Fuzzy supertertiary interactions within PSD-95 enable ligand binding

Article activity feed

1. 

   Version published to 10.7554/elife.77242 on eLife

   Sep 7, 2022
2. 

   eLife

   Aug 24, 2022

   Author Response

   Joint Public Review

   1. The structures of the PDZ domains of PSD95 have been determined and they are well-folded and stable. In addition, the PSG module has been shown to adopt a stable structure after expression and purification. The authors should cite papers, their own and those by Zeng et al. (e.g. J. Mol. Bio, 2018), to reassure readers that the protein is not destabilized by the cysteine mutations. The authors need to state how many purifications of the mutants have been done and how many replicates have been made for the FRET measurements. Did the FRET data change over time?

   We appreciate the importance of selecting labeling sites that do not disrupt protein structure and activity. There are two protein constructs in this work: full-length PSD-95 and the PSG truncation of this same protein, which have been …

   Author Response

   Joint Public Review

   1. The structures of the PDZ domains of PSD95 have been determined and they are well-folded and stable. In addition, the PSG module has been shown to adopt a stable structure after expression and purification. The authors should cite papers, their own and those by Zeng et al. (e.g. J. Mol. Bio, 2018), to reassure readers that the protein is not destabilized by the cysteine mutations. The authors need to state how many purifications of the mutants have been done and how many replicates have been made for the FRET measurements. Did the FRET data change over time?

   We appreciate the importance of selecting labeling sites that do not disrupt protein structure and activity. There are two protein constructs in this work: full-length PSD-95 and the PSG truncation of this same protein, which have been expressed hundreds of times over more than a decade in my lab. The cysteine mutations used in this work have all been validated as non-disruptive to the protein and the dyes in several ways. 1) We selected labeling sites using the available x-ray and NMR structures to ensure surface accessible residues within alpha helices or short loops to minimize tertiary structural disruption; 2) we ensured that the two point mutations don’t affect the expression and purification protocols. Misfolding or changes in conformation would be visible on elution profiles from chromatography as well as proteolytic cleavage patterns, which are sensitive to protein folding; 3) in our previous work, we measured both donor anisotropy and acceptor quantum yield for all of the variants in use here but one, which relied on existing sites in a new combination. Dyes involved in interactions with proteins or changes in dye environment would become apparent through changes in quantum yield and anisotropy. Any problematic labeling sites have been purged from the current work, which uses a small subset of the mutants from our earlier work. The repeatability of the expression and purification of all these constructs has been demonstrated in our published work and is not affected by the specific labeling mutants in use. The stability of these constructs is supported by the numerous other NMR and x-ray crystallography studies published on these robustly expressing proteins. To highlight this important issue, we have added additional discussion of the origin and validation of these mutants in the text on page 4 and in the methods section. We also included references to the tables of photophysical measurements for the library of PSD-95 cysteine mutants adapted for this study.

   We did not explicitly track the number of purifications used in this work, which spanned more than five years. We were not aware of any expectation to provide such records but will be more aware going forward. The measurements for this paper come from one or in some cases two protein expression runs, each of which generates 2 or more cell pellets. Each of these pellets generates a single affinity and ion exchange purified sample. This is then aliquoted and frozen, which may produce more than a dozen samples for fluorescent labeling. Individual labeled samples are given additional rounds of desalting and size exclusion chromatography immediately before measurements to ensure than the full length proteins are used and that there has been no aggregation or degradation. In terms of repeatability, the data shown in this manuscript involves repeat measurements of the same constructs using different FRET dye pairs, collected on different instruments at different times and still shows excellent agreement. All of the measurements involve as few as one protein expression run and a minimum of two separate labeling and purifications for two independent sets of measurements. Some variants exceeded this standard but this was not tracked during this long study.

   Regarding the agreement of experimental observables across different protein preparations, one of the variants within the existing dataset (P2-S3) was measured on two experimental setups, two years apart, using two different expression runs each with separate protein purifications and labeling reactions. Comparison of these measurements revealed that the mean FRET efficiency values measured at Clemson were 0.70 while that measured at HHU was 0.71 w mean DDA lifetimes were 2.29 and 2.4, respectively.

   1. The authors have not explained how the approach taken in this paper compares to their previous simulated annealing approach of mapping PDZ3 using FRET data in McCann et al., 2012. That study resulted in a model in which PDZ3 binds to a completely different interface, which is not mentioned in this manuscript.

   We apologize for this oversight and thank the reviewers for this reminder. The omission was an error of trimming the manuscript for brevity and we appreciate the opportunity to highlight how much our approach has improved over the intervening time. We have included commentary on our previous modeling in the revised discussion.

   1. The biochemical disulfide (DS) mapping experiments provide a useful check of predictions of the FRET and DMD conclusions. However, in order to interpret these correctly, the authors need to show data from negative controls testing cysteine pairs that are predicted NOT to interact.

