Define and visualize pathological architectures of human tissues from spatially resolved transcriptomics using deep learning

Abstract

We present RESEPT, a deep-learning framework for characterizing and visualizing tissue architecture from spatially resolved transcriptomics by reconstructing and segmenting a transcriptome mapped RGB image. RESEPT can identify the tissue architecture, and represent corresponding marker genes and biological functions accurately. RESEPT also provides critical insights into the underlying mechanisms driving the complex tissue heterogeneities in Alzheimer’s disease and glioblastoma.

MAPLE: A Hybrid Framework for Multi-Sample Spatial Transcriptomics Data

“…As shown in Figure 1, MAPLE accepts multi-sample HST data input in the form of an integrated data object, where batch correction and adjustment for technical artifacts such as sequencing depth can be accomplished using standard approaches [Hao et al, 2020, Hafemeister and Satija, 2019, Korsunsky et al, 2019]. Users may then pass data to scGNN [Chang et al, 2021b, Wang et al, 2021] to compute low-dimensional cell spot embeddings from raw gene expression data using a spatially aware graph neural network, or use other standard dimension reduction methods such as PCA. Given the resultant low-dimension cell spot embedding, MAPLE then implements a spatial Bayesian finite mixture model [Frühwirth-Schnatter and Pyne, 2010] as detailed in Materials and Methods.…”

“…We sought to compare the effect of using spatially aware gene expression features generated from scGNN for tissue architecture identification versus standard spatially unaware features such as principal components (PCs). We considered the set of 16 manually annotated human brain samples analyzed in Chang et al [2021b], and for each data set we computed multi-dimensional cell spot gene expression embeddings using PCA, SpaGCN, and scGNN, while varying the number of dimensions from 3 to 18. We quantified agreement between ground truth expert annotations and MAPLE cell spot labels using the adjusted Rand Index (ARI) [Hubert and Arabie, 1985].…”

“…However, in the context of HST, these approaches are non-trivial due to the irreconcilable differences in spatial architecture across samples. Further, while a variety of methods have been proposed for cell spot sub-population identification in spatial transcriptomics data [Dries et al, 2019, Hao et al, 2020, Pham et al, 2020, Zhao et al, 2021, Chang et al, 2021b] there are no available methods for multi-sample analysis of HST data. Specifically, no existing methods simultaneously infer spatially-informed cell spot sub-populations in each sample while sharing information across samples.…”

“…SpaGCN [Hu et al, 2021] and scGNN [Wang et al, 2021], among others, have provided deep learning-based approaches for deriving spatially informed dimension reductions of HST data. These methods each train a spatially aware graph neural network to produce low dimension embeddings of cell spots, and have shown advantages of these spatially-aware features compared to standard non-spatial embeddings like principal components analysis (PCA) [Chang et al, 2021b, Hu et al, 2021, Wang et al, 2021].…”

Analysis of community connectivity in spatial transcriptomics data

“…The proliferation of HST data has lead to the development of several computational tools for discerning cell sub-populations in HST data, while considering both gene expression and spatial information. The existing tools span a range of methodological categories, including neural networks [Chang et al, 2021, Hu et al, 2021, Canozo et al, 2022], graph clustering algorithms [Dries et al, 2019, Hao et al, 2020, Pham et al, 2020], and Bayesian statistical models [Zhao et al, 2021, Allen et al, 2021].…”