ScaiDigest Volume 4: Machine Learning Helps Deliver on the Promise of Spatial Omics

“As a reader of this newsletter, I am certain that you are aware of how single-cell multiomics has brought about a paradigm shift in our ability to understand and characterize complex biological systems. Nonetheless, conventional single-cell analyses in my opinion have a severe shortcoming, which is their inability to account for cellular spatial context. Cell-cell communication (both direct contact and paracrine signaling) play crucial roles in various disease contexts, with several immunotherapies being targeted at disrupting this class of interaction (e.g. anti-CTLA-4 and anti-PD-1/PD-L1 checkpoint inhibitors) in cancer.

While spatial ‘omics and other multiplexed imaging techniques such as Co-Detection by indEXing (CODEX) and imaging mass cytometry (IMC) are bridging this gap, enabling researchers to gain an unbiased systems-level understanding of tissue development and disease, there are several areas of active research and development in the field.

A typical image analysis pipeline requires scientists to first demarcate cells from images (segmentation), next to assign these segmented cells to a specific type or phenotype (e.g. to differentiate between tumor and responder immune cells), and finally to understand the association of the measured data with a disease or treatment outcome of interest (such as response to immunotherapy). Machine learning is being exploited in all steps of this analysis pipeline, and here I present a few recent papers which caught my eye in the field.

Constituting the first step of this pipeline, cell segmentation was recently the subject of a NeurIPS challenge, with a summary paper being published recently. In general, all approaches used deep learning, with transformer-based models performing the best. Since its introduction in 2017, the transformer architecture has taken the ML world by storm (powering Dall-E and ChatGPT amongst others), and transformer-based models demonstrated amongst the best performance in this challenge, including the best-performing algorithm (T1-osilab/MEDIAR). While this challenge evaluated the broad applicability of such algorithms across diverse image types, in the context of subcellular spatial transcriptomics (as enabled by Stereo-seq and Seq-scope), we typically also have access to the underlying measured ‘omics data (e.g. transcriptomic data), and transformer algorithms such as SCS which incorporate this data are also emerging as important players in the field, and in my opinion are likely to out-perform generalist segmentation algorithms due to their ability to exploit biological knowledge and relationships.

Once cells have been segmented, phenotyping and cell-type assignment is an important task for further downstream analyses, followed by an association of specific cell types with an endpoint of interest.