Research map / Probabilistic and scientific modeling / Molecules and proteins
Latent Retrieval Augmented Generation of Cross-Domain Protein Binders
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 RADiAnce, a retrieval-augmented latent diffusion framework for structure-based protein binder co-design, using a contrastively trained all-atom VAE to retrieve cross-domain interfaces (peptides, antibodies, protein fragments) that condition the diffusion generator.
Problem
Structure-based generative models for protein binder design condition only on the target binding site and lack rationality and interpretability, since they do not leverage known analogous interfaces.
Approach
Unifies retrieval and generation in a shared contrastive latent space: an all-atom VAE with contrastive loss embeds binding sites and binder interfaces, enabling dot-product retrieval of cross-domain interfaces, which then condition a latent diffusion model via cross-attention.
Key results
- On peptide codesign, RADiAnce reaches 39.42% AAR, 2.29 Å RMSD and 52.15% ISM, versus 38.69% AAR and 49.13% ISM for the next best baseline UniMoMo.
- On antibody CDR codesign, RADiAnce improves HCDR3 AAR to 54.66% and ISM to 71.64% (vs. UniMoMo 88.33/…); on L3 AAR reaches 76.75%.
- Contrastive VAE with cross-domain training reaches RC-5% of 67.74% for peptide retrieval and 100% for antibody retrieval.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add cross-domain training data, including peptides and protein fragments, to CVAE retrieval training (ablation) | CVAE retrieval for antibodies · Antibody retrieval | ITO (%): 28.21 -> 43.93 | Helped |
| add cross-domain training data, including antibodies and protein fragments, to CVAE retrieval training (ablation) | CVAE retrieval for peptides · Peptide retrieval | RC-5%: 25.81 -> 67.74 | Helped |
| add VAE reconstruction and KL losses to contrastive-only training (ablation) | CVAE retrieval · Antibody retrieval | RC-0.5%: 30.00 -> 96.67 | Helped |
| use top-N retrieval instead of no retrieval (ablation) | RADiAnce · HCDR3 codesign | AAR: 51.55 -> 54.66 | Helped |
| use top-N retrieval instead of no retrieval (ablation) | RADiAnce · Peptide codesign | AAR: 37.94 -> 39.12 | Helped |
| replace fixed top-N retrieval with adaptive similarity-cutoff retrieval (ablation) | RADiAnce · HCDR3 codesign | RMSD: 0.9682 -> 0.9147 | Helped |
| replace fixed top-N retrieval with adaptive similarity-cutoff retrieval (ablation) | RADiAnce · Peptide codesign | ISM (%): 52.15 -> 52.89 | Helped |
| use cross-attention instead of AdaLN-Zero conditioning (ablation) | RADiAnce · HCDR3 codesign | AAR (%): 50.42 -> 54.66 | Helped |
| use cross-attention instead of in-context prompt integration (ablation) | RADiAnce · HCDR3 codesign | RMSD: 1.019 -> 0.9443 | Helped |
| add antibody and protein-fragment templates to peptide-only retrieval (ablation) | RADiAnce · Peptide codesign | ISM (%): 49.82 -> 52.15 | Helped |
| add peptide and protein-fragment templates to antibody-only retrieval (ablation) | RADiAnce · Antibody H3 codesign | ΔΔG (kJ/mol): -5.200 -> -6.236 | Helped |
Surprising findings
- Random and reverse retrieval perform worse, and retrieving only one template gives the lowest diversity, while the no-retrieval baseline is markedly worse than any retrieval setting.
- Removing the VAE reconstruction loss (encoder-only contrastive model) substantially degrades retrieval performance.
Limitations
- Performance depends on retrieval quality, and the authors note structural descriptors could be improved.
- Antibody design without bound scaffold relies on AlphaFold3 docking and PyRosetta ΔG estimates, which are approximate.
Experimental setup
Datasets: PepBench LNR benchmark SAbDab RAbD ProtFrag
Benchmarks: LNR benchmark RAbD PepBench
Metrics: AAR RMSD ΔΔG IMP ISM ITO recall at top N% (RC) diversity DockQ FoldX ΔG
Baselines: RFDiffusion PepFlow PepGLAD UniMoMo MEAN DyMEAN DiffAb GeoAB-R GeoAB-D
Builds on: EPT (equivariant full-atom transformer) latent diffusion AdaLN-Zero DiffAb
Closest papers by meaning
- UniMoMo: Unified Generative Modeling of 3D Molecules for De Novo Binder Design (ICML 2025 · similarity 0.88)
- Interaction-based Retrieval-augmented Diffusion Models for Protein-specific 3D Molecule Generation (ICML 2024 · similarity 0.88)
- Full-Atom Peptide Design with Geometric Latent Diffusion (NeurIPS 2024 · similarity 0.86)
- Knowledge Enhanced Representation Learning for Drug Discovery (AAAI 2024 · similarity 0.86)
- Retrieval Augmented Diffusion Model for Structure-informed Antibody Design and Optimization (ICLR 2025 · similarity 0.85)
- CBGBench: Fill in the Blank of Protein-Molecule Complex Binding Graph (ICLR 2025 · similarity 0.85)
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.