Research map / Image generation / Fast diffusion sampling
PepTune: De Novo Generation of Therapeutic Peptides with Multi-Objective-Guided Discrete Diffusion
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PepTune, a multi-objective discrete diffusion model for therapeutic peptide SMILES generation using a bond-dependent masking schedule, an invalid loss, and inference-time Monte Carlo Tree Guidance over property classifiers.
Problem
De novo generation of chemically modified (non-natural, cyclic) therapeutic peptides optimized jointly for multiple properties (binding affinity, permeability, solubility, hemolysis, non-fouling), where discrete generative models struggle with conflicting objectives, gradient estimation, and sparse
Approach
Combines a masked discrete diffusion LM (MDLM) with a bond-dependent masking schedule that masks peptide-bond tokens more slowly, an invalid-SMILES loss propagated via softmax, and Monte Carlo Tree Guidance that uses Pareto-based rewards over property classifiers at inference without gradients or retraining.
Key results
- PepTune reaches validity 1.000 on peptide SMILES versus 0.450 for PepMDLM, with uniqueness 1.000 and SNN 0.486.
- Property classifiers improve over PeptideBERT: solubility F1 0.660 vs 0.597, hemolysis F1 0.846 vs 0.483, non-fouling F1 0.768 vs 0.699.
- Dual-target TfR/GLAST binders reach Vina docking scores as low as -10.5 kcal/mol (TfR) and -9.2 kcal/mol (GLAST), with all seven selected binders at or below -7.5 kcal/mol.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply PepTune multi-objective guidance instead of unconditional PepMDLM generation | PepTune · Peptide SMILES (max token length 200) | validity: 0.450 -> 1.000 | Helped |
| Apply PepTune multi-objective guidance instead of unconditional PepMDLM generation | PepTune · Peptide SMILES (max token length 200) | diversity: 0.705 -> 0.677 | Hurt |
| Apply PepTune multi-objective guidance instead of unconditional PepMDLM generation | PepTune · Peptide SMILES (max token length 200) | SNN: 0.513 -> 0.486 | Helped |
| Apply PepTune multi-objective guidance instead of unconditional PepMDLM generation | PepTune · Peptide SMILES (max token length 200) | randomness: 4.11 -> 4.12 | Helped |
| Apply PepTune multi-objective guidance instead of unconditional PepMDLM generation | PepTune · Peptide SMILES (max token length 200) | KL-divergence: 0.174 -> 0.173 | Hurt |
| Use the peptide-SMILES property predictors instead of PeptideBERT | Solubility classifier · Solubility | F1: 0.597 -> 0.660 | Helped |
| Use the peptide-SMILES property predictors instead of PeptideBERT | Hemolysis classifier · Hemolysis | F1: 0.483 -> 0.846 | Helped |
| Use the peptide-SMILES property predictors instead of PeptideBERT | Non-fouling classifier · Non-fouling | F1: 0.699 -> 0.768 | Helped |
| Use the peptide-SMILES property predictors instead of PeptideBERT | Solubility classifier · Solubility | accuracy: 0.651 -> 0.661 | Helped |
| Use the peptide-SMILES property predictors instead of PeptideBERT | Hemolysis classifier · Hemolysis | accuracy: 0.823 -> 0.846 | Helped |
| Use the peptide-SMILES property predictors instead of PeptideBERT | Non-fouling classifier · Non-fouling | accuracy: 0.873 -> 0.766 | Hurt |
Surprising findings
- Non-fouling accuracy for PeptideBERT (0.873) is higher than the authors' model (0.766), so the classifier is not uniformly better.
- PepMDLM's validity is only 0.450, indicating that much of the unconditional model's output is invalid without the bond-dependent schedule and invalid loss.
Limitations
- Docking scores from AutoDock Vina are in silico only; no experimental validation of binding is reported.
- Property predictors are trained on limited peptide SMILES data, and the PepTune diversity score (0.677) is lower than the data (0.885).
Experimental setup
Datasets: peptide SMILES corpus (11 million) PAMPA permeability data (34,853 peptides)
Benchmarks: peptide SMILES generation quality (validity, uniqueness, diversity, SNN, randomness, KL-divergence) PeptideBERT-comparison property prediction (F1, accuracy)
Metrics: validity uniqueness diversity SNN F1 accuracy Vina docking score
Baselines: PepMDLM HELM-GPT PeptideBERT semaglutide liraglutide T7 peptide
Builds on: MDLM Monte Carlo Tree Search RoFormer PeptideBERT Autodock Vina
Closest papers by meaning
- Full-Atom Peptide Design with Geometric Latent Diffusion (NeurIPS 2024 · similarity 0.88)
- A Multi-Modal Contrastive Diffusion Model for Therapeutic Peptide Generation (AAAI 2024 · similarity 0.86)
- UniMoMo: Unified Generative Modeling of 3D Molecules for De Novo Binder Design (ICML 2025 · similarity 0.86)
- Steering Masked Discrete Diffusion Models via Discrete Denoising Posterior Prediction (ICLR 2025 · similarity 0.86)
- Zero-Shot Cyclic Peptide Design via Composable Geometric Constraints (ICML 2025 · similarity 0.86)
- Reward-Guided Iterative Refinement in Diffusion Models at Test-Time with Applications to Protein and DNA Design (ICML 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.