acceptodds
Under review as a conference paper at ICLR 2027

Domain-Aware Protein Tokenization for Structure Generation

Abstract

De novo protein backbone generation is a central task in protein design, and discrete structural representations provide a multimodal generative paradigm for this task by enabling the unified modeling of protein structure, sequence, and function. However, existing methods focus primarily on residue-level local geometry and rarely use evolutionarily conserved substructures such as domains as training signals, limiting their ability to capture the modular organization of proteins. We introduce DPAR, a domain-aware discrete framework that encodes backbone geometry and domain segmentation as a sequence of discrete symbols. DPAR combines region-aware tokenization that captures both domain and global content with autoregressive generation of structure tokens and domain boundaries, using the resulting layout to guide segment-wise conditioning of a flow-matching decoder. We validate DPAR on reconstruction and generation tasks and demonstrate that it matches or outperforms existing models based on residue-level protein structure tokenizers. We show how domain layouts enable segment-wise conditioning during inference, which boosts backbone designability. Together, these properties establish DPAR as a promising framework for domain-aware discrete protein structure generation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
protein structure generation, protein tokenization, flow matching

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_YMJoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

Existing protein generation models primarily focus on residue-level local geometric features, with limited attention to protein domains, which represent an intermediate structural scale between local geometry and global topology. The lack of explicit domain-level constraints may lead to the accumulation of local prediction errors during inference, potentially degrading backbone generation quality and causing substantial structural errors in multidomain protein reconstruction and design. To address these limitations, the authors propose DPAR (Domain-aware Protein Tokenization), a discrete protein backbone generation framework that explicitly incorporates domain information and domain boundaries to improve backbone generation quality.

Critical strengths
  1. Novelty: As stated by the authors, DPAR is “the first tokenization model to explicitly incorporate domain information into discrete token-based generation.” This represents a meaningful innovation in discrete protein structure representation and generation.
  2. Significance: Existing protein structure generation models, particularly discrete token-based protein language models (PLMs), primarily focus on residue-level local geometric features and lack explicit modeling of domains as intermediate-scale structural units. This may limit their ability to capture multiscale structural information. DPAR provides a promising approach to addressing this limitation and offers a new perspective on improving protein structure representation and backbone generation quality.
Critical weaknesses
  1. Insufficient ablation studies on domain-aware modeling. The reviewer suggests directly comparing the full DPAR model with variants that (i) remove the domain-aware encoder, (ii) remove domain boundary prediction, and (iii) replace actual domain boundaries with random boundaries. Such experiments would help determine whether the observed performance improvements primarily arise from explicit domain modeling rather than other components of the framework.
  2. Insufficient evaluation on multidomain proteins. Most of the reconstruction and generation experiments presented in the paper focus on single-domain proteins. The reviewer suggests conducting additional experiments specifically on multidomain proteins and reporting within-domain errors, inter-domain placement errors, and designability. These evaluations would provide stronger evidence for the effectiveness of DPAR in reconstructing and generating complex multidomain protein structures.
Questions for authors

Please clarify the potential advantages of DPAR over continuous protein backbone generation methods, such as RFdiffusion? In particular, what unique benefits does the domain-aware discrete-token framework offer compared to continuous generative approaches? Clarifying these differences would help me better assess the broader significance of this work.

3 comments
AuthorsOpenReview ↗

Response to Reviewer YMJo

Thank you for recognizing the novelty and potential of explicitly modeling domains in discrete protein generation. We appreciate your suggestions for strengthening the ablations and multidomain evaluation.

Comment 1: Insufficient ablation studies on domain-aware modeling. The reviewer suggests directly comparing the full DPAR model with variants that.....

Response: Thank you for suggesting these comparisons. We evaluated two supplementary models on CAMEO alongside the DPAR result reported in the paper. All three use a flow-matching decoder.

The control model without domain information encodes the whole chain and removes domain annotations, domain conditioning, and auxiliary domain supervision. The model trained with fixed pseudo-boundaries retains domain-aware conditioning and supervision but replaces the original domain intervals during training, while preserving each chain’s domain count, set of domain lengths, and number of covered residues.

ModelDomain configuration during trainingCAMEO RMSD ↓ (Å)
Control model without domain informationNo domain supervision2.2195
DPAROriginal domain intervals0.9194
Model trained with pseudo-boundariesFixed pseudo-intervals1.6005

Using raw outputs for both supplementary models, the fixed pseudo-boundary model reduced mean RMSD by 0.6190 Å (27.9%) relative to the control without domain information. This suggests that domain partitioning and its associated conditioning and supervision can improve reconstruction even when training uses artificial boundaries. The DPAR result reported in the paper is a further 0.6811 Å lower than the fixed pseudo-boundary result. One possible benefit of the original boundaries is that they group residues from the same folding unit, giving the decoder more coherent information about the geometry within each domain.

