A framework for antibody validation via spatial proteo-transcriptomics using the CellScape platform

Spatial proteomics depends on antibody reliability, yet many antibodies used in routine diagnostics and research - particularly in conventional brightfield and multiplex fluorescence immunohistochemistry - are only subjectively or insufficiently validated, leading to potential false conclusions and limiting the use of spatial proteomics in large-scale clinical research studies.To address this limitation, we developed a fully automated - thus objective - deep learning-based framework for antibody validation. The in-situ staining patterns and interdependencies of various antibody clones for the same target were compared to the corresponding RNA expression pattern using a network of deep learning algorithms. For this purpose, a tissue microarray (TMA) was constructed from 50 different normal and 50 different neoplastic tissues including carcinoma entities, lymphomas, and other human neoplastic tissues from 1040 patients. All antibody clones were pre-tested in conventional brightfield IHC and assembled in 10- to 30-plex mfIHC assays using the CellScape™ Precise Spatial Proteomics platform (Bruker Spatial Biology, USA)to compare different antibody clones for the same target across 100 different human tissue types. The same TMA slide can then be used for spatial transcriptomics on the CellScape™ via HCR™ Gold (Molecular Instruments, Los Angeles, US) or on the CosMx® Spatial Molecular Imager (Bruker Spatial Biology, USA). A framework of different deep learning models (U-Net and DeepLab3 + ) was developed for analysing protein and RNA expression. This spatial proteo-transcriptomics approach allows for both (i) the direct deep learning-based comparison of different antibody clones for the same target and (ii) an interpretation in view of the mRNA expression of the corresponding gene on a single-cell level across 100 different healthy and neoplastic tissues across the human body. Through the automatic assessment of every individual antibody clone, objective metrics such as fraction of expression on different tissue-compartments (%), accordance with other antibody clones for the same target (%), as well as correspondence with the RNA expression (%) were computed. Given that linear regression analysis showed an ultra-low signal deterioration across the different cycles (<1 %) - due to the novel EpicIF for gentle bleaching of the fluorochromes on the CellScape platform - a direct comparison of antibodies from different cycles was possible without cumbersome rearrangement of the antibodies for mfIHC panel creation. These findings highlight the potential of using an automated framework for antibody validation in favour of using publicly available non-spatial RNA expression libraries.In conclusion, here we present the first objective, pathologist reviewed, and fully automated deep learning-based framework for antibody validation using spatial proteo-transcriptomics.