Counting missed cells changes the winner in a tissue-mapping benchmark
A peer-reviewed comparison of cell-mapping methods found that rankings changed when the share of recovered cells counted alongside their quality. The benchmark covered human and mouse tissues across several measurement platforms. Drawing cell boundaries also changed inferred gene activity and communication between cells, making the choice consequential for biological interpretation.
Science··Midday
Cell quality and coverage produce different rankings
A peer-reviewed benchmark found that the definition of success changed which method ranked first for mapping cells. The field, spatial transcriptomics, locates gene activity within tissue. Segmentation draws the boundaries used to assign detected gene signals to individual cells. SAM, an image-based model for separating objects, led the aggregate ranking for the quality of retained cells. Once the proportion of cells recovered was included, SCS, a method combining images and gene signals, moved into first place.[1]
The comparison spans human and mouse tissues
The benchmark comprised roughly 2.3 million cells representing 12 tissue types. It compared 11 approaches on 15 datasets generated with 5 measurement platforms. They considered cell shape, gene-expression consistency, separation of cell types and biological analyses. Filtering removed cells with insufficient gene signals. A method could therefore score well among the cells it kept while representing a smaller share of the tissue. Two image models were excluded from comprehensive comparison after producing artificial patterns along image-patch boundaries.[1]
Cell boundaries alter inferred biological connections
Segmentation choices changed apparent contamination of cell-type marker genes and inferred interactions between signalling molecules and their receptors. Comparisons were limited by variation in tissue specimens, measurement platforms and the genes included in each panel. No method ranked best throughout. The detailed gene-subset sensitivity test used one melanoma dataset. A shared Nextflow workflow, a system for running analysis steps consistently, and public code make the benchmark reusable. The study assesses computational methods rather than diagnostic accuracy in patients.[1]
Related columns
For more information on this topic, you can read the related columns.
