acceptodds
Under review as a conference paper at ICLR 2027

Towards multistate protein design with all-atom generative models

Abstract

Designing proteins that adopt defined conformations in different molecular contexts remains a central challenge in protein design. Here, we ask whether all-atom generative models can learn such context dependence and use it for multistate design. We first develop AtomDiff-X, a streamlined model for end-to-end all-atom protein design, and experimentally validate its designs. We then curate a large dataset of chameleon sequences, 6–11-residue segments that adopt different secondary structures in different protein contexts. To learn from these paired structural alternatives, we introduce a dual-track attention mechanism and fine-tune AtomDiff-X to reconstruct paired protein structures. The resulting model, AtomDiff-Chameleon, generates de novo chameleon sequences up to 48 residues long together with their structural contexts and achieves the best performance in an in silico benchmark. We further apply AtomDiff-Chameleon to design pairs of proteins with identical N-termini but different C-termini, yielding distinct structures that mimic alternative splicing to switch protein function. These results indicate that sequence–structure ambiguity in natural proteins provides scalable supervision for learning context-dependent conformations. Combined with end-to-end all-atom generation, this supervision offers a promising data-driven route towards multistate protein design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
ai for science, genrative model, diffusion, protein design

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_1SxPOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors explore the all-atom generative approach for chameleon protein design. The AtomDiff-X is first developed for end-to-end all-atom protein design, and a fine-tuned model, AtomDiff-Chameleon, is introduced to generate alternative conformations for identical amino-acid sequences in different structual contexts. In addition, a large paired-structure dataset from AFDB is constructed and utilized for finetuning and evaluation. Experimental results demonstrate that the proposed AtomDiff-Chameleon outperforms ProteinGenerator in joint generation of the shared chameleon sequences. Case study on joint motif and chameleon design illustrate the potential of designing switchable protein variants.

Critical strengths
  1. Multistate protein design is a valuable problem to investigate, and the paper identifies the challenges of that.
  2. A large paired-structure dataset is constructed, including 3M distinct sequences and 11M structural pairs, which is benefit for sequence-structure relationships across alternative conformations.
  3. The AtomDiff-Chameleon, which is finetuned based on the proposed AtomDiff-X with the constructed dataset, is introduced. Experimental results demonstrate its effectiveness and performance advantage on de novo chameleon sequence design, and a case study illustrate its potential of designing switchable protein variant.
Critical weaknesses
  1. The authors acknowledge that the designs share local sequences while their surroundings differ. A shared segment can adopt different conformations in different structual environments without establishing that one complete protein sequence can access both conformations. The structual context and tertiary interactions outside the shared segment may be decisive. The paper should frame the contribution as a framework for generating alternative structures with shared sequence segments, rather than as a practical framework for multistate protein design. The paper should clearly and consistently distinguish alternative local conformations from conformational switching by one complete protein sequence to avoid misleading.
  2. Both training set and test set are constructed from AFDB. The train/validation/test splitting procedure should be described in detail. The paper should establish that the experimental results do not rely on close homologs, duplicated structural contexts or overlapping segments, since random splitting can produce overly optimistic estimates if related sequences, motifs, or structual contexts appear across training and test sets. It would be better to report sequence identity distributions between test and training set and explain how overlapping pairs and homologous sequences are handled.
  3. AtomDiff-Chameleon contains several key components, including Token DiT, Atom DiT, and cross-branch attention. The paper would benefit from a targeted component ablation study with the same paired structural supervision.
Questions for authors
  1. How are the paired structures split across training and testing? Are sequence clusters and homologous structural contexts kept separate?
  2. How much of the improvement is attributable to each key architectural designs? Could the authors isolate these effects with more controlled ablations?
4 comments
AuthorsOpenReview ↗

Responses 1/n

We thank the reviewer for recognizing the importance and challenges of multistate protein design, the value of the Chameleon Dataset for studying how the same sequence can adopt different local structures, and AtomDiff-Chameleon’s performance in de novo chameleon design. We also appreciate the reviewer’s interest in the potential of our motif-conditioned designs. We address the reviewer’s concerns as follows:

1. Distinction between chameleon design and full-length multistate protein design.

We thank the reviewer for highlighting this distinction.

As discussed in both the Introduction and the Conclusion, we have made this distinction explicitly and explain why we take chameleon design as a starting point rather than as a demonstration of full-length conformational switching. Our approach is designed to circumvent the scarcity of multistate protein structure data. The abundance of natural chameleon sequences forms the starting point for our work. A foreseen pathway from the chameleon sequences to multistate protein design is conceived as follows:

  • First, by finetuning an all-atom generative model with paired chameleon data, we aim to obtain a model with the basic knowledge that the same local sequence fragments can correspond to different second structures within different intrachain contexts. The ability to generate de novo chameleon pairs would provide initial evidence that the model has learned this context dependence. This work is mainly focused on solving this problem.
  • Second, we envision extending the cross-branch communication mechanism developed for chameleon design to the design of multistate proteins that adopt different structures depending on their interacting partners. This extension could draw on data spanning diverse molecular interactions (protein–protein, protein–small-molecule, and protein–DNA/RNA interactions) together with a limited set of paired multistate structures. We plan to explore this direction in future work building on the current results.

We apologize that this vision was not clearly presented to the reviewer previously. We hope the additional experiments with RFD4-Proteina can convince the reviewer that this is a promising pathway worth taking (see response section 5).

