Characterizing Transient Protein-Protein Interactions by Trp-Cys Quenching and Computer Simulations

Abstract

Transient protein-protein interactions occur frequently under the crowded conditions encountered in biological environments. Tryptophan-cysteine quenching is introduced as an experimental approach with minimal labeling for characterizing such interactions between proteins due to its sensitivity to nano- to microsecond dynamics on sub-nanometer length scales. The experiments are paired with computational modeling at different resolutions including fully atomistic molecular dynamics simulations for interpretation of the experimental observables and to gain molecular-level insights. This approach is applied to model systems, villin variants and the drkN SH3 domain, in the presence of protein G crowders. It is demonstrated that Trp-Cys quenching experiments can differentiate between overall attractive and repulsive interactions between different proteins, and they can discern variations in interaction preferences at different protein surface locations. The close integration between experiment and simulations also provides an opportunity to evaluate different molecular force fields for the simulation of concentrated protein solutions.

Graphical Abstract

Biological cells are densely packed with proteins and other biomolecules at volume fractions of 20–40% in the aqueous phase¹. Frequent encounters between diffusing proteins are unavoidable at such high concentrations, impacting protein diffusion, structure, and dynamics^2–7. Varying propensities for aggregation and phase separation^{8, 9} add an additional dimension. A starting point for understanding crowding is the excluded-volume effect where diffusion is limited and where compact arrangements are favored for lack of free space¹⁰. Beyond this simplifying scenario, proteins may interact transiently to different degrees as a function of protein and concentration^11–15. Dynamic clusters that persist on nanosecond to microsecond time scales can impact diffusion properties^16–19 or interfere with ligand binding events²⁰. In fact, the slow-down in diffusion experienced by proteins in crowded environments may be primarily due to slower-diffusing transient clusters rather than increased solvent viscosity or reduced free space^{19, 21}. Experimental evidence has come primarily from the key observation that diffusion varies as a function of protein crowder species at the same volume fractions^22–24, and that rotational diffusion is slowed down as much as or more than translational diffusion¹², a hallmark of interacting particles. Other approaches have explored transient interactions more directly in specific biological systems^{15, 25–28}. However, many questions remain about the determinants of transient interactions between two arbitrary proteins under highly concentrated conditions.

A direct characterization of the transient interactions between any pair of proteins via experiment is difficult as such interactions are short-lived and lack strong preferences for specific contacts. In this work, we introduce quenching of the tryptophan (Trp) triplet state by cysteine (Cys) as a new tool for examining the details of such interactions in concentrated solutions. Trp-Cys quenching has been used for many years for investigating intramolecular dynamics of disordered polypeptide chains^29–34. A Trp residue within a folded protein is excited to a long-lived triplet state via the fluorescence singlet state. In the absence of quenchers, the lifetime of this state is ~40 μs³⁰, and it is much more efficiently quenched via electron transfer by Cys than by any other amino acid. Trp-Cys quenching is sensitive to microsecond dynamics over distances up to 10 Å, similar to typical bimolecular distances under physiological conditions⁵, when 1–3 water layers are separating macromolecules³⁵. The Trp triplet in a concentrated solution may have a shorter lifetime than in dilute solutions, but it is still measurable, so that contacts can be distinguished from close interactions. Trp-Cys quenching is thus well-suited for characterizing transient intermolecular encounters in concentrated solutions. A further advantage is that only natural amino acids are involved and additional labeling that may interfere with protein-protein interactions can be avoided. Moreover, because of the sensitivity to relatively short distances, the location of the probes on the protein surfaces provides spatial resolution that can be explored by mutating Trp or Cys residues. As Trp-Cys quenching decays in crowded conditions are difficult to interpret at the molecular level, computer simulations are needed to provide context. Molecular dynamics simulations of concentrated protein solutions in atomistic detail over microseconds are now possible^{36, 37}, and this work presents a close comparison between experiment and simulation, including an opportunity for validating the force field parameters used in the simulations. Once validated, the simulations can reveal additional information about protein-protein interactions at the molecular level. Together, this integrated experimental-computational approach provides direct, experiment-driven insights into transient protein interactions at elevated protein concentrations.

