Tuesday, Aug 5: 10:30 AM - 12:20 PM
0175
Invited Paper Session
Music City Center
Room: CC-Davidson Ballroom B
Applied
Yes
Main Sponsor
Section on Statistics in Genomics and Genetics
Co Sponsors
Biometrics Section
ENAR
Presentations
Spatial proteomic technique often measures 20-40 protein markers of each cell, using samples collected from tissue microarrays. While each tissue sample is small and the number of protein markers is limited, this approach is more scalable than other spatial omic techniques. A few studies have collected spatial proteomic data from hundreds or thousands of patients. We propose to use graph neural network to extract features from spatial proteomic data so that those features are comparable across samples. Graph neural network combines spatial information (graph connections) together with attribute of each cell (e.g., the cell type of each cell inferred from the spatial protein omic data). Such features can be used for supervised or unsupervised down-stream analysis. We will focus on unsupervised analysis to identify meaningful domains in tumor samples.
Keywords
Graph Neural Network
Spatial Proteomic
Clinical Outcomes
Co-Author(s)
Si Liu
Li Hsu, Fred Hutchinson Cancer Research Center
Wei Sun, Fred Hutchinson Cancer Center
Speaker
Wei Sun, Fred Hutchinson Cancer Center
Spatial omics technologies have revolutionized biomedical research by providing detailed, spatially resolved molecular profiles that enhance our understanding of tissue structure and function at unprecedented levels. However, the widespread application of spatial omics is hindered by its high costs. In contrast, generating hematoxylin and eosin stained histology images is significantly more cost-effective. Previous studies have shown correlations between spatial molecular patterns and histology image features, suggesting that histology images can be leveraged to design cost-effective spatial omics experiments. In this talk, I will present our recent work on utilizing histology images to select regions of interest for spatial omics experiments and how the resulting data can be trained to virtually construct spatial molecular profiles for entire tissue sections. We demonstrate that our histology image-guided design can substantially reduce experimental costs while retaining the desired spatial molecular variations.
Keywords
experimental design
spatial omics
AI
Speaker
Mingyao Li, University of Pennsylvania, Perelman School of Medical
Spatial transcriptomics technologies measure RNA expression at thousands of locations in a 2D tissue slice providing information about the spatial distribution of cell types and the spatial variation in gene expression across a tissue. However, these measurements are typically sparse with high rates of missing data. This talk will describe approaches to align and integrate spatial transcriptomics data from multiple tissues slices using optimal transport, a framework for computing maps between probability distributions. One method, DeST-OT aligns slices from pairs of timepoints from a development process, using semi-relaxed optimal transport to derive rates of cell growth/death. The second method Hidden Markov OT finds consistent low rank representations across multiple timepoints from a developmental process enabling the derivation of cell differentiation maps between cell types. Application of these methods to spatial transcriptomics data from multiple species will be described.
Keywords
Spatiotemporal Transcriptomics Data
Studying how intra-tumoral immune populations coordinate to generate anti-tumor responses can guide precise treatment prioritization. Recent genomic technologies that measure cell features at the resolution of single cells or in a spatially-resolved manner, present exciting opportunities to study the heterogeneity of cells and characterize complex interactions in the tumor microenvironment (TME). However, analyzing and integrating these data types in particular in complex patient specimens involves significant statistical and computational challenges. I will present a set of statistical machine learning methods developed to infer temporal and spatial dynamics of cells in the TME and tumor-immune interactions. I will show their application in the characterization of coordinated immune cell networks in an established adoptive cellular therapy, donor lymphocyte infusion (DLI) in relapsed leukemia, as well as checkpoint therapy in melanoma.
Keywords
computational biology
spatiotemporal dynamics
single-cell genomics
spatial transcriptomics
cancer immunology