  • Our current work on designing de novo chameleon pairs has given us new insights to move forwards: First, the model can design shared segments longer than those observed during training. Second, because conformational changes need not involve the entire protein, a variable chameleon fragment can be combined with a rigid functional motif.
  • Our experiment of changing base model to RFD4-Proteina allows us to take one step further towards full-length multistate protein design with external interacting partners. The chameleon-trained RFD4-Proteina is ready to generate de novo fold-switching proteins in the context of different binding partners. This provides preliminary computational evidence for context-dependent multistate design (see response section 4).

2. Ablations of key architectural components

We thank the reviewer for this suggestion. We conducted four ablations of the components introduced for chameleon design: token-level cross-branch attention, atom-level consensus communication, and the shared-sequence consistency loss (shared-JS). We do not ablate Token DiT and Atom DiT due to the limited resources and they are mainly related to uncondition base model. We evaluated all variants using their 10k-step EMA weights on the original 1,000 paired cases, with identical secondary-structure conditions, random seeds, and initial noise for each case.

ModelExactSuccessSuccessSS80
Full model53.20%14.30%4.10%
Without token-level communication (atom only)34.10%8.10%2.00%
Without atom-level communication (token only)0.00%0.00%0.00%
Without either communication pathway0.00%0.00%0.00%
Without shared-JS (both pathways retained)5.50%1.00%0.20%

In this setting, atom-level consensus was the key communication pathway for achieving shared-sequence agreement. The full model outperformed the atom-only variant on Exact (53.2% vs. 34.1%), Success (14.3% vs. 8.1%), and SuccessSS80 (4.1% vs. 2.0%), indicating token-level context exchange complementing atom-level consensus. Cross-branch communication allows the branches to exchange information, whereas shared-JS explicitly rewards them for using that information to make consistent residue choices. All three parts contribute to the final performance.

AuthorsOpenReview ↗

Responses 2/n

3. Sampling conditions, overlap with the training data, and sequence novelty

We thank the reviewer for raising concerns about close homologs, repeated structural contexts, and overlapping segments. In our sampling task, the model starts from noise and jointly generates the shared chameleon segment XX and its surrounding sequence and structural context in each of the two chains, A and B. We specify the length of each chain, the length of XX, its position in each chain, and the residue-to-residue correspondence between the two segments. For Table 3, we also provide local secondary-structure (SS) labels for XX in states A and B. These labels come from natural segments, and the model receives neither their amino-acid sequences nor their 3D coordinates or surrounding backbones. Below, we clarify what information each input provides and directly compare the generated sequences with the training sequences.

  1. Length of the shared segment. The length kk specifies only how many residues must have the same sequence in the two chains. In addition to the 6–10-residue segments in Table 3, we tested lengths of 12, 24, 36, and 48 residues in Figure 5. All four exceed the longest shared segment in our curated natural chameleon pairs (11 residues). Successful designs are thus not restricted to shared-segment lengths observed in the natural paired data.

  2. Position of the shared segment. We sample the start position of XX independently and uniformly from the eligible internal positions in each chain. We grouped the original 1,000 designs according to whether the midpoint of XX fell in the N-terminal, middle, or C-terminal third of each 96-residue chain. Each cell below reports the number of successful pairs divided by the number of pairs with that placement (Success rate). | Position of XX in chain A | Chain B: N-terminal third | Chain B: middle third | Chain B: C-terminal third | | :--- | ---: | ---: | ---: | | N-terminal third | 22/105 (21.0%) | 40/118 (33.9%) | 23/103 (22.3%) | | Middle third | 28/103 (27.2%) | 39/147 (26.5%) | 23/117 (19.7%) | | C-terminal third | 22/103 (21.4%) | 27/114 (23.7%) | 32/90 (35.6%) |

    We observed successful designs in all nine placement combinations, including cases in which XX lay near the N-terminus of one chain and near the C-terminus of the other. Successful design is therefore not restricted to a particular position, nor does it require XX to occupy the same position in both chains.

  3. Local SS conditions. The local H/E/C labels are the only sampling input taken directly from natural structures. Some of the natural pairs from which these labels were drawn also appear in the training index. We held the chain lengths, the length and positions of XX, the random seeds, and the initial noise fixed, removing only the SS labels on XX in both chains. | Position of XX in chain A | Chain B: N-terminal third | Chain B: middle third | Chain B: C-terminal third | | :--- | ---: | ---: | ---: | | N-terminal third | 22/105 (21.0%) | 40/118 (33.9%) | 23/103 (22.3%) | | Middle third | 28/103 (27.2%) | 39/147 (26.5%) | 23/117 (19.7%) | | C-terminal third | 22/103 (21.4%) | 27/114 (23.7%) | 32/90 (35.6%) |

    Without SS labels, 293 pairs still had identical sequences across XX and passed independent refolding for both chains.

  4. Similarity between generated and training sequences. We compared all 2,000 chains generated under the original SS-conditioned setting in Table 3 with the host sequences eligible for sampling during fine-tuning or earlier monomer pretraining. We also analyzed the 256 successful pairs separately. For each 96-residue generated chain, we compared its sequence with every contiguous 96-residue window in the training hosts and recorded the highest ungapped sequence identity over all 96 positions. | Comparison with training hosts | All designs (2,000 chains) | 256 successful pairs (512 chains) | | :--- | ---: | ---: | | Full 96-residue sequence occurs verbatim in a training host | 0/2,000 | 0/512 | | Median sequence identity to the closest 96-residue window | 27.08% | 26.04% | | Chains with ≥50% sequence identity to the closest 96-residue window | 11/2,000 | 1/512 |

    These results support the novelty of the designs: none of the 2,000 generated chains was copied verbatim from a training host, and the median identity to the closest 96-residue training window was only 27.08% overall and 26.04% among chains from successful pairs.

