PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Sophia Tang, Yinuo Zhang, Pranam Chatterjee

arXiv:2412.17780v4q-bio.BMcs.AI

TL;DR

PepTune addresses limits in searching modified peptides and in discrete guidance across conflicting objectives. It combines masked discrete diffusion with bond-dependent masking, an invalid-sequence loss, and Monte Carlo Tree Guidance; the reported experiments include diverse binders and chemically modified peptides. The authors conclude that this is a modular approach to multi-objective sequence design in discrete state spaces.

  • Problem

    Searching the large space of chemically modified peptides remains difficult, while existing peptide design approaches are limited by target-structure dependence, canonical-only vocabularies, or challenges in multi-objective discrete guidance.

    Source: §1 · p. 1

  • Method

    PepTune uses masked discrete diffusion with bond-dependent masking and an invalid-peptide loss, alongside inference-time Monte Carlo Tree Guidance to refine sequences across multiple objectives.

    Source: §1 · pp. 1–2, §2.2 · p. 2, §2.3 · p. 4

  • Results

    PepTune generated diverse peptide binders for TfR and GLAST; all seven selected binders had docking scores ≤−7.5 kcal/mol, and its unconditional generator produced valid peptides with higher average nAA frequency than experimentally validated peptide SMILES for membrane permeability and binding affinity.

    Source: §5 · pp. 8–9

  • Takeaways & Limitations

    The authors present MCTG as a modular, inference-time approach for multi-objective sequence design in discrete state spaces.

    Source: §6 · p. 9

    The bond-dependent masking design assumes peptide SMILES share the described bond-centered structure around each amino acid's central carbon.

    Source: §2.2 · p. 2

Abstract

from arXiv · show

We present PepTune, a multi-objective discrete diffusion model for simultaneous generation and optimization of therapeutic peptide SMILES. Built on the Masked Discrete Language Model (MDLM) framework, PepTune ensures valid peptide structures with a novel bond-dependent masking schedule and invalid loss function. To guide the diffusion process, we introduce Monte Carlo Tree Guidance (MCTG), an inference-time multi-objective guidance algorithm that balances exploration and exploitation to iteratively refine Pareto-optimal sequences. MCTG integrates classifier-based rewards with search-tree expansion, overcoming gradient estimation challenges and data sparsity. Using PepTune, we generate diverse, chemically-modified peptides simultaneously optimized for multiple therapeutic properties, including target binding affinity, membrane permeability, solubility, hemolysis, and non-fouling for various disease-relevant targets. In total, our results demonstrate that MCTG for masked discrete diffusion is a powerful and modular approach for multi-objective sequence design in discrete state spaces.

1. Introduction

Peptide design is constrained by the difficulty of searching chemically modified sequences and by limits in existing structure-based and sequence-based approaches. PepTune combines discrete peptide-SMILES generation with multi-objective guidance and validity-focused methods.

1. Introduction

  • Searching chemically modified peptides remains a major limitation, while existing approaches struggle with unstable target structures, canonical-only vocabularies, or conflicting objectives.Peptides made only from the 20 wild-type amino acids are susceptible to enzymatic degradation and have low membrane permeability.Source: §1 · p. 1
  • PepTune is introduced as a multi-objective-guided discrete diffusion model for de novo peptide-SMILES generation.The paper describes it as the first such model.Source: §1 · p. 1
  • The model generates peptide SMILES with non-canonical amino acids and cyclic modifications using a masked diffusion language model.Source: §1 · p. 2
  • A bond-dependent masking schedule is designed to increase the structural validity of generated peptide SMILES.Source: §1 · p. 2
  • A peptide-SMILES-filter-based invalid loss propagates penalties from discrete sequences to continuous probability distributions.Source: §1 · p. 2
  • Monte Carlo Tree Guidance iteratively refines Pareto-optimal peptide sequences across therapeutic properties without gradient estimation or retraining.The paper also trains predictors for binding affinity, membrane permeability, solubility, hemolysis, and non-fouling.Source: §1 · p. 2

2. Bond-Dependent Masked Discrete Diffusion

PepMDLM learns peptide-SMILES generation by reconstructing sequences from masked inputs. Its bond-dependent schedule masks peptide-bond tokens more slowly early in diffusion, and its training losses combine a bond-dependent diffusion objective with a filter-based penalty for invalid sequences.

