Artificial, biomimetic and hybrid enzymes: general discussion

First published on 23rd August 2024

Janko Čivić opened the discussion of the paper by Gerard Roelfes: On what methods is the tool Zymspot, which you used to predict mutation hotspots, based?

Gerard Roelfes replied: It is a proprietary software, so I don't know the details. But generally speaking, it relies on bioinformatics, evolutionary methods and protein networks.

Janko Čivić remarked: Zymspot predicted 40% of protein residues to be potential mutational hotspots. However, you included additional positions beyond those identified by the tool (https://doi.org/10.1039/d4fd00069b). Could you explain the criteria you used to select these additional positions? Additionally, do you have any insights into why these positions might have been overlooked by Zymspot?

Gerard Roelfes answered: The additional mutations were consensus mutations from the multi-sequence alignment. These are solely based on bioinformatics. The positions captured by Zymspot, in contrast, result from metrics derived from both bioinformatics and structural data, so these are selected based on distinct criteria.

Sílvia Osuna commented: According to your MD simulations, multiple conformations labeled as cis/trans are obtained. Do these differences in conformation match any available X-ray structure or NMR data?

Gerard Roelfes responded: We indeed have some X-ray structures of the enzymes now. While they need to be further analyzed, some of them indeed do show a forward orientation of R92, and some of them show a back orientation. This seems to be in agreement with the MD simulations.

Sílvia Osuna asked: You mentioned that the mutations have an effect on the network of hydrogen bond interactions; are they only affecting H-bonding or do they affect other weak non-covalent interactions? How are you computing such interactions?

Gerard Roelfes answered: We computed all the interactions (polar, non-polar, hydrophobic, hydrogen bonds) using EMDA,¹ based on spatial distances between residues (i.e., residues closer than 3 Å are considered in contact). Subsequently, we focused our analyses specifically on hydrogen bonds by means of the HydrogenBondAnalysis class within the MDAnalysis package.

1 https://github.com/MolBioMedUAB/EMDA.

Guto G. Rhys remarked: Some of the distal mutations that have a deleterious effect may lead to a more optimal backbone, where subsequent sequence optimisation of the active site might lead to even more active enzymes. Have you considered pursuing this with your examples?

Gerard Roelfes replied: That is an interesting suggestion. You are right, but we have not done that. We also do not claim that the mutants we introduce here are per se the best enzymes. It is very possible that better variants can be found if you start re-engineering the enzyme based on mutations in other positions. Here, we just show that you can do that and that shifting the conformational distribution is an important contributor to improved activity.

Joelle N. Pelletier asked: Is the main effect you have seen tightening of the core interactions?

Gerard Roelfes responded: Yes, together with a change in the orientation of residue 92.

Joelle Pelletier remarked: It may be a question of terminology, concerning how we qualify distant mutations. In your paper (https://doi.org/10.1039/d4fd00069b), the key alpha-carbon distances are 10–12 Å. Have you considered the side chains and their orientation/distance, particularly if they are pointing towards the active site? If the side chains of the relevant residues are not pointing toward the active site, it would be helpful to specify the minimum distance of the side chains. This would add weight to the assertion of the mutations being distant.

Gerard Roelfes responded: This is a good point. We based our definition of distant mutations on the available crystal structure. There, the side chains point opposite to the catalytic non-canonical amino acid, bringing their distances to 12–16 Å.

Meilan Huang remarked: It is nice to see that two remote mutations showed improved activity and stability. Sometimes higher activities can compromise thermostability. What would be a good strategy in the selection of the promising variants from the library?

Gerard Roelfes replied: Here we actually saw that increased activity and thermostability go hand in hand. We were not necessarily looking for this, but it is of course a nice additional feature. Thermostability was not a factor in the selection, so if the thermostability had decreased, we would have found active variants using the same screening.

Meilan Huang said: The three best mutations were identified and combined. Did you try to combine other positions to improve the activity further (e.g. combining a couple of remote sites with modest activities, or combining a remote site with the pore region)?

Gerard Roelfes replied: We have not done this yet. It definitely may lead to further improvement. Here we showed that, in principle, it is possible to improve artificial enzymes by exploiting distal mutations. It would indeed be interesting to take this further and include other positions. But, in view of the effort, I would prefer to do this on an enzyme that catalyzes an important synthetic transformation instead of on this model reaction.

Adrian Mulholland commented: You showed interesting activity improvement that seems to be due to selection of specific, active conformations. You also show better coupling in the F54L mutant. Do you see a change in the active-site structure or accessibility, or in the side chains of the region around residue 54?

Gerard Roelfes replied: We did not see any obvious structural change in the region surrounding residue 54.

Adrian Mulholland remarked: Allosteric effects and sites of distal mutation can also be identified by other computational methods, e.g. shortest path maps (as presented by Prof. Osuna [https://doi.org/10.1039/d3fd00156c]) and dynamical nonequilibrium molecular dynamics (D-NEMD) simulations (e.g. ref. 1). Can you comment on the timescales required for the techniques that you have used: how long do these take? How do they compare to other methods?

1 I. Galdadas, S. Qu, A. S. F. Oliveira, E. Olehnovics, A. R. Mack, M. F. Mojica, P. K. Agarwal, C. L. Tooke, F. L. Gervasio, J. Spencer, R. A. Bonomo, A. J Mulholland and S. Haider, Allosteric communication in class A β-lactamases occurs via cooperative coupling of loop dynamics, eLife, 2021, 10, e66567, DOI: 10.7554/eLife.66567.

Gerard Roelfes replied: All the mentioned methods are useful for the purpose of identifying distal hotspots having an effect on dynamics and catalysis. The potential of Zymspot is that it does not use time-consuming simulations such as molecular dynamics, so it is very effective for evaluating the entire sequence landscape. The computational pipeline runs in less than an hour and it might take the user around two days in total to prepare and evaluate the results.

Guangcai Xu said: The active site of LmrR is two-fold symmetrical. When one non-canonical amino acid (AA) is introduced on one side, another is introduced on the other symmetrical side. Evolution sometimes tends to disrupt/break the symmetry. Have you considered fusing the two dimers of LmrR to create more diversity in the active site during protein engineering?

Gerard Roelfes replied: Yes, you are absolutely right. Because of the homodimeric nature of the protein, every mutation you do twice. So far, that was never a limitation. But indeed it could be desirable to be able to mutate only one position in the pore. So, yes, we have been exploring single-chain variants of LmrR, but so far we have not (yet) succeeded. But we also have a collection of different protein scaffolds now where the pore is not part of the dimer interface, so you can engineer more “precisely”.

Alexander McKenzie asked: Does Zymspot identify a higher number of conformational hotspots in this artificial enzyme compared to native (and presumably more stable) enzymes?

Gerard Roelfes replied: Zymspot could identify a significantly higher number of hotspots for LmrR than for natural enzymes, in our experience. The reason behind this observation might be the intrinsic nature of LmrR, which is known to be highly flexible and has many allosteric inter-residue communication events. Since Zymspot captures hotspots strongly involved in the conformational dynamics of the enzyme, it is reasonable that a high number was found for LmrR.

Lubomír Rulíšek asked: In studying the activity of a series of enzyme mutants for the particular substrate, would it be difficult to also explore the substrate scope, similarly to how it is done in typical organic chemistry work? Would it bring additional important information? To me, this seems to be less effort than preparing various variants of the enzyme (which are prepared anyway). That is, once the mutants are prepared, testing various modified substrates in a particular assay seems to be beneficial. And it would add another dimension to the screening of the enzyme variants.

Gerard Roelfes answered: It would not be difficult, but it would involve a lot of additional work. And yes, it will deliver new information, but very different information. This study was really about using distal mutations to optimize the activity of an artificial enzyme for a given reaction and, thus, focuses on what happens structurally in the protein.

Donald Hilvert queried: Does the computational program you used only predict the sites to target, or specific mutations at those sites?

Gerard Roelfes answered: It predicts the sites that are important for dynamics. It does not suggest mutations. We selected the mutations based on multisequence alignments, choosing the residues most commonly found at these positions.

Donald Hilvert asked: Does the program rationalize why the targeted sites should be successful?

Gerard Roelfes answered: No. The program identifies positions that should be important for protein dynamics, not for catalytic activity. The hypothesis is that those positions are important to modulate conformational dynamics and in this way improve catalysis by favoring productive conformations or disfavoring unproductive ones.

Sean Adeoti Thompson remarked: If I'm not mistaken, you mentioned other programs that simulated fewer hits for mutation sites. Have you compared to see if these hits are similar for those programs and this one? Or are there different hits in the other programs? Do the other programs show the same sites that produced the higher k_cat values?

Gerard Roelfes replied: This could be useful, but we have not done this. For us, Zymspot worked to identify distal positions that are important for dynamics and from there we started the protein engineering.

Dominic J. Campopiano asked: Can you analyse the iminium intermediates? This would be to see if they go down the wrong pathway and get stuck. Any mechanistic insights?

Gerard Roelfes answered: What we report here is about the enzyme without substrates. This was the first step. The next step is to look at what happens with a substrate bound. You are right, this would give us valuable information. It is definitely on our to-do list.

Dominic J. Campopiano queried: Have you determined the crystal structure with the product bound?

Gerard Roelfes answered: Unfortunately, we don't have a crystal structure with product bound yet. Of course, this could give us valuable information. We are working on it.

E. Neil G. Marsh asked: Have you looked at the temperature dependence of the mutants that display higher activities? It is quite possible that some mutants show a steeper temperature dependence than others so that the most active mutants at one temperature may not be the most active at a different temperature. I think this could be very informative. I note that this is a separate matter from the thermal stability imparted by the mutations – in principle, you can deconvolute stability and activity. These two properties may not run in parallel.

Gerard Roelfes answered: Thank you for the suggestion. To date, we have not looked further at the temperature dependence, but we agree that it might be very informative.

Nicholas Turner opened the discussion of the paper by Ivana Drienovská: How do you select your scaffolds? They appear to be selected on the basis of binding one of the substrates, but they are quite different proteins, with different architectures. Please comment on the thinking. Is it only based on the fact that they bind one component of the reaction?

Ivana Drienovská replied: The choice of scaffolds is indeed a very important question in the field of artificial enzymes and there are discussions in deciding on the best course of action. This is why, in the presented study (https://doi.org/10.1039/d4fd00057a), we also tried to explore different approaches to see the variety of resulting outcomes. As you mentioned, within the natural enzymes we have chosen, we attempted to take a strategic approach toward scaffolds that are known to bind one of the components of our reaction – cinnamaldehyde. This contrasts with strategies that utilize large active sites. While exploring these, we aimed to also study if these active sites could accommodate a slightly larger chain of our amino acids (DProK) and how they would react to further evolution. We are still collecting more data to address some of these questions comprehensively.

Nicholas Turner asked: So you introduce the noncanonical amino acid and get evidence of reactivity; do you then think about changing the reaction to something that would benefit from similar catalytic machinery?