   We agree that negative controls are a critical part of the disulfide mapping experiments and thank the reviewers for this suggestion. As a negative control, we selected a cysteine pair that showed low FRET in our 2012 PNAS paper (Q374C-K591C), which was not included in this work nor was the cysteine pair involved in contact interfaces identified from simulations or modeling. This cysteine pair showed no evidence of intramolecular disulfide formation. In the manuscript, we have provide an additional supplemental figure panel to document that this negative control sample does not form disulfides.

   1. The SH3-GUK domain of PSD95 can undergo domain swap dimerization and the dimerization is promoted by binding of the synGAP PDZ-ligand to PDZ3. The authors should mention the existence of domain-swap dimerization (citing McGee [2001] and Zeng et al. [2018]) and indicate whether they tested that the FRET-labeled proteins are monodisperse. This is particularly important in light of the high variation in diffusion time for individual variants - 0.91-10.19 ms (see also #10 below). In particular, the P3-G4 FRET variant has a long diffusion time of 10.19; could it be undergoing domain swap dimerization?

   We are very interested in the prospect of domain swapping as has been suggested previously. However, we have not seen evidence for this at the concentrations used here. As reported in our 2012 PNAS paper, both full-length PSD-95 and the PSG fragment are monodisperse as judged by size exclusion chromatography, which suggests that lack of stably populated oligomeric states under these conditions at 10-5 molar concentrations. The PSG fragment runs very true to its calculated formula weight while the full-length protein does migrate faster than expected based on formula weight but not high enough to be a dimer.

   The DS mapping experiments did reveal some higher molecular weight species. However, these higher order species never accounted for more than 5% of the total input. Thus, any intramolecular interaction is transient and not well occupied under the buffer conditions and concentrations used in these studies. Our size exclusion and disulfide mapping experiments are carried out at protein concentrations that are orders of magnitude higher than used for single molecule imaging. Thus, dimerization is unlikely at the single-molecule concentrations used for the present FRET experiments. If dimerization were to occur, we would expect the appearance of additional static subpopulations in the MFD histograms. If dimerization were significant, we would also expect the appearance of an additional diffusion term in fluorescence correlation curves, which was not the case in these experiments.

   1. On page 4, line 5 the authors state: "the number and occupancy of conformational states were set as global fitting parameters". This assumes that the protein is unbiased by the labeling and that the protein behaviour is independent of the purification batch. Have the authors verified this?

   The reviewers are correct in stressing the importance of quality control in the selection of labeling sites and reproducibility in sample preparation. The PSD-95 purification has been carried out hundreds of times in the Bowen lab using different variants. The cysteine mutations used in this work have all been validated as non-disruptive to the protein and the dyes in several ways. 1) We selected labeling sites using the available x-ray and NMR structures to ensure surface accessible residues within alpha helices or short loops to minimize tertiary structural disruption; 2) we ensured that the two point mutations don’t affect the expression and purification protocols. Misfolding or changes in conformation would be visible on elution profiles from chromatography as well as proteolytic cleavage patterns, which are sensitive to protein folding; 3) in our previous work, we measured both donor anisotropy and acceptor quantum yield all of the variants in use here but one, which relied on existing sites in a new combination. We have insured that sites with poor properties are never included in our published work. Indeed, the reproducibility of sample preparation, using chromatography before and after labeling, gives confidence that the attachment of fluorescent dyes is not altering macromolecular properties. For the dyes to change the protein structure, they would have to interact competitively with the protein interfaces or disrupt local structure. These would be expected to change the dye quantum yield or the anisotropy, which were each measured in our previous work. In addition, the multiparameter fluorescence detection includes anisotropy measurements of the current samples. None of these measurements reveal aberrant fluorophore behavior (Supplemental File 3C).

   This alone does not rule out that the dyes affect the conformational ensemble. One can take additional confidence that our protein handling workflow does not affect the results from the cross-methods agreement that we demonstrate in the current work. First, between measurements of both full-length PSD-95 and its PSG truncation, using confocal and TIRF experiments boosts confidence. The labeled samples for each experiment were prepared from the same purified proteins but labeled independently with different dye pairs. The different dyes attached to the samples used for confocal and TIRF did not impact the time averaged distances between these residue pairs save for one slight outlier. Additionally, our cross-validation using disulfide mapping, which is entirely label free, provides additional confidence that the interdomain contact interfaces, observed in the data collected using the labeled proteins, are preserved when the labels are not present. Finally, independent DMD simulations of label-free PSG were in excellent agreement with regards to the predominant states identified from rigid body docking based on experimental FRET distance and the disulfide mapping.

   1. On line 6 the authors state: "Based on fitting statistics, we demonstrate that a two-state model with a small donor-only (or no FRET) population (Supplementary file 1C &D) is sufficient to fit all data.” From the average Χ2 this can be concluded, but for individual datasets sometimes a 1 state model or 3 state model seems more appropriate. The authors should explain why measuring more cys mutants justified using 'one unifying model'? How can the data contain donor-only contributions if pulsed-interleaved excitation (PIE) is used to select only molecules with active donor and acceptor fluorophores?