2.1. Masked Discrete Diffusion Model

  • PepMDLM uses the MDLM framework to reconstruct clean peptide-SMILES sequences from inputs corrupted with mask tokens.A RoFormer backbone predicts token probabilities for masked positions.Source: §2.1 · p. 2

2.2. Bond-Dependent Masking Schedule

  • Peptide-bond tokens are masked more slowly than other tokens early in diffusion, then both masking probabilities approach 1 as time approaches 1.The schedule uses log-linear masking for non-bond tokens and log-polynomial masking for peptide-bond tokens; w = 3 increased validity while maintaining diversity.Source: §2.2 · pp. 2–3

2.3. Loss Functions

  • The bond-dependent continuous-time NELBO combines negative log-loss terms for peptide-bond and non-peptide-bond tokens under their respective masking schedules.The training objective maximizes the ELBO of the distribution over dataset peptide sequences by minimizing its negative.Source: §2.3 · p. 3
  • The invalid-peptide loss penalizes argmax-generated sequences that fail the validity filter, using sampled-token softmax probabilities to scale the penalty.The filter flags invalid SMILES notation or sequences that cannot be decoded into peptides.Source: §2.3 · p. 4

3. Multi-Objective Guided Discrete Diffusion

Monte Carlo Tree Guidance (MCTG) adapts tree search to guide discrete diffusion toward a set of Pareto-optimal peptide sequences across multiple objectives. Its selection and expansion steps balance reward-guided refinement with exploration of unmasking paths.

3. Multi-Objective Guided Discrete Diffusion

  • Because objectives can trade off, a Pareto-optimal sequence need not be best across all objectives.The set approximates non-dominated sequences and their objective scores rather than a single universal optimum.Source: §3 · pp. 4–5
  • Each iteration selects a path to an unexpanded leaf, expands it into M partially unmasked children, rolls them out to clean sequences, then back-propagates rewards.The algorithm updates the Pareto-optimal set during this process.Source: §3 · pp. 6–7
  • Selection balances accumulated rewards for previously productive unmasking steps with reverse-posterior probabilities that favor sufficiently likely, less-explored steps.The exploration coefficient is c = 0.1, and reward influence grows as child nodes receive more visits.Source: §3 · p. 6
  • The authors chose Gumbel-noise sampling for expansion because it produced greater sequence diversity than top-k sampling, which instead prevented unlikely tokens.They favored the first method to allow greater exploration.Source: §3 · p. 6
  • Clean sequences are scored for binding affinity, membrane permeability, solubility, hemolysis, and non-fouling; non-dominated sequences are added to the Pareto set.Rewards compare each child’s objective scores with sequences already in the current Pareto-optimal set.Source: §3 · p. 7
  • A single MCTG run returns Pareto-optimal sequences from the Niter · M total sequences sampled across iterations.The output is a set of non-dominated sequences across the K objectives.Source: §3 · p. 7

4. Therapeutic Property Classifiers

The authors train peptide-SMILES-specific property predictors and report significantly enhanced performance relative to PeptideBERT.

4. Therapeutic Property Classifiers

  • The authors train peptide-SMILES-specific regressors for target-binding affinity and membrane permeability and classifiers for solubility, hemolysis, and non-fouling.The work addresses a stated gap in high-quality property models trained specifically on peptide SMILES data.Source: §4 · p. 7
  • The prediction models achieve significantly enhanced peptide property prediction performance compared with the PeptideBERT baseline.The comparison is reported in Table 2.Source: §4 · p. 7
  • Table 2 benchmarks solubility, hemolysis, and non-fouling prediction against PeptideBERT using XGBoost models trained on PeptideCLM SMILES embeddings.Source: Table 2 · p. 7

5. Experiments

PepTune experiments evaluate chemically modified peptide generation and multi-objective binder design. Generated binders show strong docking results for GLP-1R and dual TfR–GLAST targeting, alongside reported gains in property scores and peptide validity.

