Estimating neighboring cells by using only nuclear markers is crucial in many biological applications. Although several strategies have been used for this purpose, most published methods lack a rigorous characterization of their efficiencies. Remarkably, previously described methods are not automatic and depend only on cell-cell distance, neglecting the importance of pair-neighborhood interaction.
To develop a robust and automatic method for assessing cell local neighborhood, while analyzing the impact of the physical variables involved in this task.
We inferred neighbors from images with nuclei labeling by approximating the cell-cell interaction graph by the Delaunay triangulation of nuclei centroids. Each edge of this graph was filtered by thresholding in cell-cell distance and the maximum angle that each pair subtends with shared neighbors (pair-neighborhood interaction). Thresholds were calculated by maximizing a new robust statistic that measures the communicability efficiency of the cell graph. Using a variety of images of diverse tissues with additional membrane labeling to find the ground truth, we characterized the assessment performance.
On average, our method detected 95% of true neighbors, with only 6% of false discoveries. Even though our method’s performance and tissue regularity are correlated, it works with performance metrics over 86% in very different organisms, including Drosophila melanogaster , Tribolium castaneum , Arabidopsis thaliana and C. elegans .
We automatically estimated neighboring relationships between cells in 2D and 3D using only nuclear markers. To achieve this goal, we filtered the Delaunay triangulation of nuclei centroids with a new measure of graph communicability efficiency. In addition, we found that taking pair-neighborhood interactions into account, in contrast to considering only cell-cell distances, leads to significant performance improvements. This becomes more notorious when the number of cells is low or the geometry of the cell graph is highly complex.