   We apologize for the lack of clarity as to how we arrived at the determination that two states were present in the conformational ensemble. The fitting statistics show that there is an improvement in global fitting upon increasing the number of states in the model from one-state to multiple states. The statistics in the former Supplementary file 1C show significant improvement upon fitting with two states relative to one while adding a 3rd state marginally improves 3 variants while the remaining 9 remain unchanged or show a slightly worse fit. The former Supplementary File 1D (now 3C) provides a list of the values for each of the constants that arise from fitting the 2-state model to all datasets simultaneously and the individual fit statistic for fitting this model to the specific variant dataset. This table assigns the global population fractions and their associated donor lifetimes but was not used to assign the number of states. That there are two states is based solely on the improvement in fitting statistics with two states shown in the former Supplementary File 1C. Thus, the statistics do not justify us including an additional state. Because this is such a critical point, we have moved the former Supplementary File 1C to the main text as Table 2 and add additional discussion to the manuscript to highlight how we arrived at a 2 state model.

   The reviewer is correct that a global fit of the dataset could result in suboptimal fits for an individual FRET pairs to satisfy the global minimum. In this case, most variants were best fit by a two state model. The reason for using one unifying model is our underlying assumption that the same conformational distribution for PSD-95 is sensed differently by each labeling combination. A primary conclusion from this assumption is that all variants share a population distribution. A secondary assumption is that protein handling is not biasing this conformational ensemble, which we verify as described above. Each measurement provides part of the same story so we were only interested in models which simultaneously explained all observed FRET data, and as such enforced the single global model. A global fit also proved the best way to uniquely assign each distance to its corresponding state. Furthermore, the FRET Network Robustness analysis explicitly examined how much our model depends on any one labeling variant and found no systematic deviations. This revealed an ensemble of structures that satisfy the data without enforcing a global model for all samples simultaneously.

   We also thank the reviewer for correctly observing that we misapplied the term donor-only (DO) in the manuscript. The population we referred to is more appropriately termed a “No-FRET” or “low-FRET” population. The reviewer is correct that active, FRET-labeled molecules were selected using PIE parameters. We have corrected this in the manuscript.

   1. All variants are shown to be dynamic, but they are positioned differently on the dynamic FRET line (Fig. 1D and S3). Does the same kinetic model underly each variant? If the same state occupancies are implied, then why not the same kinetic constants, especially for distances probing the same two domains?

   While the global population fraction is shared between variants the transitions rates for Individual variants are not constrained. As such the variants do not share a single equilibrium rate constant. While the FRET data is fit to two global states, our DMD simulations showed that there is substantial fuzziness within these global states. Thus, the full kinetic network is more complex than the global 2-state transition. As our screening of DMD snapshots showed, each FRET variant is uniquely sensitive to the underlying conformational transitions. Hence, the system is underdetermined and we are not able to adequately determine forward and backward kinetic rates for each variant individually.

   It is important to recall that the data shown in multiparameter FRET histograms has been binned with millisecond time resolution, which is slower than the local conformational dynamics arising from fuzzy domain rearrangements. The position of the peak will depend on the underlying rate constants. Our Photon Distribution Analysis reveals the kinetic processes that dominate the broadening of the FRET efficiency distributions. This analysis also measures the fractions of the effectively “static” population. Fast transitions, which do not significantly contribute to changes in FRET efficiency (or broadening) on the binning timescale, will appear as static populations. Thus, the simple PDA model captures the broadening that is also present in MFD histograms, but does not adequately describe dynamics at the fastest timescales.

   1. Could the data also be explained by "fuzziness" within domains, without interdomain dynamics? The authors should discuss this given the possibility of domain swap dimerization of the SH3-GuK domain.

   In this work, we use the term fuzziness to refer to alternate residue interactions and domain orientations within a global contact basin. Using this definition, we do not expect significant structural rearrangements within the PDZ, SH3 or GuK domains. These domains are well folded and have been studied individually and in combination using x-ray crystallography and NMR, which did not reveal local distortions of the domain fold (e.g. SH3-GuK interactions). This is not to say that there are not conformational dynamics within loop regions or other small scale subdomain motions. Our rubric for selection of labeling sites is to avoid large loops to minimize the local dynamics as this conformational variability compromises the resolving power of the FRET restraint. Our DMD screening provides details as to how each FRET pair senses changes in local and global conformation. In comparison to the global changes extracted from the fluorescence lifetime decays, the intradomain dynamics are occurring rapidly on small length scale and are not expected to affect our global positioning of PDZ3. We do not observe a significant population of dimers or other multimers under the concentrations used for these experiments as discussed above.

   1. Regarding supplemental File 2: The authors should justify that PDA is an appropriate method to quantify relaxation time of Fluorophores. Dynamics being so fast, how do the authors explain that when binned in 2 ms time bins, discrete subpopulations in the PDA histograms are still clearly observed (e.g. Figure 2B, Fig. 2 supp 3)? Why would the protein move through certain very discrete states and not others? Doesn't this imply that the model is oversimplifying the actual mechanism (even though the Chi^2 is alright)? It is strange that for some mutants (fig 2 supp 3B P1G3) PDA displayed discrete states, while for others (e.g. fig 2 supp 3A P2G6) PDA histograms were smooth, implying it cannot be a low-histogram-count artifact. Or can it?