PepMDLM generation

  • PepMDLM generated diverse chemically modified and cyclic peptides, including peptides with higher average non-natural amino-acid frequency than experimentally validated peptide SMILES for permeability and binding affinity.The comparison is reported as evidence of searching a broader peptide-SMILES subspace.Source: §5 · pp. 7–8

Generative quality

  • The resulting PepTune peptide pool retained uniqueness and diversity similar to PepMDLM and the training set, while valid peptides reached 100% after 20 MCTS iterations.Table 1 reports generative-quality metrics for peptide SMILES with maximum token length 200.Source: Table 1 · p. 8, §5 · p. 8

GLP-1R binders

  • PepTune binder docking scores of -7.4 and -7.0 kcal/mol beat semaglutide (-5.7) and liraglutide (-5.1), with overlapping GLP-1R binding motifs.The authors suggest these peptides could serve as GLP-1R agonists or antagonists.Source: Figure 3 · p. 8, §5 · p. 8

Dual-target binders

  • All seven selected dual-target binders docked at ≤−7.5 kcal/mol; the top scored −10.5 kcal/mol for TfR and −9.2 kcal/mol for GLAST.The selected binders also had diverse secondary structures, positive solubility, and low hemolysis probabilities.Source: Table 3 · p. 9, §5 · p. 9

Multi-objective guidance

  • Across five-property conditioning, all scores improved over iterations, and final solubility, hemolysis, and non-fouling scores surpassed TfR-only-conditioned binders.The authors report that this strategy avoids significant property trade-offs.Source: §5 · p. 9

6. Discussion

The discussion presents PepTune as a discrete diffusion framework for generating chemically modified peptide SMILES and guiding multi-objective optimization. It emphasizes inference-time guidance, Pareto-optimal sequence diversity, and operation without target 3D structures.

PepTune

  • PepTune combines bond-dependent masking and an invalid loss with discrete diffusion to generate peptide SMILES containing non-natural amino acids and cyclic modifications.Source: §6 · p. 9

Monte Carlo Tree Guidance

  • MCTG guides discrete diffusion at inference time without additional training, balancing batched unmasking with Gumbel noise against reward-based exploitation of promising paths.Source: §6 · p. 9
  • MCTG is described as producing diverse Pareto-optimal sequences across an arbitrary number of objectives while operating strictly in discrete state space.Source: §6 · p. 9

Structure-free design

  • Unlike the cited binder-design methods, PepTune requires no 3D target structures, supporting peptide design for conformationally diverse proteins and optimization beyond local geometric interactions.Source: §6 · p. 9

Impact Statement

The impact statement presents multi-objective diffusion guidance as potentially applicable to generative tasks with simultaneous constraints. It also calls for oversight because the framework could be misused to generate peptides with harmful biological properties.

Broader applicability

  • The authors propose multi-objective diffusion guidance as a framework that could apply to generative tasks requiring simultaneous optimization of multiple constraints.Source: Impact Statement · p. 10

Potential therapeutic impact

  • PepTune focuses on therapeutically viable peptides and could advance disease treatments by enabling discovery optimized for binding, solubility, and other critical properties.Source: Impact Statement · p. 10

Responsible use

  • The authors identify potential misuse in generating peptides with harmful biological properties and encourage careful oversight and ethical application.Source: Impact Statement · p. 10

Declarations

The appendix case studies apply PepTune across established and binder-free targets, including intracellular and dual-target settings. They evaluate predicted properties and in-silico docking for generated candidates.

