Integrating Explainable Machine Learning and Transcriptomics Data Reveals Cell-Type Specific Immune Signatures Underlying Macular Degeneration

Abstract

Genome-wide association studies (GWAS) have established a key role of dysfunctional immune response in the etiology of Age-related Macular Degeneration (AMD). However, immune cells constitute a small proportion of the retina, and their role in AMD is not completely resolved. Here we develop an explainable machine learning pipeline using transcriptome data 453 donor retinas, identifying 81 genes distinguishing AMD from controls with an AUC-ROC of 0.80 (CI 0.70-0.92). These genes show enrichment for pathways involved in immune response, complement and extracellular matrix and connected to known AMD genes through co-expression networks and gene expression correlation. The majority of these genes were enriched in their expression within retinal glial cells, particularly microglia and astrocytes. Their role in AMD was further strengthened by cellular deconvolution, which identified distinct differences in microglia and astrocytes between normal and AMD. We corroborated these findings using independent single-cell data, where several of these candidate genes exhibited differential expression. Finally, the integration of AMD-GWAS data identified a common regulatory variant, rs4133124 at PLCG2, as a novel AMD-association. Collectively, our study provides molecular insights into the recurring theme of immune dysfunction in AMD and highlights the significance of glial cell differences as an important determinant of AMD progression.