Cell segmentation-free inference of cell types fromin situtranscriptomics data

Abstract

Multiplexed fluorescence in situ hybridization techniques have enabled cell class or type identification by mRNA quantification in situ. However, inaccurate cell segmentation can result in incomplete cell-type and tissue characterization. Here, we present a robust segmentation-free computational framework, applicable to a variety of in situ transcriptomics platforms, called Spot-based Spatial cell-type Analysis by Multidimensional mRNA density estimation (SSAM). SSAM assumes that spatial distribution of mRNAs relates to organization of higher complexity structures (e.g. cells or tissue layers) and performs de novo cell-type and tissue domain identification. Optionally, SSAM can also integrate prior knowledge of cell types. We apply SSAM to three mouse brain tissue images: the somatosensory cortex imaged by osmFISH, the hypothalamic preoptic region by MERFISH, and the visual cortex by multiplexed smFISH. SSAM outperforms segmentation-based results, demonstrating that segmentation of cells is not required for inferring cell-type signatures, cell-type organization or tissue domains. KeywordsIn situ transcriptomics, spatial cell-type calling, segmentation-free, multiplexed FISH, SSAM, osmFISH, mFISH, multiplexed smFISH, MERFISH, spatially resolved RNA profiling Lignell et al., 2017), or total poly-A RNA (Codeluppi et al., 2018;Moffitt et al., 2018).Accurate cell segmentation, however, is difficult to achieve due to tightly apposed or overlapping cells, uneven cell borders, varying cell and nuclear shapes, signal intensity variation, probe fluorescence emission efficiency variation, and tiling artifacts (Thomas and John, 2017). The underlying problem is that the cellular structures one would want to segment are much smaller than the resolution of a diffraction-limited microscope. Therefore, there is a need for robust segmentation-independent methods for identification of cell-type signatures, cell-type organization, and tissue domains from multidimensional mRNA expression data in complex tissues. These methods could be used for datasets lacking landmarks or to validate segmentation-based approaches and identify associated artifacts.Here we introduce a novel computational framework named Spot-based Spatial cell-type Analysis by Multidimensional mRNA density estimation (SSAM), a multi-platform segmentation-free computational framework for identifying cell-type signatures and reconstructing cell-type and tissue domain maps from both 2D-and 3D-spatially resolved in situ transcriptomics data. We apply SSAM to three mouse brain tissue images obtained by different techniques: the somatosensory cortex by osmFISH, the hypothalamic preoptic region by MERFISH, and the visual cortex by multiplexed smFISH. We demonstrate the performance of SSAM in identifying 1) cell types in situ, 2) spatial distribution of cell types, 3) spatial relationships between cell types, and 4) tissue domains (e.g., cortical layers) based on the local composition of cell types without having even segmented a single cell. ResultsThe SSAM computational fra...

Bayesian segmentation of spatially resolved transcriptomics data

“…In spatial measurements, the cells of a distinct type will give rise to small molecular neighborhoods with stereotypical transcriptional composition. This patch-like structure of the spatial transcriptomics data can be used to interpret it without performing explicit cell segmentation 19 . To perform such neighborhood composition analysis, we generated a neighbourhood composition vector (NCV) for each molecule by taking its k spatially nearest neighbors and estimating the relative frequency of different genes among the neighboring molecules (Fig.…”

“…Such approaches can characterize cell type composition of the tissue or identify distinct regions, but cannot be easily extended to many other kinds of downstream analyses 13,19 .…”

“…Baysor can take advantage of such information by using a pre-calculated segmentation as a probabilistic prior. Computational segmentation of nuclear and cytoplasmic stains, however, remains a challenge in itself 18,19 , and the pre-calculated segmentations will typically contain many errors (Fig. 4d).…”

“…Not all of the downstream analyses require cell segmentation. For instance, region segmentation or tissue cell type composition may be inferred directly from molecular data 19 . We show that a relatively simple segmentation-free approach based on the composition of local neighborhoods (NCVs) can be used to assess the quality of the dataset, estimate the number and identity of the major cell types, and effectively visualize the organization of the tissue (Fig.…”

JSTA: joint cell segmentation and cell type annotation for spatial transcriptomics

“…pciSeq assigns cell types to nuclei based on proximity to mRNA of marker genes, circumventing the need for pixel level segmentation 9 . Similarly, SSAM creates cell type maps based on RNA distributions ignoring cellular boundaries 21 . However, both pciSeq and SSAM are limited to cell type information and do not create a segmentation map for the assignment of non-cell-type marker genes.…”