Experiments:

Trp-Cys quenching experiments were carried out for villin or drkN SH3 probe molecules in the presence of protein G quenchers. Fig. 1A shows the decay curves for the K33W variant of villin at different concentrations of protein G along with their derivatives vs. log(t). Without protein G, the decay curve reflects the unquenched lifetime of the Trp triplet state (40 μs). The slightly non-exponential decay is due to low-probability quenching by non-Cys amino acids within the folded structure³⁰. Once protein G is added, decay occurs on shorter time scales due to Cys quenching. The resulting quenching rates in the presence of protein G still display non-exponential behavior as indicated by a broad, shallow minimum in the derivative of absorbance vs. log(t). When the protein G concentration is increased, both the location of the minimum in time and the depth are affected. A shift to shorter times at higher concentrations results from more frequent protein-protein contacts. The depth changes as quenching shifts from slow, but short-range intramolecular quenching by other amino acids in the villin to faster intermolecular quenching by cysteine from interacting protein G molecules.

Figure 1.

Experimental measurements of Trp triplet lifetime in villin variants and SH3 in the presence of protein G. Decays are shown as absorbance vs. time (top) and as the derivatives of the decays vs. log(t) calculated as the slope of a linear fit over a sliding window of 21 time points (bottom). (A) K33W at various concentrations of protein G. (B) Villin headpiece wild-type (black), V10W (red), K33W (blue), R15T+K30E (green) and SH3 (magenta). Curves were averaged from six independent measurements and normalized to 0.88 at 147 ns and to 0 at 369 μs. Error bars represent the standard deviation. Error bars in the derivatives represent the error of the fit.

The quenching curve for the SH3 domain is shifted significantly to longer times compared to the villin variants (Fig. 1B and Table S1), indicating that interactions between different proteins can be distinguished. On the other hand, decay curves for different villin mutants at a fixed protein G concentration have varying depths of the derivative minimum with little shift in time. This suggests that differences in local interactions with respect to the Trp location (residue 24 in wild-type villin and the R15T+K30E mutant, and residues 10 and 33 in the V10W and K33W mutants, respectively) can be probed as well. Computational modeling results will explain the origin of the varying quenching curves below.

Modeling of Trp-Cys quenching on 1D potentials:

A simple computer model was applied to interpret the experimental quenching curves. A 1D variable, the probe-quencher distance, was sampled stochastically via Monte Carlo (MC) simulation using different contact potentials (Eq. S1). Fig. 2 shows quenching curves obtained with potentials that approximate Trp-Cys distance distributions extracted from atomistic simulations described below. The shape and characteristics of the decay curves are similar to the experiments. When potentials are varied with different contact minima depths and different kinetic barriers near the contact minimum (Fig. 2A), the calculated quenching curves (Fig. 2B) shift little in time but the derivative minimum varies in depth. The depth is most strongly correlated with the difference in probability between the contact minimum and large separation distances (calculated at 10 Å, Fig. S1A). Weaker interactions with a shallower contact minimum result in a deeper derivative value, vice versa, stronger interactions with a deeper contact minimum lead to shallower derivatives. The barrier height near the contact minimum, measured as the difference in energy at the peak vs. the energy at 10 Å, has a smaller effect as higher barriers tend to shift the minimum down (Fig. S1B). In addition, smaller barriers cause a slight shift to shorter times and with higher barriers to longer times (Fig. S1D). If the potential is further modified by introducing long-range attraction or repulsion (Fig. 2C), quenching curves shift to longer time scales with a deeper minimum in the case of repulsion; with long-range attraction the derivative minimum mainly shifts upward (Fig. 2D). The shift to longer time scales in the case of repulsion is a result of less frequent contacts, whereas additional attraction does not significantly shift the time scale because most of the time is already spent in or near contact. The curves are also shifted in time when the system size is reduced (Fig. S2). A reduced system size is a model for increased concentration and the changes in the curves are similar as seen in the experiments at different protein G concentrations.

Figure 2.