In general, given the generative nature of our model, we did not conduct the conventional train/test analysis used for discriminative models. Based on our examination of the model inputs, we see no indication of data leakage or memorization.

AuthorsOpenReview ↗

Responses 3/n

4. Bridging the gap between chameleon pairs and fold-switching proteins

To further illustrate our vision for multistate protein design, we also fine-tune RFD4-proteina on our chameleon dataset using cross-branch communication. Using RFD4 as the base model, we can combine its protein–protein interface design capability with chameleon design to generate real fold-switch proteins. We extend paired chameleon design to two complexes, C:PAC:P_A and C:PBC:P_B, both designed de novo. The same full-length protein CC adopts different folds depending on whether it binds PAP_A or PBP_B. Therefore, the design objective is a full-length protein that switches folds depending on its binding partner.

To generate de novo fold-switching proteins, we sampled using the recommended RFD4 settings. We set the protein length of 32 for CC and 96 for PAP_A and PBP_B. We generated 64 pairs of protein complexes and predicted their structures using AlphaFold3(AF3). We set the in silico success criteria as follows:

  • Both complexes have ipTM>0.6 and complex-RMSD<2,
  • The two conformations of CC differ (cross-state RMSD>3).
Success rateAvg. complex RMSD (Å)Avg. ipTMAvg. cross-state RMSD of chain CC (Å)Avg. SS difference in chain CC
15.6%(10/64)0.600.919.3886%

Here we report our in-silico Success Rate, the metrics of complex formation, and the structural differences of CC within two binding contexts. Avg. metrics are calculated for all successful designs. These preliminary results strongly indicate that our chameleon-based approach can generalize to full-length multistate protein design. Together with open-source all-atom models like RFD4-proteina, we are now even closer to brigde the gap between chameleon pairs and fold-switching proteins. The visualization of our designed fold-swtich proteins are shown in link: 5 selected successful fold-switch designs

The reviewer appreciates the authors' thorough rebuttal and additional analyses. The responses has addressed most of the concerns. The reviewer therefore maintains the rating of 3 (Weak Acceptance).

Reviewer_EMSROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents a data-driven framework for multistate protein design by leveraging context-dependent sequence–structure relationships. The authors first develop AtomDiff-X, an end-to-end all-atom diffusion model that jointly generates protein sequences and structures, and validate its design capability through computational benchmarks and biochemical experiments. To address the scarcity of multistate training data, they construct a large-scale chameleon sequence dataset containing approximately 3.77 million distinct sequences and 11.76 million structural pairs, where identical sequence segments adopt different conformations in different protein contexts. Building on this dataset, they introduce AtomDiff-Chameleon, which incorporates dual-track attention and paired structural supervision to jointly generate shared sequences and their distinct structural contexts. In computational benchmarks, the proposed method achieves a 25.6% design success rate compared with 7.1% for ProteinGenerator, and demonstrates generalization to shared sequences up to 48 residues. Finally, motif-conditioned design case studies illustrate the potential to modulate functional-site accessibility through context-dependent conformational changes. Overall, the work provides a promising approach to multistate protein design, although experimental validation of the generated multistate designs and their functional switching remains an important next step.

Critical strengths
  1. Large-scale and valuable dataset contribution. The authors construct a large-scale Chameleon Dataset containing 3.77 million distinct sequences and 11.76 million structural pairs. This provides a scalable approach to extracting paired supervision from existing protein structure databases, addressing the scarcity of training data for context-dependent protein design.

  2. Comprehensive computational evaluation with preliminary experimental validation. The paper evaluates the proposed models through computational benchmarks, generalization experiments, and biochemical characterization. AtomDiff-Chameleon significantly outperforms ProteinGenerator in paired design success (25.6% vs. 7.1%), while experimental results for AtomDiff-X demonstrate promising soluble expression and monomeric behavior, providing initial evidence of the framework's practical feasibility.

Critical weaknesses
  1. Gap between the proposed task and genuine multistate protein design. The proposed method generates protein pairs that share only a local sequence segment while having different surrounding sequences. This differs from genuine multistate protein design, where the same full-length protein sequence adopts different conformations under different conditions. The authors should more clearly distinguish these settings and avoid overstating the demonstrated conformational switching capability.

  2. Insufficient ablation studies of key architectural components. The proposed framework introduces several components, including token-level cross-branch attention, atom-level consensus communication, shared sequence consistency loss, and secondary structure conditioning. However, the current ablation studies mainly focus on removing paired training data. More systematic ablations are needed to clarify the necessity of dual-track attention and the individual contributions of these components.

  3. Limited generalization to longer sequences. Although the model demonstrates the ability to generate chameleon sequences up to 48 residues, its success rate decreases substantially as sequence length increases. This raises concerns about its scalability and generalization to more challenging multistate protein design tasks.

  4. Uncertainty regarding dataset quality and structural reliability. The Chameleon Dataset is primarily constructed from AFDB-predicted structures rather than experimentally determined conformational states. Some observed secondary-structure differences may arise from prediction errors, and identical short sequences in different proteins do not necessarily indicate genuine conformational switching. Additionally, structural redundancy may reduce the effective diversity of the dataset. Validation using experimentally determined PDB structures and further analysis of structural confidence and diversity would strengthen the reliability of this contribution.

Questions for authors

see Weaknesses.

2 comments
AuthorsOpenReview ↗

Responses 1/n

We thank the reviewer for recognizing the value of the Chameleon Dataset, the improvement in paired design success over ProteinGenerator, and the preliminary biochemical characterization of our designs. We also appreciate the reviewer’s constructive feedback on the task definition, ablation studies, generalization to longer shared sequences, and dataset reliability. We address each concern in turn below.

