acceptodds
Under review as a conference paper at ICLR 2027

Towards de novo enzyme design with multistate Gibbs sampling

Abstract

Enzyme catalysis is fundamentally multistate, yet most computational design methods optimize proteins around a single catalytic geometry. We introduce Multistate Gibbs Sampling (MuGS), a framework for designing one sequence to support multiple catalytic states by alternating between motif conditioned structure prediction and shared sequence updates. We find that Gibbs sampling improves atomic motif scaffolding in the single state setting, then extend to multiple states by curating an initial benchmark of experimentally characterized enzymes (MAME) and validate on a five state serine hydrolase catalytic cycle. Across all, MuGS refines and rescues existing designs and can also generate multistate compatible sequences de novo. These results suggest a path toward enzyme design that treats catalytic trajectories, rather than isolated states, as the primary object of design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
protein, design, enzyme, multistate

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_Lcy5OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This manuscript proposes Multistate Gibbs Sampling (MUGS) for the design of protein seqs that can support multiple states of an enzyme's catalytic cycle. The method alternates between predicting structures for different catalytic sttes using a new motif-conditioned structure model and updating a shared protein seq using LigandMPNN. MUGS can both refine existing designs and do the de novo design. The authors evaluate MUGS on the single-state AME benchmark, and introduce a new multistate benchmark, and show a case study on a 5-state serine hydrolase system. The paper also explores inference-time scaling using beam search and sequence Monte Carlo, and applies the motif-conditioned structure model (MCFOLD) to approximate reaction free-energy profiles with reduced QM/MM sampling.

Critical strengths
  1. MUGS combines motif-conditioned structure prediction with multi-state seq design in a straightforward iterative procedure, which is flexible for different design settings.
  2. Promising experimental results on computational benchmarks. The method shows consistent improvements over the RFD2 baseline on AME and MAME. These results provide encouraging evidence that multistate conditioning is useful.
  3. MCFOLD provides atomic motif conditioning for sequence-based structure prediction. The ablations show improvements over standard structure prediction and template conditioning. The MAME benchmark also provides a useful starting point for evaluating multistate enzyme design methods.
Critical weaknesses
  1. The sequence update does not clearly optimize compatibility across all states. A central goal of MUGS is to find one sequence that is compatible with every catalytic state. However, in Section 2.3, the authors use tied LigandMPNN by averaging the amino acid probabilities predicted for each state. I am not convinced that this is the best way to combine information from different states. For example, an amino acid that is strongly preferred in one state but poorly supported in another may still receive a high average probability. As a result, the sequence update does not explicitly require agreement across all states. This is important because the evaluation uses the worst-performing state. There seems to be a gap between how the sequence is generated and how success is measured.

  2. MUGS independently predicts a structure for each catalytic state, conditioned on the same sequence. However, this does not ensure that the protein can physically move between these structures or that the different states have suitable relative energies. More importantly, successful catalysis requires more than structural compatibility. An enzyme must stabilize the relevant transition states while allowing substrate binding, product release, and turnover. A design that strongly stabilizes every state could even reduce catalytic efficiency by over-stabilizing the substrate or product. The current evaluation focuses on motif RMSD, which does not capture these effects. Therefore, the results demonstrate improved multistate structural compatibility, but provide weaker evidence for improved catalytic function.

  3. The main comparisons are between RFD2 and MUGS with different numbers of conditioning states. However, MUGS introduces several changes at once: a new motif-conditioned structure predictor, repeated structure-sequence updates, and additional structural information from multiple states. It is therefore difficult to determine which component is responsible for the improvement. For example, MUGS with K=3 requires three structure predictions per iteration, while K = 1 requires only one. The improvement from increasing K may partly come from the additional compute rather than the multistate information itself.

  4. The serine hydrolase experiment is directly related to the paper's motivation. However, its conditioning motifs are derived from a previously validated enzyme. I think it does not fullly establish that MUGS can design enzymes for new reactions where suitable catalytic geometries are unknown or uncertain.

Questions for authors
  1. Is structural compatibility enough for the good enzyme design? Does better multistate motif RMSD lead to more favorable catalytic properties?
  2. Why is averaging amino acid probabilities the right choice for multistate sequence design?
  3. How do the authors verify that the designed sequences support the different states without relying on motif constraints? Are the final structures evaluated using an independent predictor without motif conditioning? Have the authors tested the designs using alternative structure predictors or physical relaxation? Also, can the authors report the similarity of the generated structures to native enzymes and to structures used in model training?
  4. How accurate are the predicted reaction free-energy barriers? The paper mainly reports agreement in terms of Pearson correlation. Could the authors also report errors in activation barriers and reactant-product free-energy differences?
Reviewer_yiG3OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Multistate Gibbs Sampling (MuGS), a framework for enzyme design that explicitly considers multiple catalytic states. The method alternates between motif-conditioned structure prediction using a newly trained MCFold model and shared sequence design using tied LigandMPNN.

