acceptodds
Under review as a conference paper at ICLR 2027

MevoCraft: Motif-Scoped Evolutionary Conditioning for Multistate Protein Design

Abstract

Many proteins function by switching among multiple distinct conformational states, and designing proteins with such behavior could enable new applications in synthetic biology and therapeutics. A recent method, SwitchCraft, introduces a unified framework for multistate protein design that optimizes a sequence by gradient-based hallucination through a biomolecular structure predictor. Though promising, SwitchCraft performs every forward pass without a multiple sequence alignment (MSA), omitting the evolutionary information that is typically useful for structure prediction. As a result, the gradient used to update the sequence is computed from the inaccurate structures. Incorporating an MSA into this hallucination process is challenging because the query sequence changes at every optimization step, meaningful homologs may not exist, and the sequence is often represented as a continuous distribution rather than a discrete sequence. To address these challenges, in this paper we propose a novel method, called otif-scoped lutionary conditioning for multistate protein design (MevoCraft). MevoCraft conditions each forward pass of gradient-based hallucination with a motif MSA which is constructed by a single homology search on the fixed functional motif and injected only into states required to exhibit that motif. To the best of our knowledge, MevoCraft is the first method to incorporate an MSA into gradient-based hallucination for multistate protein design. Experiments show that MevoCraft outperforms SwitchCraft in 50 of 70 positive-allostery cases and 49 of 70 negative-allostery cases, while matching it in another 19 positive-allostery cases and 14 negative-allostery cases. In motif switching, MevoCraft produces the number of successful designs achieved by SwitchCraft.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
multistate protein design, de novo protein design, multiple sequence alignment

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_WzFwOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes MevoCraft, a method that injects state-condition motif MSA information into each forward pass of a gradient-based protein sequence optimization pipeline. The method outperforms SwitchCraft, another gradient-based approach without MSA information, across several design tasks.

Critical strengths
  1. The proposed method is plausible, as using MSA information is important in protein design tasks. It is interesting to see how such information can be injected into continuous states during the optimization process without exhausting search at each step. This approach, to my knowledge, is novel and worth further exploration in future work.
  2. The paper is clearly written and easy to follow.
Critical weaknesses
  1. My main concern is about the experiments. The paper only compares with SwitchCraft, which shares most part of the algorithm, and lacks comparisons with other protein design frameworks that incorporate MSAs into discrete states.
  2. The experimental results do not include any statistical tests. Table 2 shows that using full MSA at evaluation only can perform well and is slightly behind the proposed approach.
Questions for authors
  1. Table 1 lacks diversity and novelty metrics. Line 258-259 state that MSAs may bias structure predictors towards particular conformations. Without novelty and diversity, it is unclear how this bias affects the obtained designs?
3 comments

We thank the reviewer for the positive assessment of our work and the constructive feedback. We are encouraged that the reviewer recognizes the novelty of our approach and its potential for further exploration. We address each of your concerns below.

W1: Limited comparisons with other protein design frameworks

R1: We thank the reviewer for raising this concern. To broaden our experimental evaluation, we extended MevoCraft to an alternative structure predictor (Protenix) and an additional protein design task (motif scaffolding), demonstrating its applicability beyond the original Boltz-based multistate design setting.

Generalization to Protenix. We integrated MevoCraft into Protenix and performed the full gradient-based sequence optimization using Protenix as the structure predictor. We also reproduced SwitchCraft using Protenix under the same optimization protocol for a fair comparison. The experiments cover four motifs and five conditioning ligands for both positive and negative allostery, with 100 independently optimized designs per case.

TaskMevoCraft-PSwitchCraft-P
Positive allostery11.25% (225/2000)4.00% (80/2000)
Negative allostery5.05% (101/2000)3.80% (76/2000)

MevoCraft-P achieves 2.81× and 1.33× as many successful designs as SwitchCraft-P for positive and negative allostery, respectively. These results show that the proposed motif-scoped evolutionary conditioning strategy can be integrated into a different structure predictor. Detailed results are provided in Appendix B.8.