1. Distinction between chameleon design and full-length multistate protein design.

We thank the reviewer for highlighting this distinction.

As discussed in both the Introduction and the Conclusion, we have made this distinction explicitly and explain why we take chameleon design as a starting point rather than as a demonstration of full-length conformational switching. Our approach is designed to circumvent the scarcity of multistate protein structure data. The abundance of natural chameleon sequences forms the starting point for our work. A foreseen pathway from the chameleon sequences to multistate protein design is conceived as follows:

  • First, by finetuning an all-atom generative model with paired chameleon data, we aim to obtain a model with the basic knowledge that the same local sequence fragments can correspond to different second structures within different intrachain contexts. The ability to generate de novo chameleon pairs would provide initial evidence that the model has learned this context dependence. This work is mainly focused on solving this problem.
  • Second, we envision extending the cross-branch communication mechanism developed for chameleon design to the design of multistate proteins that adopt different structures depending on their interacting partners. This extension could draw on data spanning diverse molecular interactions (protein–protein, protein–small-molecule, and protein–DNA/RNA interactions) together with a limited set of paired multistate structures. We plan to explore this direction in future work building on the current results.

We apologize that this vision was not clearly presented to the reviewer previously. We hope the additional experiments with RFD4-Proteina can convince the reviewer that this is a promising pathway worth taking (see response section 5).

  • Our current work on designing de novo chameleon pairs has given us new insights to move forwards: First, the model can design shared segments longer than those observed during training. Second, because conformational changes need not involve the entire protein, a variable chameleon fragment can be combined with a rigid functional motif.
  • Our experiment of changing base model to RFD4-Proteina allows us to take one step further towards full-length multistate protein design with external interacting partners. The chameleon-trained RFD4-Proteina is ready to generate de novo fold-switching proteins in the context of different binding partners. This provides preliminary computational evidence for context-dependent multistate design (see response section 5).

2. Ablations of key architectural components

We thank the reviewer for this suggestion. We conducted four additional ablations to examine the contributions of token-level cross-branch attention, atom-level consensus communication, and the shared-sequence consistency loss (shared-JS). We evaluated all variants using their 10k-step EMA weights on the original 1,000 paired cases, with identical secondary-structure conditions, random seeds, and initial noise for each case.

ModelExactSuccessSuccessSS80
Full model53.20%14.30%4.10%
Without token-level communication (atom only)34.10%8.10%2.00%
Without atom-level communication (token only)0.00%0.00%0.00%
Without either communication pathway0.00%0.00%0.00%
Without shared-JS (both pathways retained)5.50%1.00%0.20%

In this setting, atom-level consensus was the key communication pathway for achieving shared-sequence agreement. The full model outperformed the atom-only variant on Exact (53.2% vs. 34.1%), Success (14.3% vs. 8.1%), and SuccessSS80 (4.1% vs. 2.0%), indicating token-level context exchange complementing atom-level consensus. Cross-branch communication allows the branches to exchange information, whereas shared-JS explicitly rewards them for using that information to make consistent residue choices. All three parts contribute to the final performance.

AuthorsOpenReview ↗

Responses 2/n

3. Generalization to longer chameleon segments

We thank the reviewer for highlighting this limitation.

We agree that success rate declines as the shared segment XX grows longer: SuccessSS80 falls from 9.25% at 12 residues to 5.75%, 2.00%, and 1.25% at 24, 36, and 48 residues, respectively. Long shared segments are scarce in our training data: 94.82% of the curated chameleon pairs share only six residues, and the longest shared segment is 11 residues.

The design problem may also become harder because of intrinsic physical constraints. Because each chain is fixed at 96 residues, extending XX from 12 to 48 residues increases the shared region from one-eighth to one-half of the chain. Thus, the decline may reflect not only limited extrapolation by the model, but also a more constrained design space: a longer sequence must satisfy both structural contexts, with fewer surrounding residues available to stabilize the distinct conformation.

4. Dataset quality and structural reliability

We acknowledge this concern and accounted for regional differences in prediction reliability during dataset construction. AlphaFold predictions can be less reliable for flexible loops, partner-dependent interfaces, and ligand-binding pockets. Our task instead focuses on ordered local structures stabilized by their intrachain environments, where predictions are generally more reliable. We require every residue in the shared segment to have pLDDT ≥80 in both structures and also filter for confidence in the surrounding regions. To control redundancy, we require the two parent proteins to have less than 30% overall sequence identity, exclude pairs with highly similar flanking sequences, and merge duplicate pairs. With these filters, we consider AFDB a suitable source of large-scale supervision for design conditioned on intrachain context.

We also recognize the value of the PDB as a source of experimentally determined structures. We chose AFDB because its consistently modeled single-chain structures align with our task definition: only the portion of the same chain outside the shared segment serves as its folding context. Conformational differences between PDB entries may also reflect differences in binding partners, ligands, or experimental conditions such as temperature and pH. Without accounting for these factors, it is difficult to distinguish the influence of the intrachain sequence context from that of external conditions. This is the main reason we did not include PDB structures in the present training set. As we extend our framework to model binding partners, ligands, and other sources of external context, we plan to use the PDB as our primary data source to study how these factors influence protein conformation.

