Digital Twins : High Resolution Disease Models for Optimized Diagnosis and Treatment

Sammanfattning: To study immune-mediated diseases, which can affect the expression of thousands of genes among many different cell types and organs, is a daunting challenge. However, for effective diagnosis and therapeutic treatment it is relevant to understand the regulatory functions of disease. In this thesis, we hypothesized that regulatory functions in complex diseases can be effectively prioritized based on so called digital twins, which are based on high-resolution single cell data in combination with network theories. More specifically, we tested if digital twins could be used on a patient-group level to prioritize cell types, genes, and/or organs based on their regulatory function in the disease progression. If this hypothesis is true, potential biomarkers and therapeutic targets can be identified for optimized diagnosis and treatment. The long-term goal is to construct digital twins for personalized medicine, to predict the optimal treatment strategies for the individual patients. Although, this is a very ambitious goal which could not be reached through this thesis, relevant steps towards it have been reached.First, we tested if high-resolution disease models based on single cell RNAsequencing (scRNA-seq) data could be used in combination with network theories, to predict and prevent disease. For this aim, we used a mouse model of antigeninduced arthritis (AIA). Based on the cell type specific genes in AIA joint, we identified a multi-cellular disease model (MCDM), including predicted cell-cell interactions. Analyzing this model, Granulocytes were identified as most central in AIA joint. The results from this centrality analysis correlated with GWAS enrichment among the cell type specific genes, as well as with the centrality analyses based on human RA, supporting our results relevance for human disease. A drug, bezafibrate, was further identified which mainly targeted shared disease modules over the central and GWAS enriched CD4+ T cells in nine of 13 analyzed human diseases. Bezafibrate treatment of our AIA mouse model resulted in a decrease in arthritis severity score as well as a decrease in T cell proliferation into the joint.Since blood is an easily available source of data, it is of interest to know it’s potential usefulness when constructing digital twins. To test if samples taken from blood are representative of the inflamed organ, we performed a meta-analysis of different samples from blood and joint of patients with rheumatoid arthritis, as well as from joint and blood Granulocytes of our AIA mouse model. Based on differentially expressed genes (DEGs) between sick and healthy samples from each dataset, we performed pathway analyses and predicted potential biomarkers and upstream regulators (URs). Comparing the lists of pathways, biomarkers, and URs between the datasets from different subsets of blood samples showed low or no similarities. However, the datasets of human bulk or mouse single cell data collected from synovial fluid or full joint showed high similarities. Furthermore, the top shared enriched pathways, predicted biomarkers, and URs from both human and mouse were to a higher degree connected to known functions of autoimmune diseases or rheumatoid arthritis, compared to the respective results from samples taken from blood. These findings indicate that inflammatory mechanisms in cells in blood and inflamed organs differ greatly, which may have important diagnostic and therapeutic implications.We next analyzed if digital twins could be used to identify the early regulatory mechanisms that are also present at the late time points. For this, we used an in vitro time series model of seasonal allergic rhinitis. Samples were taken before allergen stimulation, as well as at 12 hours, 1 day, 2 days, 3 days, 5 days, and 7 days after allergen stimulation, for scRNA-seq and MCDM construction. Multi-directional interactions including all cell types were found at all time points, even before allergen stimulation, which complicated the identification of one key regulatory cell type or gene. Instead, we found that the regulatory genes could be ranked based on their overall downstream effect over all the time points. Our top-ranked regulatory gene, PDGFB, targeted most of the cell types at all the time points, while a previously known early regulator and drug target in allergy, IL4, targeted only five cell type and time point combinations. Validation studies further showed that neutralization of PDGF-BB on allergen-stimulated PBMC from SAR patients were more effective compared to neutralization of IL-4.Finally, we tested if a digital twin including data from multiple organs could be used to understand the systemic interactional changes due to disease. For this aim, we used a systemic mouse model of arthritis, namely collagen induced arthritis (CIA). We first analyzed ten different organs, based on which we prioritized five organs with the highest number of DEGs between CIA and healthy mice, namely joint, lung, muscle, skin, and spleen. Although only joint showed signs of inflammation, many DEGs were identified in all five organs. Those changes were organized into a multi-organ multi-cellular disease model, which indicated an on/off switch of pro-/anti-inflammatory functions in joint and muscle respectively. Validation studies in human immune-mediated inflammatory diseases supported this on/off switch, where pro-inflammatory functions were mainly found in inflamed organs, while anti-inflammatory functions were found in non-inflamed organs.In conclusion, this thesis supports the potential of using high-resolution disease models for digital twin construction. Such digital twins could then be used to prioritize cell types and genes, for further prediction of diagnostic markers and therapeutic targets. Even though the identification of one key regulatory function was complicated due to multidirectional interactions, the genes could be ranked based on their relative downstream effect. For reproducible results, we found that digital twins should ideally be based on data from locally inflamed organs, while systemic models and models covering different disease stages could be useful to understand the disease progression.