A. Additional Case Studies

A.1. Targeting Receptors on the Blood-Brain Barrier

  • The TfR binding-affinity classifier score rose above 9.0, and PepTune shifted predicted-affinity distributions above PepMDLM and training sequences.This trend was observed while conditioning on binding affinity, solubility, hemolysis, and non-fouling.Source: §A.1 · p. 15, Figure 5 · p. 16
  • All selected TfR binders docked below -6.0 kcal/mol, with the top scoring -8.4 kcal/mol; the top candidates shared contacts with T7.The generated binders had distinct surface motifs and included cyclic or side-chain-modified structures.Source: §A.1 · pp. 15–16, Figure 6 · p. 17

A.2. Targeting Intracellular Proteins

  • GFAP binders were conditioned on five properties, including cell-membrane permeability, and showed docking affinities below -7 kcal/mol.The target is intracellular, so the design also sought peptides able to cross the astrocyte membrane.Source: §A.2 · pp. 16–17

A.3. Targeting Proteins Without Existing Binders

PepTune generates candidates for targets without established binders and for dual targeting, while the appendix also details its discrete diffusion and training objective. The reported evidence includes predicted affinity scores, docking results, and method descriptions.

A.3. Targeting Proteins Without Existing Binders

  • 6.708 was PepTune’s mean predicted NCAM1 binding-affinity score, versus 5.298 for PepMDLM and 5.360 for experimentally tested peptides.The highest-scoring selected binder docked at -8.6 kcal/mol; another wrapped around NCAM1 with numerous polar contacts.Source: §A.3 · p. 17, Figure 8 · p. 18
  • AMHR2 candidates had predicted Vina binding affinities below -6 kcal/mol, while predicted solubility declined over optimization iterations.The authors attribute the solubility decline to hydrophobic patches near the ligand-binding site.Source: §A.3 · p. 19, Figure 9 · p. 19

A.4. Dual-Targeting of GFAP and an E3 Ubiquitin Ligase for Target Protein Degradation

  • Three non-dominated GFAP–RBX1 candidates with predicted affinities above 7.0 showed strong post-docking binding to both targets.The candidates docked along GFAP’s rod domain and near RBX1’s Cullin interaction site rather than its Skp2 F-box adaptor site; the authors note motif conditioning could benefit clinical potential.Source: §A.4 · pp. 19–20, Table 4 · p. 19, Figure 11 · p. 21

B.1. Continuous-Time Discrete Diffusion

  • The forward discrete diffusion process applies categorical noise across sequence positions, with transitions governed by a time-dependent schedule.The appendix describes transitions that retain the current token with probability σ(t) or switch according to token distribution π.Source: §B.1 · p. 20

B.2. Continuous-Time Negative Evidence Lower Bound (NELBO)

  • The negative ELBO provides the diffusion model’s loss-minimization objective and separates reconstruction, prior, and intermediate diffusion terms.The reverse model is trained against the forward corruption process to reconstruct clean sequences from noisy states.Source: §B.2 · pp. 22–24

B.3. Guided Diffusion Models

Guided diffusion conditions reverse transitions using property scores, but existing approaches face data, discrete-gradient, and multi-objective limitations. The section motivates classifier-based guidance that avoids conditioned generative training, while noting the computational challenge of scaling it to multiple objectives.

Guided Diffusion Models

  • The desired guided reverse transition samples a previous state conditioned on the current sequence and property, with Bayes’ rule providing its decomposition.The transition is framed as sampling from q(zs|zt, y).Source: §B.3 · p. 24

Classifier-Free Guidance

  • Classifier-free guidance depends on quality-annotated data and does not scale to multiple objectives when jointly conditioned data are scarce.These limitations are especially relevant to peptide sequences.Source: §B.3 · p. 24

Classifier-Based Guidance

  • Classifier-based guidance combines an unconditional diffusion model with a classifier score to form a property-conditioned reverse distribution.The update uses the gradients of the unconditional distribution and classifier probability.Source: §B.3 · pp. 24–25
  • Gradient-based classifier guidance is not directly applicable to discrete states, and continuous-latent projection can produce tokens that do not match optimized embeddings.Classifier-based guidance avoids training a generative model on conditioned data, unlike classifier-free guidance.Source: §B.3 · p. 25

Guidance in the Discrete State Space

  • Taylor approximation avoids enumerating every reverse transition, but multi-objective guidance generally requires a joint property distribution rather than an independence-based product.The passage identifies this joint-probability computation as a scaling challenge for contradictory or dependent objectives.Source: §B.3 · p. 25

C.1. PepMDLM Training Data

PepMDLM is trained on a large, chemically varied peptide-SMILES collection and uses clustered data splits, length-aware batching, and peptide-motif tokenization. The model represents masked sequences as token-probability distributions and uses a RoFormer backbone.