AuthorsOpenReview ↗

Response to Reviewer YMJo

Comment 2: … conducting additional experiments specifically on multidomain proteins and reporting within-domain errors, inter-domain placement errors, and designability.

Response: Thank you for suggesting a dedicated evaluation of multidomain proteins. We evaluated how well DPAR preserves domain geometry and arrangement in reconstruction, and how often its generated backbones with multiple domains are designable. In reconstruction, DPAR’s main advantage was preserving the relative arrangement of domains. In generation, 279 of the 316 multidomain candidates identified in the length-only experiment met the designability threshold.

(1) Multidomain reconstruction. To evaluate reconstruction on complete multidomain chains, we selected 150 chains with TED annotations [1] from the AFDB [2] validation set. We used the DPAR weights from the submitted paper and ran Kanzi [3], APT [4], and ESM3 [5] on the same chains. DPAR achieved the lowest whole-chain RMSD of the four methods, at 0.7800 Å.

To understand the difference in whole-chain reconstruction, we measured errors in individual domain geometries and in the relative arrangement of domains. W is the RMSD over all domain residues after aligning each domain independently. J is the RMSD after jointly aligning the same domain residues. We use P = √[max(0, J² − W²)] to measure the additional error from joint alignment.

ModelWhole-chain RMSD ↓ (Å)Intradomain W ↓ (Å)Additional joint-alignment error P ↓ (Å)
DPAR0.78000.58350.3449
Kanzi1.03310.82410.4905
APT2.37342.01220.8726
ESM30.89370.31810.6473

ESM3 achieved the lowest intradomain error W. DPAR achieved the lowest additional joint-alignment error P, 46.7% below ESM3, and a 12.7% lower whole-chain RMSD. This suggests that DPAR’s advantage on these multidomain chains primarily lies in preserving the relative arrangement of domains.

(2) Multidomain generation. We made 100 generation requests at each of five lengths: 128, 160, 192, 224, and 256 residues. Each request specified only the chain length. We applied Merizo [6] to the ESMFold [7] prediction for the first designed sequence and required at least two domains, at least 20 residues per domain, and at least 0.8 total coverage.

ModelDesignability ↑scRMSD ↓ (Å)
DPAR88.29%1.516

Short-chain comparison after refolding. To assess designability under the short-chain free-generation setting of the submitted paper, we analyzed the generated results of DPAR and two baselines from the paper, APT and ESM3. We used Merizo to segment the refolded structure of the first designed sequence. We applied the same criteria to every model, requiring at least 20 residues per domain and at least 0.8 total coverage. Each model’s designability rate uses its own number of multidomain candidates as the denominator. DPAR achieved the highest designability and the lowest mean whole-chain scRMSD. It also had the lowest mean intradomain and interdomain errors, indicating that refolding better preserved its domain geometry and relative domain arrangement.

ModelDesignability among multidomain candidates ↑Mean whole-chain scRMSD ↓ (Å)Mean intradomain error ↓ (Å)Mean interdomain error ↓ (Å)
DPAR97.74%1.0210.7810.625
APT60.53%2.1621.3031.585
ESM346.58%3.3811.9402.558

References

[1] Lau, Andy M., et al. “Exploring structural diversity across the protein universe with The Encyclopedia of Domains.” Science 386(6721) (2024): eadq4946.

[2] Varadi, Mihaly, et al. “AlphaFold Protein Structure Database: massively expanding the structural coverage of protein-sequence space with high-accuracy models.” Nucleic Acids Research 50(D1) (2022): D439–D444.

[3] Dilip, Rohit, Evan Zhang, Ayush Varshney, and David Van Valen. “Flow Autoencoders are Effective Protein Tokenizers.” arXiv:2510.00351 (2025).

[4] Dilip, Rohit, Ayush Varshney, and David Van Valen. “Adaptive Protein Tokenization.” arXiv:2602.06418 (2026).

[5] Hayes, Thomas, et al. “Simulating 500 million years of evolution with a language model.” Science 387(6736) (2025): 850–858.

[6] Lau, Andy M., Shaun M. Kandathil, and David T. Jones. “Merizo: a rapid and accurate protein domain segmentation method using invariant point attention.” Nature Communications 14 (2023): 8445.

[7] Lin, Zeming, et al. “Evolutionary-scale prediction of atomic-level protein structure with a language model.” Science 379(6637) (2023): 1123–1130.

AuthorsOpenReview ↗

Response to Reviewer YMJo

Comment 3: Please clarify the potential advantages of DPAR over continuous protein backbone generation methods, such as RFdiffusion....

