Research map / Probabilistic and scientific modeling / Molecules and proteins
Interaction-based Retrieval-augmented Diffusion Models for Protein-specific 3D Molecule 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 IRDIFF, a retrieval-augmented SE(3)-equivariant 3D diffusion model that retrieves high-binding-affinity ligand references via a pre-trained affinity network (PMINet) to steer protein-specific molecule generation.
Problem
Structure-based drug design: generating 3D ligand molecules that bind specific protein pockets with high affinity, where existing diffusion models generate from scratch without reference ligands and struggle to optimize binding affinity.
Approach
Pre-train a protein-molecule interaction network (PMINet) on binding affinity, use it to retrieve top-k high-affinity ligands from a database as references, and condition the 3D diffusion model on them via retrieval augmentation and self augmentation (using predicted molecule interaction features).
Key results
- IRDIFF (n=128, k=3) achieves Avg. Vina Score -6.03 and Avg. QED 0.53 on CrossDocked2020, versus -5.67 for DecompDiff and -5.47 for TargetDiff.
- IRDIFF surpasses Pocket2Mol by 17.3% and 46.6% in Avg. and Med. Vina Score, and DecompDiff by 6.3% and 14.1%.
- On average 67.4% of IRDIFF molecules bind better than the ground-truth ligand (High Affinity), versus 64.4% for DecompDiff and 58.1% for TargetDiff.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add retrieval-augmented diffusion generation | IRDIFF · CrossDocked2020 test | Avg. Vina Score: -5.67 -> -6.03 | Helped |
| add self augmentation (ablation) | IRDIFF · CrossDocked2020 test | Avg. Vina Score: -5.04 -> -4.91 | No clear effect |
| add self augmentation (ablation) | IRDIFF · CrossDocked2020 test | Avg. QED: 0.46 -> 0.52 | Helped |
| add retrieval augmentation (ablation) | IRDIFF · CrossDocked2020 test | Avg. Vina Score: -5.04 -> -5.39 | Helped |
| add retrieval augmentation (ablation) | IRDIFF · CrossDocked2020 test | Avg. QED: 0.46 -> 0.51 | Helped |
| add both self and retrieval augmentation (ablation) | IRDIFF · CrossDocked2020 test | Avg. Vina Score: -5.04 -> -5.86 | Helped |
| augment without pretrained PMINet (ablation) | IRDIFF · CrossDocked2020 test | Avg. Vina Score: -5.04 -> -4.14 | Hurt |
| increase retrieval database size from n=32 to n=128 (ablation) | IRDIFF · CrossDocked2020 test | Avg. Vina Score: -5.67 -> -5.86 | Helped |
| increase reference count from k=1 to k=3 (ablation) | IRDIFF · CrossDocked2020 test | Avg. Vina Score: -5.86 -> -6.03 | Helped |
| replace high-affinity reference with low-affinity reference (ablation) | IRDIFF · CrossDocked2020 test | Avg. Vina Score: -5.39 -> -4.12 | Hurt |
| use high-affinity reference instead of no reference (ablation) | IRDIFF · CrossDocked2020 test | Avg. Vina Score: -5.04 -> -5.39 | Helped |
| use IRDIFF instead of DecompDiff | CrossDocked2020 generation · CrossDocked2020 test | Avg. QED: 0.45 -> 0.53 | Helped |
Surprising findings
- Augmentation without the pre-trained PMINet hurt generation performance in ablation (Avg. Vina -4.14 vs baseline -5.04).
Limitations
- Using the full training set for retrieval would improve results but increases computational burden, so only 128 ligands are used.
- QED and SA are weakly emphasized as screening metrics by the authors.
- Retrieval database is fixed during training and testing; robustness only examined in appendix.
Experimental setup
Datasets: PDBbind v2016 CrossDocked2020
Benchmarks: CrossDocked2020
Metrics: Vina Score Vina Min Vina Dock High Affinity QED SA Diversity JSD of bond/atom distance distributions
Baselines: LiGAN GraphBP AR Pocket2Mol TargetDiff DecompDiff
Builds on: TargetDiff DDPM EGNN
Closest papers by meaning
- Aligning Target-Aware Molecule Diffusion Models with Exact Energy Optimization (NeurIPS 2024 · similarity 0.89)
- Latent Retrieval Augmented Generation of Cross-Domain Protein Binders (NeurIPS 2025 · similarity 0.88)
- Controllable 3D Molecular Generation for Structure-Based Drug Design Through Bayesian Flow Networks and Gradient Integration (NeurIPS 2025 · similarity 0.88)
- Binding-Adaptive Diffusion Models for Structure-Based Drug Design (AAAI 2024 · similarity 0.87)
- PPDiff: Diffusing in Hybrid Sequence-Structure Space for Protein-Protein Complex Design (ICML 2025 · similarity 0.87)
- Prior-Guided Flow Matching for Target-Aware Molecule Design with Learnable Atom Number (NeurIPS 2025 · similarity 0.87)
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.