Critical strengths

The paper addresses an important limitation of existing enzyme design approaches, which typically focus on a single catalytic state despite the multistate nature of enzymatic reactions.

The proposed framework is intuitive and provides a natural way to incorporate multiple catalytic geometries directly into sequence design.

The improvements over RFdiffusion2 in single-state scaffolding and the consistent gains from multistate conditioning provide encouraging empirical evidence.

The MAME benchmark and the reaction free-energy sketching application are interesting contributions that could support future research.

Critical weaknesses

The main evaluations focus on structural compatibility, primarily measured by motif RMSD. While these results are promising, it remains unclear how well they translate into actual catalytic activity. Experimental validation or additional physics-based assessments would strengthen the conclusions.

The MAME benchmark contains only five motif sets, and the de novo multistate design experiment focuses on a single serine hydrolase system. A broader evaluation would help establish the generalizability of the approach.

Questions for authors

Please refer to cons.

3 comments

We thank the reviewer for their overall positive review and address their concerns below!


"evaluations focus on structural compatibility [...] remains unclear how well they translate into actual catalytic activity."

While structural compatibility cannot guarantee activity, the two are meaningfully linked as demonstrated by Lauko et al. (2025). They have wet lab validated new serine hydrolases across 5 distinct folds, and show that screening for active site preorganization across the reaction cycle substantially improved experimental success. MuGS builds directly on this by moving multistate compatibility from downstream filtering into the generation process itself.

But to the reviewer's point, motifRMSD may not be perfect. We therefore also compare the catalytic triad rotamers of MuGS K=3 serine hydrolase against Lauko's enzyme, finding strong agreement in cosine sim.

StateSer χ1His χ1His χ2Asp χ1Asp χ2
Apo0.9950.9940.9760.9991.000
TI10.9030.9840.9780.9990.991
AEI0.9640.9920.9701.0000.992
TI20.9260.9940.9761.0000.999

"Experimental validation or [...] physics-based assessments would strengthen..."

We do indeed plan to pursue wet lab validation of our designs, but hope the reviewer agrees that this is beyond the scope of an ML methods paper and that our technical contributions stand their own. We would also like to point back to Fig. 6, where the red vertical line marks all state motif RMSD of Lauko's wet lab validated enzyme (1.59 Å). When we reran RFD1 (same pipeline that produced that Lauko enzyme) and RFD3, neither reached it suggesting that much more sampling was required. MuGS K=3 and K=2 were able to reach this threshold, and these are the designs we plan to test experimentally. We hope this provides sufficient evidence of promise for catalytic activity.

Regarding physics based evals, we have now performed GFN2 xTB calculations on the catalytic active sites of our MuGS K=3 design and Lauko's validated enzyme (both its crystal structure and Boltz refolds). We relaxed the reacting atoms in each state and calculated their potential energies relative to the substrate. While these are not full profiles, we find encouraging agreement between MuGS and Lauko's relative energies.

StateLauko crystalLauko Boltz foldMuGS k=3 success
sub0.000.000.00
TI1-12.50-15.70-10.38
AEI-4.27-2.30-3.54
TI2-9.12-15.72-11.42
prod+4.00+2.17+2.30

We also now ran three rep 1 ns apo MD simulations to test whether the catalytic hydrogen bonds remain stable. MuGS K=3 best design retains the His–Asp interaction in 97.7% of frames, comparable to Lauko's validated enzyme (100%), while RFD3's best design loses it entirely. MuGS also shows higher Ser–His occupancy than Lauko (25.9% vs. 13.3%) (Appendix B.5.1, Table 7, Fig. 12). We hope this along with the previous provide convincing physical validation.

DesignSer–His H-bond (%)His–Asp H-bond (%)
MuGS K=3 (best design)25.9 ± 35.197.7 ± 2.2
Lauko (wetlab validated)13.3 ± 20.0100.0 ± 0.0
RFD3 (best design)5.3 ± 7.90.0 ± 0.0

"broader evaluation would help"

We understand the reviewer's concern about benchmark scale, but many of the problems we tackle lack established benchmarks or standard evaluation protocols. Us curating MAME relied on the fortuitous availability of multiple experimental structures capturing different catalytic states of the SAME enzyme (or close homolog). For such structures to exist, experimentalists have often had to trap short lived reaction intermediates and stabilize them for crystallization and solve. Thus, suitable multistate examples are inherently rare, and we had to identify and match these structures across M-CSA and the PDB. Similarly, our reaction free energy experiments required finding well characterized reactions with suitable reference data. We therefore believe these are true steps forward to full scale benchmarks and are substantial contributions in their own right.

