Presented By: Department of Computational Medicine and Bioinformatics DCMB
DCMB Tools and Technology Seminar by Jeongbin Park (Computational Medicine & Bioinformatics)
“Hierarchical inference of spatial tissue niches and cellular composition using linked variational autoencoders”
Abstract
Spatially resolved transcriptomics (SRT) measures gene expression in thousands to millions of spatial locations, providing information on the spatial organization of various tissues. The spatial organization for complex tissues forms a hierarchical structure, from individual genes, to cell types and co-localized cell groups, to large spatial domains. Although many computational methods identify spatial domains using SRT data, modeling the hierarchical organization still needs improvement. Moreover, the latest SRT technologies (e.g., 10x Xenium and Visium HD) measure over 0.1 million spots, further challenging the scalability of previous computational methods. We developed COMPOSITION that models the hierarchical spatial structure of a tissue profiled by the latest SRT technologies and scales to the large number of spots in latest SRT technologies. We explicitly model the conditional probability relationships between adjacent layers of the tissue hierarchy in COMPOSITION and develop a neural network structure based on variational autoencoder (VAE) that enables fast inference of the posterior probability of the model.
About the DCMB Tools & Technology Seminar Series
The DCMB Tools and Technology Seminar Series is held in Medical Science Building 1 (MS1), Room 4B700, each Thursday at 12pm EST. Each seminar highlights a computational tool, technology, or methodology that is under development or in current use and is of special interest to DCMB and University researchers. Presenters are U-M researchers and students.
These seminars are live-streamed and recorded and made available for future viewing via the DCMB YouTube Channel
Spatially resolved transcriptomics (SRT) measures gene expression in thousands to millions of spatial locations, providing information on the spatial organization of various tissues. The spatial organization for complex tissues forms a hierarchical structure, from individual genes, to cell types and co-localized cell groups, to large spatial domains. Although many computational methods identify spatial domains using SRT data, modeling the hierarchical organization still needs improvement. Moreover, the latest SRT technologies (e.g., 10x Xenium and Visium HD) measure over 0.1 million spots, further challenging the scalability of previous computational methods. We developed COMPOSITION that models the hierarchical spatial structure of a tissue profiled by the latest SRT technologies and scales to the large number of spots in latest SRT technologies. We explicitly model the conditional probability relationships between adjacent layers of the tissue hierarchy in COMPOSITION and develop a neural network structure based on variational autoencoder (VAE) that enables fast inference of the posterior probability of the model.
About the DCMB Tools & Technology Seminar Series
The DCMB Tools and Technology Seminar Series is held in Medical Science Building 1 (MS1), Room 4B700, each Thursday at 12pm EST. Each seminar highlights a computational tool, technology, or methodology that is under development or in current use and is of special interest to DCMB and University researchers. Presenters are U-M researchers and students.
These seminars are live-streamed and recorded and made available for future viewing via the DCMB YouTube Channel
Related Links
Explore Similar Events
-
Loading Similar Events...