   We apologize for this confusion but the photon distribution analysis was not used to “quantify relaxation times” of the fluorophores, which comes from fitting of the lifetime decays. Rather, PDA was used to estimate the rates of exchange between limiting states (i.e. the inter-fluorophore distances derived from fitting the fluorescence decays). Obtaining the rates is accomplished by fitting time-binned FRET efficiency histograms with a model that accounts for broadening due to exchange between limiting states.

   We agree with the reviewers that the two-state model, which is sufficient to fit the lifetime decays, is too simplified to fully describe the dynamic exchange between limiting states. To address this, we performed the FNR analysis to describe the limiting state basins within which fast dynamics occur. This extends the model beyond two discrete limiting states. Further, DMD screening shows that different FRET variants do report differently on the underlying conformational landscape. Some exhibit a degree of degeneracy showing similar FRET efficiency for different conformations making each variant insensitive to specific subsets of possible transitions.

   Using fluctuation correlation analysis to probe FRET-induced changes in intensities, we observed dynamics on the 10-5 second timescale, which is much too fast to give rise to broadening in the fluorescence observable histograms. However, these dynamic transitions did not correspond to exchange between states with large differences in FRET efficiency because, if such fast dynamics involved a large change in FRET, this would be associated with a narrow distribution about the mean in MFD histograms. We explain the appearance of distinct peaks for some variants as an increase in the relative contribution of fast dynamics within limiting ensembles compared to the slower processes of exchange between limiting ensembles. This can occur without a relative shift in forward/backward exchange rates and with only a slight shift in the overall relaxation rates on the timescales to which PDA is sensitive (~.01-1 ms).

   1. Regarding supp file 3A and Table S9: The spread on tdiff, (the average diffusion time through the confocal volume) for individual variants is very broad - 0.91-10.19 ms. Considering that the authors use global fits for many different parameters, it's surprising that they didn't use it for this parameter which should unbiasedly be the same for all the protein mutants, at least if all are well-behaved (i.e. non-aggregating). The high variation in tdiff may be a warning that the model is not accurately accounting for all dynamics. For example, might the P3-G4 variant be undergoing domain swap dimerization?

   We thank the reviewers for their observation and apologize for the confusion as to why there are differences in the diffusion time through the confocal volume for the different variants. We expect that there would be three distinct diffusion times because the samples were measured on two experimental setups using different confocal volumes and pinhole sizes. There are also two distinct protein constructs (full-length and PSG), which differ in molecular weight. The longest timescale processes included in the fFCS fits are attributed to long-timescale photophysical effects, such as blinking. As discussed above, we do not expect a significant population of dimers or other multimers at the pM concentrations used for these single molecule experiments.

   We agree with reviewers that the diffusion time for a given construct on a given instrumental setup should be a constant. In this light, we reanalyzed the filtered fFCS curves with enforced consistency for the diffusion times in measurements involving the same construct measured on the same setup. While this refitting slightly changed the values of fit parameters, none of these differences significantly affected the parameters used for modeling and therefore the conclusions of the paper have not been impacted. We have updated the manuscript to indicate the change in the fit models.

   1. In the results section, the authors state: "Summarizing the dynamics observed for the PDZ3-GuK variants, fFCS depicts three relaxation times." This is an overstatement because the authors imposed these three broad relaxation times. The authors should describe how they made these assignments. Is this common practice? Regarding Supplemental File 2 versus Supplemental File 3A: In principle, the relaxation time implied from fFCS and that from PDA should align. However, the 'Average' of fFCS and the T_R of PDA do not align. Is it possible that the dynamics analysis from PDA should have been constrained in some way by the results from fFCS? It would be useful to add error estimations for PDA here.

   We agree with the reviewer that it is an overstatement to say that the number of relaxation terms arises from the correlation analysis. We have removed this statement and instead focus on the differences in dynamics. The assignment of three relaxation terms was made to probe the extent of dynamics across decades in time as each time regime is typically associated with distinct forms of protein dynamics. We enforced these consistent timescales in order to directly compare amplitudes across different FRET variants. However, we do not enforce any assignment that dynamics arising from a particular type of exchange process occur at the same timescale.

   We also agree that obtaining agreement between PDA and fFCS is desirable. In our experience, such agreement is only obtainable for simple kinetic schemes when dynamics probed by fFCS and PDA all occur within the same relative timescales. Here, the contributions to dynamics occur across several decades in time including those obtainable only through fFCS analysis but too fast to be quantified by PDA. Using the methods we employed, we recover only the effective relaxation times rather than the absolute kinetic rate constants because the system is underdetermined. Differences for individual variants arise because the variants differ in sensitivity to specific transitions (Figure 8-Figure Supplement 1) while fFCS and PDA differentially report on the underlying kinetic scheme.

