Research map / Probabilistic and scientific modeling / Molecules and proteins
GenMol: A Drug Discovery Generalist with Discrete Diffusion
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 GenMol, a BERT-based masked discrete diffusion model over SAFE sequences that handles de novo, fragment-constrained, hit, and lead optimization tasks with one checkpoint, using fragment remasking and molecular context guidance.
Problem
Existing molecular generative models each cover only one or two drug discovery scenarios, and autoregressive SAFE-based models (SAFE-GPT) are unnatural for order-free fragment sequences, slow, and need expensive RL finetuning for goal-directed generation.
Approach
Replace autoregressive SAFE generation with masked discrete diffusion (BERT denoiser trained with MDLM objective), add fragment remasking as a fragment-level exploration move that can be viewed as Gibbs sampling, and apply molecular context guidance (autoguidance by partially re-masking the input) to calibrate predictions, all without task-specific finetuning.
Key results
- On the 23-task PMO hit generation benchmark, GenMol reaches a sum AUC top-10 of 18.362 versus 16.928 for the best baseline f-RAG, best in 19 of 23 tasks.
- In de novo generation, GenMol (N=1, tau=0.5, r=0.5) achieves 84.6% quality versus 54.7% for SAFE-GPT, with a 2.5x shorter sampling time at N=3.
- GenMol optimizes seed molecules successfully in 26 of 30 lead optimization tasks, with lower docking scores than RetMol and Graph GA in most tasks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add fragment remasking to fragment attaching (ablation) | GenMol · PMO hit generation (23 tasks) | sum AUC top-10: 17.641 -> 18.208 | Helped |
| add token remasking to fragment attaching (ablation) | GenMol · PMO hit generation (23 tasks) | sum AUC top-10: 17.641 -> 18.091 | Helped |
| add SAFE-GPT fragment remasking to fragment attaching (ablation) | GenMol · PMO hit generation (23 tasks) | sum AUC top-10: 17.641 -> 18.074 | Helped |
| use fragment-level rather than token-level remasking (ablation) | GenMol with remasking · PMO hit generation (23 tasks) | sum AUC top-10: 18.091 -> 18.208 | Helped |
| use discrete diffusion rather than SAFE-GPT for fragment remasking (ablation) | GenMol with fragment remasking · PMO hit generation (23 tasks) | sum AUC top-10: 18.074 -> 18.208 | Helped |
| add molecular context guidance (ablation) | GenMol with fragment remasking · PMO hit generation (23 tasks) | sum AUC top-10: 18.208 -> 18.362 | Helped |
| add molecular context guidance (ablation) | GenMol · Motif extension | Quality (%): 27.5~$\pm$~0.8 -> 30.1~$\pm$~0.4 | Helped |
| increase tokens unmasked per step to N>1 | GenMol · Molecule generation | generation quality | Hurt |
| increase softmax temperature and randomness | GenMol · De novo generation | Quality (%): 84.6~$\pm$~0.8 -> 63.0~$\pm$~0.4 | Hurt |
| use GenMol instead of SAFE-GPT | molecular generation · Scaffold decoration | Quality (%): 10.0~$\pm$~1.4 -> 31.8~$\pm$~0.5 | Helped |
| use GenMol instead of SAFE-GPT | molecular generation · Superstructure generation | Quality (%): 14.3~$\pm$~3.7 -> 34.8~$\pm$~1.0 | Helped |
| use GenMol instead of f-RAG | goal-directed hit generation · PMO hit generation (23 tasks) | sum AUC top-10: 16.928 -> 18.362 | Helped |
Surprising findings
- Unmasking more than one token per step (N>1) degrades generation quality, creating a quality-speed trade-off.
Limitations
- Unmasking more than one token per step (N>1) degrades quality, limiting the speed-quality trade-off.
- Hyperparameters such as the MCG guidance ratio gamma are tuned per task via search.
Experimental setup
Datasets: SAFE dataset ZINC250k ZINC UniChem DUD-E ChEMBL
Benchmarks: PMO benchmark fragment-constrained generation benchmark (linker design, scaffold morphing, motif extension, scaffold decoration, super de novo generation (validity, uniqueness, quality, diversity) DUD-E lead optimization docking
Metrics: validity uniqueness quality diversity PMO AUC top-10 docking score sampling time
Baselines: SAFE-GPT f-RAG Genetic GFN Mol GA REINVENT Graph GA RetMol JT-VAE DiGress GP BO SELFIES-REINVENT
Builds on: MDLM SAFE autoguidance classifier-free guidance D3PM
Closest papers by meaning
- FlashMol: High-Quality Molecule Generation in as Few as Four Steps (NeurIPS 2026 · similarity 0.89)
- Drug Discovery with Dynamic Goal-aware Fragments (ICML 2024 · similarity 0.88)
- DMol: A Highly Efficient and Chemical Motif-Preserving Molecule Generation Platform (NeurIPS 2025 · similarity 0.88)
- Molecule Generation with Fragment Retrieval Augmentation (NeurIPS 2024 · similarity 0.88)
- LDMol: A Text-to-Molecule Diffusion Model with Structurally Informative Latent Space Surpasses AR Models (ICML 2025 · similarity 0.87)
- DecompOpt: Controllable and Decomposed Diffusion Models for Structure-based Molecular Optimization (ICLR 2024 · 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.