5. Bridging the gap between chameleon pairs and fold-switching proteins

To further illustrate our vision for multistate protein design, we also fine-tune RFD4-proteina on our chameleon dataset using cross-branch communication. Using RFD4 as the base model, we can combine its protein–protein interface design capability with chameleon design to generate real fold-switch proteins. We extend paired chameleon design to two complexes, C:PAC:P_A and C:PBC:P_B, both designed de novo. The same full-length protein CC adopts different folds depending on whether it binds PAP_A or PBP_B. Therefore, the design objective is a full-length protein that switches folds depending on its binding partner.

To generate de novo fold-switching proteins, we sampled using the recommended RFD4 settings. We set the protein length of 32 for CC and 96 for PAP_A and PBP_B. We generated 64 pairs of protein complexes and predicted their structures using AlphaFold3(AF3). We set the in silico success criteria as follows:

  • Both complexes have ipTM>0.6 and complex-RMSD<2,
  • The two conformations of CC differ (cross-state RMSD>3).
Success rateAvg. complex RMSD (Å)Avg. ipTMAvg. cross-state RMSD of chain CC (Å)Avg. SS difference in chain CC
15.6%(10/64)0.600.919.3886%

Here we report our in-silico Success Rate, the metrics of complex formation, and the structural differences of CC within two binding contexts. Avg. metrics are calculated for all successful designs. These preliminary results strongly indicate that our chameleon-based approach can generalize to full-length multistate protein design. Together with open-source all-atom models like RFD4-proteina, we are now even closer to brigde the gap between chameleon pairs and fold-switching proteins. The visualization of our designed fold-swtich proteins are shown in link: 5 selected successful fold-switch designs

Reviewer_soWwOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work introduces AtomDiff-X, an all-atom design model that is finetuned on curated chameleon sequences to generate de novo chameleon proteins. The resulting model named AtomDiff-Chameleon performs best on their introduced Chameleon benchmark and is also appleid to an alternative splicing experiment for multi state protein design.

Critical strengths
  • Introduction of new synthetic data based on AFDB is valuable community asset
  • New protein architecture through use of unidirectional token and atom track.
  • New conditional design benchmark for chameleon sequences
Critical weaknesses

-Overall AtomDiff-X usage seems very much forced. The paper claims "Current protein design workflows generate a protein backbone and then design sequences for that backbone" however most recent work does all-atom co-design (Pallatom, La-Proteina). AtomDiff-X performs poorly on monomer design. 3 and 4x structure diversity collapsed compared to Pallatom (that shares atom14) and La-Proteina respectively and 100% codesignability is trivial with a collapsed model. Why use this model when the goal is downstream chemeleon design rather than the prior models for chameleon finetuning when La-Proteina was also validated for its expression as well [1]?

  • Given the same atom14 representation and EDM denoising as Pallatom what explains the diversity collapse? -ProteinGenerator is one of the worst performing models so the table 3 results would be better supported by producing other baselines to validate the value of AtomDiff-X model framework itself -Table 3 benchmark does not make sense since the benchmark is introduced inthis work along with the data. Wouldn't the other unconditional models have to be fine-tuned in a similar way? -Much focus is made on multistate design but really the focus in on this subset chameleon task which is similar but this task itself should be motivated. Why is this new dataset and task useful when prior referenced work liek SwitchCraft is doing multistate design directly

[1] https://www.biorxiv.org/content/10.64898/2026.09.12.751118v1

Questions for authors
  1. How is the inference speed compared to Pallatom and La-Proteina both on batch size 1 and max batch on equal compute? Similarly as results are taken from Geffner et al what is the Novelty, secondary structure content and MPNN-based designability?
  2. Why was AtomDiff-X chosen as its use of atom14 according to table 1 performs worse than Pallatom and also worse than la-proteina. It seems that it does not learn as good of a representation to denosie all-atom proteins limiting actual performance for multistate design.
  3. Given the results in Table 1 and the subsequent validation of La-Proteina what happens when it is finetuned on the same chameleon task? I am unsure if the actual AtomDiff-X model is doing anything as its structure is near 3x collapsed compared to fellow atom14 Pallatom and 4x compared to La-Proteina. Similar architecture adaptations should be possible that enabled the AtomDiff-Chamelon lift as both prior work rely on similar usage of either atom or token level DiT.

I do not expect baseline models to be retrained during the rebuttal period but this is important future work to motivate and validate teh AtomDiff framework. It is not surprising that this model beats protein generator as that prior work struggles with the general protein modality so there needs to be more extensive benchmarking otherwise it is hard to understand what modeling contributions are vs just the data itself.

4 comments
AuthorsOpenReview ↗

Responses - 1/n

We thank the reviewer for the thoughtful feedback and suggestions, which motivated the additional experiments and analyses presented below.

1. Why did we choose to develop AtomDiff-X as the base model?

We thank the reviewer for raising this question. The focus of this work is not to develop a state-of-the-art unconditional monomer generator. Our primary goal is to explore data-driven approaches towards multistate protein design within the established schemes of all-atom generative models. We developed AtomDiff-X as a fully trainable all-atom platform where we can implement and test our hypothesis for multistate design.

Why did we not use Pallatom or La-Proteina directly? As the reviewer kindly pointed out, these are two all-atom generative models developed before our work. They represent two approaches: modeling atom14 coordinates directly, as in Pallatom, or representing residue-level atomic details in latent space, as in La-Proteina. The atom14 representation, with its direct physical interpretation of atoms and residues, appeals to us first. Thus, Pallatom seems a natural starting point. But its public release does not contain a complete training pipeline that prevent us from direct reuse. We decided to build our own atom14 model AtomDiff-X, which is fully trainable with the open-source data pipeline provided by La-Proteina.

