acceptodds
Under review as a conference paper at ICLR 2027

Ligand-conditioned Discrete Diffusion Protein Co-design Model

Abstract

Proteins perform their biological functions through three-dimensional structures encoded by amino acid sequences, and ligand-binding protein co-design requires models that generate sequence–structure compatible proteins under explicit ligand constraints. Although continuous diffusion and flow-based models have enabled ligand-aware protein design in coordinate or latent feature spaces, existing discrete diffusion protein language models mainly operate over sequence or structure tokens without direct small-molecule conditioning. We introduce ProtLiD, a Protein Ligand-conditioned Discrete Diffusion model for protein sequence–structure co-design. ProtLiD jointly generates amino-acid sequence and discrete structure tokens while incorporating ligand chemical and geometric information through geometry-aware cross-attention. Trained on over one million ligand–protein complexes, ProtLiD extends masked discrete diffusion from general sequence–structure generation to ligand-aware functional protein design. For decoding, we combine MDLM-based stochastic candidate selection with confidence-margin ranking under a geometric retention budget, committing argmax predictions at retained positions and deferring the remaining proposals. Experimentally, ProtLiD improves global fold confidence over Complexa in ligand-conditioned whole-protein design, increasing TM-score from to and pLDDT from to . In ligand-binding pocket co-design, ProtLiD reduces active-site BB-RMSD from Å for FAIR and Å for PocketGen to Å, and improves ligand-aware combined pass rates over PocketGen from to and from to under increasingly stringent docking thresholds. These results demonstrate the potential of ligand-conditioned discrete diffusion as an effective token-space framework for functional protein co-design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Protein design, Discrete diffusion, Sequence–structure co-design, Ligand conditioning, Protein–ligand interactions, Generative models

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_C8bGOpenReview ↗
Rating
1
Confidence
4(+2)
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a ligand conditioned diffusion language model for co-designing the protein sequence and structure. This motivation is too natural (transferred from the continuous diffusion model), to the point where it feels lacking in novelty. But the experiment seem to show strong performance compared to previous work.

Critical strengths
  1. The experiments in this paper are somewhat rich and the claimed improvement is somewhat significant.
  2. The proposed innovation is very natural.
Critical weaknesses
  1. The writing of this paper is very poor: Abbreviations without any explanation exist in the abstract; the content of the introduction is like Related Work and the Related Work is like Preliminaries. I understand that ligand-conditioning is important for DISCRETE DIFFUSION IN PROTEIN DESIGN and there is a gap in this community, but it is hard for me to extract from the introduction what innovation or changes this paper has made for it. The summary of contributions is very abrupt and difficult to connect with the previous paragraphs.
  2. Technological innovation seems to be limited, and the most important thing is to contribute new datasets. I can hardly tell from the introduction what specific innovation this paper has made to fill the gap.
Questions for authors
  1. Please explain the abbreviation in abstract: MDLM.
  2. Can you explain in detail your algorithm or innovation details?
  3. Why is your standard deviation larger than the mean in Tables 1 and 3? Does this make sense?
10 comments
AuthorsOpenReview ↗

Clarifying methodological contributions and statistical reporting, and requesting concrete technical support for the novelty assessment

We disagree with the basis of your assessment of novelty and informativeness. You acknowledge an important capability gap and strong experimental results, yet characterize the contribution as limited because the motivation is “too natural.” That characterization does not explain which methodological additions are already established or why the demonstrated extension lacks scientific value. We address your concerns directly below.

1. The novelty criticism requires a concrete technical comparison. Our paper addresses ligand-conditioned joint sequence–structure generation within discrete diffusion protein modeling. Assessing this contribution requires examining the capabilities of prior discrete models, the conditioning mechanism we introduce, and the resulting experimental evidence.

Your review does not identify the prior methods underlying your novelty judgment or explain which specific additions you consider insufficient. A research objective can be intuitive while its implementation and empirical investigation remain meaningful contributions. The naturalness of the motivation does not, by itself, establish that the contribution is limited.

Please specify the relevant prior work and explain how it supports your assessment of our methodological contribution.

2. The contribution extends beyond dataset construction. Your characterization of the dataset as the principal contribution does not address the conditioning mechanism described in Section 3.2 and Appendix B.

