celltype identification/deconvolution

[BigOmics Analytics Team]

[Daniel email 2.3.2023]

One more question about the cell type mapping module.  I notice several cell signature databases and mapping algorithms were included in the BigOmics.  What is the major algorithm for estimating the cell types from a mixture of cells such as PBMCs, and how is the proportion of different cell types estimated?  Could you send me some references that you used for this type of analysis?  

[BigOmics Analytics Team]

Hi Daniel,

Celltype estimation/deconvolution is indeed one of the strong functionalities in Playground. We have selected some of the best performing methods (at the time of implementation). We looked for methods that are general (i.e. can take any reference dataste) and efficient. We have concentrated on bulk RNA-seq methods, rather than the ones used for single-cell. Among the R packages we used DeconRNAseq and DCQ. One of the more known methods is CIBERSORT (from Stanford) but we couldn't include it in the final product because of the licensing. We implemented I-NNLS (iterative non-negative least squares) and non-negative linear modeling (both in linear and ranked space). In fact CIBERSORT is very similar to those. Finally, meta and meta.prod are "meta" methods that combine all methods to get a more robust estimation. Each methods performs differently. If you go to Mapping and select "heatmap by method" (see below), you can see how each method estimates the celltype and which methods are similar or different. Basically, all these methods are some kind of constrained-based (e.g. non-negative) regression. Using a set of reference vectors, we can use regression to estimate the coefficients.

The proportion of different cell types is simply computing the average of the most probable cell type (or probabilities) among the different groups, e.g. in the control group or treated group. 

In the future, we will need to evaluate some new methods and also the new reference atlas based methods for single cell RNAseq.

Ivo

BigOmics Team