C.1. PepMDLM Training Data

  • The training collection contains 11 million peptide SMILES from CycPeptMPDB, SmProt, and CycloPs-generated modified peptides, with SELFIES used to check sequence integrity.CycloPs peptides include non-natural amino acids, dextro-chiral alpha carbons, N-methylated backbones, PEGylation, and attempted cyclization conformations.Source: §C.1 · p. 25
  • K-means clustering by Morgan fingerprints groups chemically similar sequences into 1,000 clusters, which are split 0.8/0.2 to retain similar diversity across partitions.The authors use this grouping to expose batches to diverse sequences across masking levels.Source: §C.1 · p. 26

C.2. Dynamic Batching

  • Dynamic batching sorts SMILES by length to improve GPU-memory use, with a maximum of 16,000 tokens per GPU.This approach handles variable-length token sequences.Source: §C.2 · p. 26

C.3. SMILES Tokenization

  • SPE tokenization with the 581-token PeptideCLM vocabulary outperformed the tested alternatives, producing precise, valid peptides with an average of four characters per token.The tokenization captures common motifs in the training data.Source: §C.3 · p. 26

D.1. Notation; D.2. Model Architecture

  • The model predicts a vocabulary-sized probability distribution for each token from the masked sequence and time, using a RoFormer backbone with rotary positional embeddings.The notation defines x0 as a clean token, zt as a state at time t, and αt(x0) as its probability of remaining unmasked.Source: §D.1 · p. 26, §D.2 · p. 26

D.3. Unconditional Generation Results

PepMDLM generates chemically modified and cyclic peptide SMILES, while the paper evaluates its generation quality and peptide-SMILES property predictors. The section also describes the validity filter, evaluation procedures, and bond-dependent masking schedule used to preserve peptide structure.

D.3. Unconditional Generation Results

  • 10% of unconditionally generated peptides contain cyclic modifications, and their non-natural amino-acid frequencies resemble those in experimentally validated membrane-permeable and binding peptides.The frequency comparison is shown in Figure 12.Source: §D.3 · p. 27, Figure 12 · p. 28
  • PepMDLM has higher uniqueness and diversity and lower nearest-neighbor similarity than HELM-GPT, while its unconditional validity is lower; MCTS guidance raises validity to 100% in 20 iterations.The comparison uses HELM-GPT’s HELM representation, whereas PepMDLM’s SMILES tokens can assemble into invalid peptides when a token is misplaced.Source: Table 7 · p. 27, §D.3 · p. 27

E.1. Protein Target-Binding Prediction

  • The binding-affinity regressor reaches Spearman correlations of 0.869 on training data and 0.633 on held-out validation data.It predicts binding scores from paired target-protein sequences and peptide SMILES, trained on 1,806 PepLand sequences.Source: §E.1 · p. 28

E.2. Cell Membrane Permeability Prediction

  • The membrane-permeability regressor reaches Spearman correlations of 0.998 on training data and 0.943 on test data.The model predicts PAMPA lipophilicity scores from peptide-SMILES embeddings.Source: §E.2 · p. 29

E.3. Solubility and Toxicity Prediction

  • The solubility, hemolysis, and non-fouling classifiers achieve strong Spearman correlations and improve on the PeptideBERT baseline.The objectives maximize positive solubility and non-fouling predictions and minimize positive hemolysis predictions.Source: §E.3 · p. 30

F.1. Peptide Validity Filter

  • The validity filter checks molecular validity, peptide bonds, central carbon atoms, bond patterns, modifications, and amino-acid side chains, and can identify cyclic peptides.It detects a library of more than 200 non-natural amino acids.Source: §F.1 · pp. 30–31