   1. Regarding the DS bond formation data, the authors state, "The α-basin variant showed slightly more DS formation than the beta-basin variant in full-length PSD-95 but the rates of DS formation were similar". It isn't clear what this means physically. It seems to suggest that there is static heterogeneity in the population, i.e. some proteins can and some proteins cannot form DS bonds. The presence of this effect may contradict the assumption that every state at some point interconverts to any other state, which underlies the FRET PDA analysis. The authors should discuss this possible inconsistency.

   We agree with the reviewer that this statement was not clear. It was never our intention that the DS formation kinetics be directly related to FRET data in this way. The goal of DS mapping experiments was to provide qualitative confirmation that supertertiary structures suggested by DMD and FRET experiments occur in solution. We meant to focus on the DS formation kinetics, which are in indication of structural proximity. The extent of DS formation comes from the fitting as a matter of course. The reactions progress to near completion (Figure 7-Figure Supplement 1). The differences in extent of disulfide formation, while real, are very small and we did not intend to highlight them. We have removed any discussion of the extent of DS formation in the manuscript.

   1. In the discussion of the DS experiments, the authors state, "We also observe significant kinetic differences when PSD-95 is truncated in agreement with FRET studies." This sentence is vague. The authors need to state more completely what they mean here. Exactly what is in agreement with the FRET studies?

   We agree with the reviewers that the claim was vague. We intended to communicate that the DS mapping is generally consistent with FRET experiments in that they confirm the proposed limiting conformational states. The formation of disulfides at these points confirms the accessibility and proximity of these sites with respect to one another within the supertertiary structure. Also, both DS mapping and fFCS observed changes when PSD-95 was truncated to the PSG fragment. However, the rates of DS formation are not directly comparable to the rates of conformational dynamics. We have removed this statement from the paper to avoid directly linking these two unrelated kinetic measurements.

   1. The text in the section on "Structural Modeling with Experimental FRET Restraints" is often unclear. The authors appear to have equated States A and B, formerly used only in the seTCSPC analysis to the alpha and beta basins extracted from the DMD snapshots. The authors should discuss whether there might be other conformations in the DMD results that would be consistent with the FRET-derived distances from seTCSPC? It seems possible that there could be, given that in Fig 6 sup 1, large discrepancies exist between simulated distances and FRET-measured distances for some of the FRET pairs. The authors should discuss explanations for the discrepancies that do not compromise the actual model.

   We apologize for the lack of clarity in our description of structural modeling with FRET restraints. We thank the reviewer for the suggestions as to how we can improve this discussion. In the course of this study, we do reach the conclusion that states A and B, obtained from modeling solely based on FRET data, are equivalent to conformations within the alpha and beta basins from DMD, respectively. Because the representative structures were obtained independently via distinct techniques, we felt that it would be premature to use the same terminology when we are introducing the FRET results.

   We agree that more than a single snapshot from DMD per basin appropriately satisfies the FRET restraints and that no one model satisfies all restraints equally. Our goal with the later FNR analysis, which explicitly incorporates FRET-derived restraints, was to identify ensembles of structural snapshots from DMD that are compatible with experimental data. Instead of finding the single best model for the full set of FRET-derived distances, each snapshot in the ensembles from FNR satisfies all distance thresholds independently. Thus, the ensembles from FNR do refer to both experiment and DMD.

   Further, the vertical lines shown in Figure 8 Figure Supplement 1 represent the distances from the initial global fit of all samples simultaneously. For some variants, this likely includes biases in certain distances due to the enforcement of this global model, which FNR seeks to alleviate. For SH3-GK FRET pairs, these deviations are most likely the result of the restraints placed on the motions of the GK domain in the DMD simulations.

   1. A weakness of the modeling approaches in this manuscript is that they are difficult to validate. Could the authors include a test of the modeling in which they show how small changes of the input FRET data would influence the final FRET-restrained model? Could they quantify their confidence in the final model, given all the limitations of the FRET data?

   We agree with reviewers as to the importance of validating structural models regardless of the data modality used in their determination. We respectfully disagree that this study is lacking in model validation. In this work, we generated models based on confocal FRET data and validated the FRET models using independent DMD simulations and disulfide mapping. We also employed smTIRF measurements using a different dye pair to independently validate the time-averaged FRET from confocal measurements. While this may fall short of complete validation of the associated dynamic information, we feel that this represents the state of the art in model validation regardless of the experimental approach. While it is difficult to validate novel methods for deriving structural models, we feel that have done so through cross-validation against other established techniques.

   As suggested, we did test the dependency of the experimental models on small changes in the input FRET data. To accomplish this, we used the same analysis framework described for FRET Network Robustness Analysis. Instead of removing datasets as in FNR, we introduced artificial error into the FRET distances for each variant and repeated the classification of DMD structures using the altered distances. For each trial, we introduced a random, artificial error on each of the FRET distances and repeated the classification of structures from DMD into the two basin ensembles. To check the dependence on the magnitude of the error, we used introduced a random error to each variant between 5 and -5% or between 15 and -15% of the original distance. Each condition was repeated 3 times with different random errors. To compare conditions, we measured the change in the center of mass of the surface distribution composed from the individual PDZ3 centers of mass identified by that screen (Figure 8-Figure supplement 2). We found that increasing the distance error did not significantly impact the classification of structures into the two ensembles. The variance in the mean ensemble positions over three repeats increased with increasing error along with small shifts in the mean positions. Notably, +/-15% is greater than the uncertainties in distances obtained via global fitting of fluorescence decays, suggesting that the intrinsic uncertainty in the FRET-derived distances from a single fit (Supplemental file 3D) does not significantly impact the ensemble assignment or their fuzziness.

