This research introduces an improved in-context learning framework for molecular analysis that overcomes limitations of current methods by grounding graph-based representations with Morgan fingerprints.

• Eliminates need for extensive pretraining and fine-tuning of large language models
• Develops more effective prompt retrieval methods for molecular tasks
• Captures both local and global molecular structures for better property prediction
• Demonstrates improved performance in molecular property prediction and captioning tasks

For biology applications, this approach enables more accurate prediction of molecular properties without computational overhead, potentially accelerating drug discovery and material science research.

Graph-based Molecular In-context Learning Grounded on Morgan Fingerprints