Modeling the structure-conditioned sequence landscape for large-scale protein design with TriFlow.

Area of Science:

• Computational biology
• Protein engineering
• Machine learning

Background:

• Generative models are key in computational protein design.
• Current methods use local context and autoregressive generation, limiting efficiency and quality.
• High-quality sequence design from backbone structure is critical.

Purpose of the Study:

• To develop an efficient and high-quality protein sequence design model.
• To improve de novo binder design capabilities.
• To explore structure-conditioned sequence landscapes.

Main Methods:

• Developed TriFlow, a model using a RoseTTAFold-like three-track architecture for global structural context.
• Employed discrete flow-matching for efficient, few-step sequence generation.
• Trained on interacting protein chains (PDB) and domains (AlphaFold DB) to learn interface properties.

Main Results:

• TriFlow shows improved performance across benchmarks, especially in de novo binder design.
• Boosted the in silico success rate of state-of-the-art pipelines like BindCraft.
• Successfully generated and validated binders for over 500 diverse protein targets.
• Highlighted functional active sites by contrasting model constraints with evolutionary profiles.
• Demonstrated systematic design of specific binders against human class I cytokines, optimizing affinity and minimizing off-target interactions.

Conclusions:

• TriFlow offers a robust framework for large-scale protein engineering.
• Provides a method for exploring fundamental principles of structure-conditioned sequence landscapes.
• Enhances computational protein design efficiency and quality, particularly for binder design.