Reproduce “Propensity Score Alignment of Unpaired Multimodal Data” on OpenResearch

About this paper

Johnny Xi, Jana Osea, Zuheng Xu, Jason Hartford

The propensity-score matching method aligns unpaired biological measurements by training separate classifiers to predict experimental labels in each modality. It compares predicted probability vectors within labels and uses optimal transport to form soft matches. On held-out CITE-seq data, it reports median FOSCTTM of 0.3049 versus 0.3953 for VAE+OT, with lower scores better; the result is limited to this benchmark.

Key insights

• Propensity scores, derived from observable perturbation labels, offer an identifiable, common, and maximally informative low-dimensional space for robustly matching unpaired multimodal data.
• The methodology effectively disentangles shared latent information from modality-specific noise by compressing the latent state to retain only information relevant to the perturbation.
• The validation cross-entropy loss of the propensity score classifiers serves as a reliable proxy for actual matching performance, providing a practical tool for model selection when ground truth is unknown.

Summary by alphaXiv.