🧬 MoleCode

An LLM-native, graph-explicit molecular language

MoleCode presents molecules as code so LLMs can read, write, edit, and reason on chemistry directly — instead of reconstructing structure from cryptic strings.

Paper (arXiv:2605.16480) · GitHub · AtomFlow

What is MoleCode?

A molecule is a graph: atoms are nodes, bonds are edges. Yet LLMs are usually fed molecules as linear strings (SMILES) where the graph is implicit — so the model must rebuild the structure from syntax before it can reason.

MoleCode makes the structure the language. Every atom and bond is a typed declaration with a persistent identifier, serialized as a Mermaid graph, and deterministically, losslessly inter-convertible with SMILES / MOL via RDKit (no learned model). One grammar spans small molecules, polymers, and Markush structures.

Install

pip install molecode

Quick start

from rdkit import Chem
from molecode import mol_to_mermaid, mermaid_to_mol, mol_to_smiles

# SMILES -> MoleCode graph
graph = mol_to_mermaid(Chem.MolFromSmiles("CC(=O)Oc1ccccc1C(=O)O"), name="Aspirin")
print(graph)

# MoleCode graph -> SMILES (lossless round-trip)
assert mol_to_smiles(mermaid_to_mol(graph)) == Chem.CanonSmiles("CC(=O)Oc1ccccc1C(=O)O")

Three domains, one grammar:

from molecode.polymer import polymer_to_mermaid, mermaid_to_psmiles   # polymers (×n repeat unit)
from molecode.markush import MoleCodeGraph, molecode_isomorphic        # Markush ({R1}/{Boc} nodes)
from molecode.prompts import MOLECULE_SYSTEM_PROMPT, MARKUSH_SYSTEM_PROMPT  # LLM system prompts
from molecode import LLMClient                                         # optional OpenAI-compatible client

Why it matters

• Generalization, not memorization. On novel molecules SMILES accuracy falls to ~20% while MoleCode holds ~76–80% (same model, same knowledge).
• Cheaper reasoning. Sub-linear chain-of-thought growth — ~5× lower total token cost per query than SMILES.
• Scales to big, repetitive objects. Full-chain SMILES collapses toward 0% as polymer chains grow; MoleCode stays flat.

For coding agents

MoleCode also ships as an Agent Skill (Claude Code auto-discovers it; Codex reads AGENTS.md) with a CLI for SMILES/PSMILES/Markush ↔ MoleCode conversion, validation, isomorphism comparison, and OCSR (molecule image → MoleCode). See the repository.

Citation

@article{yan2026molecode,
  title={MoleCode unlocks structural intelligence in large language models},
  author={Yan, Zhiyuan and Liu, Chen and Zhao, Boxuan and Lin, Kaiqing and Zhao, Jixiang and Wang, Yimi and Lv, Liuzhenghao and Li, Hao and Zhang, Shanzhuo and Yuan, Li and others},
  journal={arXiv preprint arXiv:2605.16480},
  year={2026}
}

License

MIT © 2026 AtomFlow-AI