Research map / Probabilistic and scientific modeling / Molecules and proteins
MolEditRL: Structure-Preserving Molecular Editing via Discrete Diffusion and Reinforcement Learning
Yuanxin Zhuang, Dazhong Shen, Ying Sun
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SELFIES-based generation (BioT5, MolGen) instead of graph editing | BioT5/MolGen · MolEdit-Instruct | Acc_all (tau=0.65): 0 on most tasks vs 0.342 (GSK3b) | Hurt |
| use graph diffusion + RL (MolEditRL) over strongest baselines | MolEditRL · MolEdit-Instruct single and multi-property | Acc_all (tau=0.65): 0.342 vs 0.236 (GSK3b, DrugAssist) | Improved |
| use MolEditRL on unseen properties via RL with task oracles | MolEditRL · BBBP/HIA/hERG | Acc_all (tau=0.65): 0.474 vs 0.2992 (hERG down) | Improved |
| increase number of simultaneous target properties to 3 | MolEditRL · MolEdit-Instruct | Acc_all (tau=0.15): 0.363 (3 properties) | Hurt |
| use more reverse steps in pretraining (smaller denoising step size) (ablation) | Pretrain (no RL) · MolEdit-Instruct | accuracy | Improved |
| use stepwise DDPO instead of full-trajectory KL-regularized RL (ablation) | MolEditRL · MolEdit-Instruct | validity during training | Hurt |
| add KL regularization to GDPO (ablation) | GDPO · MolEdit-Instruct | training stability | Improved |
| add KL regularization to MolEditRL RL fine-tuning (ablation) | MolEditRL · MolEdit-Instruct | stability and accuracy | Improved |
Closest papers by meaning
- Searching for High-Value Molecules Using Reinforcement Learning and Transformers (ICLR 2024 · similarity 0.87)
- pCoMole: Pareto-Constrained Molecule Editing with Discrete Flows (NeurIPS 2026 · similarity 0.87)
- Chain-of-Generation: Progressive Latent Diffusion for Text-Guided Molecular Design (NeurIPS 2026 · similarity 0.87)
- Aligning Transformers with Continuous Feedback via Energy Rank Alignment (NeurIPS 2025 · similarity 0.86)
- Domain-Agnostic Molecular Generation with Chemical Feedback (ICLR 2024 · similarity 0.86)
- DMol: A Highly Efficient and Chemical Motif-Preserving Molecule Generation Platform (NeurIPS 2025 · similarity 0.86)
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.
