PepTune: De Novo Generation of Therapeutic Peptides with Multi-Objective-Guided Discrete Diffusion

We are grateful for your insightful review.

Essential References Not Discussed:

We appreciate the reviewer’s suggestion to consider Wang et al. (2024). Although the title may suggest a close relation to our work, the scope, modeling assumptions, and core goals differ significantly. MMCD generates peptides composed of 20 natural amino acids only using a multi-modal approach. While MMCD uses the label “therapeutic peptide generation,” its evaluation is limited to docking scores and embedding-level metrics, without any testing of therapeutic properties such as solubility, hemolysis, or non-fouling. In contrast, our work directly targets real-world therapeutic peptide design by expanding the chemical space to include non-canonical residues and cyclic structures and by incorporating classifier-guided objectives tied to experimentally actionable properties. This direction aligns closely with the goals of the Application-Driven Machine Learning Track. Given these fundamental differences in scope and design goals, a direct comparison is not appropriate.

Addressing Weaknesses:

Weakness 1: We would like to push back on the assertion that our paper lacks significant novelty. Although NELBO is used in previous works, we derive a unique case of the NELBO given our novel bond-dependent masking schedule (See Appendix D.3). While the MDLM was already introduced by Sahoo et al., we make a novel extension with a bond-dependent masking schedule and derive a reverse posterior that differs from the general case in Sahoo et al. (See Appendix D.1-D.2). In addition, we introduce a novel invalidity loss that penalizes invalid peptide structure during training.

Finally, despite the use of MCTS in prior discrete models, our paper introduces the first application of MCTS to diffusion models, with our method being published in arXiv before Yu et al. (2025). In addition, our submission date to ICML was 9 Jan 2025, which was before the publication of Yu et al. (2025) on 11 Feb 2025.

Weakness 2: The baseline mentioned (Wang et al.) is designed to generate peptides consisting only of the 20 natural amino acids, while the core innovation of PepTune is to expand the search space to peptides with modified amino acids and cyclicizations which has not been achieved by prior models, which has been shown to enhance various therapeutic properties. These are two fundamentally different problems that are difficult to compare. Furthermore, the motivation behind the key components of our architecture, including our bond-dependent masking schedule and invalid penalty loss, stems from our unique goal of generating valid peptides from SMILES tokens despite the increased granularity.

Our source code is included in the arXiv version of the manuscript; however, given the anonymity of the review process, please refer to the anonymous repo link: https://anonymous.4open.science/r/peptune-86CB/README.md.

Weakness 3: We believe you may have misinterpreted the comparisons in Table 1. PepMDLM is our model trained with our bond-dependent masking schedule only without MCTS guidance. We make this comparison to demonstrate that our MCTS-guidance strategy does not significantly lower the diversity of sequences and increases validity through our iterative MCTS-guidance algorithm that rewards valid expansion steps.

Given that most of your concerns are regarding clarity and not the methodological and experimental soundness of our paper, and we have addressed your doubts on the novelty of our approach, we hope you consider raising your score in response to our answers. We will plan to introduce clarifications into the main text for the camera-ready version.

Thank you for your detailed suggestions.

Claims and Evidence:

• We conduct an ablation study as suggested. From each model, we sampled 100 sequences of token length 100 and checked validity with our SMILES2PEPTIDE decoder. We show that both the bond-dependent masking and invalid loss improve the validity of peptides.

• Due to the significant compute time needed to train additional baseline models on our 11 million peptide dataset, we decided it was not feasible to do so in this rebuttal period. Moreover, we note that Gruver et al. evaluate at most two objectives and require performing guidance updates in the latent embedding space based which can result in invalid conversion back to a discrete sequence, and Nisonoff et al. is limited to single-objective guidance. To the best of our knowledge, PepTune is the first framework for classifier-based multi-objective guidance for discrete diffusion, overcoming previous limitations via reward-based tree expansion and Pareto-aware backpropagation.

Essential References Not Discussed:

Shi et al. leverage a per-token parameter $w$, which is trained in parallel with the discrete diffusion model. In contrast, we fix $w=3$ for tokens that exist within a peptide bond to ensure the model learns to reconstruct them earlier. This construction presents a practical extension of state-dependent masking to preserve peptide structure, which could be further explored for preserving various other structured data types. In addition, we rederive the state-dependent NELBO loss in a fundamentally different manner than Shi et al. (See Appendix D.3).

Other Comments or Suggestions:

We thank you for the thoughtful suggestions. We have made the notations and figures clearer upon your request in the manuscript for the camera-ready version.

1. We notice the inconsistencies in notation that you refer to, but in Equations 1-4, we simplify the notation so that $\mathbf{x}_0$ refers to the one-hot vector of the true token at an arbitrary position in the sequence. We have added clarification and removed the $(\ell)$ superscript in the other equations to maintain consistency.
2. Yes, $w$ is a scalar greater than one. Please refer to the last sentence of Section 2.1, where the exact value is specified: “Empirically, we found that $w=3$ increased peptide validity while maintaining diversity across generated samples.”
3. The sentence is fixed to “where $\mathbf{b}$ is the vector with ones at indices of peptide bond tokens and zeroes in remaining indices.”
4. We have increased the font size for clarity and fixed Equation 11 to include the $\log$.

