Research map / Probabilistic and scientific modeling / Molecules and proteins
A Multi-Modal Contrastive Diffusion Model for Therapeutic Peptide Generation
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 MMCD, a multi-modal diffusion model that co-generates therapeutic peptide sequences and backbone structures, with inter- and intra-modal contrastive learning at each diffusion timestep.
Problem
Generating therapeutic peptides (antimicrobial and anticancer) with deep generative models that use only sequence or only structure information, under limited therapeutic peptide data.
Approach
Couples sequence (discrete) and structure (Gaussian/EGNN) diffusion models and adds per-timestep contrastive losses: Inter-CL aligns a peptide's sequence and structure embeddings, while Intra-CL separates therapeutic from non-therapeutic peptides using non-therapeutic peptides as negatives.
Key results
- On the AMP dataset, MMCD reaches 0.8810 antimicrobial score and 24.4107 similarity, beating the best baseline SimDiff (0.8560, 25.5385).
- For structure generation on AMP, MMCD improves RMSD to 1.8278 (23.3% better than DiffAB) and Ramachandran to 80.4661.
- Ablation shows removing Inter-CL or Intra-CL degrades all sequence metrics, and removing both lowers antimicrobial score to 0.8013.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add sequence-structure co-generation with Inter-CL and Intra-CL | MMCD · AMP and ACP | therapeutic score: 0.8013 -> 0.8810 (AMP); 0.6996 -> 0.7604 (ACP) | Helped |
| add Inter-CL to align sequence and structure embeddings (ablation) | MMCD · AMP and ACP | Antimicrobial / Anticancer: 0.8494 -> 0.8810 (AMP); 0.7352 -> 0.7604 (ACP) | Helped |
| add Intra-CL to distinguish therapeutic and non-therapeutic peptides (ablation) | MMCD · AMP and ACP | Antimicrobial / Anticancer: 0.8584 -> 0.8810 (AMP); 0.7513 -> 0.7604 (ACP) | Helped |
| add Inter-CL and Intra-CL (ablation) | MMCD · AMP and ACP | Antimicrobial / Anticancer: 0.8013 -> 0.8810 (AMP); 0.6996 -> 0.7604 (ACP) | Helped |
| use MMCD instead of SimDiff for sequence generation | AMP and ACP generation · AMP and ACP | Similarity: 25.5385 -> 24.4107 (AMP); 28.8245 -> 27.4685 (ACP) | Helped |
| use MMCD instead of SimDiff for structure generation | AMP generation · AMP | Docking: 1682 -> 1728 | Helped |
Limitations
- Non-therapeutic peptides serve as negatives, and augmentations of therapeutic peptides are avoided due to functional sensitivity, leaving positive-pair design limited.
- Structures for many peptides are predicted with Rosetta rather than experimentally determined.
- Docking evaluation covers only one reference peptide and one target protein.
Experimental setup
Datasets: AMP dataset (20,129 antimicrobial peptides) ACP dataset (4,381 anticancer peptides) non-therapeutic peptides from public databases
Benchmarks: AMP dataset ACP dataset
Metrics: similarity instability antimicrobial score anticancer score Ramachandran score RMSD docking score
Baselines: LSTM-RNN AMPGAN WAE-PSO HydrAMP APPTEST FoldingDiff ProtDiff DiffAB SimDiff
Builds on: denoising diffusion probabilistic models EGNN InfoNCE ProtDiff DiffAB
Closest papers by meaning
- Diffusion on Language Model Encodings for Protein Sequence Generation (ICML 2025 · similarity 0.87)
- PepTune: De Novo Generation of Therapeutic Peptides with Multi-Objective-Guided Discrete Diffusion (ICML 2025 · similarity 0.86)
- UniMoMo: Unified Generative Modeling of 3D Molecules for De Novo Binder Design (ICML 2025 · similarity 0.86)
- DPLM-2: A Multimodal Diffusion Protein Language Model (ICLR 2025 · similarity 0.86)
- Diffusion Language Models Are Versatile Protein Learners (ICML 2024 · similarity 0.86)
- Full-Atom Peptide Design with Geometric Latent Diffusion (NeurIPS 2024 · 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.