Ivana Drienovská responded: Absolutely! We chose secondary-amine-based amino acids as catalytic residues because, as small molecule catalysts, they have demonstrated their efficiency in a wide range of transformations, often proceeding through iminium ions and enamines as key intermediates. Therefore, we believe that artificial enzymes containing these amino acids could also be effectively used for a broad range of catalytic reactions. We have tested this in several C–C bond-forming reactions, and we observed catalysis, albeit usually with low reaction rates. Currently, we are working on further improving these initial outcomes and expanding the panel of different reactions tested.

Nicholas Turner queried: Do you find that the activity translates to another reaction? Or is it too early to say?

Ivana Drienovská replied: As mentioned in my previous reply, we have tested the reactivity of some of our artificial enzymes with secondary-amine-based ncAAs in a couple of reactions proceeding through enamine/iminium anion intermediates. However, we still need to explore this in more detail to observe if there are some general trends translated within.

Amanda G. Jarvis said: You used both substrate-templated and pocket-guided approaches to help choose potential scaffolds for your designs, and found in this case that the substrate template approach worked best. Do you think this will always hold true or do you think that for some applications pocket-guided methods will be useful?

Ivana Drienovská responded: I think it depends, and I cannot generalize that the substrate-guided approach is always better than the pocket-guided one. It really depends on the amino acids, the reaction itself, and the particular enzyme/protein choice. As discussed already, the dynamics/rigidity of the protein can also play significant roles in the catalysis outcome. In our experience, small active sites can offer a large number of interactions, which could potentially lead to a better acceleration of the reaction or induced selectivity, which sometimes may be a struggle in larger pockets. However, we have also worked with enzymes chosen via a pocket search that, upon the incorporation of our amino acid, could not accommodate all the components of the reaction. I believe that advancements in computational prediction will help us choose optimal scaffolds for both strategies. I believe that both approaches can yield promising candidates depending on the specific requirements of a reaction.

Artur Góra said: You have mentioned that you want to start research from a nice starting point. Surprisingly you chose to pick a scaffold selection strategy rather than an amino acid selection strategy. Could you comment on the reason for choosing this strategy? Is it more efficient at providing a better starting point?

Ivana Drienovská responded: When designing artificial enzymes, both parameters – non-canonical amino acids and the protein accommodating them – play a crucial role, often reflecting the specific reaction we aim to catalyze. In terms of choosing one strategy over another, we actually started by preparing and exploring different secondary-amine-based non-canonical amino acids first, as described in our recent study.¹ From this exploration, we select the best variant to test within different protein scaffolds in the study presented at this Faraday Discussion (https://doi.org/10.1039/d4fd00057a). In that initial study, however, we used the protein scaffold LmrR, which is already well known for its compatibility with different non-canonical amino acids.² This choice provided a reliable starting point to focus on evaluating different non-canonical amino acids, knowing that the scaffold could effectively accommodate them. This approach allowed us to first identify an amino acid that has an overall good rate of incorporation and performed best in the selected catalysis (in this case, a Michael addition reaction) and then explore this within various protein environments.

Thus, it is challenging to definitively state which strategy should come first, as both are interdependent. However, starting with a well-characterized scaffold can streamline the process, providing a stable and efficient platform to evaluate different amino acids. Therefore, I think revisiting and balancing both strategies is crucial to determine the most optimal starting point.

1 A. Gran-Scheuch, E. Bonandi and I. Drienovská, Expanding the Genetic Code: Incorporation of Functional Secondary Amines via Stop Codon Suppression, ChemCatChem, 2024, 16, e202301004, DOI: 10.1002/cctc.202301004.

2 G. Roelfes, LmrR: A Privileged Scaffold for Artificial Metalloenzymes, Acc. Chem. Res., 2019, 52, 545–556, DOI: 10.1021/acs.accounts.9b00004.

Anthony P. Green asked: Could the efficiency of your imidazolidinone-containing biocatalysts be improved by removing the bulky gem-dimethyl substituents adjacent to the catalytic amine?

Ivana Drienovská responded: In general, removing the dimethyl group could make the imidazolidinone-based ncAA more reactive. However, for that particular ncAA (reported in our previous work¹), we tried to mimic the standard McMillan catalyst as closely as possible. Indeed, the dimethyl groups primarily influence selectivity in this type of organocatalyst by aiding proper substrate orientation. However, they have also been said to influence transition-state stabilization and help avoid side reactions, therefore leading to better overall efficiency. It would indeed be interesting to compare the two ncAAs within protein scaffolds. However, we have experienced some challenges in the synthesis of these molecules, which has not yet allowed us to achieve the non-methyl-substituted variant.

1 A. Gran-Scheuch, E. Bonandi and I. Drienovská, Expanding the Genetic Code: Incorporation of Functional Secondary Amines via Stop Codon Suppression, ChemCatChem, 2024, 16, e202301004, DOI: 10.1002/cctc.202301004.

Louis Y. P. Luk asked: Regarding the choice of an unnatural amino acid (AA), would a shorter AA make it easier to design/model the enzyme?

Ivana Drienovská responded: The length of the linker within non-canonical amino acids (between the alpha carbon and the functional part of the molecule) can significantly influence the design and catalytic outcome of artificial enzymes. There are several benefits to shorter linkers, especially in design and modeling, where shorter linkers can reduce the complexity of interactions and the number of possible conformations. Shorter side chains are also likely to minimize steric hindrance within the enzyme's active site. However, in some cases, longer linkers may be necessary to accommodate the complexity of a functional group. Furthermore, the choice of linker length can also be influenced by the choice of orthogonal pairs (as the possibility of incorporating desired ncAAs is also crucial for the design process).

For example, in our experience, orthogonal pairs derived from the pyrrolysine system from different origins were more successful in incorporating non-canonical amino acids with longer linkers, such as those structurally close to pyrrolysine. When working with a diverse panel of ncAAs featuring secondary-amine functional groups, we found that those with longer linkers were successfully incorporated into proteins. Conversely, we struggled to identify a system that could efficiently incorporate ncAAs with very short linkers. This suggests that while shorter linkers may simplify design and reduce steric hindrance, longer linkers can be essential for the successful incorporation of more complex functional groups.

Amanda G. Jarvis remarked: Picking up a bit on what we heard in the last talk from Anthony (https://doi.org/10.1039/d4fd00019f), what are the translation efficiencies of the non-canonical amino acids you used? Did you maintain good protein yields, or have you encountered problems?

Ivana Drienovská responded: The translation efficiency and obtaining a good protein yield have been a bit of challenge in this project. This is due to the fact that the orthogonal pair we have worked with (pyrrolysine pair from Methanosarcina barkeri) is a wild type without further engineering toward improved efficiency for incorporation of DProK. As a result, we had to use high concentrations of amino acids (5–10 mM) to achieve a good protein yield.

We are currently optimizing this process to significantly reduce the required concentration, ideally to at least the level of other standard ncAAs (0.5 to 1 mM). This is one of the long-standing issues in the field, so it is great to see papers like Anthony Green's addressing it efficiently.

Overall, the product yield is also influenced by the protein of choice itself. In our experience, some proteins and certain positions within the protein work well, while others work poorly. Generally, positions earlier in the sequence lead to higher overall production, while later ones tend to result in poorer outcomes and more frequent recognition of stop codons. The LmrR and natural enzymes presented in this study (https://doi.org/10.1039/d4fd00057a) usually had overall good expression and protein yields of 2–30 mg upon incorporation of DproK. The scaffolds from the genome screening generally did not have good expression as wild types, which subsequently led to very poor yields when incorporating the ncAA of choice as well.

Andrew Gomm commented: Thank you for your talk. I was wondering what range of nucleophiles can be used in this reaction. Are you limited to small, hard, carbon-based nucleophiles? Do you know what ranges of pK_a can be deprotonated by the enzymatic residues?

Ivana Drienovská answered: In the field of organocatalysis with small molecules, a wide range of nucleophiles have been explored to prepare enolates for this reaction, including nitroalkanes, malonic esters, β-keto esters, and β-keto nitriles. However, we have only explored nitromethane so far in the Michael addition reaction with our artificial enzymes containing secondary-amine-based ncAAs. Notably, previous mechanistic studies suggest that the nitromethane is not deprotonated by other amino acid residues in the active site, but by the hydroxyl group in the intermediate formed between cinnamaldehyde and the secondary-amine ncAA. This definitely raises an interesting question for further exploration regarding both the substrate scope and the engineering of our enzyme, as the active site indeed should have the capability to deprotonate a wider range of pK_a values, as demonstrated in general enzymatic mechanisms.

Cathleen Zeymer opened the discussion of the paper by Anthony P. Green: In your opinion, what is currently the best strategy to evolve a completely new synthetase? Is it still positive and negative selection rounds, basically what has been done for more than 20 years now? Or would you rather go for modern methods, like continuous evolution strategies?

Anthony P. Green answered: Personally, the tried and tested positive/negative selection methods would still be my go-to method for engineering aminoacyl tRNA synthetases (aaRSs) towards new ncAAs. However, these modern methods (e.g. continuous evolution, tRNA display) do offer some interesting advantages and so we may see these becoming more widely used by the community in the coming years.

Friedrich Johannes Ehinger remarked: You successfully transplanted previously discovered beneficial mutations from the MaPylRS to its G1PylRS orthologue and by that lowered the required MeHis concentration for incorporation into proteins (https://doi.org/10.1039/d4fd00019f). Do you have a structural rationale for this improvement (MaPylRS vs. G1PylRS)? Which residues might contribute to this and how?

Anthony P. Green responded: The differences are likely very subtle, and further study is needed to tease out any structural differences that may lead to the observed activity increases.

Amanda G. Jarvis commented: You mentioned that looking back at what caused the improvements it was difficult to see if you could predict what mutations would make a difference. Can this approach be used for other amino acids and their equivalent aminoacyl-tRNA synthetases?

Anthony P. Green answered: Yes, if you are using an M. mazei or M. barkeri PylRS variant to encode an ncAA, I would suggest exploring a range of other PylRS homologs as it is possible that one of these could be more efficient. This could be particularly useful for expensive ncAAs or ncAAs that are not commercially available and require a multi-step chemical synthesis.

Amanda G. Jarvis asked: Following on, could this approach be used for the tyrosyl-tRNA synthetases?

Anthony P. Green replied: This approach would be more challenging for tyrosyl systems, as they first need to be engineered to remove their native activity towards tyrosine. Compared with PylRSs, Tyr-aaRS are also more tightly regulated, e.g. with in-built proof-reading mechanisms and activity that is dependent on the tRNA anticodon, which makes them more challenging to engineer for ncAA incorporation.

Carlos Acevedo-Rocha queried: How much engineering has been done on the tRNA binding site and have you looked at co-evolution patterns at other amino acid sites?

Anthony P. Green responded: We haven't targeted the tRNA binding site or tRNA during engineering, but some other groups have.¹ To engineer orthogonal translation components with efficiencies akin to natural systems it will likely be important to interrogate regions outside of the aaRS active site during evolution.

1. M. Amiram, A. Haimovich, C. Fan, et al., Evolution of translation machinery in recoded bacteria enables multi-site incorporation of nonstandard amino acids, Nat. Biotechnol., 2015, 33, 1272–1279, DOI: 10.1038/nbt.3372.