Generalization to Motif Scaffolding. We further extended MevoCraft to the motif-scaffolding task, evaluating five functional motifs with 100 designs per motif. For this task, we employ the motif loss and contact loss defined in our design specification, keeping the motif residues fixed while optimizing the surrounding scaffold residues to support the target motif geometry. A design is considered successful if its mean motif Cα RMSD across five structure samples is at most 1.0 Å and its mean pLDDT exceeds 0.7.

MetricMevoCraft-MSSwitchCraft-MS
Success rate ↑52.8% (264/500)42.6% (213/500)
Motif RMSD (Å) ↓0.9601.226
pLDDT ↑0.7730.758

MevoCraft-MS improves the overall success rate by 10.2%, while achieving lower motif RMSD and higher pLDDT. These results show that the evolutionary conditioning strategy is also applicable to protein design tasks beyond multistate switching.

Detailed results are provided in Appendix B.7. Thanks again for your comment.

W2: Full MSA at evaluation is slightly behind MevoCraft and lack of statistical tests

R2: We thank the reviewer for this important comment. To further investigate whether motif MSA conditioning benefits sequence optimization beyond its effect on final evaluation, we extended the comparison between MevoCraft and its evaluation-only variant (MevoCraft-eval) from a single motif in Table 2 to 70 positive-allostery cases, 70 negative-allostery cases, and 30 motif-switching cases.

The results are summarized below:

TaskMevoCraftMevoCraft-evalRelative Improvement
Positive allostery8.51%6.76%+26.0%
Negative allostery4.77%3.86%+23.7%
Motif switching5.90%4.33%+36.2%

MevoCraft consistently achieves higher success rates than MevoCraft-eval across all three tasks, with relative improvements of 26.0%, 23.7%, and 36.2%, respectively. To quantify the uncertainty of these improvements, we performed bootstrap analyses by resampling motifs, while keeping the five ligand cases within each resampled unit together. The results are summarized below:

TaskAbsolute Improvement (pp)Bootstrap 95% CI (pp)
Positive allostery+1.76[0.79, 2.89]
Negative allostery+0.91[0.03, 1.84]
Motif switching+1.57[0.37, 2.67]

These results further support the benefits of incorporating motif MSA conditioning during optimization rather than only at evaluation.

We have added the detailed per-case success-count differences and statistical analyses to Appendix C.1. Thanks again for your comment.

Q1: Add diversity and novelty metrics

R3: We thank the reviewer for this valuable suggestion. To investigate whether MSA conditioning restricts the diversity and novelty of generated proteins, we conducted additional evaluations comparing MevoCraft and SwitchCraft across all positive- and negative-allostery cases.

Sequence Diversity. For each motif–ligand case, we measure sequence diversity among the 100 generated designs using pairwise global Needleman–Wunsch alignment. For a pair of sequences (xi,xj)(x_i,x_j), let nijn_{ij} denote the number of alignment columns containing identical amino acids, and LijL_{ij} the total alignment length, including gap-containing columns. We define pairwise identity and sequence diversity as

D=1-\frac{2}{N(N-1)}\sum_{i<j}\mathrm{Identity}(x_i,x_j).$$ **Sequence Novelty.** To assess novelty relative to known natural proteins, we search each designed sequence against Swiss-Prot using MMseqs2 with two search iterations, up to 2,500 hits per query, and an E-value threshold of $10^{-3}$. For a designed sequence $x$, novelty is defined as $$\mathrm{Novelty}(x)=1-\max_{h\in\mathcal D}\mathrm{Identity}(x,h),$$ where $\mathcal D$ denotes the retained Swiss-Prot hits and identity is calculated over the MMseqs2 alignment. If no hit satisfies the search criteria, novelty is assigned a value of 1. **Results.** The following table summarizes the average diversity and novelty across all positive- and negative-allostery cases. | Task | Metric | MevoCraft | SwitchCraft | | --- | --- | ---: | ---: | | Positive allostery | Diversity ↑ | 0.612 | 0.610 | | Negative allostery | Diversity ↑ | 0.609 | 0.608 | | Positive allostery | Novelty ↑ | 0.722 | 0.723 | | Negative allostery | Novelty ↑ | 0.718 | 0.713 | The results show that **MevoCraft maintains comparable sequence diversity and novelty to SwitchCraft**, despite achieving substantially higher design success rates. This suggests that motif-scoped MSA conditioning can guide structure prediction toward desired motif conformations without substantially restricting the diversity or novelty of the resulting sequences. Furthermore, we computed the cross-method sequence similarity between designs generated by MevoCraft and SwitchCraft, obtaining an average similarity of approximately 0.385. This indicates that the two methods generate substantially different sequences even under the same design objectives, suggesting that MevoCraft does not merely reproduce the sequences explored by SwitchCraft. We have added these diversity and novelty evaluations into Sections B.4–B.6 of the revised appendix. Thanks again for your suggestion. --- We sincerely appreciate the reviewer's constructive feedback. We hope that these additional experiments and analyses adequately address your concerns and further strengthen your confidence in our work.
Reviewer_WzFwOpenReview ↗

