Efficient prediction of peptide self-assembly through sequential and graphical encoding.

Area of Science:

• Computational chemistry and bioinformatics
• Artificial intelligence in drug discovery
• Peptide science

Background:

• Peptide research is rapidly expanding due to therapeutic and commercial potential.
• Deep learning models require robust peptide encoding for accurate property prediction.
• Molecular dynamics simulations generate large datasets for training AI models.

Purpose of the Study:

• To systematically analyze the impact of different peptide encoding strategies on deep learning model performance.
• To benchmark state-of-the-art sequential and structural deep learning models for peptide self-assembly prediction.
• To provide guidance for selecting optimal peptide representations in AI-driven peptide research.

Main Methods:

• Generated a large dataset (>62,000 samples) of peptide self-assembly using coarse-grained molecular dynamics.
• Employed advanced deep learning models: sequential (RNN, LSTM, Transformer) and structural (GCN, GAT, GraphSAGE).
• Evaluated encoding methods using amino acid sequences and molecular graphs as inputs.

Main Results:

• Transformer demonstrated superior performance among sequence-encoding models for peptide self-assembly prediction.
• The study successfully predicted self-assembly properties for decapeptides.
• Benchmarking revealed significant differences in model accuracy based on encoding techniques.

Conclusions:

• Peptide encoding is crucial for enhancing deep learning prediction accuracy in peptide science.
• Transformer is the most effective sequence-based deep learning model for peptide self-assembly prediction.
• This work serves as a guide for various peptide property predictions, including isoelectric points and hydration free energy.