Gumbel-Softmax Flow Matching with Straight-Through Guidance for Controllable Biological Sequence Generation

Area of Science:

• Computational Biology
• Machine Learning
• Generative Models

Background:

• Flow matching on the simplex is promising for DNA design but limited in higher dimensions.
• Scaling generative models to peptide and protein design requires higher-dimensional simplex representations.

Purpose of the Study:

• Introduce Gumbel-Softmax Flow and Score Matching for scalable sequence generation.
• Develop Straight-Through Guided Flows (STGFlow) for efficient, training-free guidance.
• Enable controllable *de novo* generation of DNA, peptides, and proteins.

Main Methods:

• Novel Gumbel-Softmax interpolant with time-dependent temperature.
• Parameterized velocity field for Gumbel-Softmax Flow Matching.
• Gumbel-Softmax Score Matching for density gradient regression.
• Straight-Through Guided Flows (STGFlow) for classifier-based guidance.

Main Results:

• High-quality and diverse sequence generation across dimensions.
• Efficient scaling to higher-dimensional simplices.
• State-of-the-art performance in DNA promoter design, protein generation, and peptide design.

Conclusions:

• Gumbel-Softmax Flow and Score Matching provides a robust framework for *de novo* sequence generation.
• STGFlow enables efficient inference-time guidance for discrete flow methods.
• The framework demonstrates significant advancements in biological sequence design.