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

Conference: ICML 2025
arXiv: 2412.17780
Code: https://huggingface.co/ChatterjeeLab/PepTune
Area: Computational Biology
Keywords: Therapeutic peptide design, discrete diffusion, multi-objective optimization, MCTS, SMILES

TL;DR

PepTune combines a Masked Discrete Language Model (MDLM) with a Monte Carlo Tree Search (MCTS) multi-objective-guided strategy within the discrete peptide SMILES space to optimize multiple therapeutic properties (binding affinity, solubility, membrane permeability, etc.) simultaneously, enabling the de novo design of peptide drugs containing non-canonical amino acids and cyclization modifications.

Background & Motivation

Background: Peptide therapeutics (such as GLP-1 receptor agonists semaglutide/liraglutide) have achieved milestone successes in treating diseases like diabetes and obesity. Peptides are capable of binding diverse protein surfaces and disrupting protein-protein interactions; since 2000, 33 therapeutic peptides have been approved by the FDA.

Limitations of Prior Work: Designing peptides that simultaneously satisfy multiple conflicting objectives (e.g., binding affinity, solubility, and membrane permeability) remains a major challenge. Existing methods are limited to (1) continuous spaces, (2) unconditional generation, or (3) single-objective guidance. Traditional methods rely on screening random combinatorial libraries of \(10^{12}\) magnitude.

Key Challenge: Therapeutic peptides require non-canonical amino acids (nAAs) and cyclization modifications to improve stability and permeability, but existing deep learning models can only process 20 standard amino acids. Meanwhile, multi-objective guidance in discrete spaces is extremely difficult because gradients cannot be directly computed.

Goal: To build a diffusion model that can perform multi-objective conditional generation in the discrete peptide SMILES space.

Key Insight: (1) Representing peptides using SMILES instead of amino acid sequences to support nAAs and cyclization; (2) Using MCTS rather than gradient guidance to solve the challenge of discrete-space guidance.

Core Idea: MDLM is responsible for exploring valid structures within the discrete peptide space, while MCTS guides the generation process to evolve towards the Pareto-optimal direction across multiple therapeutic targets.

Method

Overall Architecture

• Input: Target protein sequence + list of therapeutic properties to be optimized
• First Stage: Pre-train PepMDLM (unconditional generator) on 11 million peptide SMILES
• Second Stage: Perform multi-objective conditional guidance on the generation process using an MCTS strategy guided by property classifiers
• Output: A set of Pareto-optimal peptide SMILES + corresponding property scores

Key Designs

1. State-dependent Masking Schedule:

   • Key Insight: Peptide bonds are the foundational structure of all valid peptides.
   • Designing a polynomial masking schedule: \(\alpha_t(\mathbf{x}_0) = 1 - t^w\) for peptide bond tokens, and \(\alpha_t = 1 - t\) for non-peptide bond tokens.
   • Peptide bond tokens are masked later in the forward process and unmasked earlier in the reverse process.
   • The loss weight of peptide bond tokens in training is scaled up by a factor of \(w\): \(\frac{w}{t} \log \langle \mathbf{x}_0, \mathbf{x}_\theta \rangle\).
   • Why: The vast majority of arbitrary SMILES are not valid peptides—letting the model first "build the backbone" before filling in side chains is crucial.

2. Global Sequence Invalidity Loss (Invalid Loss):

   • Take the argmax of predicted probabilities to obtain a discrete sequence, then check if it represents a valid peptide.
   • Invalid sequences backpropagate gradients using a penalty weighted by softmax probabilities.
   • \(\mathcal{L}_{\text{invalid}} = \sum_\ell \text{SM}(x_{\theta,k}^{(\ell)}) \cdot \mathbf{1}[\tilde{\mathbf{x}}_0 \text{ is Invalid}]\)
   • Why: Since argmax is non-differentiable, the softmax probability is used as a surrogate to bypass this bottleneck.

3. MCTS Multi-Objective Guidance:

   • Selection: Starting from the root node (fully masked), non-dominated child paths are selected based on cumulative rewards.
   • Expansion: Sample \(M\) different unmasking schemes using Gumbel noise.
   • Rollout: Greedily unmask to a complete sequence and calculate scores for \(K\) objectives using classifiers.
   • Backpropagation: The reward vector is backpropagated to update all nodes along the path.
   • Reward Definition: \(r_k(\mathbf{x}) = \frac{1}{|\mathcal{P}^*|} \sum_n \mathbb{I}[s_k(\mathbf{x}) \geq s_k(\tilde{\mathbf{x}}_n)]\) (the proportion of designs beaten in the Pareto set).
   • Penalty for invalid peptides: Deduct scores proportional to the invalidity ratio across all dimensions.
   • Why: Since the discrete space lacks gradients, MCTS achieves gradient-free multi-objective guidance through search and reward feedback.