Response: Thank you for this question. RFdiffusion [1] generates backbones by denoising continuous 3D residue frames and is designed to use geometric constraints during generation. DPAR provides a discrete representation of backbone geometry and domain boundaries for autoregressive modeling. A flow-matching decoder then maps the generated tokens back to backbone coordinates.

The domain layout is itself a generated output in DPAR. The autoregressive model predicts where domains start and end alongside the structure tokens, so the decoder can use that layout to condition on each domain separately. The token representation also supports different levels of control. A CATH class token specifies a broad structural category without requiring a template, while the structure tokens of a natural domain provide a concrete structural context for extending the chain. Both types of input can be handled by the same autoregressive model.

As discussed in our Introduction, this representation also makes structure easier to incorporate into multimodal protein language models. Backbone geometry, amino acid sequence, and functional information can all enter a model through tokens. Explicit domain boundaries additionally give the model a way to associate information with a particular domain while retaining its place in the larger chain. This makes DPAR a structural component for future joint modeling across protein sequence, structure, and function. [2]

References

[1] Watson, Joseph L., et al. “De novo design of protein structure and function with RFdiffusion.” Nature 620(7976) (2023): 1089–1100.

[2] Hayes, Thomas, et al. “Simulating 500 million years of evolution with a language model.” Science 387(6736) (2025): 850–858.

Reviewer_EVnpOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DPAR, a domain-aware protein structure tokenizer for backbone generation. DPAR incorporates domain annotations into structure encoding, jointly generates discrete structure tokens and domain boundaries using an autoregressive model, and employs domain-guided conditioning for flow-based backbone decoding. Experiments demonstrate improved reconstruction accuracy and strong generation performance relative to existing structure-tokenization approaches.

Critical strengths

Strong empirical results. DPAR achieves competitive or superior reconstruction and generation performance across multiple benchmarks, including substantial improvements in designability and structural consistency.

Biologically motivated representation. Incorporating protein domains as an intermediate level of structural organization is intuitive and potentially useful for multidomain protein generation. The domain-guided conditioning experiments provide evidence that structural segmentation can improve generation quality.

Critical weaknesses

Limited methodological novelty. The central contribution appears to be augmenting existing residue-level structure tokenization with explicit domain information, including domain-specific encoding, boundary tokens, and segment-wise conditioning. While effective, these modifications seem largely incremental relative to established discrete structure tokenizers and autoregressive generation pipelines. More importantly, the experiments do not convincingly isolate whether the improvements arise from the proposed tokenization scheme itself or simply from providing additional domain supervision and conditioning. A comparison with a similarly sized conventional tokenizer augmented with domain annotations would help establish the incremental contribution.

Dependence on external domain annotations and limited scalability analysis. DPAR relies on domain annotations during tokenizer training, which are not freely available for arbitrary protein structures. The paper uses annotations from TED and restricts the representation to at most three selected domains, raising questions about generalization to proteins with many domains, ambiguous domain boundaries, or incomplete annotations. Although domain boundaries are generated autoregressively at inference, this does not eliminate the annotation requirements during training. The practical cost and scalability of obtaining these annotations are insufficiently discussed.

Questions for authors
  1. How much of DPAR's improvement comes from the proposed tokenization architecture versus the additional domain supervision? Would a conventional structure tokenizer with domain-boundary conditioning achieve similar improvements?

  2. How does DPAR scale to proteins with more than three domains or unreliable domain annotations? What is the expected impact of annotation quality and coverage on training scalability and downstream generation?

4 comments
AuthorsOpenReview ↗

Response to Reviewer EVnp

Thank you for recognizing the biological motivation behind DPAR and its performance across the benchmarks. We appreciate your suggestions for clarifying the method’s contribution and practical requirements.

Comment 1: … these modifications seem largely incremental relative to established discrete structure tokenizers and autoregressive generation pipelines.

Response: Thank you for the opportunity to clarify. All three methods condition a continuous decoder on discrete tokens. Kanzi [1] uses residue-level tokens, and APT’s prefixes progressively describe the whole chain [2]. DPAR gives residues within a domain an independent encoding context and includes boundary markers in the structure token sequence. The autoregressive model can jointly predict structure and boundaries, and the decoder can use the generated boundaries to select conditioning prefixes for individual domains.

Comment 2: How much of DPAR's improvement comes from the proposed tokenization architecture versus the additional domain supervision? …

Response: Thank you for raising this point. Both full DPAR and the control model without domain information use a flow-matching decoder. The control encodes the whole chain and removes domain conditioning and auxiliary domain supervision. DPAR has lower reconstruction error on CAMEO, suggesting that domain-aware encoding, domain conditioning, and auxiliary domain supervision together improve reconstruction accuracy.

