Spatial omics technologies are producing an unprecedented amount of ultra-high plex in situ data that are promising to revolutionize cancer prognoses and treatments. The lack of analysis tools hinders discoveries into the pathological processes underlying cancer initiation and progression. Here we present STimage, a machine learning approach to flexibly combine transcriptome-wide spatial sequencing data with single-cell spatial phenotyping (Phenocycler-Fusion), generated from same tissue samples. As a ground-truth for cell typing, the Akoya’s single-cell spatial phenotyping technology enabled us to precisely define cell types and cell states that were used to evaluate and deploy the data integration pipeline.
With these data, STimage first maps cells in tissue sections to one common histopathological tissue image reference (H&E image). This way multiple layers of molecular data are transferred into one common framework, which can then be analysed together. Traditional pathology annotations on the H&E images are also integrated to add human understanding of morphological patterns in a cancer tissue. The integrated analysis improves cell neighbourhood identification, which allows cell-cell interaction analysis based on spatial co-localisation between cell types (using single-cell resolution protein data) and locally co-expressing ligand-receptor pairs (using transcriptome-wide spatial data).
We applied STimage in head and neck cancer and skin cancer, demonstrating the broad applicability to multiple technologies and cancer types. We show the applications in two areas, diagnosis, and prognosis. For diagnosis, ST image identified and cross-validated cell types whilst assessing the expression of markers for drug targets. We also present an in-depth case study for an oropharyngeal squamous cell carcinoma patient not responding to Nivolumab treatment (anti-PD-1). Based on Visium and PhenoCycler-Fusion data, we discovered a spatial signature for non-responsiveness. Furthermore, we used the predicted ligand-receptor interactions to rank the patient’s response potential to currently available drugs, with the top target being TF-TFRC. STimage is thus able to integrate multiple layers of spatial omics data to improve and solidify prognostic biomarkers for cancer treatments.
This study highlights the power of machine learning integration to combine multiple spatial multi-omics data, in particular PhenoCycler-Fusion and Visium, for improving diagnosis, prognosis and treatment of diverse cancers.