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

Abstract

The utilization of artificial intelligence in studying the dysregulation of gene expression in Alzheimer's disease (AD) affected brain tissues remains underexplored, particularly in delineating common and specific transcriptomic signatures across different brain regions implicated in AD-related cellular and molecular processes, which could help illuminate novel disease biology for biomarker and target discovery. Herein we developed a deep learning framework, which consisted of multi-layer perceptron (MLP) models to classify neuropathologically confirmed AD versus controls, using bulk tissue RNA-seq data from the RNAseq Harmonization Study of the Accelerating Medicines Project for Alzheimer's Disease (AMP-AD) consortium. The models were trained based on data from three distinct brain regions, including dorsolateral prefrontal cortex (DLPFC), posterior cingulate cortex (PCC), and head of the caudate nucleus (HCN), obtained from the Religious Orders Study/Memory and Aging Project (ROSMAP). Subsequently, we inferred a disease progression trajectory for each brain region by applying unsupervised dimensionality transformation to the distribution of the subjects' expression profiles. To interpret the MLP models, we employed an interpretable method for deep neural network models, obtaining SHapley Additive exPlanations (SHAP) values and identified the most significantly AD-implicated genes for gene co-expression network analysis. Our models demonstrated robust performance in classification and prediction across two other external datasets from the Mayo RNA-seq (MAYO) cohort and the Mount Sinai Brain Bank (MSBB) cohort of AMP-AD. By interpreting the models both mechanistically and biologically, our study elucidated subtle molecular alterations in various brain regions, uncovering shared transcriptomic signatures activated in microglia and sex-specific modules in neurons relevant to AD. Notably, we identified, for the first time, a sex-linked transcription factor pair (ZFX/ZFY) associated with more pronounced neuronal loss in AD females, shedding light on a novel mechanism for sex dimorphism in AD. This study lays the groundwork for leveraging artificial intelligence methodologies to investigate AD at the molecular level, which is not readily achievable from conventional analysis approaches such as differential gene expression (DGE) analysis. The transcription factor implicated in sex difference also underpins a new molecular mechanistic basis of women's greater neurodegeneration in AD warranting further study.

Fig. 1. The deep learning and interpretation framework employed in this work.

A Overview of the framework. Using the gene expression profiles from AD and control subjects and their diagnosis class as the input for supervised classification, the model was trained by a multilayer neural network. The trained network was passed forward to the profiles from the whole cohort with the resulting output manifold subject to unsupervised dimensionality transformation (UMAP) to obtain the pseudo-temporal trajectory and severity index (SI). SI was linearly correlated with phenotypic data for evaluation. Trained model was interpreted by SHAP explainer to obtain the most salient features (index genes, or IGs). Their co-expression relationship was examined for biological interpretation. The framework was applied to the three brain regions (DLPFC, PCC, and HCN) from ROSMAP cohort, respectively, and the derived data were compared across different brain regions. B Technical details of the workflow.

Fig. 2. The pseudo-temporal trajectories from the trained models for the transcriptome from three regions of ROSMAP cohort and the SI correlation with phenotypical data.

A, B DLPFC without and with OTHER samples. C, D Spider plots showing the linear correlations of SIs with phenotypical traits for all three models.

Fig. 3. The pseudo-temporal trajectories from the forward pass of ROSMAP model and mapping to the same 3D space as ROSMAP (DLPFC only), based on the transcriptome from two regions of MAYO cohort and the SI correlation with phenotypical data from all three models’ predictions.

PA pathological aging, PSP progressive supranuclear palsy. A TCX mapped to DLPFC; B CER mapped to DLPFC. C, D Spider plots showing the linear correlations of SIs with phenotypical traits from all three models.

Fig. 4. The pseudo-temporal trajectories from the forward pass of ROSMAP model and mapping to the same 3D space as ROSMAP (DLPFC only), based on the transcriptome from four regions of MSBB cohort and the SI correlation with phenotypical data from all three models’ predictions.

A–D Four regions (FP, STG, PHG, and IFG) mapped to DLPFC. E, F Spider plots showing the linear correlations of SIs with phenotypical traits from all three models.

Fig. 5. Co-expression modules resolved from the expression profiles of IGs in DLPFC region and their functional annotations.

A–E Five modules clustered by their cell type enrichment. Only hub genes are labeled. F Functional enrichment for each module.

Fig. 6. Curated co-expression plots for two modules from DLPFC region with all gene nodes labeled.

A Microglia module. B A subset from the neuron module including the hub gene SVOP and a sex-linked submodule.

Fig. 7. SVOP and ZFX (ZFY) co-expression based on the reprocessed ROSMAP data from DLPFC region.

A Regression between SVOP and ZFX, ZFY and their combined expression. B Boxplots showing ZFX, ZFY, their cumulation and SVOP expression stratified by sex.