Molecular and immune signatures, and pathological trajectories of fatal COVID-19 lungs defined by in situ spatial single-cell transcriptome analysis.

Despite intensive studies during the last 3 years, the pathology and underlying molecular mechanism of coronavirus disease 2019 (COVID-19) remain poorly defined. In this study, we investigated the spatial single-cell molecular and cellular features of postmortem COVID-19 lung tissues using in situ sequencing (ISS). We detected 10 414 863 transcripts of 221 genes in whole-slide tissues and segmented them into 1 719 459 cells that were mapped to 18 major parenchymal and immune cell types, all of which were infected by SARS-CoV-2. Compared with the non-COVID-19 control, COVID-19 lungs exhibited reduced alveolar cells (ACs) and increased innate and adaptive immune cells. We also identified 19 differentially expressed genes in both infected and uninfected cells across the tissues, which reflected the altered cellular compositions. Spatial analysis of local infection rates revealed regions with high infection rates that were correlated with high cell densities (HIHD). The HIHD regions expressed high levels of SARS-CoV-2 entry-related factors including ACE2, FURIN, TMPRSS2 and NRP1, and co-localized with organizing pneumonia (OP) and lymphocytic and immune infiltration, which exhibited increased ACs and fibroblasts but decreased vascular endothelial cells and epithelial cells, mirroring the tissue damage and wound healing processes. Sparse nonnegative matrix factorization (SNMF) analysis of niche features identified seven signatures that captured structure and immune niches in COVID-19 tissues. Trajectory inference based on immune niche signatures defined two pathological routes. Trajectory A primarily progressed with increased NK cells and granulocytes, likely reflecting the complication of microbial infections. Trajectory B was marked by increased HIHD and OP, possibly accounting for the increased immune infiltration. The OP regions were marked by high numbers of fibroblasts expressing extremely high levels of COL1A1 and COL1A2. Examination of single-cell RNA-seq data (scRNA-seq) from COVID-19 lung tissues and idiopathic pulmonary fibrosis (IPF) identified similar cell populations consisting mainly of myofibroblasts. Immunofluorescence staining revealed the activation of IL6-STAT3 and TGF-ß-SMAD2/3 pathways in these cells, likely mediating the upregulation of COL1A1 and COL1A2 and excessive fibrosis in the lung tissues. Together, this study provides a spatial single-cell atlas of cellular and molecular signatures of fatal COVID-19 lungs, which reveals the complex spatial cellular heterogeneity, organization, and interactions that characterized the COVID-19 lung pathology.

Transformer for Gene Expression Modeling (T-GEM): An Interpretable Deep Learning Model for Gene Expression-Based Phenotype Predictions.

Deep learning has been applied in precision oncology to address a variety of gene expression-based phenotype predictions. However, gene expression data's unique characteristics challenge the computer vision-inspired design of popular Deep Learning (DL) models such as Convolutional Neural Network (CNN) and ask for the need to develop interpretable DL models tailored for transcriptomics study. To address the current challenges in developing an interpretable DL model for modeling gene expression data, we propose a novel interpretable deep learning architecture called T-GEM, or Transformer for Gene Expression Modeling. We provided the detailed T-GEM model for modeling gene-gene interactions and demonstrated its utility for gene expression-based predictions of cancer-related phenotypes, including cancer type prediction and immune cell type classification. We carefully analyzed the learning mechanism of T-GEM and showed that the first layer has broader attention while higher layers focus more on phenotype-related genes. We also showed that T-GEM's self-attention could capture important biological functions associated with the predicted phenotypes. We further devised a method to extract the regulatory network that T-GEM learns by exploiting the attributions of self-attention weights for classifications and showed that the network hub genes were likely markers for the predicted phenotypes.