Modeling Type 1 Diabetes progression from single-cell transcriptomic measurements in human islets

Abstract

Type 1 diabetes (T1D) is a chronic condition in which the insulin-producing beta cells are destroyed by immune cells. Research in the past few decades characterized the immune cells involved in disease pathogenesis and has led to the development of immunotherapies that can delay the onset of T1D by two years. Despite this progress, early detection of autoimmunity in individuals who will develop T1D remains a challenge. Here, we evaluated the potential of combining single-cell genomics and machine learning strategies as a prime approach to tackle this challenge. We used gradient-boosting-based machine learning algorithms and modeled changes in transcriptional profiles of single cells from pancreatic tissues in T1D and nondiabetic organ donors collected by the Human Pancreas Analysis Program. We assessed whether mathematical modelling could predict the likelihood of T1D development in nondiabetic autoantibody-positive organ donors. While the majority of autoantibody-positive organ donors were predicted to be nondiabetic by our model, select donors with unique gene signatures were classified with the T1D group. Remarkably, our strategy also revealed a shared gene signature in distinct T1D associated models based on different cell types including alpha cells, beta cells and acinar cells, suggesting a common effect of the disease on transcriptional outputs of these cells. Together, our strategy presents the first report on the utility of machine learning algorithms in early detection of molecular changes in T1D.