Computational modeling of Trp-Cys quenching via 1D potentials. Probability distribution functions are shown on the left (A, C, E) and derivatives of the calculated quenching curves are shown on the right (B, D, F). The reference distribution in black in A-D is based on the contact probability extracted from atomistic simulations for the wild-type villin structure (E). Variations in the potential near the contact minimum are shown in A and C. Variations in long-range attraction (d=−1 in Eq. S1, red) or repulsion (d=1 in Eq. S1, green) are shown in C and D. Trp-Cys contact probability functions from atomistic simulations are shown in E (solid lines) along with fitted 1D potentials (dashed lines). Atomistic probabilities are for minimum villin-protein G Trp-Cys distances. Distances are limited to 25 Å since there is almost always a protein G molecule sufficiently close for minimum distances to be less than 30 Å due to crowding. Derivatives on the left in F were obtained with uniform Monte Carlo sampling, derivatives on the right in F resulted from sampling with diffusion-matched, distance-dependent step sizes (see Supplementary Material). Colors in E and F reflect different probes: villin wild-type (black), villin V10W (red), villin K33W (blue), villin R15T+K30E (green), SH3 (magenta).

This analysis allows a first interpretation of the experimental data. The shift to longer time scales for SH3 vs. villin, at the same concentrations, reflects long-range repulsion between SH3 and protein G relative to villin and protein G. This may be expected as both SH3 (−6) and protein G (−5) have net negative charges, whereas, villin has a net positive (+2) or slightly negative (−1) charge for the double mutant. On the other hand, the observed differences between villin variants appear to stem from different contact potentials near villin residue 24 (wild-type and R15T+K30E), near residue 10 (V10W mutant), or near residue 33 (K33W mutant), since the derivative minima are deeper for the V10W and K33W single mutants than for the wild-type and double mutant. There is also a difference between the wild-type and double mutant with the deeper minimum for the double mutant suggesting weaker interactions, presumably due to a more negative charge compared to the wild-type.

Atomistic simulations of Trp-Cys quenching:

To further interpret the experimental observations at the molecular level we applied μs-scale molecular dynamics (MD) simulations. The simulations sample the diffusion and interactions between the proteins in the same systems that were studied experimentally. The survival probabilities of the triplet state of Trp were calculated from the simulation according to Eq. 1 based on the distance between Trp and Cys residues, including time-scale corrections (see Methods). They are compared with the experimental results in Fig. 3. For wild-type villin, we considered four CHARMM force field variants c36³⁸, c36+water¹⁶, c36m³⁹, and c36mw³⁹ (Fig. 3A). All force fields resulted in somewhat faster decay curves compared to experiment. That is discussed further below. Depending on the force field, the minimum depth in the derivative of the survival probability function varies. The deepest minimum is found for c36+water, followed by c36mw³⁹, both of which decrease protein-protein interactions by scaling water interactions (Fig. S3). Shallower minima are found for c36m³⁹ and c36³⁸ where protein interactions are stronger according to larger and longer-lived transient clusters (Fig. S4). C36 and c36m perform similarly, but the agreement with the experimental data is somewhat better for c36m. c36m has also been found to perform well in other contexts^39–41. Therefore, we subsequently focus on simulations using c36m.

Figure 3.

Comparison of survival probability decays between experimental measurements and simulated data. The probability and its derivative against log-time are shown at top and bottom of each panel, respectively. Derivative values were obtained from a linear-fit of probabilities against log-time with a window size of 21. Experimental values shown as solid lines with transparent shades for standard errors were adjusted to match the simulated data at 100–500 ns since the initial decay varies in the simulations and experiments are not sensitive to quenching during the initial 100 ns after excitation. A: Results for wild-type villin obtained using different force fields c36 (orange), c36+water (lime), c36m (black), and c36mw (cyan); B: Results with c36m for wild-type villin (black), V10W (red), K33W (blue), and the R15T+K30E double mutant (green); C: Results with c36m for wild-type villin (black) and SH3 (magenta).

The simulations capture many of the features seen in the experimental quenching curves (Fig. 3B). The double mutant, R15T+K30E, has a deeper minimum in the derivative than the wild-type, presumably due to weaker local interactions. The K33W mutant shows weaker interactions than V10W as in the experiment. For both single mutants, the derivatives of the survival probabilities are in excellent agreement with experiment, but the minima of the wild-type and double mutant are both shifted downward and to shorter times compared to experiment. Comparing wild-type villin with SH3, the simulations show a similar shift to longer times and with a deeper minimum for SH3 as in experiment (Fig. 3C).

