A Unified Peptide Generative Framework via a Weakly Order-Dependent Autoregressive Language Model and Lifelong...

Area of Science:

• Computational chemistry
• Bioinformatics
• Drug discovery

Background:

• Bioactive peptides show therapeutic promise, driving demand for advanced peptide generation models.
• Current deep generative models struggle with peptide conformational flexibility and residue interactions.

Purpose of the Study:

• To introduce PepGenWL, a unified framework for peptide generation using a weakly order-dependent autoregressive language model and lifelong learning.
• To address limitations in existing models regarding conformational flexibility and residue dependencies.

Main Methods:

• PepGenWL employs an autoregressive language model with tolerance for out-of-order input.
• Mixture-of-Experts-style plugins balance memory stability and learning plasticity during fine-tuning.
• The framework was evaluated on generating antimicrobial peptides, anticancer peptides, and peptide binders.

Main Results:

• PepGenWL surpassed state-of-the-art models in generating therapeutic peptides, including antimicrobial and anticancer types.
• The model demonstrated an ability to incorporate beneficial residues for bioactivity.
• A property-guided pipeline achieved a 28.6% target binding rate for peptide binders with specificity.
• PepGenWL's applicability extends to peptide SMILES, enabling generation of modified and cyclic peptides.

Conclusions:

• PepGenWL offers a unified and flexible framework for general-purpose peptide generation.
• The model effectively handles peptide conformational flexibility and residue interactions.
• PepGenWL shows significant potential for advancing therapeutic peptide discovery and development.