Fig 1.
New gene expression data from cell types (e.g. from a tumor microenvironment) will be used to construct a new signature matrix de novo or by augmenting an existing signature matrix. First ADAPTS will rank marker genes for the cell types using the function rankByT described in Eq 4. Then ADAPTS adds marker genes in rank order using the function AugmentSigMatrix as described in Algorithm 1 resulting in a new signature matrix. This matrix may be tested for spillover between cell types using the function spillToConvergence described in Algorithm 2. Finally, ADAPTS separates cell types with heavy spillover using the hierarchical deconvolution function hierarchicalSplit described in Algorithm 3 to estimate the percentage of cell types present in bulk gene expression data.
Fig 2.
Curve showing the selection of an optimal condition number for MGSM27.
Fig 3.
Spillover matrix showing mean misclassification of purified samples for LM22. Rows show purified cell types and columns show what those samples deconvolve as. Cells are colored by percentage, such that each row adds up to 100%. For example, if the row is ‘B.cell.memory’, the column is ‘Plasma.cells’, then the color is light blue indicating that purified ‘B.cell.memory’ samples deconvolve as containing (on average) 18% ‘Plasma.cells’.
Fig 4.
LM22 converged spillover matrix.
Iterative deconvolution shows how easily confused cell types conspicuously form clusters. Rows show purified cell types and columns show what those samples deconvolve as. Cells are colored by percentage, such that each row adds up to 100%.
Table 1.
Deconvolved tumor accuracy in WBM samples.
Table 2.
Deconvolution of CD138+ purified samples.
Table 3.
Deconvolution of WBM samples.
Fig 5.
scRNAseq signature matrix construction.
Curve showing the selection of an optimal condition number for the single cell RNAseq augmented signature matrix data.
Table 4.
Deconvolution of pancreas training set.
Table 5.
Deconvolution of pancreas test set.
Fig 6.
Clustering of Top 100 gene signature matrix.
The cell type clusters identified using the signature matrix constructed from the 100 genes with the highest variance across cell types in the single cell data drawn from a normal pancreas sample. Rows show purified cell types and columns show what those samples deconvolve as. Cells are colored by percentage, such that each row adds up to 100%.
Fig 7.
Clustering of augmented gene signature matrix.
The cell type clusters identified using the augmented signature matrix that was seeded with the 100 genes exhibiting the highest variance in the normal pancreas sample. Rows show purified cell types and columns show what those samples deconvolve as. Cells are colored by percentage, such that each row adds up to 100%.