ModelCAMEO Mean RMSD ↓ (Å)
DPAR0.9194
Control model without domain information2.2195

In generation, providing the control model with the full chain token sequence yields a lower mean best scRMSD than providing a chain prefix. The mean still exceeds 12 Å, indicating substantial differences between the generated and refolded backbones.

Decoder conditioningMean best scRMSD ↓ (Å)
Whole-chain prefix15.45
Full token sequence12.44

References

[1] Dilip, Rohit, Evan Zhang, Ayush Varshney, and David Van Valen. “Flow Autoencoders are Effective Protein Tokenizers.” arXiv:2510.00351 (2025).

[2] Dilip, Rohit, Ayush Varshney, and David Van Valen. “Adaptive Protein Tokenization.” arXiv:2602.06418 (2026).

AuthorsOpenReview ↗

Response to Reviewer EVnp

Comment 3: Although domain boundaries are generated autoregressively at inference, this does not eliminate the annotation requirements during training. How does DPAR scale to proteins with more than three domains or unreliable domain annotations... What is the expected impact of annotation quality and coverage on training scalability and downstream generation...

Response: Thank you for the opportunity to clarify. Training requires domain annotations. We use TED [1] annotations to define the regions that the tokenizer encodes independently. The autoregressive model learns to predict the corresponding structure tokens and domain start and end markers. During free generation, the autoregressive model samples these markers, so no target structure annotations are required.

Domain coverage and the three-domain limit. We chose the three-domain limit based on the coverage of the training data. Under the annotation and preprocessing rules in Appendix C.3, 394,906 of the 400,936 training chains retained zero to two domains, and only 6,030 reached the three-domain limit. Thus, at most approximately 1.5% of training chains could have fewer explicitly modeled domains because of this limit. For inputs with more than three annotated domains, DPAR selects up to three for independent encoding. The remaining residues are assigned to regions outside the selected domains and reconstructed together with them.

Selected domains per chainChains
029,641
1284,758
280,507
36,030
Total400,936

Obtaining domain annotations. For training, we reuse the public TED consensus annotations [2] and perform chain ID mapping, overlapping-interval merging, and mask preprocessing offline. The processed annotations can be cached and reused across epochs, so we do not rerun Merizo on the training set.

For inputs with structure coordinates and missing or incomplete annotations, pretrained Merizo [3] can predict domain boundaries automatically. Merizo reports mean segmentation times of 0.112 seconds per chain on GPU and 1.095 seconds per chain on CPU on CATH-663, providing a reference for the cost of annotating a single chain.

When reconstruction or structure-conditioned generation needs a partition for an unannotated input, the cost is incurred before encoding. In unconditional generation, the autoregressive model samples the layout directly. Any subsequent use of Merizo to identify candidate domains is part of evaluation.

Predicted boundaries at inference. We use Merizo to partition each input structure and fix that partition before reconstruction and scoring. Residues not assigned to a domain remain in the regions outside the domains. The mean TM-score exceeds 0.94 on CAMEO, CASP14 [4], and CASP15 [5], showing that automatically predicted boundaries support high-fidelity backbone reconstruction on these test sets.

DatasetMean RMSD ↓ (Å)Median RMSD ↓ (Å)Mean TM-score ↑
CAMEO0.9780.8230.9608
CASP141.2660.7580.9641
CASP151.7500.8820.9480

References

[1] Lau, Andy M., et al. “Exploring structural diversity across the protein universe with The Encyclopedia of Domains.” Science 386(6721) (2024): eadq4946.

[2] Lau, Andy, et al. “The Encyclopedia of Domains (TED) structural domains assignments for AlphaFold Database v4.” Zenodo, v5 (2024). [Dataset].

[3] Lau, Andy M., Shaun M. Kandathil, and David T. Jones. “Merizo: a rapid and accurate protein domain segmentation method using invariant point attention.” Nature Communications 14 (2023): 8445.

[4] Kryshtafovych, Andriy, et al. “Critical assessment of methods of protein structure prediction (CASP)—Round XIV.” Proteins 89(12) (2021): 1607–1617.

[5] Kryshtafovych, Andriy, et al. “Critical assessment of methods of protein structure prediction (CASP)—Round XV.” Proteins 91(12) (2023): 1539–1549.

Reviewer_EVnpOpenReview ↗

Reply to Authors

Thanks for the response. The inference results using Merizo as an auto-annotator of domain boundaries are good, but this shifts the reliance of domain annotations to reliance of a high-quality domain annotator, raising the question how much your model is bootstrapping off another deep learning model.

Thank you for raising this concern. Merizo's role in our supplementary reconstruction experiment is to provide domain boundaries for test structures without existing annotations. DPAR then computes the structure tokens and reconstructs the backbone using its own encoder and decoder. During unconditional generation, DPAR jointly predicts domain boundaries and structure tokens. We only apply Merizo after sampling to identify multidomain candidates for evaluation. Thus, free generation proceeds without calling Merizo to supply boundaries.

