Abstract: Variations of cell-type proportions within tissues could be informative of biological aging and disease risk. Single-cell RNA-sequencing offers the opportunity to detect such differential abundance patterns, yet this task can be statistically challenging due to the noise in single-cell data, inter-sample variability and because differential abundance (DA) patterns are often characterized by small effect sizes. Here we present a novel DA-testing paradigm called ELVAR, which, unlike the popular Louvain clustering method, takes cell attribute information into account when inferring cell-states within the high-dimensional single-cell manifold. We validate ELVAR using both simulated and real single-cell and single-nucleus RNA-Seq data, demonstrating improved inference over the popular Louvain algorithm and competing DA-testing methods. In lung tissue, ELVAR detects a decrease in the naïve Cd4 + T-cell proportion with age, as well as a shift of alveolar macrophages towards an M2 polarization program. In colon tissue, ELVAR predicts increased stem-cell and T-regulatory fractions in polyps preceding adenoma. In summary, leveraging cell attribute information when inferring cell communities can denoise single-cell data and help retrieve more robust cell states for subsequent DA-testing. ELVAR is available as an open-source R-package.