Generative Flows on Discrete State-Spaces: Enabling Multimodal Flows with Applications to Protein Co-Design

Area of Science:

• Machine Learning
• Generative Models
• Computational Biology

Background:

• Generative models often struggle to combine discrete and continuous data effectively.
• Existing flow-based models lack a unified approach for multimodal data.
• Protein co-design requires simultaneous generation of structure and sequence.

Purpose of the Study:

• Introduce Discrete Flow Models (DFMs) for handling multimodal continuous and discrete data.
• Develop a unified flow-based modeling framework for diverse data types.
• Advance the field of protein co-design through improved generative modeling.

Main Methods:

• Developed Discrete Flow Models (DFMs), a novel flow-based generative model for discrete data.
• Utilized Continuous Time Markov Chains to achieve discrete flow matching.
• Integrated DFMs into a multimodal framework for joint data generation.
• Applied the framework to protein structure and sequence co-design.

Main Results:

• DFMs provide a unified approach for multimodal continuous and discrete data problems.
• The method offers improved performance over existing diffusion-based approaches.
• Achieved state-of-the-art results in protein co-design.
• Demonstrated flexible generation of protein sequence or structure using the same model.

Conclusions:

• Discrete Flow Models represent a significant advancement in generative modeling for multimodal data.
• DFMs offer a versatile and high-performing solution for complex data generation tasks.
• The developed framework has strong implications for biological sequence and structure design.