SciGroveBeta
Cheminformatics

How Well Can Frontier Large Language Models Generate Structures? High Quality Prediction of Molecular Geometries with Help from Fine-Tuning

Joseph M. Cavanagh, Jonathan B. Arnold, Gian-Bautista Alteri, Andrew Gritsevskiy, Teresa Head-Gordon

Featured August 1, 2026

Original Paper

This analysis was generated by SciGrove. Upload your own PDFs or enter a DOI — and get the same AI breakdown on any paper.

Get started

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

Teaching smart computer programs to understand molecule shapes using Z-matrix instructions helps them accurately predict 3D structures and diverse forms, while still letting them answer normal questions.

In depth
The authors demonstrate that Large Language Models (LLMs) can be fine-tuned to accurately predict 3D molecular geometries and diverse conformers from 2D chemical structures. They show that representing molecular geometries as Z-matrices (internal coordinates) is superior to Cartesian coordinates for LLM adaptation, and a pseudorehearsal strategy allows the model to retain its original natural language abilities.

Key Takeaways

  • 1
    Fine-tuning LLMs with Z-matrix representations enables high-quality prediction of molecular geometries and conformational ensembles, outperforming specialized deep learning models.
  • 2
    The GeomLlama model, a fine-tuned Llama-3.1-8B-Instruct, achieves excellent coverage and precision in generating conformers for small organic and drug-like molecules.
  • 3
    A pseudorehearsal strategy, involving mixing natural language data during fine-tuning, successfully mitigates catastrophic forgetting, allowing GeomLlama to retain its general language modeling capabilities.

Conceptual Flow

HIGH LEVEL
1
Methodology: Teaching LLMs Molecular Geometry

The paper teaches a smart computer program to understand molecule shapes by showing it many examples of molecule names and their 3D building instructions.

Molecule Name
Teach Program
3D Building Instructions
2
Results: High-Quality 3D Shapes and Retained Language Skills

The new method creates much more accurate 3D molecule shapes than old methods, especially when using special building instructions, and still understands normal questions.

Old Way
New Way
Compare Results
Less Accurate Shapes
More Accurate Shapes + Understands Questions