Nevertheless, we have greatly expanded our experiments throughout the rebuttal thanks to all the reviewers constructive comments. We have now benchmarked RFD3 across all MAME and the serine hydrolase (shows improvements in single state design do not necessarily translate to multistate compatibility (Fig. 6)). We also evaluated MuGS design novelty against the entire PDB, ran MD and xTB physical simulations, and extended our reaction free energy experiments with a better Hamiltonian for BsCM and even added a second enzyme, DhlA (Appendix Figs. 10–11).


We hope to have addressed the reviewer's concerns, but are happy to answer any more during the remaining week of rebuttal! If all concerns have been addressed, we would appreciate the reviewer's consideration of a higher score.

We wanted to kindly follow up before the discussion period closes today. We hope our rebuttal, including the new short MD simulations, expanded RFD3 benchmarks, and additional reaction free energy experiments has addressed your concerns. If you have any remaining questions or suggestions for further experiments, we're happy to address them today. We really appreciate your positive review and would love to hear your thoughts on the new results. Thank you again!

Thank you for the detailed response and the additional experiments, particularly the physics-based evaluations. My concerns have been largely addressed. I will maintain my positive assessment and keep my score at 3.

Reviewer_qpB9OpenReview ↗
Rating
3(+2)
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

Enzyme catalysis of chemical reactions undergoes multiple mechanistic steps, each associated with a different structural state of the enzyme, yet most computational design methods optimise enzyme design around a single state. The authors propose Multistate Gibbs Sampling (MuGS), a framework for designing one enzyme sequence conditioned on multiple catalytic site states. To that end, they alternate between motif conditioned structure prediction and shared sequence updates. The authors show that Gibbs sampling improves single-state atomic motif scaffolding when seeded on RFD2 designs. Then, they proceed to the multi-state setting by curating an initial benchmark of experimentally characterised enzymes (MAME) of 5 multi-state motif sets and show that MuGS rescues failed RFD2 designs. MuGS is then validated on the de novo design (masked sequence outside of catalytic site) of a five-state serine hydrolase and shown to improve the single-state design rate. The authors the validate their MCFold model on the reaction free energy sketching task for the chorismate mutase.

Critical strengths

The main strength of this work is the novel implementation of an explicitly multi-state Gibbs sampling framework for a protein design task. The authors extend recent protein design models (e.g. ProteinHunter, LASErMPNN) that use off-the-shelf structure prediction and inverse folding models in an alternating way - starting from a masked/random sequence and a condition specifying the design task - to generate de novo single-state proteins. In the case of enzyme design, the design task is scaffolding of some of the (fixed motif) catalytic site atoms in a full enzyme structure. To that end, the single-state conditional structure prediction step is replaced with multiple parallel prediction channels, one for each catalytic state condition. This results in an ensemble of enzyme structural states being predicted, which then necessitates a multi-state inverse folding approach (tied MPNN) to design one enzyme sequence that accommodates all states. The authors explore further inference time scaling in width by rollouts of multiple MuGS trajectories in parallel, using beam search and Feynman-Kac steering-like approach.

Critical weaknesses

The main weakness of this work is the poor case that is made for the utility of multi-state Gibbs sampling for de novo enzyme design. MuGS can be presented at best as a useful tool for saving failed RFD2 designs. In Sec 3.1, the authors find that single-state Gibbs sampling improves the AME benchmark results of RFD2 when used as an optimisation tool taking the RFD2 design sequences as input and saving over 1000 failed designs. However, when used for de novo design, single-state MuGS cannot consistently outperform RFD2 across the AME benchmark.

The authors curate a new multi-state benchmark of catalytic motifs. The size of the benchmark is very small: only 5 motifs with 3 states each. Again, the authors show that multi-state MuGS can significantly improve success rates of RFD2 when used as an optimisation model starting from the RFD2 design sequences, but no de novo design results are shown on this benchmark. The authors present, however, de novo design results on a five-state serine hydrolase and show improvements with respect to the original design campaign in Lauko et al. However, that campaign used RFD-1, a protein design model that is suboptimal for latest standards. I recommend the authors use a state of the art model such as RFD-3. The fact that multi-state MuGS was analysed on a single system heavily detracts from the credibility of MuGS being a useful multi-state de novo tool.

