Hu, Hongru

ABSTRACT Recent development of single-cell technology across multiple omics platforms has provided new ways to obtain holistic views of cells to study disease pathobiology. Alzheimer’s disease (AD) is the most common form of dementia worldwide, yet the detailed understanding of its cellular and molecular mechanisms remains limited. In this study, we analyzed paired single-cell transcriptomic (scRNA-seq) and chromatin accessibility (scATAC-seq) data from the Seattle Alzheimer’s Disease Brain Cell Atlas (SEA-AD) Consortium to investigate the molecular mechanisms of AD at a cell-subpopulation-specific resolution focusing on glial cells. We benchmarked various multi-omics integration methods using diverse metrics and built an analytic workflow that enabled effective batch correction and cross-modality alignment, creating a unified cell state space. Through integrative analysis of 26 human brain samples, we uncovered AD-associated gene expression and pathway changes in glial subpopulations and highlighted important transcriptomic and epigenomic signatures via functional inference and interpretable machine learning paradigms, discovering the profound involvement of the Solute Carrier proteins (SLC) family genes in multiple glial cell types. We also identified glial cell-specific regulatory programs mediated by key transcription factors such as JUN and FOSL2 in astrocytes, the Zinc Finger (ZNF) family genes in microglia, and the SOX family of transcription factors in oligodendrocytes. Our study provides a comprehensive workflow and a high-resolution view of how glial regulatory programs are disrupted in AD. Our findings offer novel insights into disease-related changes in gene regulation and suggest potential targets for further research and therapy.

Abstract Single-cell RNA sequencing (scRNA-seq) enables high-resolution exploration of cellular diversity and gene regulation, yet analyzing such data remains challenging due to technical and methodological limitations. Existing task-specific deep generative models like Variational Auto-Encoder (VAE) and its variants struggle to incorporate external biological knowledge, while transformer-based foundational large Language Models (LLMs or large LaMs) face limitations in computational cost and applicability to tabular gene expression data. Here, we introduce sciL-aMA (single-cell interpretable Language Model Adapter), a novel representation learning frame-work that bridges these gaps by integrating static gene embeddings from multimodal LLMs with scRNA-seq tabular data through a paired-VAE architecture. Our approach generates context-aware representations for both cells and genes and outperforms state-of-the-art methods in key single-cell downstream tasks, including batch effect correction, cell clustering, and cell-state-specific gene marker and module identification, while maintaining computational efficiency. sciL-aMA offers a computationally efficient, unified framework for comprehensive single-cell data analysis and biologically interpretable gene module discovery. Source code is available at https://github.com/microsoft/sciLaMA