ProtLiD² jointly denoises amino-acid and structure tokens under explicit ligand conditioning. Ligand chemical representations and coordinate embeddings enter the denoising network through geometry-aware cross-attention. Protein-token hidden states produce learned proxy coordinates; distances between those proxies and ligand atoms are transformed into radial-basis-function features and head-specific attention biases. This mechanism allows ligand geometry to influence both sequence and structure predictions throughout denoising.

The dataset supports learning this conditional model. The model formulation, conditioning mechanism, and empirical investigation are additional contributions. For example, active-site backbone self-consistency RMSD is 1.97 Å for ProtLiD², compared with 3.46 Å for FAIR and 3.40 Å for PocketGen. These are computational self-consistency results, and we interpret them within that scope.

We explicitly acknowledge the existing tokenizer, ligand encoder, and diffusion formulation on which we build. Assessing the contribution requires evaluating what our extension enables and demonstrates; identifying its existing components is not sufficient.

3. Presentation concerns and methodological novelty require separate assessments. MDLM denotes a masked diffusion language model. Its formulation appears in Section 2.2. We agree that the abbreviation should also be expanded at its first occurrence in the abstract, and we will make that correction.

We will also streamline the introduction and separate mathematical preliminaries from related work. These changes will make the connection between the gap, the conditioning mechanism, and the experimental evidence more explicit.

However, difficulty extracting a contribution from the introduction does not establish that the contribution is technically limited. Your review repeatedly makes the former observation without providing the technical comparison needed to substantiate the latter. The method section and appendix describe the proposed mechanisms; please identify which of those mechanisms you consider insufficient and explain why.

4. A standard deviation larger than the mean is mathematically valid. In response to your statistical question: yes, this can occur for nonnegative measurements. There is no requirement that their standard deviation be smaller than their mean. For example, 1,1,1,1,111,1,1,1,11 have a mean of 33 and a population standard deviation of 44.

The reported pattern occurs across the compared methods. Table 1 reports global backbone RMSD of 12.07±13.1612.07 ± 13.16 Å for ProtLiD² and 10.35±11.2410.35 ± 11.24 Å for Complexa. In Table 3, global backbone RMSD has a standard deviation exceeding the mean for all three methods.

The large dispersion warrants attention to performance variability, and we will clarify the aggregation procedure. SD exceeding the mean does not, by itself, indicate an invalid calculation or invalidate the comparison.

We welcome specific criticism of the formulation, evaluation, and interpretation. However, we find the clear-rejection recommendation insufficiently substantiated by the reasoning currently provided. Please explain which methodological additions you consider inadequate, which prior comparisons support that judgment, and which experimental conclusions you dispute. These concrete technical points are necessary to make the basis of your assessment assessable.

AuthorsOpenReview ↗

Request for substantive follow-up and justification of the clear-rejection rating

Dear Reviewer C8bG,

We are following up on our response posted on 8 November, which remains unanswered two days later. We provided explanations addressing your questions about MDLM, the conditioning mechanism, and the reported standard deviations. We request a substantive response identifying which concerns remain unresolved.

Please address the following points directly:

  1. What specifically supports your novelty assessment? Please identify the relevant prior work or explain which methodological additions you consider insufficient. Describing the motivation as “too natural” does not explain why the proposed extension and its empirical findings constitute an inadequate contribution.

  2. Which parts of our response are insufficient? If our explanations do not address your questions, please identify the remaining issues precisely. Without this clarification, we cannot determine what further explanation or evidence you consider necessary.

  3. What are the decisive grounds for the rating of 1 (“clear rejection”)? Your review acknowledges the capability gap and empirical improvements and describes the soundness concerns as “minor or fixable.” Novelty and presentation can independently affect the overall assessment, but please explain which specific weaknesses you consider decisive and how our response affects or fails to affect that judgment.

Our request for the technical basis of your assessment remains unanswered. If you maintain the clear-rejection recommendation, please explain why the methodological details, experimental evidence, ablation studies and our response are insufficient to address the concerns underlying that recommendation.

We request a prompt, point-by-point reply during the discussion period so that we have an opportunity to address any remaining substantive concerns. We also ask the Area Chair to help facilitate this follow-up, ensuring that both the authors’ response and the reviewer’s reasoning are available for assessment.

Dear Authors,

The decisive reason for my rating of 1 is the poor presentation. I did not criticize the novelty of this paper; on the contrary, I praised the naturalness of the idea in the Strengths section. I fully recognize the effort every author put into this work, and I am generally tolerant in my reviews—including this one. I sincerely offer the following writing suggestions. If any come across as offensive, please take them as well-intentioned advice.

