Title: Utilizing Flow-Based Density Ratio Estimation for Intractable Distributions in Genomic Applications

Abstract: A fundamental challenge in probabilistic modeling involves estimating density ratios between pairs of distributions that are difficult to tract. This capability facilitates rigorous comparisons of sample likelihoods under varying data-generating processes across different conditions. Although normalizing flows and other exact-likelihood models present a viable pathway for this estimation, standard approaches are often hindered by high computational costs and susceptibility to discretization errors, as they necessitate the independent simulation of each distribution’s likelihood. To address these limitations, our study employs condition-aware flow matching to establish a unified dynamical framework for monitoring density ratios throughout generative trajectories. We validate our approach through competitive results on simulated benchmarks for closed-form ratio estimation. Furthermore, we illustrate the method’s utility in single-cell genomics, where likelihood-based assessments of cellular states across experimental settings allow for the estimation of treatment effects and the evaluation of batch correction efficacy.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC