An effective self-supervised framework for learning expressive molecular global representations to drug discovery
Problem
Producing expressive molecular representations for drug discovery is hard because labeled data is scarce and supervised models generalize poorly.
Why it matters
Better molecular representations can accelerate drug discovery where experimental labeled data is scarce and expensive.
What the AI did
MPG pre-trains the MolGNet graph neural network with self-supervised node- and graph-level strategies on 11 million unlabeled molecules, then fine-tunes it with one output layer for downstream tasks.
Results
Pre-trained on 11 million unlabeled molecules; 'state-of-the-art' models claimed, no concrete metrics visible in abstract
Impact
Provides a pre-trained framework claimed to yield state-of-the-art models across drug discovery tasks.
Record
| Venue | Briefings in Bioinformatics |
|---|---|
| Published | 2021 |
| Work type | article |
| Authors | Pengyong Li, Jun Wang, Yixuan Qiao, Hao Chen, Yihuan Yu, Xiaojun Yao, Peng Gao, Guotong Xie et al. |
| Citations | 151 (OpenAlex, at import time) |
| References | 88 |
| Model or method | MPG / MolGNet (graph neural network) |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1093/bib/bbab109 |
| Open access | closed |
Importance rating
significant — thousands to millions affected
Expressive molecular representations are fundamental to AI-driven drug discovery where labeled data is scarce; abstract claims state-of-the-art across 14 benchmark tasks after pre-training on 11M molecules, but no impact scale stated.