   Read the original source
3. 

   eLife

   Jun 8, 2022

   Evaluation Summary:

   This paper is of broad interest to investigators studying the function and regulation of protein scaffolds, dynamic protein structure, and the regulation of the postsynaptic density at excitatory synapses. The authors develop an integrated approach using fluorescence-based biochemical methods, disulfide mapping, and discrete molecular dynamic simulations to study the dynamic supertertiary conformation of the synaptic scaffold protein PSD95.

   (This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. Reviewer #1, Reviewer #2 and Reviewer #3 agreed to share their name with the authors.)

   Read the original source
4. 

   eLife

   Jun 8, 2022

   Joint Public Review:

   PSD95, localizes at excitatory synapses, where it plays a scaffolding role, linking different functionally-related synaptic proteins together. It is representative of the complexity of multi-domain proteins found in signal transduction. The flexibility of its domain-domain interactions makes its dynamic supertertiary structure difficult to characterize.

   The authors have a decade-long history of applying smFRET to PSD95, and so from their portfolio of FRET-labeled mutants they specifically drew 12 double-cysteine versions to probe the supertertiary orientation of two domains of PSD95, PDZ3, and the interlinked SH3 and GuK domains, usually referred to as SH3-GuK. They worked with both full-length PSD95 and a truncated version containing only PDZ3 and the SH3-GuK domain (PSG). They and other investigators have shown …

   Joint Public Review:

   PSD95, localizes at excitatory synapses, where it plays a scaffolding role, linking different functionally-related synaptic proteins together. It is representative of the complexity of multi-domain proteins found in signal transduction. The flexibility of its domain-domain interactions makes its dynamic supertertiary structure difficult to characterize.

   The authors have a decade-long history of applying smFRET to PSD95, and so from their portfolio of FRET-labeled mutants they specifically drew 12 double-cysteine versions to probe the supertertiary orientation of two domains of PSD95, PDZ3, and the interlinked SH3 and GuK domains, usually referred to as SH3-GuK. They worked with both full-length PSD95 and a truncated version containing only PDZ3 and the SH3-GuK domain (PSG). They and other investigators have shown that the PSG supramodule remains well-folded when the first two PDZ domains are removed. Motions of the PDZ3 domain in relation to the SH3-GUK domain were studied with a combination of techniques that probe protein dynamics in complementary ways. First, interactions between the PDZ3 domain and the SH3-GUK domain were probed by Multiparameter fluorescence measurements (MFM) in which single-molecular Fluorescent Resonance Energy Transfer (FRET) data were acquired by sub-ensemble Time-Correlated Single Photon Counting (seTCSPC) and used to determine a range of "FRET-distances." The authors then used Total Internal Reflectance Fluorescence microscopy (smTIRF) to compare fluorescence measurements from truncated PSG supramodules to those from full-length PSD-95. The FRET measurements were further analyzed by Photon Distribution Analysis (PDA) to extract two limiting energy states governing interdomain interactions. To resolve fast dynamics, they performed filtered Fluorescence Correlation Spectroscopy (fFCS). To fully map and refine the conformational energy landscape of the supramodule, they performed replica exchange Discrete Molecular Dynamics (DMD). Conformations predicted from the DMD data were verified by disulfide (DS) mapping after the introduction of pairs of cysteine mutations. Finally, the authors used the FRET distance restraints to simulate the accessible volume (AV) for dye pairs at each labeling site and map the conformational dynamics within the two limiting energy states.

   For MFM, the authors used 12 variants of full length PSD-95 each of which contains a FRET pair between distinct defined sites in the PSG supramodule. They used seTCSPC to measure the dynamics of interactions between each pair. They plotted the FRET efficiency of each pair against the average donor fluorescence lifetime and found that all variants showed dynamic rather than static conformations. A simultaneous global analysis of all of the variants indicated that a two-state model is sufficient to fit all the data. Their analysis shows that the PSG module moves between two non-overlapping limiting states (A [46%] and B [54%]). Variants involving a FRET pair between PDZ3 and the rest of the module showed broad, irregular distributions indicating heterogeneity in their conformations. FRET variants spanning only the SH3-GUK domain showed less heterogeneity.

   They used smTIRF to show that the truncated PSG module results in broader FRET distributions than when it is contained within full-length PSD95. They used donor-acceptor cross-correlation amplitudes to show a uniform increase in FRET transitions in the truncated compared to full-length variants. The authors refined the seTCSPC data from truncated and full-length variants using PDA. That analysis suggests that the slowest exchanges between conformations were slowed further by the truncation. The results from PDA agreed well with measurements made with smTIRF. A global fit of seTCSP for the truncated PSG module demonstrated two states (A and B) similar to the full-length protein, but slightly reduced the occupancy of state B (48%).

