DeCompress

Welcome to DeCompress - a semi-reference-free method to deconvolve targeted panels of mRNA expression into tissue compartment. A tissue compartment is a group of cells of similar type or biological function (i.e. immune or stroma or tumor compartments). Please cite Bhattacharya et al 2020 if you use our package. Visit our documentation page.

Installation

You can install DeCompress using devtools::install_github('bhattacharya-a-bt/DeCompress'). Make sure to have dependencies installed as well!

Using DeCompress

DeCompress requires two input data matrices:

• the target matrix (k X n) with k genes and n samples, and
• the reference matrix (K X n) with K genes and N samples with K > k.

DeCompress also requires a priori knowledge of the number of tissue compartments. If you don't know the number of compartments you wish to deconvolve to, you can use the findNumberCells() function to get an estimate from an singular value decomposition of the target matrix.

To illustrate DeCompress, let's generate some fake data. Here, pure is a matrix of compartment-specific expression profiles, reference_props and target_props are different $200 \times 4$ matrices of proportions, and reference and target gives us the input expression matrices.

pure = matrix(abs(rnorm(1e4*4)),ncol=4)
rownames(pure) = paste0('Gene',1:1e4)
reference_props = apply(matrix(abs(rnorm(200*4)),ncol=4),
        1,function(x) x/sum(x))
target_props = apply(matrix(abs(rnorm(200*4)),ncol=4),
        1,function(x) x/sum(x))

reference = pure %*% reference_props
target = pure[sample(1:1e4,400),] %*% target_props

Step 1: Feature selection

Step 1 of DeCompress is to feature select a set of $K'$ genes from the reference that are compartment-specific. We have a wrapper function for this:

compSpec = findInformSet(yref = reference,
                         method = 'variance',
                         n_genes = 1000,
                         n.types = 4)

The method = variance option calls TOAST for this feature selection, a method we have seen to be best suited for this task. The method = linearity uses the linseed method's mutual linearity assumption to select compartment-specific genes. findInformSet() returns the reference matrix, reduced to these $K'$ genes.

Step 2: Train the compressed sensing matrix

Step 2 of DeCompress is to train the compressed sensing model that projects the feature space of target matrix to the $K'$ compartment-specific genes. This is done with the trainCS() function.

csModel = trainCS(yref = reference[rownames(target),],
                    yref_need = compSpec,
                    seed = 1111,
                    method = c('lar',
                               'lasso',
                               'enet',
                               'ridge',
                               'l1',
                               'TV',
                               'l2'),
                    par = T,
                    n.cores = 4,
                    lambda = .1)

Here, yref takes in the reference matrix subsetted to the $k$ genes on the target and yref_need takes in the reference matrix subsetted to the $K'$ compartment-specific genes. The method option allows for various predictive models: lar is least angle regression, lasso, enet, and ridge fits the glmnet methods, and l1, TV, and l2 are non-linear optimization methods to different penalties (see the R1Magic package). You can also parallelize by toggling par and setting the number of cores to n.cores. lambda is a penalization parameter for the non-linear methods defaulted to 0.1. The three glmnet methods are fastest and work as well as (if not better than) the other methods.

Once the compressed sensing model is fit, the compression matrix is extracted with csModel$compression.matrix and the DeCompressed expression is calculated with dcexp = expandTarget(target,csModel$compression.matrix).

Step 3: Ensemble deconvolution

Step 3 of DeCompress is to run ensemble deconvolution on the DeCompressed expression data.

csModel = bestDeconvolution(yref = dcexp,
                            n.types = 4,
                            known.props = target_props,
                            methods = c('TOAST',
                                        'linseed',
                                        'celldistinguisher'))

Here, yref takes in the DeCompressed matrix with n.types taking in the number of compartments. known.props can be passed a proportion matrix if that is known; if left set to NULL, the function will output the deconvolution with the smallest reconstruction error. The method option allows for various deconvolution methods. Feel free to use your favorite reference-free method to deconvolve dcexp.