Interpretable deep learning framework for understanding molecular changes in human brains with Alzheimer’s disease: implications for microglia activation and sex differences

INTRODUCTION The objective of this study is to characterize the dysregulation of gene expression in AD affected brain tissues through an interpretable deep learning framework. METHODS We trained multi-layer perceptron models for the classification of neuropathologically confirmed AD vs. controls using transcriptomic data from three brain regions of ROSMAP study. The disease spectrum was then modeled as a progressive trajectory. SHAP value was derived to explain model predictions and identify significantly implicated genes for subsequent gene co-expression network analysis. RESULTS The models achieved excellent performance in classification and prediction in two external datasets from Mayo RNA-seq cohort and Mount Sinai Brain Bank cohort. SHAP explainer revealed common and specific transcriptomic signatures from different brain regions. DISCUSSION We identified common gene signatures among different brain regions in microglia and sex specific modules in neurons implicated in AD. This work paves the way for utilizing artificial intelligence approaches in studying AD at the molecular level. Research-in-Context Systematic review: Postmortem brain transcriptomes have been analyzed to study the molecular changes associated with Alzheimer’s disease, usually by a direct contrast approach such as differential gene expression analysis. Nuanced gene regulatory networks thus cannot be easily pinpointed from convoluted data such as those from bulk-tissue profiling. We applied a novel interpretable deep learning approach to dissect the RNA-seq data collected from three different brain regions of a large clinical cohort and identified significant genes for network analysis implicated for AD. Interpretation: Our models successfully predicted neuropathological and clinical traits in both internal and external validations. We corroborated known microglial biology in addition to revealing novel sex chromosome-linked gene contributing to sex dimorphism in AD. Future directions: The framework could have broad utility for interpreting multi-omic data such as those from single-cell profiling, to advance our understanding of molecular mechanisms of complex human disease such as AD. Highlights We applied novel interpretable deep learning methods to postmortem brain transcriptomes from three different brain regions We interpreted the models to identify genes most strongly implicated in AD Network analysis corroborated known microglial biology and revealed novel sex specific transcriptional factors associated with neuronal loss in AD