Loss & Training

• Total Loss: \(\mathcal{L} = \mathcal{L}_{\text{NELBO}}^\infty + \mathcal{L}_{\text{invalid}}\)
• RoFormer Backbone: 8 layers, 768 hidden dimensions, 8 attention heads
• Training Data: 11 million peptide SMILES (synthetic data from SmProt + CycloPs)
• PeptideCLM SPE Tokenization: 581 tokens, averaging 4 characters/token
• 8×A6000 GPUs, 1600 GPU hours, AdamW, lr=3e-4
• MCTS: 128 iterations, 50 children/expansion

Key Experimental Results

Main Results

Target Protein	Property	Best Docking Score	Baseline/Control
TfR (Blood-brain barrier)	Binding + Solubility + Non-hemolytic	-8.4 kcal/mol	T7 peptide: -8.4 (but PepTune is shorter)
GLP-1R (Diabetes)	Binding + Solubility + Non-hemolytic	-7.4 kcal/mol	Semaglutide: -5.7 (longer)
GFAP (Alexander disease)	Binding + Permeability + Solubility + Non-hemolytic	-8.5 kcal/mol	No known peptide binder
TfR+GLAST (Dual-target)	Dual-target binding + Solubility	TfR: -10.5, GLAST: -9.2	First dual-target peptide design

Ablation Study

Configuration	Validity↑	Diversity	Description
PepMDLM (No Guidance)	45% (len=100)	0.705	Baseline unconditional model
PepTune (MCTS-Guided)	100% (after 20 iter)	0.677	MCTS boosts validity to 100%
HELM-GPT (Control)	83.9%	0.595	HELM representation, does not support nAAs
Without state-dependent mask	~30%	-	Validity drops significantly
Without invalid loss	~35%	-	Validity drops

Key Findings

• MCTS guidance achieves a 100% valid peptide generation rate after just 20 iterations.
• The generated peptides are rich in non-canonical amino acids (average of 2.94 per peptide) and cyclization structures.
• Multi-objective guidance does not significantly sacrifice diversity (dropping by only 0.03 compared to unconditional generation).
• Dual-target peptide experiments demonstrate that PepTune can optimize binding affinities for two proteins simultaneously.
• The docking score of the GLP-1R binder outperforms semaglutide with a shorter sequence.

Highlights & Insights

1. First Discrete-Space Multi-Objective Guided Diffusion: The integration of MCTS and masked diffusion is a significant contribution to the field.
2. Advantages of SMILES Representation: Supports nAAs and cyclization, greatly expanding the designable space of peptides.
3. Physical Intuition of State-Dependent Masking: The generation order of "backbone first, side chains second" aligns well with chemical intuition.
4. Clinical Relevance: Case studies on multiple real-world therapeutic targets demonstrate promising prospects for downstream experimental validation.

Limitations & Future Work

1. Reliance on synthetic peptide data (CycloPs); rare nAAs may increase the difficulty and cost of synthesis.
2. The quality of property classifiers is a bottleneck—properties like membrane permeability lack external validation.
3. MCTS sampling speed is relatively slow (128 iterations × 50 children × rollout).
4. Lack of wet-lab validation (only computational docking has been performed).

Related Work & Insights

• MDLM (Sahoo et al.) provides the foundation for masked discrete diffusion.
• Structure-based models like RFpeptides require the target's 3D structure, whereas PepTune only needs the sequence.
• Inspiration: The MCTS guidance strategy can be extended to other discrete biological sequence generation tasks such as protein design and DNA sequence optimization.

Rating

• Novelty: ⭐⭐⭐⭐⭐ The combination of MCTS + discrete diffusion + peptide SMILES is highly innovative.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ 5 single-target + 2 dual-target case studies, with detailed ablation details.
• Writing Quality: ⭐⭐⭐⭐ Rich content but somewhat lengthy.
• Value: ⭐⭐⭐⭐⭐ Significant practical value for drug design.

Related Papers

• [NeurIPS 2025] Uncertainty-Aware Multi-Objective Reinforcement Learning-Guided Diffusion Models for 3D De Novo Molecular Design
• [ICML 2025] Latent Imputation before Prediction: A New Computational Paradigm for De Novo Peptide Sequencing
• [NeurIPS 2025] De novo generation of functional terpene synthases using TpsGPT
• [ICML 2025] UniMoMo: Unified Generative Modeling of 3D Molecules for De Novo Binder Design
• [ICML 2025] Kinetic Langevin Diffusion for Crystalline Materials Generation