First, avoid using bold emphasis in large blocks of text. Overuse of emphasis dilutes its effect and obscures your intended message.

Second, in lines 67–76 of the manuscript, you should briefly inform readers of the specific methodological choices you made to introduce ligand conditioning. It does not need to be overly detailed, but leaving it blank forces readers to guess based solely on the phrase “ligand condition.”

Third, please draw clearer boundaries between the Introduction, Related Work, and Preliminaries. Their roles currently overlap, which confuses the narrative.

Finally, while I personally know what MDLM stands for, you cannot assume all readers do. Introducing abbreviations without definition violates standard writing conventions.

If the presentation had been adequate from the outset, I would have assigned a rating of 3 with a confidence of 3. As it stands, however, the necessary revisions are too extensive for the manuscript to be considered publishable in its present form.

Reviewer_3JJROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ProtLiD2, a ligand-conditioned discrete diffusion framework for joint protein sequence-structure generation. It integrates ligand chemical and geometric information through geometry-aware cross-attention and introduces MCM-ReMask decoding to improve generation quality. Trained on over one million protein-ligand complexes, ProtLiD2 demonstrates improved structural consistency, binding-pocket accuracy, and ligand-aware design success rates compared to existing methods

Critical strengths
  • Extends discrete diffusion protein language models to ligand-conditioned sequence-structure co-design through geometry-aware cross-attention.
  • Constructs a dataset of over one million protein-ligand complexes with sequence-based leakage filtering.
  • Demonstrates improved pocket geometry, structural consistency, and ligand-aware pass rates over existing baselines, supported by ablation studies.
Critical weaknesses
  • The framework primarily combines existing components, including GCP-VQVAE, Uni-Mol, masked discrete diffusion, and confidence-based decoding, with relatively limited architectural innovation.
  • Evaluation relies primarily on computational metrics from ESMFold, AlphaFold3, and AutoDock Vina, without experimental verification of protein folding or ligand binding.
  • The framework relies on a frozen structure tokenizer and does not explicitly model full-atom side chains or ligand flexibility, potentially limiting binding accuracy.
Questions for authors

please refer to weakness part

8 comments

We thank the reviewer for recognizing the value of ligand-conditioned sequence–structure co-design and for identifying these limitations. We address these concerns below. Manuscript revisions are marked in red.

[W1]. The framework primarily combines existing components, including GCP-VQVAE, Uni-Mol, masked discrete diffusion, and confidence-based decoding, with relatively limited architectural innovation.

Response. We thank the reviewer for this assessment. We would like to clarify that our central contribution is establishing the feasibility and practical value of ligand-conditioned joint protein sequence–structure generation in discrete token space. The contribution comprises the modeling capability, the training data needed to realize it, and empirical evidence that it can achieve competitive results under a modest main-model training budget. We explain these contributions below.

First, we address a missing capability in discrete protein generation. Prior discrete diffusion models, such as DPLM-2, established sequence–structure co-generation and conditioning on protein sequences, structures, and motifs [1]. To our knowledge, an effective counterpart that jointly generates discrete sequence and structure tokens under explicit small-molecule conditioning had not been established. ProtLiD$^2$ addresses this gap by incorporating ligand chemistry and geometry into joint token denoising. Whether this representation can retain sufficient ligand-relevant information to produce compatible sequences and backbones is an empirical question. The availability of a structure tokenizer, a ligand encoder, and a discrete diffusion model does not establish that their combination will solve this task.

Second, we construct the training data needed to investigate this question. We curated 1,026,766 protein–ligand complexes through integration of multiple data sources, consistent preprocessing, quality filtering, and sequence-based separation from the evaluation set. This required resolving heterogeneous representations and removing unsuitable complexes, including invalid inputs and severe protein–ligand clashes. The resulting dataset provides the training basis for learning the relationship between ligand context and joint sequence–structure tokens. Dataset construction is therefore an enabling contribution to the formulation.