TED itself incorporates Merizo predictions into its consensus annotations, together with predictions from Chainsaw and UniDoc [1]. Our use of Merizo for structures without existing annotations therefore follows the automated domain-parsing approach underlying TED. TED-based data augmentation also has demonstrated practical utility. For example, Proteina-Complexa uses TED boundaries to construct the Teddymer dataset from AFDB monomers for generative pretraining [2]. These applications support using additional domain annotations as a well-founded and practically useful way to augment AFDB-based training data.

Regarding how much the benefit depends on accurate domain annotations, our previous response also reported a control trained with fixed artificial boundaries. This control retained partition-based conditioning and supervision while replacing the original training boundaries with pseudo-boundaries. It achieved a lower mean CAMEO RMSD than the control without domain information (1.6005 vs. 2.2195 Å). This indicates that, in this training setting, partition-based conditioning and supervision can still provide some benefit even with artificial boundaries. Together with the Merizo reconstruction results, this provides evidence that DPAR can exploit both artificial training partitions and automatically predicted reconstruction boundaries.

References

[1] Lau, Andy M., et al. “Exploring structural diversity across the protein universe with The Encyclopedia of Domains.” Science 386 (2024): eadq4946.

[2] Didi, Kieran, et al. “Scaling Atomistic Protein Binder Design with Generative Pretraining and Test-Time Compute.” ICLR (2026).

Reviewer_yn2YOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Existing protein structure tokenizers primarily focus on residue-level local geometry, without explicitly modeling domain-level structural organization, which may limit structural consistency and generation quality for multidomain proteins. This paper proposes DPAR, which incorporates domain information into discrete protein structure representations. The paper demonstrates competitive reconstruction and generation performance, but does not sufficiently establish the independent contribution of domain-aware tokenization or the practical value of the two-stage generation framework.

Critical strengths
  • Well-motivated and novel domain-aware protein tokenization. Explicitly incorporating domain segmentation into discrete structure representations provides a biologically meaningful extension to existing protein tokenization methods.
  • Competitive reconstruction and generation performance. The method achieves strong RMSD and TM-score results across multiple reconstruction benchmarks. Its unconditional generation designability and scRMSD substantially outperform existing discrete token-based baselines.
Critical weaknesses
  • Insufficient evaluation of multidomain protein generation. Although the paper emphasizes domain-level consistency and inter-domain structural organization, the generation evaluation is primarily limited to short, single-chain proteins. Systematic evaluations of domain count, inter-domain arrangement, and inter-domain consistency are lacking.
  • Insufficient validation of the independent contribution of domain-aware tokenization. The reconstruction improvements may primarily stem from the additional flow-matching decoder rather than the domain-aware structure representations. Controlled comparisons using decoders of comparable capacity, with and without domain-aware tokenization, would help isolate its contribution.
  • Unclear necessity and practical advantages of two-stage generation. DPAR adopts a two-stage pipeline for protein structure generation but does not sufficiently discuss potential error accumulation between stages. This design may also introduce additional challenges for downstream applications such as motif scaffolding and binder design.
Questions for authors
  • Is the flow-matching decoder also used in the reconstruction experiments?
  • What is the independent contribution of domain-aware modeling? Does DPAR genuinely improve multidomain protein generation?
  • One important motivation for discrete representations is the unified modeling of protein sequence, structure, and function. Why does this work focus exclusively on structure modeling?
4 comments
AuthorsOpenReview ↗

Response to Reviewer yn2Y

We are glad the reviewer appreciated the strength of our experimental results and the novelty of our approach. We’d like to address the primary weaknesses the reviewer raised.

Comment:

Comment 1: Is the flow-matching decoder also used in the reconstruction experiments?

Response: Thank you for the opportunity to clarify. Yes. DPAR uses its flow-matching decoder to recover backbone coordinates in both reconstruction and generation. For reconstruction, we encode the input backbone into structure tokens and obtain the domain layout by partitioning the input structure. For generation, the autoregressive model jointly samples the structure tokens and domain layout.

Comment:

Comment 2: What is the independent contribution of domain-aware modeling? Does DPAR genuinely improve multidomain protein generation?

Response: Thank you for raising this point. Both full DPAR and the control model without domain information use a flow-matching decoder. The control encodes the whole chain and removes domain conditioning and domain supervision. DPAR has lower reconstruction error on CAMEO, suggesting that domain-aware encoding and domain supervision together improve reconstruction accuracy.

ModelCAMEO Mean RMSD ↓ (Å)
DPAR0.9194
Control model without domain information2.2195

In generation, providing the control model with the full chain token sequence yields a lower mean best scRMSD than providing a chain prefix. The mean still exceeds 12 Å, indicating substantial differences between the generated and refolded backbones.

