MolLingo: Molecule-Native Representations for LLM-Powered Scientific Agents
MolLingo, presented in arXiv:2605.27853, introduces a multi-agent system for automated molecular design. It coordinates three agents—Literature Agent, Chemist Agent, and Orchestrator—through a shared memory module, each equipped with domain-specific tools. The key innovation is BRICS-based Fragment Enumeration (BFE), a synthesis-aware fragmentation method that decomposes molecules into chemically meaningful building blocks represented as block-based SMILES paired with common chemical names. This representation bridges molecular structure and LLM semantic space, enabling block-level reasoning and editing that is difficult with raw SMILES alone. As a case study in early-stage therapeutic design, MolLingo grounds the Chemist Agent's reasoning in binding site geometry and residue-level protein context. The system addresses limitations of existing LLM approaches that lack multi-agent coordination and shared memory for iterative, evidence-driven reasoning across the molecular design pipeline.
Enables LLMs to reason about molecules at a chemically meaningful level, improving automated drug design.