The shifts in time scale are not explained by differences in translational diffusion (Table S2). Diffusion is similar for all villin variants and SH3, even although SH3 is heavier than villin (6811.5 vs. 4214.9 g/mol), essentially because of less cluster formation with SH3 (Fig. S5). To further understand these results, we turn again to the 1D model introduced above. Trp-Cys distance probabilities extracted from the atomistic simulations have a clear contact minimum around 3.5–4 Å for villin, and at 4–5 Å for SH3, followed by a barrier at varying locations at 4.5–8 Å (Fig. 2E). There is also a longer-range decline of distance probabilities with a plateau for villin mutants reached at 15 Å whereas the decline continues to longer distances for SH3. A depletion beyond 10 Å is generally expected as protein G favors interactions with villin due to crowding but may interact in many arrangements that do not place protein G’s C10 near the villin Trp. MC sampling was carried out on 1D potentials fitted to reproduce the atomistic contact probabilities (Fig. 2E/F). When only the short-range potential (<10 Å) is considered, quenching occurs on similar time scales and the derivative minima are ordered mostly as in the experiment (i.e. the double mutant is lower than wild-type villin and V10W and K33W are lower than the wild-type) (Fig. S6). When the long-range decay of the Trp-Cys contact probability is added, the villin variants shift slightly in time, while maintaining their order, whereas the decay for SH3 shifts more significantly to slower times because of longer-range repulsion, again qualitatively similar to experiment (Fig. 2E/F). This analysis explains the experimental data directly from distance contact probabilities extracted from the atomistic simulations.

We now turn to the shift to shorter times in the quenching curves for wild-type villin and the double mutant in the atomistic simulations. We found that although contact potentials are similar, diffusion in the Trp-Cys distance differs significantly (Fig. S7). At a distance around 10 Å, diffusion is about twice for wild-type villin and about five-fold for the double mutant and SH3 compared to diffusion in the V10W and K33W mutants at that distance. MC sampling with a fixed maximum step size does not reproduce such differences in kinetics (Fig. S7B). However, by introducing variable MC step sizes for different systems (see Methods), diffusion kinetics similar to the atomistic simulations can be obtained via MC sampling (Fig. S7D). Once the 1D potentials were resampled with the modified pseudo-kinetics, quenching curves for wild-type villin, the double mutant, and SH3 were shifted to shorter time scales relative to V10W and K33W as in the atomistic simulations (Fig. 2F). The depth of the minima also changed slightly, with the wild-type curves now lower than V10W as in the atomistic simulations indicating that altered kinetics also affect the quenching curve derivative minima.

This leaves the question why the Trp-Cys distances fluctuate more rapidly with respect to Trp at the villin 24 position (probed by the wild-type and double mutant) vs. Trp at the 10 and 33 positions in the simulations but not in the experiment. As stated above, overall protein diffusion is not significantly different between the villin variants. Trp-Cys contact lifetimes based on contact correlation analysis are also similar between wild-type and the single mutants (Fig. S8). However, conformational fluctuations of the Trp residue itself vary depending on where it is located in the villin structure (Fig. S9). With or without protein G, Trp (or Tyr when substituted in V10W and K33W) is more dynamic at the 24 position vs. Trp at the 10 and 33 positions. The differences become significant at time scales beyond about 10 ns when protein G is present (Fig. S9). Since the systems studied via simulation are more concentrated than in experiment, villin-protein G interactions persist long enough for Trp fluctuations to affect the quenching rates. More specifically, it appears that the faster quenching rates for wild-type villin and the double mutant relative to V10W and K33W result from long associations between villin and protein G in arrangements such as the ones shown in Fig. S10 and Movie S1 where the dynamics in the Trp-Cys distance is due in part to fluctuations of W24. For comparison, Movie S2 shows quenching at V10W where the Trp is more rigid. To further test this idea, we calculated hypothetical quenching curves relative to the residue at position 24 (Tyr for V10W and K33W) based on the V10W and K33W simulations. The resulting quenching curves are shifted also for V10W and K33W (Fig. S11). Therefore, we conclude that the dynamics at the 24 position of villin is determining the faster kinetics in the simulations instead of other differences due to the V10W and K33W mutations.