Decoder conditioningMean best scRMSD ↓ (Å)
Whole-chain prefix15.45
Full token sequence12.44
AuthorsOpenReview ↗

Response to Reviewer yn2Y

Comment 3: … the generation evaluation is primarily limited to short, single-chain proteins. Systematic evaluations of domain count, inter-domain arrangement, and inter-domain consistency are lacking.

Response: Thank you for this suggestion. We expanded our evaluation to assess generation quality on longer chains and characterize the designability and structural consistency of multidomain candidates.

(1) Longer-chain generation. We compared DPAR with APT [1], DPLM-2 [2], and ESM3 [3] on designability and whole-chain scRMSD for lengths of 128–256 residues. DPAR achieved the highest designability and the lowest mean and median whole-chain scRMSD. Its backbones more often met the designability threshold, and the refolded structures were closer to the generated backbones.

ModelDesignability ↑Mean scRMSD ↓ (Å)
DPAR86.0%1.463
APT68.50%2.153
DPLM-268.00%2.280
ESM329.41%4.944

(2) Multidomain generation. To measure how often free generation produces designable multidomain candidates, we made 100 generation requests at each of five lengths: 128, 160, 192, 224, and 256 residues. Each request specified only the chain length. We applied Merizo [4] to the ESMFold [5] prediction for the first designed sequence and required at least two domains, at least 20 residues per domain, and at least 0.8 total coverage.

ModelMultidomain Designability ↑scRMSD ↓ (Å)
DPAR88.29%1.516

To compare multidomain candidates under the short-chain free-generation setting of the submitted paper, we analyzed the generated results of DPAR and two baselines, APT and ESM3. We used Merizo to segment the refolded structure of the first designed sequence. We applied the same criteria to every model, requiring at least 20 residues per domain and at least 0.8 total coverage. Each model’s designability rate uses its own number of multidomain candidates as the denominator. DPAR achieved the highest designability and the lowest mean whole-chain scRMSD. It also had the lowest mean intradomain and interdomain errors, indicating that refolding better preserved its domain geometry and relative domain arrangement.

ModelDesignability among multidomain candidates ↑Mean whole-chain scRMSD ↓ (Å)Mean intradomain error ↓ (Å)Mean interdomain error ↓ (Å)
DPAR97.74%1.0210.7810.625
APT60.53%2.1621.3031.585
ESM346.58%3.3811.9402.558

References

[1] Dilip, Rohit, Ayush Varshney, and David Van Valen. “Adaptive Protein Tokenization.” arXiv:2602.06418 (2026).

[2] Wang, Xinyou, et al. “DPLM-2: A Multimodal Diffusion Protein Language Model.” ICLR (2025).

[3] Hayes, Thomas, et al. “Simulating 500 million years of evolution with a language model.” Science 387(6736) (2025): 850–858.

[4] Lau, Andy M., Shaun M. Kandathil, and David T. Jones. “Merizo: a rapid and accurate protein domain segmentation method using invariant point attention.” Nature Communications 14 (2023): 8445.

[5] Lin, Zeming, et al. “Evolutionary-scale prediction of atomic-level protein structure with a language model.” Science 379(6637) (2023): 1123–1130.

AuthorsOpenReview ↗

Response to Reviewer yn2Y

Comment 4: … does not sufficiently discuss potential error accumulation between stages. This design may also introduce additional challenges for downstream applications such as motif scaffolding and binder design.

Response: Thank you for giving us the opportunity to clarify how the two stages interact. A sampled structure token can provide misleading geometric cues to the flow-matching decoder. DPAR limits this source of error in two ways. It selects which token contents to pass to the decoder and then gradually reduces the influence of those contents during decoding. Both operations preserve the residue positions and the sampled domain layout.

Selecting token conditions before decoding. During training, we retain a random prefix within each annotated domain and mask all structure token content with 5% probability. The decoder learns to recover the entire backbone even when some or all token content is missing. At inference, segment-wise entropy selection retains a prefix in each region. We can also mask the token content of particular regions so that any misleading cues from those tokens do not enter the decoder. In the single-domain ablation, masking the pre-domain token content while retaining the selected domain prefix gave the best results. Designability rose from 79.6% to 97.1%, and scRMSD fell from 1.472 to 0.803 Å. Random and intradomain dropout retain the same number of tokens, which allows a comparison at the same token budget.