Answers to Questions:

1. In equation 3, the first $\mathbf{x}_0$ is a one-hot vector is used to extract the probability associated with the unmasking the true token by taking an inner product. The second $\mathbf{x}_0$ indicates that if the current token $\mathbf{z}_t$ is a mask token, it has a $q(\mathbf{z}_s=\mathbf{x}_0|\mathbf{z}_t=\mathbf{m}, \mathbf{x}_0)$ probability of being unmasked to $\mathbf{x}_0$ and a $q(\mathbf{z}_s=\mathbf{m}|\mathbf{z}_t=\mathbf{m}, \mathbf{x}_0)$ probability of remaining masked. For clarity, we have changed the transpose to an inner product: $$q(\mathbf{z}_s|\mathbf{z}_t, \mathbf{x}_0)=\begin{cases} \left\langle \left(\frac{s}{t}-\frac{s^w}{t^w}\right)\mathbf{b}+\frac{t-s}{t}\mathbf{1}, \mathbf{x}_0\right\rangle \mathbf{x}_0+\left\langle\left(\frac{s^w}{t^w}-\frac{s}{t}\right)\mathbf{b}+\frac{s}{t}\mathbf{1}, \mathbf{x}_0\right\rangle\mathbf{m}&\mathbf{z}_t=\mathbf{m}\\\\ \mathbf{z}_t&\mathbf{z}_t\neq \mathbf{m} \end{cases}$$

2. $K$ represents the token vocab size. $\tilde{\mathbf{x}}\in \{0, 1\}^K$ denotes a $K$-dimensional one-hot vector where the index of the token is indicated with 1. 3. In HELM-GPT, the peptides are represented in HELM notation [1], where each token is a valid, synthesizable amino acid. In contrast, PepMDLM is trained on SMILES tokens that decompose amino acids into smaller tokens that can be pieced together into valid amino acids during generation. While this enables us to represent a greater diversity of non-natural amino acids, even a single token generated in the wrong position can result in an invalid peptide sequence, resulting in a lower validity rate. However, we note that our MCTS guidance strategy increases the validity rate to 100% due to its iterative unmasking process that is rewarded on high-scoring and valid peptides. We appreciate the reviewer’s constructive feedback and hope the new clarifications, ablations, and improved notation address your concerns. Given the novelty, empirical strength, and thorough revisions, we hope you will consider raising your score to demonstrate your support for our work.

We are very grateful for the reviewer’s thoughtful review.

Essential References Not Discussed:

We request that the reviewer refer to Appendix C, which provides an extensive analysis of prior discrete diffusion and classifier-based and classifier-free guidance methods. We also discuss at the end of Appendix C.3 the limitations of prior discrete guidance methods and challenges in multi-objective guidance.

Addressing Weaknessses:

To ensure that our model can scale to an arbitrary number of objectives, we leverage a classifier-based guidance approach which doesn’t take the protein sequence explicitly as input to the diffusion model but instead injects the target protein information at each denoising step by computing multi-objective rewards that are back-propagated through the MCTS tree. This means that unmasking steps that produce high-scoring peptides are explored further, limiting the “wasted” peptides. We also note that this exploration method enables us to generate a batch of Pareto-optimal sequences in a single run of the algorithm, which is much faster than generating only one sequence per run like prior models.

In addition, we show that our binding affinity and property classifiers generalize well on the validation data (Table 6 and Figure 14). We have to accept that only limited protein-peptide binding quantification data is available, and downstream in silico validation like VINA docking can be used to narrow down candidate peptides to be synthesized for experiments.

Answers to Questions:

• PepMDLM is trained on SMILES tokens that decompose amino acids into smaller tokens that can be pieced together into valid amino acids during generation. While this enables us to represent a greater diversity of non-natural amino acids, even a single token generated in the wrong position can result in an invalid peptide sequence, resulting in a lower validity rate, which we calculate by decoding the SMILES sequences with regular expression matching. However, we note that our MCTS guidance strategy increases the validity rate to 100% due to its iterative unmasking process that is rewarded on high-scoring and valid peptides.

• For the experiments, we generate peptides of token length 50 (7 amino acids) up to 200 (30 amino acids). This aligns with existing therapeutic peptides in clinical use, where the average length is about 20 amino acids or less [1].

• Yes, our MCTS-guidance algorithm is a modular multi-objective guidance framework that can be adapted for any type of discrete data (e.g. protein sequences, promoter and enhancer DNA, RNA, small molecule drugs, natural language, etc.) and any pre-trained classifier in a plug-and-play fashion. Furthermore, our bond-dependent masking schedule can be extended to preserve structural components of other data types (e.g. functional motifs in proteins, punctuation and grammar in natural language, etc.).

• In our work, we use “de novo” to refer to the generation of novel peptide sequences from scratch—i.e., not based on templates or fragments from known binders—and guided by pre-trained classifiers that predict binding affinity, solubility, and immunogenicity of these peptides. The model does not rely on known peptide backbones or motifs but instead generates sequences purely from noise through discrete diffusion sampling.