Based on the simulations, we find that overall villin-protein G interaction preferences are similar for the three variants (Fig. 4F). Interactions between protein G and villin residue 24 are slightly preferred over interactions near villin residues 10 or 33 (Fig. 4F and S12), but strong preferences for the vicinity of residue 24 only emerge when focusing specifically on quenching interactions (Fig. 4E), meaning when the vicinity of protein G C10 is interacting with villin. In more detail, it appears that quenching near residue 24 is highly specific to the immediate vicinity of residue 24 (Fig. 4A,B,E), whereas quenching near V10W and K33W is broader (Fig. 4C,D,E and Fig. S13). This may be expected since the surface near 10 and 33 is flatter compared to the 24 site pointing into solvent at the tip of the villin structure. From the perspective of protein G, the vicinity is of C10 is highly preferred for interactions with villin but there is little difference when interacting with different villin variants (Fig. S14). This suggests that the Trp-Cys quenching experiments are a highly sensitive tool for mapping out residue-specific transient interactions involving both the probe and quencher and that the integration with simulations can provide additional details.

Figure 4.

Residue-wise contacts between protein G and villin variants, projected onto the villin surface (A-D) and as a function of villin residue index (E, F). Surface projections are shown for Trp-Cys quenching contact (left) and for all contact positions (right) for wild-type villin (A), the R15T+K30E mutant (B), the V10W mutant (C), and the K33W mutant (D). The location of the Trp residue is indicated by arrows. Contacts per frame vs. residue index are shown at the time of quenching contact (E) and at any time of contact (F) with different variants colored as in Fig. 1. Shaded areas indicate standard errors. Contacts were defined by residue pairs whose inter-atomic distances were closer than 5 Å.

Trp-Cys quenching is a highly sensitive tool to investigate close-range interactions on the time scales in which transient interactions occur under concentrated conditions. The comparison between villin and SH3 shows that Trp-Cys quenching can distinguish interaction strengths between different proteins, and mutants of villin show relative preferences over the surface of the protein. Note that these results were validated despite the relatively low concentration of protein G in the experiments compared to the simulations. Moving the location of the Trp to additional points on the villin or SH3 surface or moving the Cys to other locations on the protein G surface should provide a more complete map of interaction preferences. A key part of this work was the interpretation of the experimental data by computational modeling and atomistic simulations. We show that we can come to similar conclusions based on either experiment, interpreted via the 1D model, or based on simulations, after appropriate time scale corrections. This validates the simulations, but more importantly it confirms the interpretation of the experiments via the 1D modeling.

The results suggest a number of ways to move forward with this technique. On the experimental side, a comparison of decay curves for different mutants of any two proteins will highlight the relative strength of bimolecular interactions. The need for Trp and Cys residues on the protein surfaces places some limitations on the systems that can be studied, but the main requirements are that only one protein has at most a small number of Trp residues on the surface and that there are no Cys residues on the probe protein near the Trp to avoid self-quenching. Analysis of quenching curves with a single Trp residue is more straightfoward, but quenching with multiple Trp residues may be deconvoluted via modeling. Since natural abundance of Trp is low, the requirements can be easily satisfied for many proteins either with wild-type sequences of via a small number of mutations. Protein solubility, especially at concentrations approaching biological cellular conditions, is a significant problem for in vitro experiments. We could not reach higher concentrations for the systems studied here, but we expect that this will be possible with alternate constructs, for different systems, or by using condensates. Our previous work on condensation of folded proteins and RNA shows that such condensates are possible⁴². There is otherwise no fundamental reason why Trp-Cys quenching could not be applied at higher concentrations. To support this point, we calculated unscaled quenching curves from simulations at 33 and 66 g/L, more than ten times higher than in the experiments (Fig. S15). According to the simulations, the majority of the absorbance decay occurs between 100 ns and 1 μs, well above the instrument limit of around 20 ns. However, even if higher concentrations cannot be reached in the experiments, the present study demonstrates how the combination with simulations allows extrapolation to higher concentrations.