Conditioning strategyMean retained token countDesignability ↑scRMSD ↓ (Å)scTM ↑
No additional dropout19.7 ± 7.679.6%1.4720.897
Before-domain token dropout12.7 ± 4.897.1%0.8030.951
After-domain token dropout18.9 ± 7.179.6%1.4520.898
Intradomain token dropout7.8 ± 6.583.9%2.0650.885
Random token dropout7.8 ± 6.581.0%2.0480.881

Reducing token guidance during decoding. The selected tokens guide the early decoding steps. As the backbone takes shape, we reduce the weight on the difference between the decoder predictions with and without token content. The two predictions use the same residue positions and domain layout. This lets the learned decoder prior play a larger role in refining the coordinates and reduces the influence of errors in the retained tokens. With the selected tokens held fixed, α = 0.25 increased designability by 37.6 percentage points and reduced scRMSD by 1.867 Å relative to constant conditioning.

Conditioning scheduleDesignability ↑scRMSD ↓ (Å)scTM ↑
Constant conditioning58.3%2.6990.850
α = 1.0072.7%1.8920.892
α = 0.5088.3%1.3740.925
α = 0.3593.3%1.0520.941
α = 0.2595.9%0.8320.951

Motif scaffolding. To assess a capability relevant to motif scaffolding, we tested whether DPAR could preserve a complete natural domain while extending the chain. We provide the natural domain’s tokens as conditioning and generate the C-terminal extension. In all three examples, the conditioned region in the refolded structure retained the input domain’s CATH H classification [1]. The table reports the domain RMSDs.

Natural domain sourceWhole-chain scRMSD ↓ (Å)Domain RMSD ↓ (Å)CATH H classification shared by the input and generated domains
2l49_B0.8331.4981.10.260.40
2f2e_A1.4861.8911.10.10.10
6fgj_A1.8811.8683.30.460.10

Binder design limitations. Binder design for a specified target was outside the scope of our evaluation. In the revised limitations section, we will state that it remains to be tested whether the current domain-level conditioning can produce the intended binding interface under target constraints. The current approach covers a subset of design tasks. Future models could improve performance by integrating information at multiple levels.

References

[1] Sillitoe, Ian, et al. “CATH: increased structural coverage of functional space.” Nucleic Acids Research 49(D1) (2021): D266–D273.

AuthorsOpenReview ↗

Response to Reviewer yn2Y

Comment 5: … Why does this work focus exclusively on structure modeling?

Response: Thank you for this question. We agree that a major advantage of discrete representations is that they make it possible to model protein sequence, structure, and function jointly, as illustrated by ESM3 [1]. In sequence-first approaches such as ProGen [2], functional information guides amino acid sequence generation, and structures can then be predicted from the generated sequences.

We want to explore a different starting point for function-guided protein generation. In the modeling approach we envision, function and structure are both represented explicitly. A functional annotation associated with a domain would specify a desired role, while the structural representation would describe that domain’s geometry and its arrangement relative to other domains. Linking these two kinds of information at the domain level would let us ask whether functional information about a particular domain can guide the generation of the backbone that contains it.

DPAR takes the first step on the structural side. Before adding functional information, we wanted to establish whether a discrete representation could capture both the geometry within domains and their arrangement along the chain, and whether a generative model could use that representation to produce new backbones. That is why the present study focuses on structure. Our experiments use structural domain annotations. In future work, we hope to follow the reviewer’s suggestion and develop a unified multimodal model of protein sequence, structure, and function.

References

[1] Hayes, Thomas, et al. “Simulating 500 million years of evolution with a language model.” Science 387(6736) (2025): 850–858.

[2] Madani, Ali, et al. “Large language models generate functional protein sequences across diverse families.” Nature Biotechnology 41 (2023): 1099–1106.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Response to Evaluation on multidomain proteins

Thank you for your suggestion. We added direct measurements of domain arrangement and generation success rates over all requests.

Direct measurement of domain arrangement. On the same 150 multidomain chains, we measured deviations from the reference structures in centroid distances and relative rotations. These metrics quantify how well reconstruction preserves the spacing and relative orientation of domains, respectively.

ModelMean centroid-distance error ↓ (Å)Mean relative rotation error ↓ (°)
DPAR0.3142.64
Kanzi0.4663.58
APT0.9047.12
ESM30.4675.60

DPAR achieved the lowest error on both metrics, supporting more accurate preservation of inter-domain distances and relative orientations, consistent with its lower P.

Joint success over all 500 requests. Evaluation is complete for all four models. Joint success is the number of candidates that are both multidomain and designable divided by all 500 requests, including failed requests.

ModelMultidomain candidates (share of all requests)Multidomain and designableJoint success rate over all requests ↑Designability among candidates ↑
DPAR316 (63.2%)27955.80%88.29%
APT291 (58.2%)19639.20%67.35%
DPLM-2306 (61.2%)19238.40%62.75%
ESM3266 (53.2%)489.60%18.05%

DPAR achieved a joint success rate of 55.80%, exceeding APT and DPLM-2 by 16.6 and 17.4 percentage points, respectively. This measure accounts for both how often a model produces multidomain candidates and how often those candidates are designable, enabling comparison over the same number of requests.