Third, we demonstrate competitive computational performance within a practical training budget. Our 370M-parameter model was trained for 100,000 steps on eight NVIDIA RTX PRO 6000 GPUs over approximately 11 days. For context, Complexa reports training its small-molecule flow model on 48 A100-80GB GPUs, followed by ligand-conditioned fine-tuning on 96 A100-80GB GPUs [2]. Our whole-protein benchmark reports a TM-score of 0.802 versus 0.672 and pLDDT of 73.00 versus 64.55 for ProtLiD$^2$ and Complexa, respectively, while Complexa retains advantages in RMSD and mean docking score. These results demonstrate that competitive ligand-conditioned co-design is achievable in discrete token space with a substantially smaller GPU allocation for main-model training.

Fourth, our component analyses help explain the behavior of the resulting model. With the trained checkpoint and tokenizer fixed, we evaluated 600 designs per setting:

SettingTM-score ↑pLDDT ↑
Full model0.53640.61
Without geometric attention bias0.52039.29
Without Uni-Mol pair-feature bias0.53340.29
Without ligand conditioning0.38526.03

Removing ligand conditioning produces the largest deterioration, while removing the explicit geometric bias produces a smaller decrease. Uni-Mol pair-feature refinement has limited incremental influence in this analysis. These inference-time removals characterize the trained model's dependence on its inputs and components.

The scientific contribution lies in turning a plausible modeling direction into an implemented and evaluated capability. Establishing that capability required the dataset, the ligand-conditioned architecture, training, and systematic evaluation.

References

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

[2] Didi, K., et al. “Scaling Atomistic Protein Binder Design with Generative Pretraining and Test-Time Compute.” arXiv:2603.27950 (2026), Appendix G.3.

[W2]. Evaluation relies primarily on computational metrics from ESMFold, AlphaFold3, and AutoDock Vina, without experimental verification of protein folding or ligand binding.

Response. We agree that wet-lab validation is essential for establishing the biological utility of generated proteins, particularly whether they fold and bind their intended ligands. We regard this as an important next step and acknowledge that the present study provides computational evidence.

First, computational evaluation has established precedent in methodological protein-design studies at major machine-learning conferences. Relevant examples include:

  • DPLM-2 (ICLR 2025) evaluates joint sequence–structure generation, folding, inverse folding, and motif scaffolding computationally. The paper explicitly acknowledges that its motif-scaffolding results have not been experimentally verified [1].
  • FAIR (NeurIPS 2023) evaluates ligand-conditioned pocket co-design using amino-acid recovery, structural RMSD, docking scores, and other computational analyses. Its limitations section identifies wet-lab assessment of designed pockets as future work [2].
  • Multiflow (ICML 2024) evaluates protein sequence–structure co-design through computational self-consistency, diversity, and novelty, including refolding with ESMFold and AlphaFold2 [3].

These studies illustrate how computational evaluation can establish a generative method's capabilities while leaving experimental validation as a subsequent stage. Our use of computational benchmarks follows this methodological approach, with the same distinction between predicted design quality and verified biological function.

Second, our contribution is a machine-learning methodology for biological design, and its evidence should be assessed against that stated objective. ICLR's call for papers explicitly includes generative models, datasets and benchmarks, and applications in biology [4]). Its reviewer guidelines emphasize the objective of the work, its contribution of new knowledge, and whether the evidence supports its claims [5].

A particularly relevant precedent is Proteina-Complexa (ICLR 2026), which evaluates protein and small-molecule binder design using computational structure-prediction confidence and alignment metrics. Appendix A explicitly states that the paper's evaluations are limited to in-silico success metrics and identifies experimental validation of generated binders as future work [6]. This provides a closely related example of a methodological contribution evaluated computationally, with biological validation remaining a separate next step.

Our central question is whether explicit ligand conditioning can support effective joint sequence–structure generation in discrete token space. The model, curated training dataset, computational comparisons, and component analyses address this question. We therefore position the present contribution around computational feasibility, structural self-consistency, and candidate prioritization.

Third, we plan a focused experimental study around a practical application. A rigorous wet-lab campaign requires substantial resources and time for protein production, assay development, and folding and binding measurements. In follow-up work, we plan to select a target ligand with a clear practical need and an appropriate experimental assay, use ProtLiD$^2$ to generate and prioritize a panel of candidate sequences, and then test their expression, folding or stability, and ligand binding with appropriate controls. This application-driven study will assess biological utility and help determine how well the computational screening criteria predict experimental outcomes.

References

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

[2] Zhang, Z., et al. “Full-Atom Protein Pocket Design via Iterative Refinement.” NeurIPS (2023).