Louis Y. P. Luk asked: Did you ever find methylated histidine accidently bound to an electrophile when purified?

Anthony P. Green responded: No, we have deliberately functionalized Me-His nucleophiles to solve the structures of inhibited enzymes or catalytic intermediates, but have not observed unintended Me-His modifications upon protein purification.

Ivana Drienovská asked: In your opinion, what concentration of effective incorporation of a non-canonical amino acid must be achieved to ensure it works well with its direct biosynthesis?

Anthony P. Green replied: This is a complex question and is dependent on many factors. As a general principle, engineering synthetases that work efficiently at low ncAA concentrations is important to avoid the need to biosynthesize high ncAA concentrations, which might be technically challenging or detrimental to cell growth.

Gerard Roelfes said: Having worked with N-methyl histidine (NMH) and being very aware of the high cost of the amino acid, I appreciate the high efficiency of your new/evolved orthogonal translation system (OTS) system. Transplanting the mutations to the homologue on the one hand is very subtle, but can also give rise to differences. In your new OTS, you report misincorporation of phenyl analine. You never observed misincorporation of lysine?

Anthony P. Green responded: No, we have only observed Phe misincorporation.

Amanda G. Jarvis opened a general discussion of the papers by Gerard Roelfes, Ivana Drienovská and Anthony P. Green: I'd like to ask the panel, where do you see the future for using non-canonical amino acids?

Ivana Drienovská answered: Overall, I believe that there is a lot of potential for non-canonical amino acids (ncAAs) in biocatalysis, and what we have seen so far has still just scratched the surface. In general, ncAAs give rapid access to features that are not found in nature, or, if they are, make access to them simpler. Recent advancements in this field, particularly in the last five years, have shown us that ncAAs not only bring the opportunity to translate synthetic chemistry features into a biological scaffold, but also pave the way to discover new types of transformations. Within the field of artificial enzymes, whether through conjugation or by making use of inherent organocatalytic, metal-binding, or photosensitizing characteristics, the incorporation of ncAAs has allowed the creation of artificial enzymes with catalytic functionalities not available in nature, expanding the biocatalytic repertoire of enzymes. Our goal is to continue in this direction and bring in even more functionalities to expand the repertoire further, even allowing for transformations that are not yet conceivable by either bio- or chemo-catalysis. Notably, this still has to be supported by further advancements in the genetic incorporation technology as well.

Gerard Roelfes added: I think there is a great future for using non-canonical amino acids in enzymes, if it can make possible chemistry that is not achievable in any other way. This is how organic chemists in the end started accepting biocatalysis, because it allowed for transformations that cannot be done (well) in the conventional way. The same will hold for these enzymes containing non-canonical amino acids.

Amanda G. Jarvis remarked: It's been seen with quite a few artificial enzymes that these systems don't have the turnover numbers or rates of natural enzymes. Do you think we will see designed enzymes reach the same level as natural enzymes?

Gerard Roelfes answered: I agree that these artificial enzymes do not yet have the same activity as natural enzymes. But as a field we are making steps and we are getting closer. I think that now also including distal mutations will help to bridge this gap.

Anthony P. Green replied: Yes, some very efficient ‘artificial’ enzymes have been developed already by the community, and with sufficient engineering and/or more effective design methods, it should be possible to develop designed enzymes with efficiencies akin to natural enzymes.

Ivana Drienovská responded: Yes, I definitely believe that in the future, we will see more artificial enzymes reaching the turnover rates of natural enzymes. We already have several examples where this has been achieved, and I think many more will follow. Of course, the choice of the type of artificial enzyme will play a role in this direction. Those that can be directly utilized for directed evolution studies in cells or cell-free extracts (such as those with non-canonical amino acids) would be preferred. These types are more likely to achieve such improvements from low rates of activity compared to more complex artificial enzymes, which require additional purification and modification steps within their preparation.

Nicholas Turner remarked : I’m excited by the different approaches, but it is not obvious what the best strategy is. Should you start with design and then tune? Or how do you choose the initial scaffold? Do you want something highly adaptable? Some of the interesting aspects get evolved out. Is the evolvability of the scaffold the most important property?

Gerard Roelfes responded: The choice of the scaffold is one of the most important choices. There are a few “privileged” scaffolds, including LmrR, that seem to be very good for enzyme design. Personally, I do think that structural flexibility is one of the most important properties for a good scaffold.

Anthony P. Green replied: This is a complex question as there are numerous factors that need to be considered when selecting starting scaffolds for enzyme design/engineering. Scaffold evolvability is certainly an important consideration – where possible I would prioritise scaffolds that have previously been shown to be highly adaptable and tolerant of large numbers of mutations.

Stefan Lutz asked: On the question of evolvability, does anybody know of an example of an enzyme or protein that is not evolvable? Isn't it more about setting clear and realistic objectives and intermediate targets to walk any enzyme towards engineering into a successful biocatalyst?

Gerard Roelfes answered: Every enzyme/protein can be mutated. But not every enzyme can be necessarily improved in this manner. Or the level of improvement is limited. I want to reiterate that the choice of scaffold in the beginning of the project is very important and that dynamics and evolvability play a key role in the success. We see the success stories in the literature. But I also know of many artificial enzyme projects that were stopped after a while because the “wrong” protein was chosen in the beginning, so improving them proved very difficult, or not possible at all.

Ivana Drienovská replied: I agree that, in general, evolution theory suggests that all proteins are evolvable when subjected to evolutionary pressure, although this may be to various degrees. However, this may indeed be a bit more of a complex question if we would like to consider every single example, as it certainly is also influenced by some parameters such as the evolutionary direction.

Anthony P. Green added: In principle I agree that all proteins are evolvable to some degree – however, they are not all equally evolvable. In some cases, very large numbers of mutations may be required to achieve the desired fitness gains, and these sequences are not always experimentally accessible within given screening constraints.

Stefan Lutz queried: When we are searching for suitable starting proteins for evolution of a tailored biocatalyst, is it fair to assume that there are privileged proteins given their ability to match or closely mimic at least some of the desired traits? What trait would you prioritize in your choice of starting protein?

Anthony P. Green responded: This is a complex question as there are many factors that contribute to starting-scaffold selection, and it very much depends on the project/target activity in question which of these factors becomes the highest priority. As a general principle for enzyme design with ncAAs, I would select scaffolds that are readily produced in E. coli with high levels of soluble expression.

Donald Hilvert asked: A related question is what is the optimal scaffold for a particular reaction? Some scaffolds may lend themselves to evolutionary optimization, but we still don't know the rules. Regarding rigidity, we recently published a paper on a collaborative project with the Baker lab to equip a computationally designed and extremely stable beta-barrel with aldolase activity.¹ An array of functional groups was embedded in the scaffold to give comparatively high initial catalytic activity, but subsequent evolutionary optimization proved difficult, likely because the scaffold was too rigid.

1 Y. Kipnis, A. O. Chaib, A. A. Vorobieva, G. Cai, G. Reggiano, B. Basanta, E. Kumar, P. R. E. Mittl, D. Hilvert and D. Baker, Design and optimization of enzymatic activity in a de novo β-barrel scaffold, Protein Sci., 2022, 31, e4405, DOI: 10.1002/pro.4405.

Gerard Roelfes answered: The choice of scaffold is one of the key design choices. Looking at successful examples of artificial enzymes, you see that there are a few “privileged” protein scaffolds that lend themselves well for the creation of artificial enzymes that can be evolved. I think structural dynamics is a very important characteristic for a successful protein scaffold that can also be evolved and I think our results with LmrR support this.

Ivana Drienovská added: As discussed in some of my previous questions and by other panel members, the choice of scaffold remains a challenge. One lesson we can learn from biocatalysis is that while the stiffness or rigidity of enzymes is beneficial in some cases, dynamic enzymes that undergo conformational changes during catalysis are needed in others, depending on the type of reaction. Therefore, I believe that we must continue to explore a diverse range of scaffolds and study the effects of engineering on their catalytic improvements. This exploration will allow us to better understand their importance and the patterns needed to enhance specific reactivities, which could hopefully in future allow for their better design.

Guto G. Rhys queried: Are there some small molecules that could only ever be accessed by biocatalysis using non-canonical amino acids or abiological cofactors?

Gerard Roelfes responded: I am convinced that there are. I think it is important that the field moves away now from proof-of-principle studies and starts addressing such challenges. This will be key to this approach being applied in the future.

Anthony P. Green answered: I can say that the enzyme-design community has developed biocatalysts with activities that are currently unknown in nature and are not easily replicated by existing synthetic catalysts. It difficult to know what discoveries and technology developments will come about in the future, and so I would be reluctant to speculate much further than this.

Ivana Drienovská added: Overall, I believe there are still numerous functionalities that are highly desirable for the preparation of small molecules that are currently missing from the repertoire of enzymes. These functionalities could potentially be provided by non-canonical amino acids or artificial cofactors. However, it is indeed debatable whether these small molecules could never be accessed via traditional biocatalysis, as this field also continues to advance and discover more reactivities or successfully engineer novel ones.

Adrian Mulholland remarked: You have shown nice examples of effective enzyme design and development. We have seen bioinspired design. There are aspects of natural enzymes that typically are not used in the design and development of new biocatalysts, however. Many natural enzymes are dimers. Enzymes are larger (e.g. across a series of TIM barrel enzymes) when the chemistry that catalyses is more difficult.¹ What are we missing and how do we get there?

1 V. L. Arcus, E. J. Prentice, J. K. Hobbs, A. J. Mulholland, M. W. Van der Kamp, C. R. Pudney, E. J. Parker and Louis A. Schipper, On the Temperature Dependence of Enzyme-Catalyzed Rates, Biochemistry, 2016, 55, 1681–1688, DOI: 10.1021/acs.biochem.5b01094.

Gerard Roelfes replied: I would like to add to this that this field is still relatively young and we are just starting. We are starting to move beyond “simple” proof-of-principle studies and starting to take some of these advanced properties into account. For example, we are now starting to appreciate the role and importance of structural dynamics and conformational equilibria for designer enzymes. As we go on, many of these additional features that you mention will become part of the design process as well, I am sure.

Anthony P. Green responded: I think it is important to remember that natural enzymes have complex evolutionary histories that are not necessarily evolved for optimal catalytic parameters – for example, enzymes often have in-built regulation features that are important in a cellular context, but are a complication that does not need to be considered for typical biocatalysis/enzyme design applications. So I believe there is still an open question of how much of the complexity observed in natural scaffolds we will need to replicate in designer enzymes to achieve ‘enzyme-like efficiencies’. Having said that, it is clear that many features of enzymes that are thought to contribute to their catalytic activities (e.g. dynamics, electric fields, etc.) are not considered during computational enzyme design. If we want to design more active enzymes in the future, we will likely need to build in some of these additional features. One question is whether we really have an accurate understanding of how these factors contribute to catalytic efficiency and, if so, do we have sufficiently accurate design methodologies to allow us to embed these features as intended.