We hope these additional experiments clarify DPAR's ability to generalize to proteins with more complex domain organization.

AuthorsOpenReview ↗

Response to Ablation studies on domain-aware modeling

Thank you for pointing out the need to isolate the contributions of the individual domain-aware components. We added two CAMEO reconstruction controls and clarify how we construct the domain partitions.

Training domain partitions and pseudo-boundaries. We map TED annotations to residue intervals in each chain, merge highly overlapping intervals, order the intervals from the N to the C terminus, and select at most three domains per chain. Remaining residues form the residual region. The pseudo-boundary training run uses the same domain-aware architecture and supervision. For each chain, we construct non-overlapping pseudo-intervals offline at different positions while matching the number of domains, the multiset of domain lengths, and the total number of covered residues. The pseudo-partition is fixed throughout training. This matches the count, sizes, and coverage of domains across the two partitioning schemes.

Component ablations under a fixed domain partition. DPAR first encodes the full chain, then replaces the representations of residues within each domain with the outputs of a separate encoding of that domain. The chain-only control uses the same domain boundaries, decoder conditioning, and auxiliary domain-crop supervision, and uses the whole-chain encoder outputs for every residue. The control without the domain-crop loss retains independent domain encoding and the original domain boundaries and removes only the auxiliary domain-crop loss.

ModelMean RMSD ↓ (Å)Mean TM-score ↑
Control without domain information2.21950.8291
Chain-only encoder with domain supervision1.99240.8562
Control without the domain-crop loss1.98850.8448
DPAR (reported in the paper)0.91940.946
Model trained with fixed pseudo-boundaries1.60050.8896

The results show where domain information can help. Compared with the control without domain information, the chain-only encoder with domain conditioning and auxiliary supervision reduces mean RMSD by 0.2271 Å. This suggests that explicit domain information is useful even with whole-chain encoding. The control that retains independent domain encoding and domain conditioning but removes the domain-crop loss still lowers mean RMSD by 0.2310 Å, suggesting that this combination remains useful with the auxiliary loss disabled. The model trained with fixed pseudo-boundaries reduces mean RMSD by 0.6190 Å, showing that a fixed artificial partition can also provide useful conditioning and supervision for reconstruction.

AuthorsOpenReview ↗

General Response & Summary of Rebuttal

Dear Reviewers, ACs and SACs,

We sincerely thank all reviewers for their careful evaluation and constructive suggestions. We have provided additional experiments and clarifications to address the concerns raised during the review and discussion. Our main responses include:

  • Added component ablations of independent domain encoding and domain-crop supervision, together with a fixed pseudo-boundary control, to examine the contributions of domain-aware components and partition-based modeling, as suggested by all reviewers.
  • Expanded complete-chain reconstruction comparisons across different domain counts and added direct measurements of inter-domain distances and relative orientations, addressing the multidomain evaluation concerns raised by Reviewers YMJo and yn2Y and the request for direct geometric metrics from Reviewer YMJo.
  • Added multidomain generation comparisons with APT, DPLM-2, and ESM3, reporting candidate counts, designability, success relative to all generation requests, and analyses by chain length and domain count, as suggested by Reviewers YMJo and yn2Y.
  • Added a separate evaluation of longer-chain generation, comparing designability and structural self-consistency with existing baselines, as suggested by Reviewer yn2Y.
  • Added reconstruction experiments using automatically predicted Merizo boundaries and clarified the roles of annotations in training, reconstruction, and generation. We also discussed TED consensus annotations and the use of additional domain annotations to augment AFDB training data, addressing annotation availability and annotator dependence raised by Reviewer EVnp.
  • Added an audit of training annotation coverage and clarified the domain-selection limit and treatment of residues outside selected domains, addressing coverage and scalability concerns raised by Reviewer EVnp.
  • Clarified DPAR’s encoding architecture and decoder usage, its contribution relative to existing discrete tokenizers and continuous backbone generation methods, and its structural focus as a basis for future multimodal modeling, addressing all reviewers.
  • Explained how token dropout and conditional annealing address error propagation between autoregressive sampling and coordinate decoding, supported by the ablations in the submitted paper, as requested by Reviewer yn2Y.
  • Clarified the natural-domain extension experiments and their relevance to scaffolding, and discussed target-conditioned binder design as a direction for further evaluation, as requested by Reviewer yn2Y.

We again thank the reviewers, ACs, and SACs for their time and thoughtful feedback. These discussions have helped us clarify DPAR's contribution and strengthen the evaluation of its multidomain modeling capabilities. We welcome any further questions during the discussion phase.

Best regards,

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.