8 February 2024
Introduces Dirichlet flow matching for DNA sequence generation
Avoids pathologies of linear flow matching on simplices
Implements guidance and distillation
Evaluated on complex DNA datasets
Outperforms baselines in generative metrics and sequence design
Dirichlet flow matching for DNA sequence generation
This paper introduces a new generative modeling approach called Dirichlet flow matching to generate DNA sequences. It is based on transporting noise to data on a probability simplex using Dirichlet distributions. This avoids issues with linear flow matching on simplices. The model allows guidance and distillation. It is evaluated on complex DNA sequence datasets where it achieves superior performance over baselines in distributional similarity and sequence design tasks.
No comments yet, be the first to start the conversation...
Sign up to comment on this paper