MolLingo: Molecule-Native Representations for LLM-Powered Scientific Agents

Researchers introduce MolLingo, a multi-agent AI system that automates molecular design by coordinating specialized agents through shared memory and domain-specific tools. The system uses BRICS-based Fragment Enumeration to represent molecules in chemically meaningful ways that LLMs can reason about effectively, achieving superior performance on drug design benchmarks compared to frontier models like GPT-5.

MolLingo represents a significant advancement in applying large language models to scientific discovery, specifically molecular design and drug development. The system addresses a critical limitation of existing LLM approaches by implementing multi-agent coordination with specialized tools rather than treating language models as standalone generative systems. This architecture enables iterative, evidence-driven reasoning that mirrors how chemists actually approach molecular optimization problems.

The introduction of BRICS-based Fragment Enumeration is the technical innovation that bridges the gap between molecular chemistry and LLM semantic space. By decomposing molecules into chemically meaningful building blocks paired with natural language names, the system allows language models to reason at a block level rather than character-level SMILES strings, fundamentally improving interpretability and reasoning quality. Grounding the system in molecular docking data and protein binding site geometry further anchors the AI's reasoning in biological reality rather than abstract chemical space.

The performance results demonstrate substantial practical impact: a fourfold improvement in docking scores over GPT-5 using the same underlying model indicates that the representational framework, not just raw model capability, drives performance gains. Consistent improvements across multiple LLM backbones and state-of-the-art results on specialized benchmarks suggest the approach generalizes effectively. For the pharmaceutical and biotechnology industries, these capabilities could accelerate early-stage drug design workflows and reduce computational screening costs.

The open-source release of MolLingo signals potential for broader adoption in academic and industrial research settings. Future development likely focuses on scaling to larger chemical libraries, integrating additional biological constraints, and validating designs through experimental validation rather than computational scoring alone.

• →MolLingo's multi-agent architecture with shared memory enables iterative, evidence-driven molecular design that outperforms frontier LLMs on drug discovery tasks.
• →BRICS-based Fragment Enumeration bridges molecular chemistry and LLM semantic space by representing molecules as chemically meaningful building blocks rather than raw SMILES strings.
• →The system achieves fourfold docking score improvements over GPT-5 using identical underlying models, indicating representational framework drives performance gains.
• →Integration of molecular docking and protein binding site geometry grounds LLM reasoning in biological context for optimized target binding.
• →Open-source release enables broader adoption in pharmaceutical research and could accelerate early-stage drug design workflows.