Reply

Hi, thank you for your response and for conducting the additional experiments, which have addressed most of my concerns regarding the experimental results. I have therefore decided to raise my rating to 4.

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

MevoCraft extends SwitchCraft by retrieving MSAs for fixed functional motifs and applying them selectively to motif-preserving states during optimization. Experiments report substantial gains in positive allostery, negative allostery, and motif switching. However, the evaluation does not clearly separate improved sequence design from effects of state-dependent predictor conditioning.

Critical strengths
  • Substantial computational gains. Allostery experiments cover 22 motifs and five ligand contexts, with 100 designs per setting. Motif switching yields 177 successful designs versus 47 for SwitchCraft, approximately a 3.8-fold improvement.
  • Useful controls and clear presentation. Evaluation-only and all-states MSA ablations probe relevant alternative explanations. Pseudocode, optimization schedules, and explicit success criteria clarify the procedure.
Critical weaknesses
  • Conditioning confound. State-specific MSAs may induce predicted conformational differences without the sequence encoding ligand-responsive switching. Table 2 addresses this concern for only one motif; benchmark-wide evaluation under matched MSA conditions, preferably with an independent predictor, is needed. Motif RMSD alone also does not establish functional allostery.
  • Performance-based subset selection. Table 1 highlights the five largest improvements, and motif-switching experiments use three motifs selected for success under MevoCraft. Although full allostery results appear in the appendix, aggregate statistics, confidence intervals, and switching tests on prespecified or held-out motifs would strengthen generalizability.
  • Limited component validation. Ablations cover only positive allostery with 1prw and omit success rates and uncertainty estimates. Broader ablations, including depth-matched shuffled-MSA or repeated-query controls, would clarify whether retrieved evolutionary information drives the gains.
  • Incomplete reproducibility and efficiency reporting. Retrieval filters, database versions, and per-motif MSA depth and diversity are unspecified. Runtime and GPU-memory comparisons are also missing.
  • Inconsistent headline results. The abstract and conclusion report 50/70 positive-allostery and 49/70 negative-allostery wins, whereas Section 4.1 and Figure 3 report 53/70 and 46/70. These discrepancies require correction.
Questions for authors

See Weaknesses.

4 comments

We sincerely thank the reviewer for the time and effort devoted to reviewing our paper and the constructive feedback. To address the concerns raised, we have conducted additional experiments, as detailed below.

W1: Conditioning confound

R1: We thank the reviewer for the insightful suggestion. Following your suggestion, we extended the comparison in Table 2 to 70 positive-allostery cases, 70 negative-allostery cases, and 30 motif-switching cases under matched MSA conditions. MevoCraft-eval denotes the evaluation-only variant in Table 2.

TaskMevoCraftMevoCraft-evalRelative Improvement
Positive allostery8.51%6.76%+26.0%
Negative allostery4.77%3.86%+23.7%
Motif switching5.90%4.33%+36.2%

MevoCraft consistently achieves higher success rates than MevoCraft-eval across all three tasks, with relative improvements of 26.0%, 23.7%, and 36.2%, respectively. To quantify the uncertainty of these improvements, we performed bootstrap analyses by resampling motifs, while keeping the five ligand cases within each resampled unit together. The results are summarized below:

TaskAbsolute Improvement (pp)Bootstrap 95% CI (pp)
Positive allostery+1.76[0.79, 2.89]
Negative allostery+0.91[0.03, 1.84]
Motif switching+1.57[0.37, 2.67]

These results further support the benefits of incorporating motif MSA conditioning during optimization rather than only at evaluation. We have added the detailed per-case success-count differences and statistical analyses to Appendix C.1. Thanks again for your comment.

