Reproduce “Uni-Mol3: A Multi-Molecular Foundation Model for Advancing Organic Reaction Modeling” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Lirong Wu, Junjie Wang, Zhifeng Gao, Xiaohong Ji, Rong Zhu, Xinyu Li, Linfeng Zhang, Guolin Ke and 1 more
Uni-Mol3 predicts reaction products using molecular tokens that encode atom features, bonds, and 3D coordinates. It is pretrained first on individual molecules and then on reactions, with prompts for downstream tasks. On scaffold-split SMol-Reactions-FP it reaches 38.7% Top-1 accuracy, versus 37.2% for T5Chem. This benchmark result supports prediction performance, not demonstrated understanding of reaction mechanisms.
Key insights
- Integrating explicit 3D structural information via the Uni-Tokenizer significantly enhances molecular representations, improving performance across diverse reaction tasks, particularly for condition generation.
- A hierarchical pre-training strategy, moving from single-molecular grammar to multi-molecular reaction mechanisms, effectively builds foundational understanding and improves model generalizability.
- Multi-task fine-tuning can improve the validity of generated molecules and, for complex tasks like retrosynthesis and condition generation, can lead to performance gains compared to single-task approaches.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.