AtomDiff-X has allowed us to develop and test chameleon-specific communication mechanisms within the atom14 scheme. The effectiveness of our cross-branch attention trained with the paired chameleon dataset (presented as the main results in this work) further motivated us to test with latent representations. With the RFD4-Proteina (released on Oct. 20 during ICLR review process), which is fully trainable with its open-source data pipeline, we show that paired supervision for chameleon design is also feasible with a latent diffusion model (see reponse section 7).

2. How should we interpret AtomDiff-X’s monomer diversity and its gap to Pallatom?

We thank the reviewer for pointing out AtomDiff-X's potential lack of diversity. In the standard benchmark, AtomDiff-X generates less structurally diverse monomers than La-Proteina and Pallatom(Table 1). This performance leaves room for improvement, but so far it does not support the reviewer's conclusion that AtomDiff-X has undergone diversity collapse. With due respect, we find it quite an overstatement to claim it as a "collapsed model".

We have multiple evidences showing that AtomDiff-X fulfilled its role as a qualified all-atom generative platform for our proposed multistate exploration:

  1. As illustrated in Figure 10, AtomDiff-X designs a range of monomer structures that can be independently refolded in silico with a high success rate. Among 8 all-atom generative models, it ranks second in quality and third in diversity. To further examine its performance, we compared AtomDiff-X's inference speed, structural novelty, secondary-structure composition, and ProteinMPNN-based designability with Pallatom and La-Proteina (see reponse section 3).

  2. More importantly, the biochemical results of purified proteins provide ground-truth evidence that supports AtomDiff-X's all-atom design capability: 14 confirmed designs cover a diverse range of length and fold family (structures illustrated in link: Structures of the 24 monomer designs selected for experimental testing)

  3. In the downstream multistate chameleon generation task, AtomDiff-X' relative lower diversity in unconditional monomer generation benchmark did not compromise its ability for generating diverse multistate chameleon structures (see reponse section 4).

To follow up on the reviewer's request of explaining the diversity gap between AtomDiff-X and Pallatom, we analyze the different criteria used by two models on selecting training data. The different criteria change the relative weight of protein folds during training and, in turn, influence the distribution of generated structures.

  • As stated in its published version [1], Pallatom filters training samples by secondary-structure composition, radius of gyration, and maximum continuous loop length, and removes structural redundancy using a TM-score threshold of 0.8 and a coverage threshold of 0.9.
  • Following La-Proteina, AtomDiff-X's training data are based on the Foldseek-clustered AFDB resource [2].

[1] https://www.biorxiv.org/content/10.1101/2024.08.16.608235v4.full.pdf

[2] https://arxiv.org/pdf/2507.09466

AuthorsOpenReview ↗

Responses - 2/n

3. How does AtomDiff-X compare with Pallatom and La-Proteina in inference speed, structural novelty, secondary-structure composition, and ProteinMPNN-based designability?

We thank the reviewer for suggesting these comparisons. We benchmarked the inference speed of all three methods on NVIDIA H20 GPUs and used the samples from Table 1 to evaluate other aspects of generation quality.

  1. Inference speed

    We measured generation time for proteins of 100, 200, 300, 400, and 500 residues, with each benchmark using a single H20 GPU. We report both the time to generate one protein at batch size 1 (Table 1a) and the time per protein at the largest supported batch size (Table 1b). For the latter, we divide the elapsed time for the full batch by the number of proteins generated.

    AtomDiff-X used 500 noise levels. Pallatom used its native 200-step sampler, and La-Proteina used 400 sampling steps.

Table 1a. Generation time at batch size 1 (seconds per protein).

Method100 aa200 aa300 aa400 aa500 aa
AtomDiff-X25.7836.7773.76128.70196.89
Pallatom19.5834.1054.6180.06OOM
La-Proteina3.124.295.938.5110.37

Table 1b. Generation time at the largest supported batch size (seconds per protein; batch size in parentheses).

Method100 aa200 aa300 aa400 aa500 aa
AtomDiff-X9.26 (348)33.63 (92)72.45 (39)128.36 (23)200.21 (15)
Pallatom19.58 (1)34.10 (1)54.61 (1)80.06 (1)OOM
La-Proteina0.69 (1201)1.87 (287)3.52 (151)5.51 (72)8.34 (50)
  1. Structural novelty, secondary-structure composition, and ProteinMPNN-based designability

    For structural novelty and secondary-structure composition, we used the subset that passed the all-atom co-designability criterion. For each structure, we searched the PDB and AFDB reference sets separately, recorded the highest TM-score in each, and averaged these nearest-neighbor scores across structures. A lower score indicates greater structural novelty. For each backbone, we redesign sequences using ProteinMPNN refolded each sequence with ESMFold.

Table 2. Additional generation-quality metrics. Values for Pallatom and La-Proteina are from La-Proteina, Table 1.

MethodPDB nearest-neighbor TM-score ↓AFDB nearest-neighbor TM-score ↓Helix (%)Strand (%)MPNN-1 (%) ↑MPNN-8 (%) ↑
AtomDiff-X0.770.83531683.293.8
Pallatom0.720.81561744.457.9
La-Proteina0.750.8272582.693.8

4. Effect of the base model on chameleon design.

We thank the reviewer for raising the question of testing other all-atom generative models, as we share the curiosity to explore how the base model choice affects the downstream multistate chameleon design. To test whether our approach extends beyond AtomDiff-X, we applied our cross-branch communication to finetuning RFD4-Proteina. As a latent-space all-atom generative model derived from La-Proteina, RFD4-Proteina has no atom-level communication pathway and our implementation uses token-level communication alone. We fine-tuned the extended RFD4-Proteina with the Chameleon Dataset and evaluated it on the chameleon design benchmark introduced in Table 3. The updated results are shown below:

MethodExact (%)↑Success (%)↑SuccessSS80 (%)↑Diversity↑
ProteinGenerator100†7.13.90.71
AtomDiff-Chameleon75.125.614.30.245
RFD4-Proteina + Chameleon fine-tuning72.349.535.20.18

† ProteinGenerator enforces sequence agreement on the shared segment by construction.

