Research map / Probabilistic and scientific modeling / Molecules and proteins
UniMoMo: Unified Generative Modeling of 3D Molecules for De Novo Binder Design
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 UniMoMo, a single generative model that designs peptide, antibody, and small-molecule binders by representing molecules as graphs of blocks and running latent diffusion over an iterative full-atom VAE.
Problem
De novo design of target-specific binders is usually handled by domain-specific generative models for small molecules, peptides, or antibodies; there is no single framework covering all of them or exploiting cross-domain transfer.
Approach
Unify peptides, antibodies, and small molecules as graphs of blocks (amino acids or principal-subgraph fragments), compress each block into a latent point via an iterative full-atom VAE, and run an E(3)-equivariant latent diffusion model conditioned on the binding site, trained jointly across domains.
Key results
- On peptide design, UniMoMo (all) reaches 39.45% AAR and 40.86% IMP versus 38.62% AAR and 18.28% IMP for the best baseline PepGLAD.
- On antibody CDR-H3 with 100 generations, UniMoMo (all) achieves 52.34% AAR and 65.00% IMP, the best among compared methods.
- On small-molecule CBGBench, UniMoMo (all) ranks first with a weighted score of 10.38 versus 9.62 for MolCRAFT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train on peptide, antibody, and small-molecule data instead of peptide-only data (ablation) | UniMoMo · LNR peptide test | ΔG: -28.72 -> -34.35 | Helped |
| train on peptide, antibody, and small-molecule data instead of antibody-only data (ablation) | UniMoMo · CDR-H3, 100 generations | AAR: 48.78% -> 52.34% | Helped |
| train on peptide, antibody, and small-molecule data instead of small-molecule-only data (ablation) | UniMoMo · Crossdocked2020 test | Weighted Score: 9.03 -> 10.38 | Helped |
| add extra noise to latent coordinates before decoding (ablation) | UniMoMo · Crossdocked2020 test | Vina score energy: -4.62 -> -5.72 | Helped |
| increase latent-coordinate KL weight to λ₂=1.0 without extra noise (ablation) | UniMoMo · Crossdocked2020 test | Vina score energy: -4.62 -> -4.42 | Hurt |
| increase decoder iterations from 1 to 10 (ablation) | UniMoMo · Peptide | JSD_bb: 0.393 -> 0.205 | Helped |
| increase decoder iterations from 1 to 10 (ablation) | UniMoMo · Antibody | JSD_bb: 0.376 -> 0.224 | Helped |
| increase decoder iterations from 1 to 10 (ablation) | UniMoMo · Small molecule | JSD_BL: 0.4216 -> 0.3223 | Helped |
Surprising findings
- Multi-domain training improves performance over single-domain training across peptides, antibodies, and small molecules.
- A small molecule generated for a GPCR pocket mimics an arginine side chain, suggesting transfer of interaction patterns from amino acids.
Limitations
- Diversity decreases slightly relative to some baselines due to the fidelity-diversity trade-off.
- Model-predicted inter-block bonds can produce invalid valencies or 2D-3D inconsistencies that require post-hoc filtering.
- Antibody and peptide AAR is regarded as unreliable and is only reported for completeness.
Experimental setup
Datasets: PepBench ProtFrag LNR SAbDab RAbD CrossDocked2020 ChEMBL
Benchmarks: LNR RAbD CrossDocked2020 CBGBench
Metrics: AAR C-RMSD L-RMSD pyRosetta binding energy IMP Clash rate JSD dihedral diversity RMSD Vina score
Baselines: RFDiffusion PepFlow PepGLAD MEAN dyMEAN DiffAb GeoAB LIGAN 3DSBDD GraphBP Pocket2Mol TargetDiff
Builds on: PepGLAD GET principal subgraph algorithm GeoDiff CBGBench
Closest papers by meaning
- BindGPT: A Scalable Framework for 3D Molecular Design via Language Modeling and Reinforcement Learning (AAAI 2025 · similarity 0.90)
- Full-Atom Peptide Design with Geometric Latent Diffusion (NeurIPS 2024 · similarity 0.89)
- Unified Guidance for Geometry-Conditioned Molecular Generation (NeurIPS 2024 · similarity 0.89)
- Latent Retrieval Augmented Generation of Cross-Domain Protein Binders (NeurIPS 2025 · similarity 0.88)
- CBGBench: Fill in the Blank of Protein-Molecule Complex Binding Graph (ICLR 2025 · similarity 0.88)
- Navigating the Design Space of Equivariant Diffusion-Based Generative Models for De Novo 3D Molecule Generation (ICLR 2024 · similarity 0.88)
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.