Ivana Drienovská replied: This is indeed a fascinating aspect of enzyme catalysis to consider. The complexity of enzymes plays a crucial role in various aspects, such as the necessity in more complex reactions, which may involve loops covering active sites or additional areas for binding desired cofactors. I believe as our understanding of these enzyme characteristics improves, our computational simulations and design strategies will likely continue to improve as well. This will hopefully enable us to utilize these complexities in the design of active sites of artificial enzymes in the future.

Peter Stockinger opened the discussion of the paper by Sílvia Osuna: Which simulation system would you recommend for analysis with SPM: one with a substrate and cofactor, protein only, or both?

Sílvia Osuna answered: You can compute SPM using MD simulations in the apo state, substrate-bound, cofactor-bound, or substrate and cofactor-bound; there is no restriction. We have actually computed SPM in all of these cases, and the generated paths will differ but provide complementary information.

Peter Stockinger said: A follow-up question: when identifying hotspots for activity engineering, is it beneficial to simulate both open and closed states? One strategy could involve mutating consensus hotspots occurring in both types of simulation systems, while another might focus on positions that differentiate between the two conformational ensembles. Have you had any experience exploring these strategies?

Sílvia Osuna responded: This is something we evaluated in this Faraday Discussions contribution (https://doi.org/10.1039/d3fd00156c). We checked how the generated SPMs differ when considering only open or closed states of the COMM domain of TrpB, or considering the whole MD trajectory. The SPMs obviously differ and it was interesting to find that the evolved 0B2-PfTrpB variant, which according to our MD simulations is able to properly close the COMM domain, contained a higher number of positions identified in the closed SPM. This was not observed in wt-PfTrpB. It should also be noted that in terms of capturing directed evolution mutations, the SPM considering the whole trajectory identifies a higher number of positions.

Donald Hilvert remarked: Rama Ranganathan has used a statistical coupling analysis to identify correlated sets of amino acids in proteins that he calls “sectors”, which are physically connected in the tertiary structure, each with a distinct functional role linked to activity, allostery, and the like.¹ Do you know whether the shortest path maps (SPMs) that you calculate are related to or correlate with Ranganathan's protein sectors?

1 N. Halabi, O. Rivoire, S. Leibler and R. Ranganathan, Protein Sectors: Evolutionary Units of Three-Dimensional Structure, Cell, 2009, 138, 774–786, DOI: 10.1016/j.cell.2009.07.038.

Sílvia Osuna answered: This is something we have not done, and I thank you for the suggestion. I completely agree with you that comparing the SPM positions with Ranganathan “sectors” could provide useful information.

Donald Hilvert said: It might be interesting to apply your SPM method to some of the proteins he has studied (DHFR, serine proteases, SH2 and SH3 and PDZ domains, etc.) to see whether your pathways overlap with his sectors.

Sílvia Osuna replied: Yes, indeed, we have run many MD simulations and SPM on serine proteases, so this could be a good starting point.

Artur Góra asked: I tried your tools and saw some changes caused by a single mutation (which was very nice, since we could not evaluate differences by other methods). However, I am wondering, how sensitive is your method to the environment of the MD simulations? How big an impact on the results do the different force fields or water models that you select for your MD have? What would be your recommendation?

Sílvia Osuna responded: In this Faraday Discussions paper (https://doi.org/10.1039/d3fd00156c), we applied two different combinations of force field and water model to evaluate the changes in the open/closed structures of TrpB. It is true that we only included the SPM resulting from the ff19SB + OPC combination as it provided a better description of the open state of TrpB. The SPM using ff14SB + TIP3P combination shows some differences, but still many regions are conserved between both. My recommendation is to run SPM on the combination of force field and water model that better describes the system under study; in the case of TrpB, we found it to be ff19SB + OPC.

Artur Góra addressed Sílvia Osuna and Shina Caroline Lynn Kamerlin: Both your methods are very interesting as we are looking at how water is transported through the interior of proteins. It is very difficult to search for experimental tools that can verify the behavior of water inside a protein. Perhaps we could also use other computational tools, like yours that show differences for MD simulations using different water models, and perhaps such cross-validation could suggest which combination of the force field and water model is more suitable for protein-interior description.

Sílvia Osuna replied: I agree, I think we should try to combine both approaches.

Shina Caroline Lynn Kamerlin answered: Water is so important, and I very much appreciate your interest in it. Tools like ours focus on the behavior of the protein. We are exploring approaches to describe water inside the protein as well, but I agree with your suggestion that cross validation using different water models/force fields would be really helpful to address this question.

Aditya Gopalakrishna Rao remarked: I would like to thank Prof. Osuna, Prof. Moliner and Prof. Kamerlin for their interesting talks: this session shows very well how molecular simulations of several different types can contribute to enzyme design and engineering, and understanding the evolution of catalytic activity. I agree that molecular dynamics simulations can help in the design and development of novel and natural enzymes. In our work, we are computationally designing photoenzymes by introducing photosensitizer binding sites¹ into de novo protein scaffolds.² We include molecular dynamics (MD) simulations as part of the workflow and design cycle, and these simulations can accelerate the process, improving design success. For instance, we are using MD simulations to assess the stability, and ligand binding ability, of our designs. Specifically, we use MD to identify successful designs where the ligand is tightly bound at the designed site.¹ These simulations are an effective tool to screen thousands of designs before experimental testing, highlighting the importance of dynamics in computational design.³

Prof. Moliner highlighted combined quantum mechanics/molecular mechanics (QM/MM) methods. QM/MM methods allow modelling of reactions in enzymes, and also calculations of electronic properties. To enhance our design strategy, we are also currently implementing a QM/MM approach that utilizes a quantum chemical method for the chromophore and a MM force field for the protein environment. Using this approach, we aim to calculate the absorption spectrum and electronic properties of the protein-bound chromophores. For photoenzyme design, accurate design algorithms could pave the way for solar-driven enzymes that sustainably target a range of critical transformations such as electricity production, N₂ fixation, H₂ production and CO₂ capture.

1 H. A. Bunzel, J. A. Smith, T. A. A. Oliver, M. R. Jones, A. J. Mulholland and J. L. R. Anderson, Photovoltaic enzymes by design and evolution, bioRxiv, 2022, preprint, DOI: 10.1101/2022.12.20.521207.

2 G. H. Hutchins, C. E. M. Noble, H. A. Bunzel, C. Williams, P. Dubiel, S. K. N. Yadav, P. M. Molinaro, R. Barringer, H. Blackburn, B. J. Hardy, A. E. Parnell, C. Landau, P. R. Race, T. A. A. Oliver, R. L. Koder, M. P. Crump, C. Schaffitzel, A. S. F. Oliveira, A. J. Mulholland and J. L. Ross Anderson, An expandable, modular de novo protein platform for precision redox engineering, Proc. Natl. Acad. Sci. U. S. A., 2023, 120, e2306046120, DOI: 10.1073/pnas.2306046120.

3 H. A. Bunzel, J. L. R. Anderson and A. J. Mulholland, Designing better enzymes: Insights from directed evolution, Curr. Opin. Struct. Biol., 2021, 67, 212–218, DOI: 10.1016/j.sbi.2020.12.015.

Jeremy N. Harvey addressed Aditya Gopalakrishna Rao: Thanks for the interesting complementary perspective. I guess your comment relates to Prof. Osuna's paper in one respect, the timescale on which important conformational fluctuation and relaxation processes happen in proteins. In Prof. Osuna's talk (https://doi.org/10.1039/d3fd00156c), I was struck by the improved free-energy landscape obtained when running multiple 30 ns MD runs instead of 10 ns. I wonder if the experience with the SPM method indicates whether the tens of ns timescale frequently shows up as one on which significant structural changes occur?

Aditya Gopalakrishna Rao responded: Bunzel and coworkers have recently used the SPM method to analyse designer enzymes and observed that longer MD runs of around 100 ns provide the most meaningful results for systems in which conformational fluctuations are integral.¹ Our current design work is aimed at designing photosensitizer binding pockets. This is similar to another work by Bunzel and coworkers that aimed to design a photoenzyme for use in bio-hybrid solar cells.² In this work, we use MD to identify failed designs that do not bind the photosensitizer. Interestingly, we have observed that longer MD simulations of 50–100 ns are required to observe ligand unbinding, while shorter simulations often do not capture the unbinding event. In our case, running 100 ns simulations has offered a quick way to eliminate failed designs and provide better hits for experimental testing. The length of the MD simulations will typically depend on the size of the system and the timescales of its conformational changes; replicate simulations should be performed to test the significance of the observations, as noted by all three speakers.

1 H. A. Bunzel, J. L. R. Anderson, D. Hilvert, V. L. Arcus, M. W. van der Kamp and A. J. Mulholland, Evolution of dynamical networks enhances catalysis in a designer enzyme, Nat. Chem., 2021, 13, 1017–1022, DOI: 10.1038/s41557-021-00763-6.

2 H. A. Bunzel, J. A. Smith, T. A. A. Oliver, M. R. Jones, A. J. Mulholland and J. L. R. Anderson, Photovoltaic enzymes by design and evolution, bioRxiv, 2022, preprint, DOI: 10.1101/2022.12.20.521207.

Sílvia Osuna added: I should mention here that our recommendation is to run SPM using multiple replica MD simulations with timescales of hundreds of ns (250–550 ns). However, in the article included in this Faraday Discussion (https://doi.org/10.1039/d3fd00156c), we used a different strategy: instead of taking the X-ray structure (or AF2 model) and following the protocol described above, we applied our recently developed template-based AF2 approach to generate a set of diverse starting structures presenting open, partially closed, and closed conformations of the COMM domain of TrpB, which were then subjected to 10–50 ns MD simulations. We found that the latter approach was able to more efficiently sample the conformational space, and the generated SPM captured a higher number of directed evolution hotspots.

Marc Van der Kamp said: In the Conclusions section of your paper (https://doi.org/10.1039/d3fd00156c), you mentioned your approach of combining template-based AlphaFold for different conformations, MD simulation and SPM analysis to (1) redesign enzymes and (2) rank new variants. Some of the SPMs you showed indicate >100 positions. How would you go about using this information for redesign? I.e., how would you narrow down which positions are most beneficial to target? For using this approach in the ranking of (computationally suggested) variants, would you need to do additional simulations on such variants, or would you already be able to get this ranking from the ‘original’ SPM?

Sílvia Osuna replied: I agree that with the SPM approach we identify a large number of positions; that's why we usually combine SPM with multiple-sequence alignment, for instance, to reduce this number further. However, this strategy still identifies too many positions and, at the end, we select a subset of positions and run MD simulations to assess whether the mutations introduced change the reconstructed free-energy landscape and stabilise the catalytically relevant states.

Marko Hanzevacki commented: You mentioned you used AlphaFold to obtain starting conformations. Standard AlphaFold gives us several structural models by default, which are usually very similar. How do you force it to give you more diverse conformations, such as open and closed states?

Sílvia Osuna responded: In the developed template-based AF2 approach, we force AF2 to give more diverse conformations by changing the templates used for the prediction, i.e., providing either closed or open structures coming from either X-ray data or conformations from MD simulations.