[3] Campbell, A., et al. “Generative Flows on Discrete State-Spaces: Enabling Multimodal Flows with Applications to Protein Co-Design.” ICML (2024).

[4] International Conference on Learning Representations. “ICLR 2027 Call for Papers.”

[5] International Conference on Learning Representations. “ICLR 2027 Reviewer Guidelines.”

[6] Didi, K., et al. “Scaling Atomistic Protein Binder Design with Generative Pretraining and Test-Time Compute.” ICLR (2026). See Appendix A, “Limitations and Future Work.”

[W3]. The framework relies on a frozen structure tokenizer and does not explicitly model full-atom side chains or ligand flexibility, potentially limiting binding accuracy.

Response. We thank the reviewer for highlighting these limitations. We agree that explicit side-chain modeling and ligand flexibility could improve atomistic binding accuracy. At the same time, prior work provides methodological and experimental support for generating protein sequences without explicitly representing side-chain conformations.

ProteinMPNN designs sequences from backbone geometry and has been validated experimentally [1]. Similarly, ABACUS-R uses neighboring amino-acid identities without their side-chain conformations, enabling rotamer-free sequence design with experimental validation that includes five X-ray crystal structures [2]. More directly relevant to ligand-conditioned design, LigandMPNN reports that augmenting training with side-chain atoms from 2–4% of protein residues as additional context did not significantly increase sequence recovery (Supplementary Fig. 1b) [3].

These studies support backbone-based representations as a practical foundation for sequence generation. ProtLiD2^2 jointly generates amino-acid identities and backbone-structure tokens under ligand conditioning. The generated amino-acid identities specify side-chain chemistry, although the model does not explicitly generate side-chain rotamers. Freezing GCP-VQVAE fixes the tokenizer parameters and representation, rather than the generated protein backbone: the diffusion model can still generate different ligand-dependent backbone structures.

Ligand flexibility remains a separate limitation. The current generator conditions on a supplied ligand conformation and does not jointly sample ligand conformational changes. Downstream complex prediction and docking provide additional computational assessments, but do not establish that the generator captures these degrees of freedom. We will clarify this scope and discuss explicit side-chain modeling, and ligand-conformer conditioning as future extensions.

References

[1] Dauparas, J., et al. “Robust deep learning–based protein sequence design using ProteinMPNN.” Science, 378, 49–56 (2022).

[2] Liu, Y., et al. “Rotamer-free protein sequence design based on deep learning and self-consistency.” Nature Computational Science, 2, 451–462 (2022).

[3] Dauparas, J., et al. “Atomic context-conditioned protein sequence design using LigandMPNN.” Nature Methods, 22, 717–723 (2025).

Reviewer_5QBUOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ProtLiD$^2$ jointly generates amino-acid sequences and backbone structure tokens conditioned on ligand chemistry and coordinates. It combines a frozen GCP-VQVAE tokenizer, Uni-Mol ligand features and a masked diffusion transformer trained on over one million complexes. The reported benchmarks show better pocket backbone self-consistency than FAIR and PocketGen, and higher whole-protein TM-score and pLDDT than Complexa.

Critical strengths
  • Explicit ligand conditioning usefully extends the joint discrete sequence-structure generation setting studied by DPLM-2. The network components and their connections are clearly described.
  • The dataset combines several sources with explicit quality filters.
  • Active-site backbone self-consistency improves: mean BB-RMSD is 1.97 Angstrom, versus 3.46 Angstrom for FAIR and 3.40 Angstrom for PocketGen (Table 3). The rotation/recanonicalization study and retention sweep provide additional sensitivity evidence.
Critical weaknesses
  • As written, Algorithm 1 retains every sampled non-mask proposal at the reported r=1 setting, so confidence ranking cannot affect the output and the transition reduces to ordinary MDLM sampling. The abstract instead specifies argmax, the implementation and source of the decoding gains remain unclear.
  • Pocket RMSD measures agreement with ESMFold, and the composite binding-related criteria also require that agreement. PocketGen has the better mean Vina score, so higher composite pass rates do not establish better affinity.
  • Table 5 removes components from the same checkpoint and reports no uncertainty, leaving their benefit after retraining unresolved. Table 12 reports rates distinct from the main tables without target counts.
Questions for authors
  • Which sampler produced the results, and how does it compare with vanilla MDLM and an argmax-only variant under equivalent settings?
  • Can whole-protein performance be reproduced using a coordinate frame computed from the ligand alone?
  • Can shuffled-ligand controls test target specificity, and a fixed-backbone LigandMPNN comparison assess the value of backbone redesign?
  • Can you add retrained ablations with paired uncertainty estimates and provide target lists, counts and evaluation settings for Tables 2, 4 and 12?
14 comments

