D-Flow: Multi-modality Flow Matching for D-peptide Design

Area of Science:

• Computational biology
• Biotechnology
• Drug discovery

Background:

• Proteins are vital for biological functions, with therapeutic peptides showing promise.
• D-peptides offer advantages like proteolysis resistance and enhanced in vivo stability.
• Limited D-protein data hinders deep learning model application in D-peptide design.

Purpose of the Study:

• To introduce D-Flow, a novel full-atom flow-based framework for de novo D-peptide design.
• To address the challenge of limited D-protein data using innovative computational strategies.
• To enhance D-peptide design by integrating structural information with protein language models.

Main Methods:

• D-Flow utilizes structural representations (backbone frames, side-chain angles, amino acid types) for receptor binding.
• A mirror-image algorithm converts L-receptor chirality to overcome D-protein data scarcity.
• Protein language models are integrated with structural awareness via a lightweight adapter for improved learning.

Main Results:

• D-Flow effectively designs D-peptides with improved sequence identity (10.2% over baseline) and high affinity scores (top 24.31%).
• Generated D-peptides exhibit closer alignment with native sequences and structures.
• The framework demonstrates successful transition from general to targeted binder design.

Conclusions:

• D-Flow shows significant potential for advancing D-peptide design.
• The framework facilitates the development of bioorthogonal, stable molecular tools and diagnostics.
• D-Flow offers a viable solution for overcoming data limitations in D-peptide discovery.