As shown in the table, using RFD4-Proteina as the base model slightly lowers the Exact sequence match (from 75.1% to 72.3%). But Success in refolding increases from 25.6% to 49.5%, and SuccessSS80 from 14.3% to 35.2%. Thus, the choice of base model does make a difference: using the RFD4-Proteina as the base model produced substantially more pairs that both refolded successfully and satisfied the prescribed secondary structures. These results support the conclusion that our approach to multistate chameleon design is generally applicable to both atom14-based and latent-space all-atom generative models.

We also measured the structural diversity of the generated chameleon designs using Foldseek. ProteinGenerator obtained the highest diversity score of 0.71. AtomDiff-Chameleon achieved a diversity score of 0.245, compared with 0.180 for fine-tuned RFD4-Proteina. According to this side-by-side comparison, the higher paired success rate does not gurantee a higher structural diversity. This result also supports that AtomDiff-X' relative lower diversity in unconditional monomer generation benchmark does not compromise its ability for generating diverse multistate chameleon structures.

AuthorsOpenReview ↗

Responses - 3/n

5. Why did we choose ProteinGenerator, and how does AtomDiff-Chameleon compare with SwitchCraft?

We thank the reviewer for questioning our choice of using ProteinGenerator as the baseline, and suggesting SwitchCraft as a potentially stronger baseline. We take this as an opportunity to explain our standards for baseline choice.

ProteinGenerator matches our definition of multistate design where two protein states can be explicitly specified on both sequence and structure levels. Its strong experimental results and easy-to-use multistate setup make it our top choice for benchmarking chameleon design.

  • The original work demonstrated multistate design (Figure 5 in Multistate and functional protein design using RoseTTAFold sequence space diffusion, Nature Biotechnology, 2025) with strong wet lab experimental results including biochemical protein purification, CD spectra validating the designed secondary structure, and NMR ensembles confirming the two distinct states. With these experimental results, we cannot agree with the reviewer's statement that ProteinGenerator is "one of the worst performing models".
  • With its multistate consensus sampling, ProteinGenerator allows users to specify positions with shared sequence identities and secondary structure (SS). This setup is ready to carry out multistate chameleon design without any additional adaptation, providing a fair and stable benchmark performance to compare with.

SwitchCraft relies on the hallucination of a folding model (Boltz2) to search for multistate sequences. Hallucination methods without sequence redesign lack widely accepted wet-lab validation. Its inference setting does not support specification on residue level.

  • A large-scale experimental test on AlphaFold2(AF2) strongly indicates that hallucinated sequences directly from a folding model cannot stand up to wet-lab validation (Hallucinating symmetric proteinassemblies, Science, 2022).
  • More importantly, SwitchCraft workflow optimizes one full-length sequence across multiple states, preventing user-specified sequence and structural states at individual positions. Using unmodified SwitchCraft for chameleon design would impose an unecessary full-length sequence constraint, resulting in an unfair comparison.

Without giving up ProteinGenerator as our baseline method, we agree with the reviewer that it would strengthen our work to include SwitchCraft in the benchmark for broad comparison and discussion. With the above-mentioned caveats, we modified SwitchCraft for multistate chameleon design and compared it with AtomDiff-Chameleon.

  • We modified the original SwitchCraft method to support shared sequence at selected segment XX to perform multistate chameleon design. We retained SwitchCraft’s original ConfChangeLoss and applied it to the X×XX \times X residue pairs, for which correspondence between states is defined. While the positions in XX share amino acid logits and receive gradients from both states, the positions outside XX are optimized independently.

Following the same evaluation settings in Table 3, we obtained:

MethodLocal SS conditioningExactSuccessTime per pair
SwitchCraft (only XX shared)No100.0%†27.80%260.4 s
AtomDiff-Chameleon (Full 20k, Table 3 setting)Yes75.10%25.60%41.6 s

† SwitchCraft enforces sequence agreement on the shared segment XX by construction through shared amino-acid logits.

Our result indicates that AtomDiff-Chameleon achieved a success rate comparable to SwitchCraft’s on this paired design task while AtomDiff-Chameleon generated 6.3 times as many pairs per unit time as SwitchCraft.

AuthorsOpenReview ↗

Responses - 4/n

6. Why study chameleon design and build this dataset when SwitchCraft can already perform multistate design?

We are fully aware that SwitchCraft can directly optimize a single sequence for multiple target states. While it represents a successful framework of performing multistate design without paired structural data, SwitchCraft does not invalidate all the other efforts, including ours, to build new datasets and explore alternative approaches.

Our approach is designed to circumvent the scarcity of multistate protein structure data. The abundance of natural chameleon sequences forms the starting point for our work. A foreseen pathway from the chameleon sequences to multistate protein design is conceived as follows:

  • First, by finetuning an all-atom generative model with paired chameleon data, we aim to obtain a model with the basic knowledge that the same local sequence fragments can correspond to different second structures within different intrachain contexts. The ability to generate de novo chameleon pairs would provide initial evidence that the model has learned this context dependence. This work is mainly focused on solving this problem.
  • Second, we envision extending the cross-branch communication mechanism developed for chameleon design to the design of multistate proteins that adopt different structures depending on their interacting partners. This extension could draw on data spanning diverse molecular interactions (protein–protein, protein–small-molecule, and protein–DNA/RNA interactions) together with a limited set of paired multistate structures. We plan to explore this direction in future work building on the current results.