On the computational side, the 1D model is an inexpensive way to visualize the potentials of mean force between proteins. Hypothetical potentials can be convolved with the quenching mechanism to produce decay curves that may be compared with experiments. It remains to be explored whether this can be taken a step further by developing a protocol in which an interaction potential is extracted directly from experimental data under certain assumptions, such as fixed concentrations and with a focus on relative differences between proteins in comparison to a known reference.

Atomistic simulations provide a more complete picture, such as the role of side chain fluctuations in accelerating quenching rates under crowded conditions. While we have compared different force fields for the systems studied here, additional experiments on a wider site of systems will allow a more comprehensive force field assessment in future work. In another direction, the present work may also facilitate the development and assessment of predictive CG models to avoid the large computational costs of atomistic simulations. The degree to which various types of simulation, modeling and experiment are deployed to understand protein-protein interactions will depend on the questions that need to be addressed.

EXPERIMENTAL AND COMPUTATIONAL METHODS

Systems:

Sequence variations of the HP35 fragment of the villin headpiece (see Table S3) and the T22G mutant of the drkN SH3 domain, each with one Trp, were studied in the presence of the B1 domain of protein G, mutated to contain one Cys (K10C) and no Trp (W43Y).

Experiments:

Villin and SH3 concentrations were held constant at 30 μM. Concentrations of protein G were varied up to 0.84 mM. The decay of the Trp triplet state was measured with a two-laser pump-probe setup: a 10 ns pulse at 289 nm excites the triplet state while a continuous wave laser at 450 nm probes the population of the triplet state by transient absorption. A background due to additional excitation of four tyrosines in the concentrated protein G was subtracted (Fig. S16).

In Trp-Cys quenching, the Trp triplet decays exponentially with the quencher distance as

$$q(r)=k_{0}exp(-\beta (r-a))$$ (1)

with a=3.5 Å (distance of closest approach), k₀=4.2×10⁹ s⁻¹ (decay rate at a), and β = 4 Å ⁻¹ (decay constant). With the large value of β, decay rates remain significant up to distances of about a+5 Å. Interactions within this range or diffusion in and out of close interactions on time scales shorter than the triplet state lifetime, i.e. <40 μs, result in non-exponential decay.

Monte Carlo sampling of 1D potentials:

1D potentials were constructed with strong repulsion at close distances and weak attraction at a contact distance. Additional Gaussian functions were used to model kinetic barriers and kinetic traps at different distances (Fig 2). The potential was sampled via MC sampling to create diffusive pseudo-dynamics without a meaningful time scale, but desired diffusion kinetics can be imposed by varying the MC step size.

Atomistic simulations:

MD simulations in atomistic detail were carried out for the experimentally studied systems., but at fivefold higher concentrations because of computational cost. Multiple simulations on μs time scales were carried out with different force fields (Table S4). Simulation results were adjusted in time-scale to compare with the experimental results. Differences in protein concentrations were accounted for by extrapolation guided by coarse-grained MD simulations (Fig. S17). Additional time-scale corrections were applied to correct for underestimated solvent viscosity and artefacts due to periodic boundaries estimated based on protein diffusion calculated from mean square displacements (Fig. S18 and Table S2).

Calculation of triplet state survival probability:

From sampled probe-quencher distance time series, survival probability curves were calculated from Eq. 2 using the quenching function in Eq. 1 for the distances sampled via MC or MD.

$$S_{quenching}(t)=\langle \exp \left(-{\int }_0^{t}q\left(t^{\prime }\right)d{t}^{\prime }\right)\rangle =\frac{1}{t_{max}-t}{\sum }_0^{t_{max}-t}\exp \left(-{\int }_0^{t}q\left(t^{\prime }\right)d{t}^{\prime }\right)$$ (2)

The triplet-to-singlet transition due to auto-bleaching was considered via Eq. 3 with an auto-bleaching rate constant (k₀) of 2.3 × 10⁻⁵ / ns ³⁰.

$$S(t)=S_{quenching}(t)\times e^{-k_{0}t}$$ (3)