Scaling SMILES-Based Chemical Language Models for Therapeutic Peptide Engineering

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Therapeutic peptides offer high specificity but face computational challenges in drug discovery.
• Current models (protein or chemical) inadequately represent complex peptide chemistry.
• Existing methods rely on limited descriptors or complex, dataset-specific pipelines.

Purpose of the Study:

• To bridge the computational gap for therapeutic peptides.
• To introduce PeptideCLM-2, a novel suite of chemical language models.
• To enable native representation of complex peptide chemistry.

Main Methods:

• Trained chemical language models on over 100 million molecules.
• Developed PeptideCLM-2 for native representation of peptide chemistry.
• Benchmarked performance against existing methods.

Main Results:

• PeptideCLM-2 effectively represents complex peptide chemistry.
• Demonstrated strong performance in predicting key development endpoints.
• Outperformed prior methods for membrane diffusion, biological function, and half-life prediction.

Conclusions:

• PeptideCLM-2 expands machine learning capabilities for therapeutic peptide research.
• This approach offers a more effective computational toolkit for peptide drug discovery.
• The model shows promise for accelerating the development of peptide-based therapeutics.