Research map / Probabilistic and scientific modeling / Molecules and proteins
Structure-based molecule design with diffusion
Generate pocket-conditioned and synthesizable ligands with diffusion on 3D structure.
48 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2019–2026) · an approach in Molecules and proteins, keyword cluster “molecules · drug · chemical”
State History Tried Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
15 papers in 2023–24 and 31 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 3.1% to 3.3%.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2019 · Learning Vine Copula Models for Synthetic Data Generation (AAAI 2019 · 48 citations)
- 2021 · Mol2Image: Improved Conditional Flow Models for Molecule to Image Synthesis (CVPR 2021 · 22 citations)
- 2023 · DecompDiff: Diffusion Models with Decomposed Priors for Structure-Based Drug Design (ICML 2023 · 26 citations)
- 2024 · Deep Confident Steps to New Pockets: Strategies for Docking Generalization (ICLR 2024 · 28 citations)
- 2025 · DiffMS: Diffusion Generation of Molecules Conditioned on Mass Spectra (ICML 2025 · 10 citations)
- 2026 · Graph Diffusion Transformers are In-Context Molecular Designers (ICLR 2026 · 1 citation)
Earliest papers here
- Learning Vine Copula Models for Synthetic Data Generation (AAAI 2019 · 48 citations)
- Mol2Image: Improved Conditional Flow Models for Molecule to Image Synthesis (CVPR 2021 · 22 citations)
- DecompDiff: Diffusion Models with Decomposed Priors for Structure-Based Drug Design (ICML 2023 · 26 citations)
What has been tried
25 papers here report 260 measured changes (a component added, removed or swapped, compared with a baseline): helped 79%, no clear effect 3%, hurt 17%, mixed 1%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| add Inter-CL to align sequence and structure embeddings | AMP and ACP · Antimicrobial / Anticancer | 0.8494 -> 0.8810 (AMP); 0.7352 -> 0.7604 (ACP) | Helped | A Multi-Modal Contrastive Diffusion Model for Therapeutic Peptide Generation (AAAI 2024 · 21 citations) |
| Add phase-two correction | ChEBI-20 test · Validity | 0.789 -> 0.871 | Helped | Text-Guided Molecule Generation with Diffusion Language Model (AAAI 2024 · 20 citations) |
| remove iterative cross-hierarchy interaction | CrossDocked2020 test · Vina Score Avg. | -5.92 -> -5.04 | Hurt | Binding-Adaptive Diffusion Models for Structure-Based Drug Design (AAAI 2024 · 11 citations) |
| add spectrum-to-fingerprint encoder pretraining | NPLIB1 Top-1 · accuracy | 4.36% -> 8.34% | Helped | DiffMS: Diffusion Generation of Molecules Conditioned on Mass Spectra (ICML 2025 · 10 citations) |
| optimize QED as the single objective | Single-objective QED optimization · QED (Avg.) | 0.48 -> 0.52 | Helped | DecompOpt: Controllable and Decomposed Diffusion Models for Structure-based Molecular Optimization (ICLR 2024 · 9 citations) |
| add fragment remasking to fragment attaching | PMO hit generation (23 tasks) · sum AUC top-10 | 17.641 -> 18.208 | Helped | GenMol: A Drug Discovery Generalist with Discrete Diffusion (ICML 2025 · 4 citations) |
| Add contrastive encoder pre-training | ChEBI-20 generation · Validity | 0.019 -> 0.941 | Helped | LDMol: A Text-to-Molecule Diffusion Model with Structurally Informative Latent Space Surpasses AR Models (ICML 2025 · 3 citations) |
| Add property prediction loss | QM9 · Molecule stability (%) | 82.0 -> 85.7 | Helped | UniGEM: A Unified Approach to Generation and Property Prediction for Molecules (ICLR 2025 · 3 citations) |
| adding SBE energy guidance (contrastive-trained energy, active compounds as negatives) to TargetDiff diffusion | CrossDocked2020 · delta score mean (lower better in Table 1 sense; higher in Table 2 reported) | 0.335 -> 0.552 | Helped | Rethinking Specificity in SBDD: Leveraging Delta Score and Energy-Guided Diffusion (ICML 2024 · 2 citations) |
| add hard fragment retrieval to the soft+GA variant | PMO benchmark (23 tasks) · sum AUC top-10 | 15.231 -> 16.928 | Helped | Molecule Generation with Fragment Retrieval Augmentation (NeurIPS 2024 · 2 citations) |
| add E to -log P_Retro in route ranking | RetroBench test · Top-1 exact match accuracy (%) | 42.0 -> 44.5 | Helped | Preference Optimization for Molecule Synthesis with Conditional Residual Energy-based Models (ICML 2024 · 1 citation) |
| Add Synthesis Cost Guidance | sEH SMALL · synthesis cost | 37.1 \std{1.0} -> 23.7 \std{4.0} | Helped | Scalable and Cost-Efficient de Novo Template-Based Molecular Generation (NeurIPS 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 14 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- ChEMBL · 2 of 14 papers
- PepBench · 2 of 14 papers
- ProtFrag · 2 of 14 papers
- SAbDab · 2 of 14 papers
- RAbD · 2 of 14 papers
- ChEBI-20 · 2 of 14 papers
- PubChem · 2 of 14 papers
- QM9 · 2 of 14 papers
Baselines
- RFDiffusion · 3 of 14 papers
- PepFlow · 2 of 14 papers
- PepGLAD · 2 of 14 papers
- MEAN · 2 of 14 papers
- dyMEAN · 2 of 14 papers
- DiffAb · 2 of 14 papers
- MolT5-large · 2 of 14 papers
- EDM · 2 of 14 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 46 papers; from OpenAlex and the papers' affiliations).
- Tsinghua University · 7 papers
- Chinese Academy of Sciences · 7 papers
- Peking University · 6 papers
- Korea Advanced Institute of Science and Technology · 5 papers
- ByteDance · 4 papers
- Massachusetts Institute of Technology · 4 papers
- University of Chinese Academy of Sciences · 3 papers
- NVIDIA · 3 papers
- Institute of Automation， CAS，China · 3 papers
- Peng Cheng Laboratory · 2 papers
Use it
Download BibTeX of 26 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Structure-based 3D molecule generation (107 papers)
- Diffusion for molecular conformations (67 papers)
- Machine-learned interatomic potentials (74 papers)
- Graph models for retrosynthesis (10 papers)
Elsewhere in the topic
- Generative protein structure design (142 papers)
- Prediction-powered and conformal inference (129 papers)
- Statistical analysis of ML practice (122 papers)
- Neuro-symbolic probabilistic layers (120 papers)
- Sparse graphical and tensor models (97 papers)
- Flow matching generative modeling (97 papers)
- Machine learning for physical simulation (96 papers)
- Multimodal trajectory prediction (95 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).