Novel Pipelines to Extract Differences in Proteome Dynamics Based on Health Status

Understanding dynamics and co-regulatory patterns in the human proteome is a promising path for unraveling the molecular basis of health and disease. Nevertheless, there remains an open challenge in extracting concise information from high-throughput proteomic data that can effectively characterize and predict health. We develop novel statistical and computational pipelines to tackle this problem in a longitudinal saliva proteomics data set collected throughout the awakening response in six healthy controls and six subjects with severe mitochondrial disease (MitoD), a clinical condition caused by genetic mitochondrial defects that affects cellular energy transformation and alters multiple dimensions of health. We undertook three independent unsupervised approaches to characterize proteome dynamics and assessed their ability to separate MitoD individuals from controls. First, we designed a permutation test to detect the global difference in the proteomic co-regulation structure between healthy and unhealthy subjects. Second, we performed non-linear embedding and cluster analysis on elasticity to capture a more complicated relationship between health and the proteome. Third, we developed a machine learning algorithm to extract low-dimensional representations of the proteome dynamic and use them to cluster subjects into healthy and unhealthy groups without any knowledge of their true status. All three methods showed clear differences between MitoD individuals and controls. Our results revealed a significant and consistent association between MitoD status and the saliva proteome at multiple levels during the awakening response, including its dynamic change, co-regulation structure, and elasticity. This connection is not restricted to a few MitoD-specific proteins but spreads over a wide range of proteins from many body functions and pathways. Pipelines such as those shown here are the first step toward establishing interpretable and accurate prediction rules for health based on proteome dynamics.