Generative β-hairpin design using a residue-based physicochemical property landscape.

Area of Science:

• Computational biology
• Biophysics
• Machine learning in drug discovery

Background:

• De novo peptide design is crucial for biological and biomedical applications.
• Existing methods often rely on sequence homology, limiting novelty and overlooking essential physicochemical properties for protein folding.
• Generative machine learning offers a path to create unique peptide sequences beyond evolutionary constraints.

Purpose of the Study:

• To develop and evaluate a custom peptide generative adversarial network (GAN) for designing novel peptide sequences.
• To specifically target the design of peptides capable of folding into the beta-hairpin secondary structure.
• To lay the groundwork for generative models that incorporate physicochemical and conformational properties for peptide design.

Main Methods:

• Developed a custom generative adversarial network (beta-GAN) tailored for peptide sequence generation.
• Incorporated physicochemical properties (e.g., hydrophobicity, residue volume) and conformational features of amino acids.
• Utilized structure-specific sequence data from the Protein Data Bank (PDB) for training.
• Assessed the model's ability to distinguish beta-hairpin structures from alpha-helix and intrinsically disordered peptides.

Main Results:

• The beta-GAN achieved up to 96% accuracy in distinguishing beta-hairpin structures from other secondary structures.
• Generated artificial beta-hairpin peptide sequences with low sequence identities (31% vs. PDB, 50% vs. non-redundant databases).
• Demonstrated the model's capability to generate novel sequences distinct from existing databases.

Conclusions:

• Generative models anchored by physicochemical and conformational properties show significant potential for de novo peptide design.
• This approach can expand the sequence-to-structure landscape beyond evolutionary limitations.
• The developed beta-GAN provides a foundation for future advancements in designing peptides with specific structural and functional properties.