DrugSAGE: Self-Evolving Agent for Efficient Drug Discovery
The DrugSAGE (Self-evolving Agent Experience) framework is designed to efficiently create cutting-edge drug discovery models by gathering and reapplying experiences across various tasks. It retains a cross-task memory that includes validated skills, statistical data on successful strategies, and documentation of frequent mistakes along with their resolutions. In certain instances, DrugSAGE can directly apply a successful solution without needing to search during testing. In a comparison of 33 molecular property prediction tasks, DrugSAGE secured the top position among nine state-of-the-art agents in a single-task environment. By leveraging memory from 16 smaller tasks, DrugSAGE demonstrates an average enhancement in performance.
Key facts
- DrugSAGE is a Self-evolving Agent Experience framework for drug discovery.
- It accumulates and reuses experience across tasks.
- It maintains cross-task memory of verified skills, statistical evidence, and error records.
- In some cases, it transfers working solutions without test-time search.
- Tested on 33 molecular property prediction tasks.
- Ranked first among nine SOTA agents in single-task setting.
- Memory accumulated from 16 smaller tasks improves performance.
- The framework aims to reduce search costs for building SOTA models.
Entities
—