We have added the detailed per-case success-count differences and statistical analyses to Appendix C.1. Thanks again for your suggestion.

W2: Performance-Based Subset Selection

R2: We thank the reviewer for raising this important concern. We would like to clarify the rationale behind our selection of motifs in Table 1 and the motif-switching experiments. We have also conducted additional experiments to further assess the generalizability of MevoCraft.

Selection of Motifs in Table 1. The purpose of Table 1 is not to demonstrate the overall superiority of MevoCraft by selectively reporting favorable cases, but rather to provide a more detailed analysis of how motif MSA conditioning affects design quality. We therefore selected five motifs exhibiting substantial improvements to investigate changes in prediction confidence, motif RMSD, and conformational differences between target states.

Importantly, MevoCraft also maintains a clear advantage over SwitchCraft across the full allostery benchmark, including all 22 motifs and five conditioning ligands:

TaskMevoCraftSwitchCraft
Positive allostery5.42%2.00%
Negative allostery3.04%0.94%

These results correspond to 2.71× and 3.23× the success rates of SwitchCraft for positive and negative allostery, respectively. Therefore, the overall performance improvements are not limited to the five motifs highlighted in Table 1.

Table 1 serves as a complementary analysis to the benchmark-wide results, helping us better understand how motif MSA conditioning influences the resulting designs. In particular, the larger motif RMSD gaps observed in the selected cases suggest that MSA conditioning helps the structure predictor better distinguish the desired apo and holo conformations.

Selection of Motifs for Motif Switching. Our motif selection was motivated by computational constraints and followed the selection strategy adopted in SwitchCraft. The full benchmark contains 22 motifs and five conditioning ligands, which would yield 2,310 motif-pair–ligand combinations. Evaluating all combinations would require substantial computational resources, with each design setting requiring tens of GPU-hours. Moreover, motif switching is a particularly challenging design task, as a single sequence must satisfy two distinct motif requirements across different ligand-conditioned states. We therefore initially selected motifs with relatively high success rates in the simpler positive- and negative-allostery tasks, following the selection strategy used by SwitchCraft. This choice enables meaningful comparisons on settings where successful switching designs are more likely to be obtained.

To further address this concern, we expanded the motif-switching experiments by including an additional motif, 1prw, and evaluating its combinations with the three previously selected motifs. The expanded set covers motif lengths of 13, 15, 24, and 40 residues, providing greater diversity in motif sizes than the previous evaluation. The results are shown below:

Motif Pair (Apo → Holo)MevoCraftSwitchCraft
1prw → 3ixt3/5000/500
1prw → 4zyp0/5000/500
1prw → 6e6r_long1/5000/500
Overall4/15000/1500

Although these additional combinations are particularly challenging, MevoCraft still produces four successful designs, while SwitchCraft fails to generate any successful designs across all 1,500 attempts. This demonstrates MevoCraft's ability to tackle challenging motif-switching tasks.

W3: Limited component validation

R3: We thank the reviewer for the valuable suggestion. We further investigate the impact of perturbing MSA information on four motifs including 3ixt, 4zyp, 6e6r_long, and 6e6r_med. We selected these motifs because Figure 3 shows substantial improvements from MSA conditioning, making them useful for examining the effects of disrupting the retrieved evolutionary information. For each motif, we evaluated five conditioning ligands with 100 designs per motif–ligand combination. We compared four variants including(1) no MSA (SwitchCraft), (2) repeated-query MSA, which replaces all retrieved homolog rows with repeated copies of the query, (3) column-shuffled MSA, which shuffles the MSA column to disrupt the evolutionary information, (4) real MSA (MevoCraft). The results are summarized below:

MethodSuccessful DesignsSuccess Rate
No MSA (SwitchCraft)148/20007.40%
Repeated-query MSA155/20007.75%
Column-shuffled MSA84/20004.20%
Real MSA (MevoCraft)364/200018.20%

We can find that repeated-query MSA achieves a success rate comparable to the No-MSA baseline, indicating that merely increasing MSA depth by duplicating the query does not account for the improvements. Moreover, column-shuffled MSA performs even worse, indicating that disrupting the relationships among homologous motif residues can negatively affect design performance. Overall, the results show that the evolutionary information in the retrieved motif MSAs plays an important role in improving design performance.