  Furthermore, the target protein is not restricted to the training set of the classifier. Both the target protein and de novo peptide binder sequence are converted into feature-rich embeddings and passed through a classifier trained to understand binding patterns for accurate prediction. Our classifiers generalize well across unseen sequences, as shown by the high validation Spearman correlation coefficients (Table 6 and Figure 14).

We appreciate the reviewer’s thoughtful feedback and hope the clarifications provided demonstrate the uniqueness and adaptability of our approach. Our response highlights how PepTune addresses key limitations in prior work through a scalable, modular guidance framework, efficient multi-objective generation via MCTS, and strong generalization to arbitrary discrete data types and guidance objectives. Given our detailed answers, we hope you can consider raising your score to demonstrate continued support for our work.

[1] Dougherty, Patrick G et al. “Understanding Cell Penetration of Cyclic Peptides.” Chemical reviews vol. 119,17 (2019): 10241-10287. doi:10.1021/acs.chemrev.9b00008

We thank the reviewer for the constructive feedback.

First, we would like to respectfully disagree with the assertion that the guidance strategy is inadequately evaluated. Given that the paper is on Application-Driven ML Track, we chose case studies that provide robust evaluation to prove efficacy in peptide-based therapeutic design. This is far more impactful and applicable to society and medicine than classical benchmarks. We ask that the reviewer reconsider their review of the paper with this in mind.

Addressing Concerns Experimental Designs Or Analyses:

• We conduct an ablation study as suggested. From each model, we sampled 100 sequences of token length 100 and checked validity with our SMILES2PEPTIDE decoder. We show that both the bond-dependent masking and invalid loss improve the validity of peptides.

• In HELM-GPT, the peptides are represented in HELM notation [1], where each token is a valid, synthesizable amino acid. In contrast, PepMDLM is trained on SMILES tokens that decompose amino acids into smaller tokens that can be pieced together into valid amino acids during generation. While this enables us to represent a greater diversity of non-natural amino acids, even a single token generated in the wrong position can result in an invalid peptide sequence, resulting in a lower validity rate. However, we note that our MCTS guidance strategy increases the validity rate to 100% due to its iterative unmasking process that is rewarded on high-scoring and valid peptides.

• We selected three target classes to demonstrate PepTune’s robustness: (1) proteins with known binders to benchmark against known binders, (2) proteins without known binders, and (3) dual-target cases. This is a rigorous evaluation for application in therapeutic peptide design.

  While previous guidance methods for discrete diffusion are limited to optimizing 1-2 objectives, we evaluate guidance across 4-5 non-trivial objectives that are highly relevant to therapeutic peptide design.

• We appreciate the reviewer’s suggestion to conduct best-of-N comparison. However, we believe that our comparisons between PepTune and PepMDLM are sufficient in demonstrating the efficacy of our guidance strategy for the following reasons:

  • In Pareto optimization, scalarizing or ranking sequences across conflicting objectives without access to the Pareto frontier leads to suboptimal trade-offs, making it difficult to rank best-of-N.
  • Despite exploring more sequences, PepTune achieves comparable runtime by reusing high-reward unmasking paths within the MCTS tree. Rather than starting from scratch, PepTune expands only unexplored nodes, reducing redundant computation. Empirically, we find similar runtimes for generating 100 valid sequences using PepMDLM (403 sec) and PepTune (347 sec for top 100 sequences with $M=10$ and $N_{\text{iter}}=128$), while PepTune yields substantially higher property scores (Figures 4–9).

• We also appreciate the suggestion to conduct a benchmark against ParetoFlow. However, ParetoFlow is designed for continuous flow-matching models, whereas our approach is designed specifically for discrete diffusion. Due to the fundamental differences in model class and data type, a direct comparison would not be meaningful; to our knowledge, PepTune is the first to apply Pareto-aware guidance for discrete diffusion, and we hope this novelty is considered.

• Given that our method supports robust guidance beyond 2-3 objectives, it is impossible to visualize the high-dimensional Pareto frontier. Instead, our graphs demonstrate that our approach simultaneously raises the scores across 5 objectives (see Figure 9) to a plateau, indicating that no further optimization is possible without sacrificing another objective.

• Each MCTS run explores $M\times N_{\text{iter}}$ sequences. The sampling cost scales linearly with respect to both of these parameters. We compared $M=\{10, 50, 70, 100\}$ and found similar guidance curves and set $N_{\text{iter}}=128$ constant.

Responses to Questions

• In this experiment, we optimized 5 properties: binding affinity to GFAP, membrane permeability, solubility, hemolysis, and non-fouling, all of which are not related to the values in Table 1. In Table 1, we aim to show that our MCTS-guidance algorithm does not significantly lower the diversity of sequences and increases validity through our iterative MCTS-guidance algorithm that rewards valid expansion steps.

We thank you for your suggestions and hope you reconsider raising your score in light of these responses and the broader goals of the Application-Driven ML Track.

[1] Zhang et al. Helm: A hierarchical notation language for complex biomolecule structure representation. http://dx.doi.org/10.1021/ci3001925