Learning the rules of peptide self-assembly through data mining with large language models.

Area of Science:

• Biomolecular self-assembly
• Computational chemistry
• Materials science

Background:

• Peptides are crucial biomolecules that self-assemble into various structures.
• Existing research on peptide self-assembly lacks a consolidated dataset for uncovering global rules.
• Understanding peptide self-assembly is vital for designing novel materials and therapeutics.

Purpose of the Study:

• To create a comprehensive database of peptide self-assembly experimental data.
• To develop machine learning models for predicting peptide assembly phases.
• To enhance peptide literature mining using advanced language models.

Main Methods:

• Curating a database of over 1000 experimental data entries on peptide sequences, conditions, and assembly phases.
• Utilizing manual processing by experts and large language model-assisted literature mining.
• Developing and training machine learning models for assembly phase classification.
• Fine-tuning a GPT model for efficient peptide literature information extraction.

Main Results:

• A curated database of over 1000 peptide self-assembly entries was established.
• Machine learning models achieved over 80% accuracy in classifying peptide assembly phases.
• A fine-tuned GPT model significantly improved information extraction from peptide literature.

Conclusions:

• The developed workflow enhances the efficiency of exploring self-assembling peptide candidates.
• The study provides a foundation for guiding experimental work in peptide self-assembly.
• This approach deepens the understanding of the fundamental mechanisms governing peptide self-assembly.