We have added the results to Appendix C.2 of the revised manuscript. Thanks again for your suggestion.

W4: Incomplete reproducibility and efficiency reporting

R4: We thank the reviewer for highlighting the need for additional implementation details and computational efficiency analysis. We have expanded the supplementary material to address these concerns.

Reproducibility. We construct motif MSAs using both structure-based and sequence-based retrieval. For structure-based retrieval, we use the public Foldseek web server with the following databases: afdb50, afdb-swissprot, afdb-proteome, pdb100, cath50, mgnify_esm30, gmgcl_id, bfmd, and BFVD. All search parameters follow the server defaults. For sequence-based retrieval, we use the ColabFold MMseqs2 API to search UniRef30. Retrieved sequences are ranked by E-value, and up to 32 MSA sequences are retained for conditioning.

We have released the environment configuration, MSA construction scripts, and MevoCraft design implementation in our anonymous repository. We have also added the retrieval details to Appendix A.1.

Computational Efficiency. We additionally benchmarked MevoCraft and SwitchCraft on an NVIDIA RTX A6000 GPU at sequence lengths of 100, 200, and 300 residues. Both methods use the same 240-step optimization protocol. The following table reports runtime and GPU memory usage for a single design run.

LengthMethodRuntime (min)GPU Memory (GB)
100SwitchCraft2.56.63
100MevoCraft2.56.64
200SwitchCraft33.39.21
200MevoCraft33.39.26
300SwitchCraft91.316.84
300MevoCraft91.416.91

The results show that MevoCraft introduces negligible computational overhead compared with SwitchCraft. Even at a sequence length of 300 residues, MevoCraft increases runtime by only 0.1 minutes and GPU memory usage by 0.07 GB. This indicates that motif-scoped evolutionary conditioning can improve multistate design performance with minimal additional computational cost.

We have included these results in Appendix B.7. Thanks again for your suggestion.

W5: Inconsistent headline results

R5: We thank the reviewer for pointing out this inconsistency. We apologize for the typographical errors in the abstract and conclusion. The correct numbers are 53/70 wins for positive allostery and 46/70 wins for negative allostery, as reported in Section 4.1 and Figure 3. We have corrected these discrepancies in the revised manuscript.


We thank the reviewer again for the valuable feedback. We hope that the additional experiments adequately address your concerns. We would appreciate it if the reviewer could re-assess our work.

Thank you for the feedback. My major concerns are addressed and I will raise my score to 3.

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

The authors propose a multistate protein design algorithm named MevoCraft to control the formation of functional motifs across multiple conformational states. The method retrieves homologous sequences and injects motif-scoped MSA information during optimization process. MevoCraft is validated on several experiments and the experiments show its the efficacy.

Critical strengths

The method is well-motivated. The experimental results seem good and show substantial improvement over SwitchCraft. The paper is well-written and easy to follow.

Critical weaknesses

The method is only experimented with Boltz. The modification to SwitchCraft seem minor. The idea is not novel to leverage MSA to manipulate structure.

Questions for authors

Do you test your method on other folding model? How sensitive is the performance to MSAs, by changing its depth and some subsampling perturbation?

2 comments

We sincerely thank the reviewer for the time and effort devoted to reviewing our paper and the constructive feedback. To address the concerns raised, we have conducted additional experiments and provided further clarification of MevoCraft's contributions, as detailed below.

W1 & Q1: Generalization to other folding models

R1: We thank the reviewer for this important question. To investigate whether MevoCraft generalizes beyond Boltz, we extended our framework to Protenix, an alternative protein structure predictor.

We performed the full gradient-based sequence optimization of MevoCraft using Protenix as the structure predictor. For a fair comparison, we also implemented SwitchCraft using Protenix under the same optimization protocol. We evaluated four functional motifs across five conditioning ligands for both positive and negative allostery, generating 100 independent designs for each motif–ligand combination.

