A-G.C.19: Conformal Inference for Gene-Specific Cell Subsets in Single-Cell Differential Expression
Christof Seiler
University Hospital Zurich, University of Zurich
Keywords
conformal prediction, single-cell data, clustering
[sponser-meet-now-chat][/sponser-meet-now-chat]
Differential expression analysis in single-cell data requires predefining cell groups in which expression is assumed to be homogeneous. Standard methods cluster cells using nearest-neighbor graphs before testing for differential expression. An alternative is to infer structure after differential analysis and identify groups with shared responses, without pre-defined cluster labels. This is relevant in cancer, where tumor heterogeneity leads to subpopulations with shared treatment responses. Pre-clustering can miss such signals.
We build on the lemur R-package. This method avoids prior clustering and models gene expression through a low-dimensional latent representation of cells with covariate-specific structure. It estimates individual treatment effects (ITEs) at the single-cell level. For each gene, it identifies a cell subset forming a region in latent space with coherent differential signal. These neighborhoods can capture patterns which might be missed by clustering before differential testing.
We extend lemur in two directions. First, we build prediction intervals for lemur ITEs estimates using conformal prediction. This method turns point predictions into sets with finite-sample, distribution-free coverage under exchangeability. Second, we construct prediction sets for each cells membership within a gene neighborhood. This quantifies uncertainty for cluster assignments. Singleton sets indicate low uncertainty and sets of size two reflect higher uncertainty by retaining multiple plausible outcomes under the same coverage guarantee.
We provide visualizations, theoretical guarantees, and empirical validation on simulated and real datasets. We combine latent-space differential modeling with conformal uncertainty quantification to reveal gene-specific cell structure potentially missed by standard clustering.
Co-authors: Christof Seiler
Contact Attendee
Warning: Attempt to read property "user_email" on string in /home/1276969.cloudwaysapps.com/ydbgzhdjeq/public_html/wp-content/plugins/my-conference-now/functions.php on line 2826