A spatio-temporal single-cell type map of the human proteome based on transcriptomics, high-resolution antibody-based imaging and artificial intelligence

For a fundamental understanding of human health, molecular medicine and targeted treatment, it is necessary to map processes unique to each tissue and cell type. The major collection of bioimages in the Human Protein Atlas (HPA) database, comprising 10 million high-resolution images of stained tissues and cells, constitutes an important resource for studying protein expression in relation to tissue and cell morphology. While the staining of the most common cell types and structures have been manually annotated and is presented as protein expression levels, the immunohistochemical data has limitations in terms of relative protein abundance, subjectivity and challenges distinguishing specific subsets of cells or rare cell types. In an effort to increase the resolution of the tissue-based protein expression data, we here aimed to set up a stringent workflow for mapping human tissues at the single-cell type level, and utilized this workflow to create a high-resolution spatio-temporal maps of >1,000 tissue or cell type-specific proteins in eight different tissues. Based on single-cell RNA sequencing data, we identified tissue-specific cell types and subsets thereof, some of which cannot be distinguishable by regular immunohistochemistry. Using a multiplex immunofluorescence pipeline, we then built antibody panels specifically outlining each of these cell types. The fixed antibody panels were stained together with the candidate protein of interest, one at a time, to pinpoint the exact protein localization at a cellular and subcellular level. Using artificial intelligence, the high-resolution images were quantified using an automated image analysis workflow. The integrated data allowed us to study temporal mRNA and protein expression gradients along with maturation processes and identify which mRNAs that are consistently translated into proteins from those that vary from a spatio-temporal aspect. We were also able to assign presumed functions to numerous uncharacterized proteins. The established pipeline presented will be further used for other human tissues, decoding the complexity of human cells, tissue heterogeneity and homeostasis, and link quantitative data with tissue morphology. The data is freely available on www.proteinatlas.org and the detailed data will likely contribute to valuable insights into protein function and molecular processes linked to disease.