F.2. Metrics; F.3. Docking

  • Generation quality is assessed with validity, uniqueness, diversity, nearest-neighbor similarity, and novelty metrics; docking is used to assess valid non-dominated candidates.Novelty and nearest-neighbor calculations sample 1,000 training-data batches because of memory and CPU limits; docking uses AutoDock Vina on selected candidates.Source: §F.2 · pp. 31–32

G.1. Bond-Dependent Masking Schedule

  • Peptide-bond tokens are masked more slowly early in diffusion and receive higher loss weight, biasing reverse diffusion toward unmasking structural bonds earlier.The schedule uses w = 1 for non-peptide bonds and w = 3 for peptide bonds; the loss weighting is higher for larger w at earlier steps.Source: §G.1 · pp. 32–33, Figure 15 · p. 33

G.2. Derivation of Bond-Dependent Reverse Posterior

The appendix derives reverse transitions and the loss for bond-dependent masking, then describes invalid-loss gradients, hyperparameter choices, and a time-dependent guidance case study. In that case study, property scores rise at different stages as guidance is introduced for additional objectives.

G.2. Derivation of Bond-Dependent Reverse Posterior

  • The bond-dependent reverse posterior distinguishes peptide-bond from non-peptide-bond tokens, while preserving already unmasked tokens under the carry-over unmasking scheme.The forward process masks tokens according to whether the clean token is a peptide bond.Source: §G.2 · pp. 33–35

G.3. Derivation of Bond-Dependent NELBO Loss

  • The continuous-time NELBO decomposes into negative log-losses for non-peptide-bond tokens under a log-linear schedule and peptide-bond tokens under a log-polynomial schedule.The two token classes use their respective masking schedules in the derived objective.Source: §G.3 · pp. 36, 38, 39

G.4. Gradient Flow of Invalid Loss

  • The invalid loss reduces the probability of sampled tokens that produce invalid peptide SMILES and proportionally raises probabilities for other tokens.Its gradient update is minimized near probabilities of 0 and 1, pushing predictions toward higher confidence.Source: §G.4 · p. 40

H. Hyperparameter Selection; I.1. Case Study for Time-Dependent Multi-Objective Guidance

  • The authors select M = 50 children and Niter = T = 128; SPE lowers perplexity while maintaining precision, and the time-dependent case study is a proof of concept rather than the setup used throughout.The paper says all its results use peptides without time-dependent guidance.Source: H. Hyperparameter Selection · pp. 40–41, §I.1 · p. 42

I.1. Case Study for Time-Dependent Multi-Objective Guidance

  • After permeability-only guidance for 50 iterations, adding the other objectives is followed by increases in GFAP binding and solubility scores and slight increases in hemolysis and non-fouling scores.The plots show average scores for 50 sampled sequences; permeability rises during its first 50 guided iterations.Source: §I.1 · p. 41, Figure 16 · p. 42

I.2. Ablation Studies

The ablation reports that bond-dependent masking and invalid loss are both critical to valid peptide-SMILES generation; the algorithms describe PepMDLM training and PepTune search procedures.

I.2. Ablation Studies

  • Both bond-dependent masking and invalid loss are critical for generating sequences that the decoder classifies as valid peptides.The ablation evaluates validity on unconditionally generated SMILES; Table 13 describes sampling 100 sequences per model.Source: §I.2 · p. 42, Table 13 · p. 42

J. Algorithms

  • PepMDLM training samples masking levels with a bond-dependent schedule and adds an invalid-sequence penalty to its NELBO objective.The training algorithm uses bond-mask information in the masking schedule and includes the invalid loss in the total loss.Source: J. Algorithms · p. 43
  • PepTune expands a selected partial sequence, rolls child sequences to completion, scores them, and backpropagates rewards while maintaining Pareto-optimal outputs.The algorithm returns sequences and classifier-score vectors from the Pareto front.Source: J. Algorithms · pp. 44, 47
  • The bond-mask routine marks ester, N-methyl, and peptide-bond spans, while SMILES2PEPTIDE checks SMILES format and peptide-bond patterns to identify peptides.The decoder is also described as a validity filter.Source: J. Algorithms · pp. 48–49

Related papers

Loading 2412.17780v4…