Thank you for the careful review and for recognizing the value of explicit ligand conditioning and the improvement in pocket backbone self-consistency. Your comments identified inconsistencies in our sampler and preprocessing descriptions and limitations in the interpretation of our evaluations. We have corrected these descriptions and narrowed the claims. Manuscript revisions are marked in red.

[Q1] Which sampler produced the results, and how does it compare with vanilla MDLM and an argmax-only variant under equivalent settings?

Response. Thank you for identifying this inconsistency. At the reported setting of r=1r = 1, Algorithm 1 retains all sampled non-mask proposals. Specifically, the number of retained proposals is obtained by multiplying the number of sampled non-mask proposals by rr and rounding down. Consequently, confidence ranking has no effect at this setting, and the update, as written, reduces to vanilla MDLM sampling.

Although the decoding strategy was not listed among the main contributions in the Introduction, we acknowledge that the name MCM-ReMask and its presentation in a dedicated subsection and algorithm could imply a distinct methodological contribution at the reported setting. We appreciate the reviewer drawing attention to this ambiguity.

To resolve this ambiguity, we have revised Section 3.3, now titled "Evaluation and Selection of Unmasking Strategies," to describe our systematic comparison of decoding policies using the same trained ProtLiD$^2$ checkpoint. To examine deferred unmasking, we augment vanilla MDLM sampling with a retention parameter r∈(0,1]r \in (0,1], yielding MDLM-R, as specified in Algorithm 1. This parameter provides a practical control over proposal acceptance rather than a separate methodological contribution. At reverse step jj, let TjT_j denote the masked positions receiving non-mask proposals. MDLM-R retains the kj=⌊r∣Tj∣⌋k_j = \lfloor r |T_j| \rfloor proposals with the largest top-1/top-2 probability margins, preserving their sampled token values. Unselected positions remain masked for subsequent resampling, while previously visible tokens remain unchanged. Thus, r<1r < 1 defers some updates as the surrounding sequence–structure context develops. When r=1r = 1, all sampled non-mask proposals are accepted, confidence ranking has no effect, and the update recovers the standard MDLM reverse transition. In the reported comparison, vanilla MDLM achieves the strongest mean backbone self-consistency at most evaluated protein lengths, motivating its adoption as our default unmasking procedure.

We have updated the method labels in Figure 2 and Table 7 to MDLM, consistent with the revised description of the reported r=1r = 1 configuration. We retain Table 6 as an empirical sensitivity analysis of MDLM-R. Among the evaluated retention settings, r=1r = 1 achieves the highest mean TM-score and lowest mean backbone RMSDs, whereas r=0.5r = 0.5 yields the highest mean pLDDT. These results show a trade-off between backbone self-consistency and prediction confidence, without establishing a consistent benefit from deferred unmasking over vanilla MDLM. Accordingly, we treat retention as an optional practical control. The revised abstract identifies MDLM as the sampling procedure, while Algorithm 1 explicitly distinguishes proposal sampling, confidence-based retention, and assignment of the original sampled token values.

The existing unmasking-strategy comparison also includes LLaDA-ReMask, an argmax-based decoding baseline evaluated under the shared-checkpoint protocol. In our implementation, this strategy selects the highest-probability token at each currently masked position and retains predictions according to their confidence, leaving lower-confidence positions masked for subsequent steps. It therefore combines argmax-based token selection with confidence-based retention and can be viewed as an argmax-only variant.

[Q2] Can whole-protein performance be reproduced using a coordinate frame computed from the ligand alone?

Response. Thank you for raising this question. The 370M-parameter model reported in this manuscript uses different coordinate-centering procedures during training and inference, a distinction that was not clearly described in the manuscript. During training, both protein and ligand coordinates were translated using the protein C-alpha centroid. During whole-protein inference, the reference protein structure is unavailable, so the ligand coordinates were instead centered using the ligand’s own centroid. Although both procedures used a ligand-derived PCA rotation, their translation origins differed.