TaskMevoCraft-PSwitchCraft-PRelative Improvement
Positive allostery11.25%4.00%+181.3%
Negative allostery5.05%3.80%+32.9%

Under Protenix-based optimization, MevoCraft achieves 2.81× and 1.33× as many successful designs as SwitchCraft for positive and negative allostery, respectively. These results show that motif-scoped evolutionary conditioning can be integrated into a different structure prediction architecture and improve gradient-based multistate sequence design beyond the original Boltz-based implementation.

We have added the experimental results and detailed comparisons to Appendix B.8 of the revised manuscript. Thanks again for your comment.

Q2: Sensitivity to MSA depth and perturbations

R2: We thank the reviewer for the valuable suggestion. To investigate the sensitivity of MevoCraft to MSA configurations, we conducted additional experiments on MSA depth and perturbations to the retrieved homologs.

Influence of MSA Depth. We first investigate the effect of MSA depth on the 6e6r_med motif by varying the number of injected homolog sequences from 0 to 64. The results are summarized below:

MSA Depth048163264
Success Rate11.2%18.8%21.4%20.8%21.2%20.4%

The success rate increases substantially as MSA depth grows from 0 to 8, and remains relatively stable between 20.4% and 21.4% for depths of 8–64. These results show that the performance of MevoCraft is relatively insensitive to depth beyond 8.

Influence of MSA Perturbations. We further investigate the impact of perturbing MSA information on four motifs including 3ixt, 4zyp, 6e6r_long, and 6e6r_med. We selected these motifs because Figure 3 shows substantial improvements from MSA conditioning, making them useful for examining the effects of disrupting the retrieved evolutionary information. For each motif, we evaluated five conditioning ligands with 100 designs per motif–ligand combination. We compared four variants: (1) no MSA (SwitchCraft), (2) repeated-query MSA, which replaces all retrieved homolog rows with repeated copies of the query, (3) column-shuffled MSA, which shuffles the MSA column to disrupt the evolutionary information, (4) real MSA (MevoCraft). The results are summarized below:

MethodSuccessful DesignsSuccess Rate
No MSA (SwitchCraft)148/20007.40%
Repeated-query MSA155/20007.75%
Column-shuffled MSA84/20004.20%
Real MSA (MevoCraft)364/200018.20%

We can find that repeated-query MSA achieves a success rate comparable to the No-MSA baseline, so merely increasing MSA depth by duplicating the query does not account for the improvements. Moreover, column-shuffled MSA performs even worse, indicating that disrupting the relationships among homologous can negatively affect the performance. Overall, these results show that MevoCraft is relatively robust to MSA depth and the evolutionary information in the retrieved motif MSAs is important in improving design performance.

We have added the results to Appendix C.2 of the revised manuscript. Thanks again for your suggestion.

W2: Clarify the novelty of MevoCraft

R2: We thank the reviewer for this comment. We agree that using MSAs to guide structure prediction toward different conformations is not new. However, our contribution is to incorporate evolutionary information into gradient-based multistate protein design, rather than simply using MSAs to manipulate predicted structures.

In gradient-based protein design, the sequence is optimized using gradients obtained from a structure predictor at each iteration. Therefore, the quality of structure prediction directly affects the optimization process and ultimately the designed sequences. While MSAs are known to improve structure prediction, using them during gradient-based design is challenging. The design sequence changes at every step, is represented by continuous sequence logits, and may not have meaningful full-length homologs.

MevoCraft addresses these challenges by retrieving homologs from fixed functional motifs and injecting the resulting MSAs into the appropriate states during optimization. These components are simple but effective and can be seamlessly integrated into gradient-based multistate protein design. Although we build on the optimization framework of SwitchCraft, our method makes it possible to use evolutionary information throughout the optimization of a changing sequence, without repeated homology searches.

We believe this is an important distinction from previous work that uses MSAs to steer structure predictions for fixed protein sequences.


We thank the reviewer again for the valuable feedback. We hope that the additional experiments adequately address your concerns. We would appreciate it if the reviewer could re-assess our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.