Jeremy N. Harvey remarked: In your study (https://doi.org/10.1039/d3fd00156c), you make use of short MD simulations initiated from multiple different structures in order to get an overview of the free-energy landscape populated by the system being studied. Crucial to this, in the case of conformational changes, is to have initial structures corresponding to each of the important conformational states. You generate these in a number of different ways, including using AlphaFold based on different structural templates. Can you comment more generally on how much of a challenge you have found it to be to obtain a good range of initial structures reflecting the conformational diversity?

Sílvia Osuna answered: In the study where we developed the template-based AF2 (tAF2) approach,¹ we tested the possibility of providing as initial structures (templates) either X-ray structures or conformations extracted from MD simulations. Interestingly, we found that the range of structures provided by tAF2 in both cases (either X-ray or MD templates) presented good pLDDT scores. More importantly, we applied principal component analysis taking all available TrpB X-ray structures and projected the tAF2 output structures. We found that most of the tAF2 output structures were fitting well in the generated PC space based on experimental X-ray data, thus confirming the quality of the tAF2 structures.

1 G. Casadevall, C. Duran, M. Estévez-Gay and Sílvia Osuna, Estimating conformational heterogeneity of tryptophan synthase with a template-based Alphafold2 approach, Protein Sci., 2022, 31, e4426, DOI: 10.1002/pro.4426.

Marc Van der Kamp commented: Your work includes the very interesting, detailed comparison between SPMs obtained based on different distance matrices. You indicate the high similarity between some of these, e.g. the ‘default’ SPM of 0B2-PfTrpB and the SPM obtained from the distance matrix generated from the individual contributions of the closed and open states, which do indeed appear clear from the image provided (Fig. 7A and D in your paper [https://doi.org/10.1039/d3fd00156c]). I also note a similarity between the ‘default’ SPM of PfTrpB and the SPM of OB2-PfTrpB based on the distance matrix of the closed structures (Fig. 6A and 7B in your paper). It would be useful to quantify such a similarity. Have you considered formulating/calculating a similarity score, and could such a similarity comparison be affected by small changes in correlation that can lead to slightly different path residue selections in SPM? Would, for example, a dot product between the underlying dynamical cross-correlation matrices be useful to help quantify the similarity?

Sílvia Osuna replied: Thank you for the question. This is something we are currently working on. I completely agree that developing a metric to compare the similarity between the obtained SPM networks would be highly useful. However, as you point out, the similarity comparison and also the SPM graphs are affected by small changes in correlation. One of the key points is how the trajectories are aligned, as the method of alignment freezes some positions, thus affecting the computed dynamical cross-correlation matrix.

Agata Raczyńska opened the discussion of the paper by Vincent Moliner: You showed in your work (https://doi.org/10.1039/d4fd00022f) that ligand conformation RC2 represents superior protein–ligand interactions, with an activation free energy for the acylation step that is around 10 kcal mol⁻¹ lower than that of the RC1 starting structure. At the same time, the kinetics of activation of the catalytic serine are essentially equivalent for both conformations.

I was wondering what causes such a significant difference in the acylation step. Upon examining these conformations, it seems that in the favourable RC2 orientation, the interactions between the oxygen leaving group and the protonated Nε nitrogen atom of the catalytic histidine are more feasible due to their closer proximity, which facilitates the stabilisation of the tetrahedral intermediate. Do you think that this improved stabilisation of the oxygen leaving group in the tetrahedral intermediate could be responsible for this substantial difference in energy?

Vicent Moliner replied: Thank you very much for your question. This is a very good question: what is the origin of the different activation free energies in the second step of the reaction when starting from different orientations of Impranil in the active site of the enzyme? Yes, despite both conformations appearing to be stable, and with favorable interactions with the protein in the reactant complex state, after the first step takes place, the relative orientation of the key residues and the substrate is much more favorable for the second step to proceed in the RC2 orientation. Not only is the tetrahedral intermediate more stable in the RC2 path, but the relative orientation of protein and substrate facilitates the transfer of the proton back from His to the ester oxygen atom of the substrate (the oxygen leaving group), as you mention.

Katarzyna Szleper said: In the QM/MM calculations, the QM region had to include the catalytic water molecule. However, considering that the active-site pocket is exposed to the solvent, could the inclusion of additional surrounding water molecules potentially lower the energy barrier by providing stabilization or acting as proton shuttles? Additionally, were the positions of water molecules considered when selecting the representative frame for the QM/MM calculations?

Vicent Moliner answered: Thank you for the question. This is a very interesting question: to select the size of the QM region in QM/MM calculations. The problem of increasing this size is that the calculations become much more expensive. In addition, it does not always provide better results. Anyway, in our case, the size of the QM subset of atoms allows a proper description of the reactive fragments. A water molecule is necessary in the hydrolysis step when the first product is released from the active site, and this is the reason we include a water molecule in the QM region. We do not expect, for this particular reaction, an effect of increasing the number of water molecules. As you comment, because the active site is exposed to the solvent, it was easy to include one water molecule in the active site.

Alexander McKenzie remarked: In your paper (https://doi.org/10.1039/d4fd00022f) you mentioned many “polyurethanes” that are in fact polyesterases. Are there any enzymes that can actually attack the carbamate group in these polymers? And is this necessary for polyurethane degradation?

Vicent Moliner answered: Thank you very much. This is a very good point. Indeed, as mentioned in the Introduction section of our paper, most of the enzymes coined as “polyurethanes” are in fact polyester-hydrolases, which break the ester groups of polyester-polyurethane chains. Apparently they do not break the carbamate groups in polyether-polyurethanes. However, depending on the kind of polyurethane sample, this kind of enzyme can be efficient enough for polyurethane degradation. We are actually working on other enzymes, like the one recently reported by Badenhorst, Bornscheuer and co-workers,¹ that are capable of breaking the carbamate group, thus opening the possibilities of degrading a wide range of polyurethane samples.

1 Y. Branson, S. Söltl, C. Buchmann, R. Wei, L. Schaffert, C. P. S. Badenhorst, L. Reisky, G. Jäger and U. T. Bornscheuer, Urethanases for the Enzymatic Hydrolysis of Low Molecular Weight Carbamates and the Recycling of Polyurethanes, Angew. Chem., Int. Ed., 2023, 62, e202216220, DOI: 10.1002/anie.202216220.

Uwe T. Bornscheuer remarked: Regarding your presentation about the polyurethane esterase PueA, it must be made more clear that esterase and urethanase do not act on the same chemical bond. In mixed polyester/polyurethanes the ester bond is cleaved by the esterase (like PueA here), and the carbamate/urethane bond is cleaved by an urethanase.

Vicent Moliner responded: Thank you very much for your comment. Yes, I agree with you that it is important to stress out this point. In fact, we mention this problem in the Introduction of our paper (https://doi.org/10.1039/d4fd00022f).

Bruce R. Lichtenstein commented: You mentioned this concept of engineering the electric field for accelerating a reaction on the basis of supporting an incipient charge in the transition state of an enzymatic reaction. I have always found this idea to be somewhat overemphasized and misconceived – the electrostatic field in a protein is naturally dynamic, and rather than focus on a static image of the transition state, which lasts for some finite time much longer than the fluctuations in the protein electric field, it would seem rather more important that the protein can accommodate electrostatic frustration that is introduced when charges change over the course of a reaction path. One example of an inadvertent means to approach this comes from the PETase field, where a group performing machine learning (ML)-guided redesign introduced a charge-transfer relay on the surface of the protein,¹ which would naturally accommodate charge build-up over the course of a reaction. Do you think this is something we can approach more rationally?

1 H. Lu, D. J. Diaz, N. J. Czarnecki, C. Zhu, W. Kim, R. Shroff, D. J. Acosta, B. R. Alexander, H. O. Cole, Y. Zhang, N. A. Lynd, A. D. Ellington and H. S. Alper, Machine learning-aided engineering of hydrolases for PET depolymerization, Nature, 2022, 604, 662–667, DOI: 10.1038/s41586-022-04599-z.

Vicent Moliner answered: Thank you for your question. This is a very good comment. We agree that focusing on the electric field can be a bit risky, just because the vectorial character of this magnitude can fluctuate significantly within the dynamics of the protein. Alternatively, we focused on the electrostatic potential generated by the protein, which is a more robust magnitude. In addition, we analyze not only the transition states (TSs), but the intermediates. In this particular case, a positive charge is increasing in the first step of the reaction, from reactant complex (RC) to the first intermediate, not only in the TS1. Thus, if stabilizing the intermediate (and the TS), the free-energy barrier would decrease, which would be in agreement with the Hammond postulate. We are now working on this aspect.

Artur Góra remarked: Impranil is very attractive as a substrate for initial activity testing, but we are not sure if this is a good substrate for enzyme activity evaluation. The problem is that the Impranil composition is not provided by the producer, and the first description of its structure is definitely wrong; a recently published one is closer to what we see in our NMR experiments.¹ Despite the problems with the structure itself, there are no any standards established to link the results of the experiments (a transfer from a milky to translucent solvent) with the degradation of a particular type of bonds and their number. In other words, we don't know how many bonds have to be cleaved, and which types, to provide effective change of solvent transparency. Is it enough to break chains to 10-mer fragments, or perhaps dimers or trimers? Also, it is extremely difficult to analyze the products of hydrolysis since not all products are solubilized, and thus it is hard to evaluate the Impranil degradation level. Summarizing, we need to develop good standard approaches to evaluate the degradation of polymers.

1 J. Fuentes-Jaime, M. Vargas-Suárez, M. J. Cruz-Gómez and H. Loza-Tavera, Concerted action of extracellular and cytoplasmic esterase and urethane-cleaving activities during Impranil biodegradation by Alicycliphilus denitrificans BQ1, Biodegradation, 2022, 33, 389–406, DOI: 10.1007/s10532-022-09989-8.

Vicent Moliner answered: Thank you for your comment; I agree with you: We need to develop good standard approaches to evaluate the degradation of polymers, and the study with a PU-like compound such as Impranil can be considered just a first step in the full project. Indeed, we selected this substrate because it was the one used by Howard and co-workers with PueA.¹ However, it is true that we could not compare our predicted kinetic data with their experimental data because they did not report rate constants. Our study suggests that this compound can bind to PueA, and the bound ester can be hydrolyzed. Future studies need to explore not only other protein scaffolds, but other substrates, including PU oligomers, that could be directly compared with experiments, as we recently did in a study of PET degradation.²

1 R. V. Stern and G. T. Howard, The polyester polyurethanase gene (pueA) from Pseudomonas chlororaphis encodes a lipase, FEMS Microbiol. Lett., 2000, 185, 163–168, DOI: 10.1111/j.1574-6968.2000.tb09056.x.

2 K. Świderek, S. Velasco-Lozano, M. À. Galmés, I. Olazabal, H. Sardon, F. López-Gallego and V. Moliner, Mechanistic studies of a lipase unveil effect of pH on hydrolysis products of small PET modules, Nat. Commun., 2023, 14, 3556, DOI: 10.1038/s41467-023-39201-1.