With this vision, we consider our dataset of 3.77 million chameleon sequences a valuable asset to the community, and the task of de novo generation of chameleon pairs worth achieving. We apologize that this vision was not clearly presented to the reviewer previously. We hope the additional experiments with RFD4-Proteina can convince the reviewer that this is a promising pathway worth taking (see response section 7).

  • Our current work on designing de novo chameleon pairs has given us new insights to move forwards: First, the model can design shared segments longer than those observed during training. Second, because conformational changes need not involve the entire protein, a variable chameleon fragment can be combined with a rigid functional motif.
  • The experimental results of changing base model to RFD4-Proteina (see response section 4) allow us to take one step further towards full-length multistate protein design with external interacting partners. The chameleon-trained RFD4-Proteina is ready to generate de novo fold-switching proteins in the context of different binding partners. This provides preliminary computational evidence for context-dependent multistate design (see response section 7).

7. Bridging the gap between chameleon pairs and fold-switching proteins

We extend paired chameleon design to two complexes, C:PAC:P_A and C:PBC:P_B, both designed de novo. The same full-length protein CC adopts different folds depending on whether it binds PAP_A or PBP_B. Therefore, the design objective is a full-length protein that switches folds depending on its binding partner.

Design Methods. As described in response section 4, we implemented our cross-branch communication attention, and finetuned RFD4-Proteina with the Chameleon Dataset. To generate de novo fold-switching proteins, we sampled using the recommended RFD4 settings. We set the protein length of 32 for CC and 96 for PAP_A and PBP_B.

Preliminary Results. We generated 64 pairs of protein complexes and predicted their structures using AlphaFold3(AF3). We set the in silico success criteria as follows:

  • Both complexes have ipTM>0.6 and complex-RMSD<2,
  • The two conformations of CC differ (cross-state RMSD>3).
Success rateAvg. complex RMSD (Å)Avg. ipTMAvg. cross-state RMSD of chain CC (Å)Avg. SS difference in chain CC
15.6% (10/64)0.600.919.3886%

Here we report our in-silico Success Rate, the metrics of complex formation, and the structural differences of CC within two binding contexts. Avg. metrics are calculated for all successful designs. These preliminary results strongly indicate that our chameleon-based approach can generalize to full-length multistate protein design. Together with open-source all-atom models like RFD4-proteina, we are now even closer to brigde the gap between chameleon pairs and fold-switching proteins. The visualization of our designed fold-swtich proteins are shown in link: 5 selected successful fold-switch designs

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

We sincerely thank all reviewers and the AC for their time and constructive feedback. We are especially grateful to Reviewer 1SxP for engaging with our rebuttal and confirming that it addressed most of their concerns. The reviewers' suggestions motivated substantial new experiments.

New results

  1. Key-component ablation (EMSR, 1SxP). Removing atom-level communication drops Success to 0%; removing token-level communication drops it from 14.3% to 8.1%; removing the shared-sequence loss drops it to 1.0%. All components contribute.
  2. Detailed unconditional-generation analysis (soWw). Comparison with Pallatom and La-Proteina on inference speed, structural novelty, secondary-structure composition, and ProteinMPNN designability, with an analysis of the diversity gap.
  3. New baseline and base-model study (soWw). We adapted SwitchCraft to chameleon design (27.8% vs 25.6% Success, ~6× slower), and fine-tuned a second base model, RFD4-Proteina (Success 25.6% → 49.5%).
  4. Input-leakage and overlap analysis (1SxP). The model receives only lengths, shared-segment position, residue correspondence, and local SS labels. None of 2,000 generated chains occurs verbatim in the training data (median identity to the closest training window: 27%).

Our goal is multistate design, not state-of-the-art monomer generation

Several concerns, notably from Reviewer soWw, centered on AtomDiff-X's unconditional monomer performance relative to Pallatom and La-Proteina. We want to be clear about what this work claims. AtomDiff-X is a fully trainable all-atom platform that we built to develop and test chameleon-specific components (paired supervision, cross-branch communication, shared-sequence loss). It is nonetheless a competent platform: among 8 all-atom models it ranks 2nd in quality and 3rd in diversity, its ProteinMPNN designability matches La-Proteina, and 14 of 24 tested designs were experimentally confirmed.

What we claim is that paired chameleon supervision enables de novo chameleon design. The same recipe applied to RFD4-Proteina, a different latent-space all-atom model, nearly doubles Success and yields the fold-switch designs below.

Highlight: binder-driven de novo fold-switch design

Reviewers EMSR and 1SxP rightly noted that shared-segment chameleon pairs differ from full-length multistate design. To take a concrete step beyond it, we fine-tuned RFD4-Proteina, a base model with PPI-design capability, on our Chameleon dataset. The resulting RFD4-Chameleon designs a full-length protein that adopts different folds when bound to two different de novo binders.

These results are preliminary and in silico. Still, since fold-switch design is very hard and paired training data are scarce, they suggest that a data-driven path from natural chameleon sequences to multistate de novo design is feasible.

We have worked to address every concern raised, and we would be grateful if Reviewers soWw and EMSR could read our responses and tell us whether any concerns remain. In light of the new ablations, baselines, analyses, and the fold-switch results above, we respectfully ask the Area Chair and reviewers to take these updates into account in their final assessment. We are happy to answer any further questions before the discussion closes.

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.