Beyond Global Embeddings: A New Approach to Single-Cell Representations
Track: Full Paper Track
Keywords: Single-cell, transcriptomics, representation learning, dimensionality reduction, k-nearest neighbor graph, topological data analysis, local embeddings, biological data analysis, manifold learning, cellular heterogeneity, scRNA-seq clustering, transcriptional profiling, graph-based learning, high-dimensional data, biological machine learning, computational biology
TL;DR: We introduce a method that stitches local embeddings into a global structure, capturing fine-grained cellular diversity and enabling high-resolution analysis of large, heterogeneous single-cell datasets.
Abstract: Uncovering the latent structure of high-dimensional data is a fundamental challenge in single-cell analysis. While many methods seek to structure single-cell data, most rely on a single global embedding space, which can obscure fine-grained variation. Here, we introduce Connectorama, a locally adaptive framework that constructs neighborhoods by aggregating information across overlapping local patches, allowing similarity metrics to adapt to local covariance structures. Applying this approach to large single-cell RNA sequencing datasets, we recover biologically meaningful subpopulations that global methods fail to resolve, including distinct immune cell subsets and hepatocyte populations with specialized gene expression signatures. By reframing single-cell representation as an ensemble of local views rather than a single projection, Connectorama offers a powerful framework for studying cellular diversity at scale.
Submission Number: 79
Loading