Jeremy N. Harvey asked: Your paper (https://doi.org/10.1039/d4fd00022f) addresses a number of challenging aspects of modelling, one of which is the difficulty in generating a sensible docked structure for the substrate, which is flexible and tends to fold back on itself under some docking protocols. In your paper, you say that some preliminary MD simulations were performed as part of this docking part of the work in which the water solvent was omitted, and only the Impranil molecule was allowed to move. This is a slightly unusual way to proceed, and I was wondering if you could explain a little bit more why you decided to use this approach.

Vicent Moliner replied: Thank you very much for your question. Yes, you are right: the docking of the substrate to the enzyme is a very sensitive step of the study of a enzyme-catalyzed reaction, particularly relevant in this case where we are trying to use a non-natural substrate: a model of a synthetic polymer. We explore the docking without water to favor the protein–substrate interactions. Then we run some optimizations and short MD simulations, to get promising and stable initial protein–substrate complex structures. Then we add the box of solvent waters and ions, and we start the long MD simulations to confirm whether our proposed structure is stable.

Adrian Mulholland said: Substrate docking is a challenge for enzymes of this type, with large and flexible substrates, and adaptation of the enzyme to substrate binding. We've found that interactive molecular dynamics in virtual reality (iMD-VR, as demonstrated in the poster sessions here) can an effective technique for flexible docking, e.g. to create enzyme–substrate complexes.^1,2 Might this iMD-VR technique be useful in enzyme design and development applications such as the one you presented? Also, as you know, it is likely that the active-site electric field at important stages in the reaction is likely to be related to catalytic activity, e.g. for deacylation (e.g. ref. 3). Have you examined electric fields from your QM/MM simulations? Would that be a potentially useful tool for designing increased esterase activity in this and other enzymes?

1 H. M. Deeks, R. K. Walters, J. Barnoud, D. R. Glowacki and A. J. Mulholland, Interactive Molecular Dynamics in Virtual Reality Is an Effective Tool for Flexible Substrate and Inhibitor Docking to the SARS-CoV-2 Main Protease, J. Chem. Inf. Model., 2020, 60, 5803–5814, DOI: 10.1021/acs.jcim.0c01030.

2 H. M. Deeks, R. K. Walters, S. R. Hare, M. B. O’Connor, A. J. Mulholland and David R. Glowacki, Interactive molecular dynamics in virtual reality for accurate flexible protein-ligand docking, PLoS One, 2020, 15(3), e0228461, DOI: 10.1371/journal.pone.0228461.

3 H. Jabeen, M. Beer, J. Spencer, M. W. van der Kamp, H. A. Bunzel and A. J. Mulholland, Electric Fields Are a Key Determinant of Carbapenemase Activity in Class A β-Lactamases, ACS Catal., 2024, 14, 7166–7172, DOI: 10.1021/acscatal.3c05302.

Vicent Moliner responded: Thank you very much for your question and your suggestion. Yes, I think iMD-VR techniques, as the one developed in your group, can be useful in enzyme design and development applications. In particular, the binding step is crucial in designing enzymes to degrade synthetic polymers, because before the chemical reaction proceed, the enzyme:polymer must for a stable reactant complex, which is not evident for these kind of systems. Consequently, the better we explore this physical step, the closer we will be to designing an efficient enzyme capable of breaking synthetic polymer chains. We have not examined the electrostatic effects (an electric field in particular directions, or the electrostatic potential generated on a specific position) on this system yet. However, we have previously done it for several systems, including methyltransferases or lipases like CALB or Bs2, and we have obtained an excellent correlation between the electrostatic potential and the magnitude of the activation energy barrier.^1–4 Anyway, we think the electrostatic potential can be a more robust magnitude to be used as a guide, because the flexibility of the proteins makes the vectorial electrostatic field a more sensitive feature.

1 K. Świderek, I. Tuñón, I. H. Williams and V. Moliner, Insights on the Origin of Catalysis on Glycine N-Methyltransferase from Computational Modeling, J. Am. Chem. Soc., 2018, 140, 4327–4334, DOI: 10.1021/jacs.7b13655.

2 M. À. Galmés, E. García-Junceda, K. Świderek and V. Moliner, Exploring the Origin of Amidase Substrate Promiscuity in CALB by a Computational Approach, ACS Catal., 2020, 10, 1938–1946, DOI: 10.1021/acscatal.9b04002.

3 M. À. Galmés, A. R. Nödling, L. Luk, K. Świderek and V. Moliner, Combined Theoretical and Experimental Study to Unravel the Differences in Promiscuous Amidase Activity of Two Nonhomologous Enzymes, ACS Catal., 2021, 11, 8635–8644, DOI: 10.1021/acscatal.1c02150.

4 M. À. Galmés, A. R. Nödling, K. He, L. Y. P. Luk, K. Świderek and V. Moliner, Computational design of an amidase by combining the best electrostatic features of two promiscuous hydrolases, Chem. Sci., 2022, 13, 4779–4787, DOI: 10.1039/d2sc00778a.

Adrian Mulholland commented: You showed an effective combination of protein design tools with QM/MM simulations for mechanistic modelling. Do you foresee other applications of QM/MM calculations in enzyme design? For example, where electronic properties (which of course QM/MM calculations can predict) are of interest?

Vicent Moliner answered: Yes, indeed. As mentioned, the presented results based on QM/MM calculations can be considered as the staring point towards the (re)design of new enzymes to degrade enzymes. Because we have observed how a positive charge is developed in the catalytic histidine residue in the first and rate-limiting steps of the reaction, we are now analyzing what mutations we could introduce to increase the negative electrostatic potential generated by the protein in this residue, which presumably would stabilize the first intermediate and, in turn, the first TS. This strategy would be similar to the one presented in our recent study published in Chemical Science.¹

1 M. À. Galmés, A. R. Nödling, K. He, L. Y. P. Luk, K. Świderek and V. Moliner, Computational design of an amidase by combining the best electrostatic features of two promiscuous hydrolases, Chem. Sci., 2022, 13, 4779–4787, DOI: 10.1039/d2sc00778a.

Marc Van der Kamp opened the discussion of the paper by Shina Caroline Lynn Kamerlin: Your tools Key Interaction Networks (KIN) and Key Interactions Finder (KIF) provide a very nice way to analyse changes in interaction networks between the Class A beta-lactamases, as you demonstrate based on the ancestrally reconstructed versions PNCA, GNCA and ENCA, for example (Fig. 8 in your paper [https://doi.org/10.1039/d4fd00018h]), moving to a more ‘specialist’ substrate scope. You also did further analysis on existing ‘modern’ Class A beta-lactamases, which are largely considered ‘specialist’. Some of these, such as the KPC variants, are actually much more ‘generalist’ than, e.g., TEM-1, as they can deal with a wide range of beta-lactam substrates, thus conferring broad-spectrum antibiotic resistance. Is this a signature you can pick up from KIN/KIF analysis of simulations of these modern, more general, beta-lactamases, or do you think the changes in interactions here are too subtle?

Shina Caroline Lynn Kamerlin responded: Thank you for this, this is a good question.

We chose the specific trajectory from PNCA to TEM-1 because we knew from prior work by Banu Ozkan that conformational dynamics is really important in facilitating the transition from generalist to specialist enzymes in these specific enzymes (see ref. 1). Our prior work originally presenting KIN looked at all modern β-lactamases as a group,² but also in the case of KPC-2, for instance, substrate specificity is linked to dynamics of the Ω-loop. So conceptually, dissecting the differences in interactions that lead to differences in loop flexibility could provide insight into the origins of the specificity. Overall, I do agree though that these are really subtle effects and very non-trivial to track.

1 T. Zou, V. A. Risso, J. A. Gavira, J. M. Sanchez-Ruiz and S. Banu Ozkan, Evolution of Conformational Dynamics Determines the Conversion of a Promiscuous Generalist into a Specialist Enzyme, Mol. Biol. Evol., 2015, 32, 132–143, DOI: 10.1093/molbev/msu281.

2 D. Yehorova, R. M. Crean, P. M. Kasson and S. C. L. Kamerlin, Key interaction networks: Identifying evolutionarily conserved non-covalent interaction networks across protein families, Protein Science, 2024, 33, e4911, DOI: 10.1002/pro.4911.

Marc Van der Kamp said: I was very interested in the concept of the specialization score you introduce in your work. You indicate that some of the largest specialization scores are for loops that are known to be quite flexible, i.e., the Ω-loop and the 214–220 loop, which thus indicates differences in the interactions that can be related to substrate specificity. It may also be possible, in general, that such highly flexible regions could lead to large specialization scores where this is likely not relevant to enzyme function and specificity, as is seen, for example, in the scores for the C-terminus (Fig. 3 in your paper [https://doi.org/10.1039/d4fd00018h]). Can you assign a significance measure to the specialization score to help distinguish this, and if so, how?

Shina Caroline Lynn Kamerlin responded: I agree completely with your concern above; we looked at this in the context of loop RMSF in a protein tyrosine phosphatase, YopH, recently (see ref. 1).

The simplest way to do this would be to do something similar to the YopH paper: compare the two scores used to make the similarity score, and do pairwise t-tests with a Benjamini–Hochberg correction to avoid false positives, and hope that gives a reasonable statistic.

To be more rigorous, this would need to be run over multiple simulation replicas, with the average value from each used to obtain a summary statistic to compare (to try to minimize running into issues with time-correlated data). If you do that though, you would need to compare between two or more different proteins at the same residue locations. Assuming you have two sets of related proteins, you would then expect the similarity scores to be high for important loops and irrelevant locations at approximately the same frequency, and then you wouldn't expect the differences to show up as significant (unless one protein has a drastically different function from the other despite an almost identical structure).

So basically you want a distribution of specialization cores associated with structural components that can be used to calculate a test statistic (to see how abnormal your specialization score actually is…). That would be challenging but could be done in a coarse way by computing three numbers for individually defined structural components (say helices, sheets, N/C-terminus residues, loops (maybe of differing lengths), etc.), and then you could create a mapping function (like how most statistical tests work) that would indicate whether a specialization score calculated for a particular residue is out of the ordinary for that type of structure. Ideally, you wouldn't want to have discrete distributions like this because you could run into major issues with edge-cases, but I don't know off the top of my head a better way to do this that would allow for the important loops to maintain a significant score while preventing C-terminus residues from lighting up as well.

(Note that I discussed all this with a graduate student in my team, Alfie-Louise Brownless, who is a QBioS PhD student with a strong background in math, and this is in large part her advice).

1 R. M. Crean, M. Corbella, A. R. Calixto, A. C. Hengge and S. C. L. Kamerlin, Sequence – dynamics – function relationships in protein tyrosine phosphatases, QRB Discov., 2024, 5, e4, DOI: 10.1017/qrd.2024.3.

Lubomír Rulíšek said: In 2010, you published a paper with Prof. Arieh Warshel entitled “At the dawn of the 21st century: Is dynamics the missing link for understanding enzyme catalysis?”¹ In the first part of that paper, you mostly discussed what is meant by the dynamical effects in enzymes catalysis (by various authors). Still, after reading the paper, I felt the answer to the title question was mostly negative. Have you, since then, somewhat changed your opinion on the influence of the enzyme dynamics on the enzyme catalysis?

