Systematic Exploration of Small-Molecule Binding via a Large Language Model Trained on Textualized Protein-Ligand...

Area of Science:

• Computational chemistry
• Biophysics
• Artificial intelligence

Background:

• Large Language Models (LLMs) demonstrate broad task performance.
• Harnessing LLMs for biophysical applications requires converting 3D chemical data into 1D language-like formats.
• A key challenge is the effective transformation and tokenization of molecular data for LLM processing.

Purpose of the Study:

• To develop a method for transforming molecular data into language-like representations for LLM utilization in biophysics.
• To train and validate a model using a known protein-ligand complex.
• To enable LLMs to assess chemical properties, identify binding similarities, and discover related drugs.

Main Methods:

• Development of a novel method to convert 3D molecular data into 1D language-like sequences.
• Tokenization of the generated language-like data for LLM input.
• Training and validation of the model on a protein-ligand complex dataset.
• Utilizing the pre-trained model for analysis of chemical properties, binding interactions, and drug relationships.

Main Results:

• The model successfully transformed molecular data into a processable language format for LLMs.
• Validation with a protein-ligand complex confirmed the model's ability to assess chemical properties and binding characteristics.
• The model identified shared binding properties and structures, and revealed related drugs.
• The developed language and model uncovered previously unreported protein-protein networks influenced by ligand interactions.

Conclusions:

• The study presents a viable method for applying LLMs to biophysical data by creating a molecular language.
• The model demonstrates potential in drug discovery by identifying novel interactions and related compounds.
• This approach opens new avenues for understanding complex biological networks and ligand-mediated effects.