SMDD-Bench: Benchmarking LLM Agents for Drug Design
SMDD-Bench has been launched by researchers as a benchmark for assessing LLM agents in practical small molecule drug design challenges. This benchmark features 502 task instances that are guaranteed to be solvable, categorized into five distinct types: Scaffold Hopping, 2D Pharmacophore Identification, Interaction Point Discovery, Lead Optimization, and Fragment Assembly. It covers an extensive chemical space and includes 102 different protein targets. Successfully addressing the benchmark demands robust reasoning in chemistry and biology, along with a solid understanding of 3D concepts.
Key facts
- SMDD-Bench is a multi-turn, long-horizon agentic benchmark for LLM agents.
- It consists of 502 task instances across 5 task types.
- Tasks involve 102 unique protein targets.
- The benchmark covers 2D Pharmacophore Identification, Interaction Point Discovery, Scaffold Hopping, Lead Optimization, and Fragment Assembly.
- All tasks are guaranteed solvable.
- The benchmark aims to standardize evaluation of LLM agents on small molecule drug design.
- Current evaluation methods are ad hoc, too simple, limited in scale, or single-turn.
- Solving the benchmark requires chemical and biological reasoning and 3D intuition.
Entities
Institutions
- arXiv