1 S. C. L. Kamerlin and A. Warshel, At the dawn of the 21st century: Is dynamics the missing link for understanding enzyme catalysis?, Proteins, 2010, 78, 339–1375, DOI: 10.1002/prot.22654.

Shina Caroline Lynn Kamerlin answered: There has been a lot of controversy about the role of conformational dynamics during the chemical step of catalysis. It should be non-controversial that dynamics are functionally important, and can be fine-tuned by evolution. The role of conformational dynamics in the evolution of new enzyme functions was laid out really nicely by Dan Tawfik in his so-called “New View” of enzyme catalysis,¹ as well as subsequent articles and reviews by his group. There has been really nice subsequent work by many researchers including Nobu Tokuriki, Colin Jackson, Sílvia Osuna and others, showcasing the importance of conformational dynamics. I refer also to two recent(ish) reviews by my group: ref. 2 and 3, discussing the role of conformational (and in particular loop) dynamics in enzyme evolution, as well as it's implications for engineering. I think by now the importance of conformational dynamics to enzyme evolution and selectivity is pretty much established!

1 L. C. James and D. S. Tawfik, Conformational diversity and protein evolution – a 60-year-old hypothesis revisited, Trends Biochem. Sci., 2003, 28, 361–368, DOI: 10.1016/S0968-0004(03)00135-X.

2 R. M. Crean, J. M. Gardner and S. C. L. Kamerlin, Harnessing Conformational Plasticity to Generate Designer Enzymes, J. Am. Chem. Soc., 2020, 142, 11324–11342, DOI: 10.1021/jacs.0c04924.

3 M. Corbella, G. P. Pinto and S. C. L. Kamerlin, Loop dynamics and the evolution of enzyme activity, Nat. Rev. Chem., 2023, 7, 536–547, DOI: 10.1038/s41570-023-00495-w.

Lubomír Rulíšek opened a general discussion of the papers by Sílvia Osuna, Vicent Moliner and Shina Caroline Lynn Kamerlin: Can you somehow use the results obtained in your dynamical simulations and analyses to compute or estimate the changes in the activation entropies between various mutants (ΔΔS^‡)? There might be some experimental data available for some enzymes which can be used for calibrating the protocol (if available).

Shina Caroline Lynn Kamerlin replied: Thank you! Activation entropies would require calculating free energies for the reaction, using an approach that can reliably capture important conformational fluctuations of the system along the chemical reaction coordinate. Johan Åqvist has done really nice work in this space; I would recommend reading ref. 1 as a starting point. This can then be coupled with simulations of the conformational changes to describe the overall conformational dynamics of the system.

1 J. Åqvist, M. Kazemi, G. V. Isaksen and B. O. Brandsdal, Entropy and Enzyme Catalysis, Acc. Chem. Res., 2017, 50, 199–207, DOI: 10.1021/acs.accounts.6b00321.

Vicent Moliner added: Thank you for your question. In our case, in order to get the contribution of entropies to the activation free energies we could run simulations at different temperatures, despite it requiring long MD simulations to get a converged result. Analysis of the long MD simulations we run for the reactant complex can give you a qualitative idea of the entropies and the exploration of the conformational space, but not the contribution to the activation free energies.

Donald Hilvert remarked: Judith Klinman and others have suggested that quantum tunnelling is important for some enzymatic transformations. tunnelling efficiency should depend on barrier height and barrier width. Do you think that the barrier width could be controlled by protein conformational dynamics? Has anyone looked at this computationally?

Shina Caroline Lynn Kamerlin replied: I think your question relates to the notion of promoting vibrations; there's been really nice work done in this space by, for instance, Steve Schwartz, and Nigel Scrutton (ref. 1 and references cited therein), but it makes intuitive sense indeed that the barrier width would also be fine-tuned by the conformational ensemble!

1 V. L. Schramm and S. D. Schwartz, Promoting Vibrations and the Function of Enzymes. Emerging Theoretical and Experimental Convergence, Biochemistry, 2018, 57, 3299–3308, DOI: 10.1021/acs.biochem.8b00201.

Vicent Moliner added: We have some experience in estimating tunnelling efficiency using variational transition-state theory with QM/MM simulations.^1-6 According to our results, the reduction in the energy barrier due to tunnelling is not so large (ca. 1 kcal mol⁻¹). Enzymes enrich the population of reactive conformations, which can involve those favoring the tunnelling effects. On the other hand, including tunnelling in the simulations has a dramatic impact in isotopic effects: Without tunnelling, we cannot reproduce experimental data.

1 N. Kanaan, S. Ferrer, S. Martí, M. García-Viloca, A. Kohen and V. Moliner, Temperature Dependence of the Kinetic Isotope Effects in Thymidylate Synthase. A Theoretical Study, J. Am. Chem. Soc., 2011, 133, 6692–6702, DOI: 10.1021/ja1114369.

2 L. Y. P. Luk, J. J. Ruiz-Pernía, W. M. Dawson, M. Roca, E. J. Loveridge, D. R. Glowacki, J. N. Harvey, A. J. Mulholland, I. Tuñón, V. Moliner and R. K. Allemann, Unraveling the role of protein dynamics in dihydrofolate reductase catalysis, Proc. Natl. Acad. Sci. U.S.A., 2013, 110, 16344–16349, DOI: 10.1073/pnas.1312437110.

3 L. Y. P. Luk, J. J. Ruiz-Pernía, A. S. Adesina, E. J. Loveridge, I. Tuñón, V. Moliner and R. K. Allemann, Chemical Ligation and Isotope Labeling to Locate Dynamic Effects during Catalysis by Dihydrofolate Reductase, Angew. Chem., Int. Ed., 2015, 54, 9016–9020, DOI: 10.1002/anie.201503968.

4 J. J. Ruiz Pernía, E. Behiry, L. Y. P. Luk, E. J. Loveridge, I. Tuñón, V. Moliner and R. K. Allemann, Minimization of Dynamic Effects in the Evolution of Dihydrofolate Reductase, Chem. Sci., 2016, 7, 3248–3255, DOI: 10.1039/c5sc04209g.

5 I. Gurevic, Z. Islam, K. Świderek, K. Trepka, A. K. Ghosh, V. Moliner and A. Kohen, Experimental and Computational Studies Delineate the Role of Asparagine 177 in Hydride Transfer for E. coli Thymidylate Synthase, ACS Catal., 2018, 8, 10241–10253, DOI: 10.1021/acscatal.8b02554.

6 J. J. Ruiz-Pernía, I. Tuñón, V. Moliner and R. K. Allemann, Why are some Enzymes Dimers? Flexibility and Catalysis in Thermotoga Maritima Dihydrofolate Reductase, ACS Catal., 2019, 9, 5902–5911, DOI: 10.1021/acscatal.9b01250.

Donald Hilvert replied: Thanks for the clarification.

Adrian Mulholland commented: As you say, quantum tunnelling is important in some enzyme-catalysed hydrogen transfer reactions. This is shown, e.g., by large primary H/D kinetic isotope effects found for some enzymes. An example is aromatic amine dehydrogenase (AADH), which catalyses deamination of tryptamine with an experimentally measured H/D kinetic isotope effect (KIE) of around 55. QM/MM calculations in a variational transition-state theory (VTST) framework, with small curvature tunnelling corrections, show comparable large KIEs, and also give effective energy barriers in good agreement with experiment.^1–3 These calculations allow dissection of different contributions to the reaction. They show that tunnelling effectively reduces the barrier to proton transfer by 4–5 kcal mol⁻¹, compared to the reaction without tunnelling. As expected, the tunnelling contribution for the deuterated substrate is smaller, around 3 kcal mol⁻¹. These calculations (and other applications) show that enzyme-catalysed reactions involving significant quantum tunnelling can be described well by TS theory. They also show no role of long-range dynamics in promoting tunnelling. There is no need to invoke long-range promoting motions to explain the experimental observations. Short-range motions of groups in the active site will have an effect, e.g., by modulating the barrier, and electrostatic effects affect the tunnelling probability by changing the reaction energy. It's also worth pointing out that the calculations show two possible reaction pathways, with proton transfer to one or other carboxylate oxygen of the catalytic aspartate, with different barriers and tunnelling contributions, which may both contribute to the unusual temperature dependence of the KIE. So tunnelling is certainly important in the reaction in this enzyme with this substrate. The question of how much tunnelling contributes to catalysis, i.e., by how much does tunnelling increase the reaction rate, depends on what you compare to. The uncatalysed proton transfer reaction between tryptamine and a carboxylate base in solution would also involve some tunnelling, though probably less than in the enzyme. So yes, tunnelling does contribute somewhat to catalysis in this case, an effective contribution to barrier lowering of less than 4 kcal mol⁻¹. Other contributions (particularly electrostatic interactions, and preorganization of the active site) are more important in catalysis (i.e., rate acceleration) by the enzyme. QM/MM VTST calculations have analysed tunnelling in other enzymes.⁴ Some other enzymes, such as methylamine dehydrogenase (MADH), also have relatively large KIEs for proton transfer with some substrates, and notable but somewhat smaller tunnelling contributions.⁵ The tunnelling contribution to hydride transfer in dihydrofolate reductase is small and is not changed in the heavy (isotopically substituted) enzyme.⁶ Kinetic data, and the temperature dependence of KIEs for AADH, MADH, DHFR and soybean lipooxygenase-1 (SLO-1) can be fit by a TS theory model, also allowing analysis of contributions from quantum tunnelling.⁷

1 L. Masgrau, A. Roujeinikova, L. O. Johannissen, P. Hothi, J. Basran, K. E. Ranaghan, A. J. Mulholland, M. J. Sutcliffe, N. S. Scrutton and D. Leys, Atomic Description of an Enzyme Reaction Dominated by Proton Tunnelling, Science, 2006, 312, 237–241, DOI: 10.1126/science.1126002.

2 L. Masgrau, K. E. Ranaghan, N. S. Scrutton, A. J. Mulholland and M. J. Sutcliffe, Tunnelling and Classical Paths for Proton Transfer in an Enzyme Reaction Dominated by Tunnelling: Oxidation of Tryptamine by Aromatic Amine Dehydrogenase, J. Phys. Chem. B, 2007, 111, 3032–3047, DOI: 10.1021/jp067898k.

3 K. E. Ranaghan, W. G. Morris, L. Masgrau, K. Senthilkumar, L. O. Johannissen, N. S. Scrutton, J. N. Harvey, F. R. Manby and A. J. Mulholland, Ab Initio QM/MM Modeling of the Rate-Limiting Proton Transfer Step in the Deamination of Tryptamine by Aromatic Amine Dehydrogenase, J. Phys. Chem. B, 2017, 121, 9785–9798, DOI: 10.1021/acs.jpcb.7b06892.

4 D. G. Truhlar, Transition state theory for enzyme kinetics, Arch. Biochem. Biophys., 2015, 582, 10–17, DOI: 10.1016/j.abb.2015.05.004.

