HELM-GPT: de novo macrocyclic peptide design using generative pre-trained transformer

Area of Science:

• Biochemistry
• Computational Biology
• Drug Discovery

Background:

• Macrocyclic peptides show therapeutic potential for targeting intracellular proteins due to their binding capabilities and cell membrane permeability.
• Existing research focuses on macrocyclic peptides as inhibitors for proteins like KRAS, crucial in cancer.
• Computational methods for de novo macrocyclic peptide design are underdeveloped.

Purpose of the Study:

• To introduce HELM-GPT, a novel computational method for de novo macrocyclic peptide design.
• To demonstrate HELM-GPT's ability to generate valid macrocyclic peptides and optimize their properties.
• To develop a strategy for co-optimizing peptide permeability and target binding affinity.

Main Methods:

• Utilized the Hierarchical Editing Language for Macromolecules (HELM) representation combined with Generative Pre-trained Transformer (GPT) models.
• Employed reinforcement learning (RL) for macrocyclic peptide generation and property optimization.
• Introduced a contrastive preference loss and a step-by-step optimization strategy for enhanced performance.

Main Results:

• HELM-GPT successfully generated valid macrocyclic peptides with optimized properties.
• The contrastive preference loss improved the RL optimization process.
• The proposed strategy effectively co-optimized peptide permeability and KRAS binding affinity.

Conclusions:

• HELM-GPT is an effective computational method for de novo macrocyclic peptide design.
• The method can identify novel macrocyclic peptides for targeting intracellular proteins, including KRAS.
• The developed approach holds promise for advancing cancer therapeutics.