231965, 231966. Structures have been deposited to the Protein Data Bank (PDB) with accession codes 9CIC, 9CID, 9CIE, 9CIF, and 9CIG. NMR data have been deposited to the Biological Magnetic Resonance Data Bank with accession codes 31182, 31183, 31184, 31185, and 31186. The data and scripts used in this study have been deposited in Dryad (59) and GitHub (https://github.com/amyguo1997/dynamic_protein_design). All other relevant data are available in the main text or supplementary materials.