Flow Matching for Count Data

Area of Science:

• Computational Biology
• Statistical Modeling
• Machine Learning

Background:

• High-dimensional count data are prevalent in single-cell RNA sequencing (scRNA-seq) and neural spike train analysis.
• Existing methods struggle with large count ranges, either treating counts categorically or transforming them into continuous spaces.
• Deep generative models show promise but require adaptation for discrete count data.

Purpose of the Study:

• To introduce count-FM, a flow-matching framework for modeling count-valued data.
• To enable efficient mapping between count distributions for various data analysis tasks.
• To improve upon existing methods in terms of sample quality and modeling efficiency.

Main Methods:

• Developed count-FM, a framework based on a continuous-time birth-death process with local unit jumps.
• Employed simulation-free training of conditional transition rates for efficient learning.
• Applied the framework to scRNA-seq and neural spike-train data.

Main Results:

• Count-FM demonstrated superior sample quality compared to baseline methods in simulations.
• The framework achieved these results with significantly fewer model parameters.
• Count-FM enabled effective unconditional generation, transport, and conditional generation on real-world datasets.

Conclusions:

• Count-FM provides an efficient and effective method for analyzing high-dimensional count data.
• The framework offers improved modeling efficiency and interpretable transport paths.
• Count-FM represents a significant advancement for applications involving count-valued data, including scRNA-seq and neuroscience.