   To resolve fast conformational dynamics, the authors performed fFCS. They then used global fitting to assign three principal decay times representing 1) fast local motions, 2) slower domain reorientations that alter interdomain interaction interfaces, and 3) the slowest large-scale translational transitions between the two energy basins (A and B). They concluded that the major contributions to dynamics came from the fast local motions, and from the slowest large-scale conformational transitions between states A and B.

   The authors next performed DMD simulations to map the conformational energy landscape of the PSG supramodule. They found that PDZ3 primarily adopted a docked medium conformation (α) with a mean Rg of 27.6 Å and a more compact conformation (β) with a mean Rg of 23.4 Å. Further analysis revealed a 2D free energy profile with a broad low energy basin corresponding to conformation α with PDZ3 closer to SH3 and localized near the HOOK insertion. A second shallower energy basin shows PDZ3 in the β conformation localized closer to the GuK domain. The two basins are separated by a 2.0 kCal /mol energy barrier. The authors conclude that α and β-basins correspond to states A and B deduced from FRET analyses.

   The authors performed disulfide mapping with discrete cysteine substitutions to test the predictions of close pairwise interactions between residues in the two basins. The α-basin variant pair showed slightly more disulfide formation than the β-variant pair but the rates of formation were similar. In truncated variants, the rate of disulfide formation increased by 30% for the α-basin variant but decreased six-fold for the β-basin variant. The disulfide cross-linking data is consistent with the conclusions reached from FRET analyses and DMD, including the finding that truncation of PSD-95 to remove PDZs 1 and 2 produces significant differences in the conformational kinetics of the PSG supramodule.

   The authors next modeled the structures of the PSG supramodule in the two limiting states (A and B) using FRET distances as restraints. Rigid body docking and screening against DMD simulations were performed using the FRET Positioning and Screening (FPS) software. The best fit models for state A showed PDZ3 adopting positions near the HOOK insertion, and, less often, extended without interdomain contacts. The best fit models for state B were more tightly clustered with PDZ3 positioned near the nucleotide binding pocket of the GuK domain.

   To independently corroborate these docking models, the authors calculated the accessible volume (AV) for all the structures from the DMD trajectories. The analyses shows that the interdomain interactions between PDZ3 and SH3 (the α-basin) are "fuzzier" or more varied, than the interactions between PDZ3 and the GuK domain (the β-basin).

   One protein that interacts with the PDZ3 domain in vivo is neuroligin. Because some of the conformations of PDZ3 in the α-basin would block access to the PDZ-ligand binding site of PDZ3, the authors placed fluorescein at the N-terminus of a 10 residue C-terminal neuroligin peptide containing the PDZ ligand. They used the peptide to measure binding to PDZ3 alone, and to the PSG supramodule by fluorescence anisotropy. They found that binding of the neuroligin peptide to PDZ3 alone is strongly pH dependent being reduced at pH 7; but in the PSG supramodule binding is slightly stronger at pH 7. They identified stabilizing electrostatic interactions of histidines in the neuroligin peptide with acidic residues in PDZ3 that are predicted to occur in several of the α-basin conformations. They conclude that the supertertiary context of PDZ3 enables the recognition of this critical physiological ligand.

   The overall research strategy and the transparent sharing of complicated structure/dynamics results should serve as a textbook example to the field. The addition of FRET-unbiased DMD and, later on, integration of DMD and FRET data in FRET-restrained modeling is a much-needed step to translate smFRET data to structural models and this paper paves the way to enable such investigations for other researchers. Finally, disulfide mapping is an excellent strategy to confirm predictions of the smFRET/DMD results. Nonetheless, there are some potential inconsistencies and issues that need to be clarified.

