Large Language Molecular Representation and Learning

Explore the potential of molecular machine learning for enhancing property prediction and drug discovery in this comprehensive lecture. Delve into the challenges of obtaining labeled molecule data and the debate surrounding molecular representation methods. Learn about MolCLR, a self-supervised learning framework utilizing graph neural networks to process millions of unlabeled molecules. Examine the effectiveness of textual representations for polymers, Metal Organic Frameworks, catalysis systems, and organic molecules compared to graph representations. Discover how pre-trained language models like BERT and RoBERTa can be leveraged for molecular property prediction. Gain insights into multimodal learning approaches for molecules and their potential to improve the learning process. The lecture covers topics such as molecular property prediction, ablation studies, transformer models, MOFormer, textual format input exploration, prompt engineering, and concludes with a Q&A session.

- Intro
- Molecular property prediction
- Ablation Study
- Improving MolCLR
- Transformers
- MOFormer
- Exploring features through textual format input
- Prompt engineering
- Conclusions
- Q+A