Sampling and ranking spatial transcriptomics data embeddings to identify tissue architecture

Yang, Lin; Wang, Yan; Liang, Yanchun; Yu, Hanry; Li, Jingyi; Ma, Qin; He, Fei; Xu, Dong

doi:10.3389/fgene.2022.912813

Cited by 1 publication

(1 citation statement)

References 36 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…These results also suggested an information saturation limit present in HST data, wherein the gained in performance from using spatially aware features relative to PCA diminishes as the number of dimensions used increases. The choice of number of dimensions used for embeddings is often ad hoc (56), and presents an opportunity to the field for future investigation.…”

Section: Discussionmentioning

confidence: 99%

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

Allen

Chang

et al. 2022

Preprint

Self Cite

View full text Add to dashboard Cite

The advent of high throughput spatial transcriptomics (HST) technologies has allowed for characterization of spatially and genetically distinct cell sub-populations in tissue samples -- an analysis known as tissue architecture identification. However, existing methods do not allow for simultaneous analysis of multiple tissue samples to assess the effect of factors like disease or treatment status on tissue architecture. Moreover, standard tissue architecture identification approaches do not accompany cell sub-population inference with uncertainty measures, thus over-simplifying biologically complex phenomena such as cell states differentiation. Finally, no existing frameworks have organically integrated the feature embedding power of deep learning with the rigor and interpretability of Bayesian statistical models in HST data analyses. We developed MAPLE: a hybrid deep learning and Bayesian modeling framework for detection of spatially informed cell sub-populations, uncertainty quantification, and inference of group effects in multi-sample HST experiments. It is the first approach to combine the recent advancements in graph neural networks for feature embedding and Bayesian spatial mixture modeling for interpretable cell sub-population identification. MAPLE also adopts a robust cell sub-population membership regression model to explain cell sub-population abundance in terms of available sample-level covariates such as treatment group, disease status, or tissue region, thereby allowing for more biologically interpretable sub-populations. It is also the first method to provide associated measures of uncertainty with cell sub-population labels, which enables transparent identification of ambiguous labels. We demonstrated the capability of MAPLE to achieve accurate, comprehensive, and interpretable tissue architecture inference through four case studies that spanned a variety of organs in both humans and animal models. An R package maple is available at https://github.com/carter-allen/maple.

show abstract

Section: Discussionmentioning

confidence: 99%