Finally, the reaction free energy sketching is an unconvincing application of MCFold. It is again only run on one system (chorismate mutase) and is only presented as a tool that improves free energy sketching given an amount of QM/MM simulations. However, the reference simulation (125 ps) seems arbitrary and does not guarantee a correct recovery of the ground truth free energy reaction profile. Moreover, QM/MM tools need a rich knowledge of the reaction catalysis mechanism (e.g. hydrogen bonding, bond forming/breakage, etc) and are known to be tricky to use for enzyme catalysis due to the large size of the system that does not allow full quantum mechanical modelling.

Questions for authors

Why did you not show de novo results for the multi-state MuGS?

How did you establish the reference QM/MM simulation of 125 ps for the free energy sketching and why did you not test MCFold on more enzyme reactions?

Comments on the paper

3 comments

I thank the authors for the comprehensive rebuttals and I apologise for the overdue response to their rebuttal. The new results that were added to the Appendix (Table 6) strongly demonstrate that MuGS is not only a pipeline useful for saving failed designs of state-of-the-art enzyme design models (seeded designs), but it consistently outperforms these enzyme design models (RFD2 and RFD3) on the unseeded (fully de novo) task as well. This result alone is worth a raising the score to a Weak Accept (3). I also advise the authors to move Table 6 to the main text as this is in my opinion the strongest result in this work.

Moreover, I accept your definition regarding "de novo" design.

My remaining concerns that prevent me from raising the score to a Clear Accept are:

  • the small size of the MAME benchmark (only 5 motifs)
  • the unclear utility of MCFold in your design pipeline, which looks disconnected from the rest of the paper and dilutes the message regarding multi-state enzyme design. I recommend the authors to demonstrate the utility of such free energy sketching in downstream filtering of MuGS designs based on the calculated free energy barriers.

We sincerely thank the reviewer for taking the time to consider our rebuttal and for being open to reconsidering their score. We really appreciate their acknowledgment of our new experiments and are glad that they helped to address original concerns. Regarding the two remaining concerns, we just wanted to briefly respond and highlight some of our ongoing and future work.


"small size of the MAME"

We appreciate the reviewer's point and are actively working to expand MAME by searching M-CSA annotated reaction mechanisms and finding corresponding multistate examples in the PDB, though we would like to emphasize that this is nontrivial, as it depends on the availability of such experimentally characterized structures across multiple catalytic states (difficult to isolate and crystallize). We completely agree that expanding this benchmark is important and consider it a central goal of our future work. Nevertheless, we believe the additional experiments provided in our rebuttal substantially strengthen our results and give us greater confidence in our designs, to the point that we are now trying to pursue wet lab validation of selected candidates.

“The unclear utility of MCFold in your design pipeline, which looks disconnected from the rest of the paper and dilutes the message regarding multi-state enzyme design.”

We would just like to clarify that MCFold is in fact a central component of MuGS, serving as the motif-conditioned structure predictor (structure step) at every iteration of the design process. Developing MCFold was necessary to enable MuGS, as existing structure prediction models could not directly accommodate our atomic motif conditioning (App. A.4).

As an added benefit, we separately also realized that such a motif-conditioned structure predictor could generate plausible protein environments around catalytic sites at different points along a reaction coordinate, enabling approximate RFE sketches through subsequent energetic evaluation. We understand the reviewer's concern that this application is distinct from the primary goal of multistate enzyme design, but believe it illustrated an additional utility of the same model developed specifically for MuGS.

Regarding the suggestion of using these energy sketches to filter MuGS designs--this is indeed a goal of ours. We are actively trying to explore this, though demonstrating whether such filtering actually improves design prioritization is difficult without wet lab activity measurements. We agree this would be a valuable next step and will discuss it in the revised manuscript.


Once again, we would really like to thank the reviewer for all their constructive criticism and for being open to such a discussion. This has been a very fruitful review process and we believe the paper has benefited greatly from it.

I thank the authors for clarifying that MCFold is the structure predicting model as part of MuGS pipeline. I revise my earlier concern that MCFold is disconnected from the rest of the paper to the more focused concern that the specific application of MCFold to enabling RFE sketches does not have a role yet in the MuGS pipeline and can be confusing to readers. I do understand that wet lab validation is the only way to prove the utility of the obtained RFE sketches for design filtering, but I still believe this experiment should be moved to the Appendix.

I appreciate the prompt response of the authors and I would like to express my moderate support again for the Acceptance of MuGS at this venue.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.