We acknowledge that this mismatch could affect generation performance. We will revise the manuscript to describe the training and inference procedures accurately and explicitly discuss this limitation.

In our newer 640M-parameter model, which is still being trained, we use consistent preprocessing across training and inference. During training, both protein and ligand coordinates are centered at the conditioning ligand’s centroid and transformed using the same ligand-derived PCA rotation, preserving their relative geometry. The same ligand preprocessing is used during inference. Consequently, both the origin and orientation of the coordinate frame are determined entirely from the ligand in both stages. Once training is complete, we will evaluate its whole-protein generation performance under the same benchmark protocol and compare it with the reported 370M-model results.

The reported whole-protein results already use a coordinate frame computed from the ligand alone at inference. The experimental comparisons and reported performance can be reproduced using the same ligand-only preprocessing, trained checkpoint, and evaluation settings.

[Q3] Can shuffled-ligand controls test target specificity, and a fixed-backbone LigandMPNN comparison assess the value of backbone redesign?

Response. Thank you for suggesting shuffled-ligand controls to assess target specificity. We report an exploratory analysis of ten generated designs spanning 130–845 residues. All designs satisfied ESMFold pLDDT > 60, backbone RMSD < 10 Å, and TM-score > 0.6.

For each designed sequence, we compared AF3 predictions obtained with its intended ligand against predictions obtained with a different ligand assigned through a random permutation. The protein sequence, cached MSAs, and AF3 settings were identical within each pair.

For these ten examples, mean ipTM was 0.834 with the intended ligand and 0.709 with the shuffled ligand, corresponding to a mean difference of 0.125. Intended ligands achieved higher ipTM for eight designs, while two designs were tied. Differences in the table are calculated as intended-ligand ipTM minus shuffled-ligand ipTM.

TargetLengthIntended-ligand ipTMShuffled-ligand ipTMDifference
7rtc_A_13310.800.52+0.28
1ua3_A_14960.970.71+0.26
5l6v_A_18450.960.75+0.21
2r1r_A_17330.800.62+0.18
3qcp_A_14100.950.79+0.16
6pe3_A_13720.890.75+0.14
4m1j_A_17130.940.93+0.01
5bvb_A_11300.620.61+0.01
4xph_A_15340.830.830.00
5eno_A_15770.580.580.00

Within these ten selected examples, eight designs achieved higher ipTM with the intended ligand, whereas two had equal reported scores under the two conditions. For the latter designs, this comparison provides no evidence of a preference for the intended ligand under AF3 interface-confidence scoring. However, ipTM measures confidence in the predicted interface rather than binding affinity or selectivity, so equal scores should not be interpreted as proof that a protein lacks target specificity.

We agree that target specificity is an important objective in protein design. The present work focuses on training a base model for ligand-conditioned protein sequence and structure generation, without explicitly optimizing binding affinity, target specificity, thermal stability, or developability. We acknowledge that generating structurally plausible designs does not guarantee these functional and practical properties. In the next stage, we plan to investigate reinforcement learning (RL) post-training to incorporate these properties as optimization objectives.

Thank you for suggesting the comparison with LigandMPNN. Currently, our model was trained for ligand-conditioned joint sequence-structure generation. In principle, its sequence-structure token representation could be adapted to inverse folding by conditioning on backbone structure tokens and ligand information while generating sequence tokens. Similarly, it could be adapted to structure prediction by conditioning on sequence tokens while generating structure tokens. However, the training used in this study did not explicitly include either conditional task. We therefore do not claim that the current model provides validated fixed-backbone inverse-folding capabilities comparable to LigandMPNN. Extending the model to these conditional tasks and conducting matched comparisons with fixed-backbone methods are directions for future work.

Comments on the paper

5 comments
AuthorsOpenReview ↗

Request for a Formal Apology and Clarification of the Confidence Increase

Dear Reviewer C8bG,

We ask you to acknowledge the personal remarks in your November 18 comment and formally apologize to us for them. You identified presentation as the decisive reason for clear rejection. After revising the manuscript in response to all four of your suggestions, we asked which writing concern remained. That was a legitimate question about the basis for your recommendation.

