Embeddings in Integrative Analysis of Spatial ’omics

It used to be the case that, to obtain measurements of a biological system, one had to be something of a specialist in the particular sensing technology. Even if a technology produced a very high-dimensional description of a system, you would only typically have one of these types of data to have to analyze at a time.

Now, in contrast, a broad array of sensing technologies are easily accessible, and it’s not uncommon for a researcher interested in a system to obtain several complementary views using several technologies. This is the starting point for integrative analysis.

This package describes analysis done for the workshop Mathematical Frameworks for Integrative Analysis of Emerging Biological Data Types. The original data and questions are described here. These analysis formed the basis of a short talk.

The running theme across the four vignettes included here is that each measurement technology gives a slightly different view of the same overall system. Our goal is to measure the extent to which these different views give redundant vs. orthogonal information. Which combinations of sensing technologies are likely to give the most diverse set of findings, or the strongest evidence for conclusions, about a single underlying system?

The primary methodological device used across the vignettes is the reduction of raw data into directly inspectable features. These representations are easier to work with than raw pixel values, say, and they facilitate statistical analysis of the relationship between data sources.

The different vignettes have different emphases,

  • Deriving Graph and Image Features shows how to build meaningful features manually, based on source MIBI-TOF raster data.
  • Interactive Visualization using Linked Views considers the use of visualization interfaces to streamline inspection of learned cell atlases.
  • Disentangling Composition and Ecological Effects is focused on the extent to which ecological features of a sample can be recovered from the cells that it contains.
  • Mapping between Cell-Level Embeddings considers the analogous problem at the cellular level.

Running the code

The goal of packaging these analysis is to make them easy to rerun with minimal effort. You don’t even need to download R – just pull the docker image and enter the vignettes. Specifically, pull and run the docker image using,

docker run -e PASSWORD=<make_up_a_password_here> -p 8787:8787 krisrs1128/birs2020_scproteomics_embeddings

and then navigate to localhost:8787 and enter username = rstudio and password = your password. In the console, install all dependencies using remotes::install_deps(dependencies = TRUE, repos = BiocManager::repositories()). Then, you can then browse to the vignettes folder and run all the analysis on this site.