CNTools: A computational toolbox for cellular neighborhood analysis from multiplexed images

Recent studies show that cellular neighborhoods play an important role in evolving biological events such as cancer and diabetes. Therefore, it is critical to accurately and efficiently identify cellular neighborhoods from spatially-resolved single-cell transcriptomic data or single-cell resolution tissue imaging data. In this work, we develop CNTools, a computational toolbox for end-to-end cellular neighborhood analysis on annotated cell images, comprising both the identification and analysis steps. It includes state-of-the-art cellular neighborhood identification methods and post-identification smoothing techniques, with our newly proposed Cellular Neighbor Embedding (CNE) method and Naive Smoothing technique, as well as several established downstream analysis approaches. We applied CNTools on three real-world CODEX datasets and evaluated identification methods with smoothing techniques quantitatively and qualitatively. It shows that CNE with Naive Smoothing overall outperformed other methods and revealed more convincing biological insights. We also provided suggestions on how to choose proper identification methods and smoothing techniques according to input data. Cellular neighborhoods (CNs), defined as cell regions with similar cell type composition, are attracting more and more attention because of their unique influence on biological processes in many diseases. However, a reliable method that can identify biologically meaningful CNs under different data settings is missing. Therefore, we provide such a method named Cellular Neighbor Embedding (CNE) with Naive Smoothing, which overall outperforms state-of-the-art methods on three real-world datasets. In addition, we make an easy-to-use toolbox that supports multiple CN identification pipelines and various downstream analyses, which can help researchers compare CN results and pursue more biological insights form CNs.