   1. The structures of the PDZ domains of PSD95 have been determined and they are well-folded and stable. In addition, the PSG module has been shown to adopt a stable structure after expression and purification. The authors should cite papers, their own and those by Zeng et al. (e.g. J. Mol. Bio, 2018), to reassure readers that the protein is not destabilized by the cysteine mutations. The authors need to state how many purifications of the mutants have been done and how many replicates have been made for the FRET measurements. Did the FRET data change over time?
   2. The authors have not explained how the approach taken in this paper compares to their previous simulated annealing approach of mapping PDZ3 using FRET data in McCann et al., 2012. That study resulted in a model in which PDZ3 binds to a completely different interface, which is not mentioned in this manuscript.
   3. The biochemical disulfide (DS) mapping experiments provide a useful check of predictions of the FRET and DMD conclusions. However, in order to interpret these correctly, the authors need to show data from negative controls testing cysteine pairs that are predicted NOT to interact.
   4. The SH3-GUK domain of PSD95 can undergo domain swap dimerization and the dimerization is promoted by binding of the synGAP PDZ-ligand to PDZ3. The authors should mention the existence of domain-swap dimerization (citing McGee [2001] and Zeng et al. [2018]) and indicate whether they tested that the FRET-labeled proteins are monodisperse. This is particularly important in light of the high variation in diffusion time for individual variants - 0.91-10.19 ms (see also #10 below). In particular, the P3-G4 FRET variant has a long diffusion time of 10.19; could it be undergoing domain swap dimerization?
   5. On page 4, line 5 the authors state: "the number and occupancy of conformational states were set as global fitting parameters". This assumes that the protein is unbiased by the labeling and that the protein behaviour is independent of the purification batch. Have the authors verified this?
   6. On line 6 the authors state: "Based on fitting statistics, we demonstrate that a two-state model with a small donor-only (or no FRET) population (Supplementary file 1C &D) is sufficient to fit all data.". From the average chi^2 this can be concluded, but for individual datasets sometimes a 1 state model or 3 state model seems more appropriate. The authors should explain why measuring more cys mutants justified using 'one unifying model'? How can the data contain donor-only contributions if pulsed-interleaved excitation (PIE) is used to select only molecules with active donor and acceptor fluorophores?
   7. All variants are shown to be dynamic, but they are positioned differently on the dynamic FRET line (Fig. 1D and S3). Does the same kinetic model underly each variant? If the same state occupancies are implied, then why not the same kinetic constants, especially for distances probing the same two domains?
   8. Could the data also be explained by "fuzziness" within domains, without interdomain dynamics? The authors should discuss this given the possibility of domain swap dimerization of the SH3-GuK domain.
   9. Regarding supplemental File 2: The authors should justify that PDA is an appropriate method to quantify relaxation time of Fluorophores. Dynamics being so fast, how do the authors explain that when binned in 2 ms time bins, discrete subpopulations in the PDA histograms are still clearly observed (e.g. Figure 2B, Fig. 2 supp 3)? Why would the protein move through certain very discrete states and not others? Doesn't this imply that the model is oversimplifying the actual mechanism (even though the Chi^2 is alright)? It is strange that for some mutants (fig 2 supp 3B P1G3) PDA displayed discrete states, while for others (e.g. fig 2 supp 3A P2G6) PDA histograms were smooth, implying it cannot be a low-histogram-count artifact. Or can it?
   10. Regarding supp file 3A and Table S9: The spread on tdiff, (the average diffusion time through the confocal volume) for individual variants is very broad - 0.91-10.19 ms. Considering that the authors use global fits for many different parameters, it's surprising that they didn't use it for this parameter which should unbiasedly be the same for all the protein mutants, at least if all are well-behaved (i.e. non-aggregating). The high variation in tdiff may be a warning that the model is not accurately accounting for all dynamics. For example, might the P3-G4 variant be undergoing domain swap dimerization?
   11. In the results section the authors state: "Summarizing the dynamics observed for the PDZ3-GuK variants, fFCS depicts three relaxation times." This is an overstatement because the authors imposed these three broad relaxation times. The authors should describe how they made these assignments. Is this common practice? Regarding Supplemental File 2 versus Supplemental File 3A: In principle, the relaxation time implied from fFCS and that from PDA should align. However, the 'Average' of fFCS and the T_R of PDA do not align. Is it possible that the dynamics analysis from PDA should have been constrained in some way by the results from fFCS? It would be useful to add error estimations for PDA here.
   12. Regarding the DS bond formation data, the authors state, "The α-basin variant showed slightly more DS formation than the beta-basin variant in full-length PSD-95 but the rates of DS formation were similar". It isn't clear what this means physically. It seems to suggest that there is static heterogeneity in the population, i.e. some proteins can and some proteins cannot form DS bonds. The presence of this effect may contradict the assumption that every state at some point interconverts to any other state, which underlies the FRET PDA analysis. The authors should discuss this possible inconsistency.
   13. In the discussion of the DS experiments, the authors state, "We also observe significant kinetic differences when PSD-95 is truncated in agreement with FRET studies." This sentence is vague. The authors need to state more completely what they mean here. Exactly what is in agreement with the FRET studies?
   14. The text in the section on "Structural Modeling with Experimental FRET Restraints" is often unclear. The authors appear to have equated States A and B, formerly used only in the seTCSPC analysis to the alpha and beta basins extracted from the DMD snapshots. The authors should discuss whether there might be other conformations in the DMD results that would be consistent with the FRET-derived distances from seTCSPC? It seems possible that there could be, given that in Fig 6 sup 1, large discrepancies exist between simulated distances and FRET-measured distances for some of the FRET pairs. The authors should discuss explanations for the discrepancies that do not compromise the actual model.
   15. A weakness of the modeling approaches in this manuscript is that they are difficult to validate. Could the authors include a test of the modeling in which they show how small changes of the input FRET data would influence the final FRET-restrained model? Could they quantify their confidence in the final model, given all the limitations of the FRET data?

   Read the original source
5. 

   Version published to 10.1101/2022.02.23.481591v1 on bioRxiv

   Feb 24, 2022