You replied, “I am a reviewer, not your advisor,” and “Reviewing is voluntary; I am not an employee you hired.” You then wrote, “Perhaps your paper would be more suited to being reviewed by AI reviewers.” We respect the time you volunteered and accept responsibility for the incorrect heading in our earlier response, which we corrected. The remarks about employment and AI reviewers nevertheless turned a discussion about the manuscript into a personal exchange. We consider them dismissive and inappropriate.

You also wrote, “In light of the above, I will increase my confidence level to 4.” The review form defines level 4 as “I have substantial direct research experience on this topic.” Please explain the basis for changing your confidence from 2 to 4 under that criterion. Dissatisfaction with our correspondence provides no basis for changing an assessment of your research experience.

Sincerely,
The Authors of Submission 27634

AuthorsOpenReview ↗

Request for clarification on the revised manuscript and confidence change, and for a formal apology

Dear Reviewer C8bG,

We ask you to address three points arising from your November 18 comment.

  1. Which passage in the revised manuscript remains unclear? You wrote that “the revised manuscript remains poorly written.” We revised the paper in response to each of your four presentation suggestions. Please identify one representative passage by section or line number, explain what a reader would struggle to understand, and indicate how that issue bears on your clear-rejection recommendation. A concrete example would help us improve the paper and help the Area Chair assess your reasoning.

  2. What was the basis for changing your confidence from 2 to 4? After discussing our correspondence, you wrote, “In light of the above, I will increase my confidence level to 4.” The review form defines level 4 as “I have substantial direct research experience on this topic.” Please clarify what “the above” refers to and why it led you to change your assessment under that criterion.

  3. We request a formal apology for the personal remarks directed at us. In response to our request for clarification, you wrote, “I am a reviewer, not your advisor” and “Reviewing is voluntary; I am not an employee you hired.” You also wrote, “Perhaps your paper would be more suited to being reviewed by AI reviewers.” We acknowledge and have corrected the misplaced heading in our earlier response. We consider these remarks dismissive and inappropriate in a scientific review, and we ask you to acknowledge them and formally apologize to the authors.

Sincerely,
The Authors of Submission 27634

AuthorsOpenReview ↗

Summary of Rebuttal Improvements

We sincerely appreciate the reviewers and the Area Chair, Senior Area Chair, and Program Chairs for their time and consideration of our submission. The reviewers’ comments helped us identify ambiguities in the methodological description, improve the manuscript’s presentation, and clarify the scope of the experimental conclusions. In response, we revised the manuscript and reported additional analyses and evaluation details. The corresponding manuscript revisions are marked in red.

Summary of Rebuttal Improvements

We thank the reviewers for their constructive feedback. Our principal revisions and additional evidence are summarized below.

[Reviewer 5QBU]: We corrected the sampler description to vanilla MDLM at r=1r=1, clarified ligand-only inference while disclosing the training–inference centering mismatch, and added evaluation counts and cohort explanations. We narrowed binding-related claims to computational screening outcomes. An exploratory ten-design shuffled-ligand analysis yielded mean AF3 ipTM of 0.834 versus 0.709 for intended and shuffled ligands, respectively; this measures interface confidence, not validated selectivity. On November 18, the reviewer explicitly supported weak acceptance of the clarified computational contribution.

[Reviewers 5QBU and 3JJR]: We added sample standard deviations for component-removal experiments involving 600 designs per setting. Removing ligand conditioning caused the largest deterioration. These results characterize inference-time sensitivity.

[Reviewer 3JJR]: We clarified the contribution as ligand-conditioned joint sequence–structure generation in discrete token space, supported by the curated dataset and computational evaluation. We explicitly acknowledged the established components, computational-only evidence, and limitations concerning side-chain modeling and ligand flexibility.

[Reviewer C8bG]: We reduced excessive emphasis, explained ligand conditioning in the Introduction, clarified the organization of introductory and background material, and defined MDLM upon first use.

[Reviewer 3JJR—Discussion engagement]: We submitted point-by-point responses to all three concerns on November 15, followed up repeatedly on November 16–18. We have received no substantive follow-up assessing these responses or identifying remaining concerns. The ICLR 2027 Reviewer Guidelines call for active discussion and explaining what did or did not change the recommendation. ICLR 2027 Reviewer Guidelines. We respectfully ask the AC and SAC to obtain an explicit assessment of our rebuttal and any remaining concerns, ensuring that the final decision considers the revised manuscript and the evidence provided during discussion.

We sincerely thank the reviewers again for their time, constructive guidance, and careful consideration of our revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.