5 K. E. Ranaghan, L. Masgrau, N. S. Scrutton, M. J. Sutcliffe and A. J. Mulholland, Analysis of Classical and Quantum Paths for Deprotonation of Methylamine by Methylamine Dehydrogenase, ChemPhysChem, 2007, 8, 1816–1835, DOI: 10.1002/cphc.200700143.

6 L. Y. P. Luk, J. J. Ruiz-Pernía, W. M. Dawson, M. Roca, E. J. Loveridge, D. R. Glowacki, J. N. Harvey, A. J. Mulholland, I. Tuñón, V. Moliner and R. K. Allemann, Unraveling the role of protein dynamics indihydrofolate reductase catalysis, PNAS, 2013, 110, 16344–16349, DOI: 10.1073/pnas.1312437110.

7 D. R. Glowacki, J. N. Harvey and A. J. Mulholland, Taking Ockham’s razor to enzyme dynamics and catalysis, Nat. Chem., 2012, 4, 169–176, DOI: 10.1038/nchem.1244.

Donald Hilvert addressed asked: Are these effects expressed in the pre-exponential factor or in the kinetic isotope effect? How would narrowing the barrier affect the pre-exponential factor?

Adrian Mulholland replied: Quantum tunnelling affects the pre-exponential factor. tunnelling is an important factor, but not the only factor, in determining kinetic isotope effects for reactions involving hydrogen transfer. Enzyme reactivity (including reactions with significant quantum tunnelling) can be described well by transition-state theory. When treating tunnelling, though, it is important to account for its temperature dependence. We have shown that a simple transition-state theory (TST) model accounts for kinetic isotope effects (KIEs), and their temperature dependence, in enzyme-catalysed reactions.^1–3 This TST model fits data for a range of enzymes, using a small number of physically reasonable parameters. In some cases, but not all, two conformations with different reactivity have to be included. For the tunnelling contribution, we use a parabolic function developed by Truhlar et al. for the transmission coefficient, which is temperature dependent.⁴ It is important to note that deuteration affects other contributions to the rate also, not just tunnelling. These effects combine to give the unusual temperature dependence of KIEs observed in many enzymes. Simple models based solely, e.g., on a donor–acceptor distance cannot reproduce these observations. Finally, I would also note that the effect of protein dynamics on the reaction rate is small (and so is not a significant contribution to catalysis), but can be measured in some cases; there can be a small dynamical effect on the rate.³

1 D. R. Glowacki, J. N. Harvey and A. J. Mulholland, Taking Ockham’s razor to enzyme dynamics and catalysis, Nat. Chem., 2012, 4, 169–176, DOI: 10.1038/nchem.1244.

2 D. R. Glowacki, J. N. Harvey and A. J. Mulholland, Protein dynamics and enzyme catalysis: the ghost in the machine?, Biochem. Soc. Trans., 2012, 40, 515–521, DOI: 10.1042/BST20120047.

3 L. Y. P. Luk, J. J. Ruiz-Pernía, W. M. Dawson, M. Roca, E. J. Loveridge, D. R. Glowacki, J. N. Harvey, A. J. Mulholland, I. Tuñón, V. Moliner and R. K. Allemann, Unraveling the role of protein dynamics indihydrofolate reductase catalysis, PNAS, 2013, 110, 16344–16349, DOI: 10.1073/pnas.1312437110.

4 R. T. Skodje and D. G. Truhlar, Parabolic tunnelling calculations, J. Phys. Chem., 1981, 85, 624–628, DOI: 10.1021/j150606a003.

Shina Caroline Lynn Kamerlin replied: With the caveat that I could be wrong, and would be interested in your thoughts on this, I think the pre-exponential factor would likely not be affected, but the KIEs report information on the barrier width, so conversely an altered barrier width would change the KIE?

Vicent Moliner answered: Thank you very much for your question. Yes, indeed: According to the VTST, both the recrossing and the tuning effects are incorporated as transmission coefficients that appear in the pre-exponential factor of the rate equation. Accordingly, the effect of the tunnelling in increasing the speed of a chemical reaction is not as dramatic as those coming from reducing the activation free-energy barrier. Incorporating the tunnelling in the simulations has a dramatic effect on the KIEs, but does not have as much of an effect on the phenomenological activation free energy. You are right, narrowing the barrier would increase the tunnelling, but this is an effect that can also appear in the reaction in solution. In any environment, if we think about a light particle transfer between a donor and accepting atoms, the closer these two are, the lower the barrier would be. But the question is whether the role of the enzyme is in creating an active site that favors the population of reactive conformations or increasing the tunnelling effect. I would say the former.

Jeremy N. Harvey addressed Shina Caroline Lynn Kamerlin and Sílvia Osuna: It is striking that the papers by your two groups (https://doi.org/10.1039/d4fd00018h and https://doi.org/10.1039/d3fd00156c) address similar challenges using related methods, yet there are also significant differences in the approaches you use. Could you both comment on where you see similarities and differences between your approaches?

Sílvia Osuna responded: I think all tools are really complementary. SPM searches for correlated movements of residues, which of course indirectly depend on the changes in non-covalent interactions. These changes in interactions are nicely traced by KIN/KIF, and also one of the key points I like of KIF/KIN is that they can trace down interactions not only within the protein scaffold, but also with the substrate/ligand/product. This is a nice complement; the three tools combined can provide useful information in identifying the key interactions for favouring a particular conformational change, which is useful for enzyme redesign.

Shina Caroline Lynn Kamerlin responded: Thank you for your question. I am actually a huge fan of Dr Osuna's SPM approach, and my group have used this approach a lot, both in the past and ongoing to dissect allostery in biomedically relevant systems. It's a really valuable tool. Key differences between SPM and KIF (Key Interactions Finder) and KIN (Key Interaction Networks) are that (1) KIF can provide information at the level of individual non-covalent interactions (not just residues in a network), as well as filtering them by interaction type, and also KIF can calculate differences in interactions between states so it's particularly well suited for tracking protein conformational changes, which is what it was designed for; and (2) KIN calculates residue interaction networks and can basically do what KIF does, but instead of looking at differences in interactions between states, it looks at evolutionary conservation of interactions across families of proteins, filtering again by interaction type and prevalence. Note that KIF can also calculate residue correlation information that can be passed on to tools such as SPM, so KIF can interface with SPM. I think that all three methods provide useful information, exactly which you use depends on what kind of questions you are asking/what level of detail you are after, and also you can (and we do) use all three in conjunction with each other to obtain different information about the same system.

Amulyasai Bakshi remarked: Whenever I try to understand the characteristics (solvent accessibility, flexibility) of my enzyme using online tools that take the AA sequence as the input, I get different results from different websites. How do I know which tool is reliable, or is there a way to know which of the tools is best for the properties I am looking for?

Sílvia Osuna answered: Thanks for the question. I don't think there is a unique general and reliable tool that should be applied, and the reason for that is that many of the properties are hard to describe using a single descriptor. For instance, I always discuss the need of these multi-scale approaches for describing enzymatic catalysis: one needs to understand the chemical steps in detail (which can be described with quantum mechanics), but also the associated conformational changes (which can be described with MD simulations). There is no general tool able to properly capture all of those, and at the end the user needs to evaluate different properties using multiple tools, which all provide relevant and complementary information. We always compare the result of the tools with experimental data, and then make predictions based on the identified descriptors to validate the generated model.

Vicent Moliner responded: Thank you very much for your question. I am afraid your question does not have a straightforward answer. When we deal with small molecules in the gas phase, we can reproduce whatever calculation at whatever level of theory from any laboratory around the world. When dealing with such complex, large and flexible systems such as enzymes, it is very difficult to compare results from different simulations obtained from different laboratories. And predicting the result is even more difficult. Regarding your specific question, understanding the characteristics of enzymes using online tools that take the AA sequence as the input is too risky, in my opinion. I am afraid that a lot of experience is required and, at the end of the day, comparison with experimental data is the only reasonable reference.

Shina Caroline Lynn Kamerlin responded: This is unfortunately a common problem; while the core physics underlying the different approaches is either the same or overlaps heavily, the fine details of implementation can vary (sometimes even just subtly) enough that there will be core information that they will all pick up, but then there will also be more fine details that some web servers will pick up but not others (for example, specific interactions, etc.).

My recommendation for you if you are a newcomer to these tools would be to:

(1) Check the literature – is this a well-used established resource that others in the community are using, and has provided reasonable results in other contexts, and (2) since many of these resources are likely to be well-used established resources, depending on how fine-grained a level of detail you need, you could run a few and take a consensus between them; if it's something they all pick up on then it's likely to be important.

(Bear in mind also that depending on what property you are calculating there can be some stochasticity there as well, so multiple runs using the same tool/web server can give subtly different results, but there again a consensus approach is helpful).

Adrian Mulholland addressed all: Ms Amulyasai Bakshi raises an important question: what is the experimentalist to make of these various computational methods, and what advice can we give on which methods to apply for a particular enzyme design or engineering application? For analysing allosteric effects and effects of distal residues, I would highlight nice recent work by Stefano Serapian, Giorgio Colombo, et al. who worked with Prof. Osuna and us, comparing several methods based on equilibrium and non-equilibrium molecular dynamics simulations.¹ This work showed that different methods reveal different, complementary allosteric features. It is important to remember that they are all, in different ways, reporting on physical properties of the enzyme. There are several nice methods (e.g. SPM, D-NEMD, etc.) and I would encourage people to try them. For mechanistic questions, QM/MM calculations are a good approach, as Prof. Moliner showed.

1 M. Castelli, F. Marchetti, S. Osuna, A. S. F. Oliveira, A. J. Mulholland, S. A. Serapian and G. Colombo, Decrypting Allostery in Membrane-Bound K-Ras4B Using Complementary In Silico Approaches Based on Unbiased Molecular Dynamics Simulations, J. Am. Chem. Soc., 2023, 146, 901–919, DOI: 10.1021/jacs.3c11396.

Gerard Roelfes responded: For us, Zymspot worked well to find hotspots of dynamics, which then proved a good starting point for further engineering. But there are many nice tools available or under development and I think these will only get more powerful. At the moment, it is hard to advise someone on a particular method for the design and engineering of enzymes: they all have their own strengths and limitations. What is important, I think, is for us to keep experimentally validating these designs; this will provide the feedback needed to further perfect these computational tools.

Sílvia Osuna replied: I completely agree with all of you. Ideally we would like to have a single universal tool for describing highly complex problems, such as enzymatic catalysis, regulation, etc. However, in practice, the description of many of the enzymatic properties require the use of multiple complementary tools, as shown, for instance, for elucidating allostery as Prof. Mulholland mentioned.

Vicent Moliner added: I agree with you. Different computational methods can be more efficient at improving different steps of the full catalytic cycle: the binding step (to increase the population of reactive conformations of the protein that, in turn, can be induced by the substrate) or the chemical steps (to reduce the activation free energy of the chemical reaction(s) taking place in the active site of the enzyme).

Conflicts of interest

Stefan Lutz declares the following competing financial interest(s): his work has been included in patent application(s) and he is an employee and shareholder of Codexis Inc. There are no other conflicts to declare.

This journal is © The Royal Society of Chemistry 2024