Finding Underlying Morphology Trends and Significant Genes across Mental Illnesses using Dimensionality Reduction

Abstract

Mental health diagnoses can be difficult to make due to the high comorbidity and overlapping symptoms across different diseases. It has been suggested that rather than focusing on a standard checklist, diagnoses be based on a hierarchy that can align to one or more disorders, some tiers of which can be identified by neuroimaging. The ENIGMA consortium has collated a group of standardized MRI case-control studies and provided this data in the ENIGMA toolbox, a Python package. In this study, we used this package to retrieve effect sizes from seven mental disorders and analyzed them using two dimensionality reduction methods: principal component analysis (PCA) and uniform manifold approximation and projection (UMAP) to uncover the latent dimensions underlying the cortical thickness commonalities. We compared PCA results with and without prior z-score normalization, as well as 2D with 3D UMAP results. We then correlated the first component of a PCA and a UMAP reduction with the Allen Human Brain Atlas (AHBA) to identify relevant genes, which we annotated to determine their biological contexts. We have found that z-scored PCA had stronger correlation across disorders. We also found that 2D and 3D UMAP are equivalently suitable for finding latent dimensions. We also determined that genes that correspond with these first components are differentially expressed in the brain, reproductive organs, hormone glands, fibroblasts, lymphocytes, liver, heart, pancreas, whole blood, and muscular skeletal system. In addition, analysis of overrepresented GO gene sets has shown that surviving genes are expressed in neurological contexts, particularly synapse signaling.