Abstract
Motivation To improve cancer immunotherapy response, one crucial step is to study the immune/stromal cell composition in the tumor microenvironment. The fraction of different cell types in the tumor microenvironment can be estimated via deconvolution algorithms from bulk transcriptomic profiles. One class of such algorithms, known as reference-based, requires as input a reference signature matrix containing the gene expression measurements of different cell types. The limitation of these algorithms is that problems might arise when the transcriptomic profiles of different cell types in solid tumors deviate from the reference, leading to poor estimation performance. A more flexible alternative is given by reference-free methods which can perform the simultaneous estimation of cell-type fractions and cell-type gene expression from the data. However, most of these algorithms rely on factor modeling which unfortunately suffers from interpretability issues, as the labeling of different factors into cell types is often problematic.
Results To overcome these limitations, we propose BayesDeBulk, a novel reference-free Bayesian model which flexibly leverages existing information of known cell-type specific markers and performs the simultaneous estimation of cell-type fractions and cell-type gene expression. Specifically, BayesDebulk imposes a novel Repulsive prior distribution on the mean of cell-type specific markers to ensure the upregulation of cell-type specific markers in a particular component. Using this prior specification, each component of the mixture model is identifiable and automatically assigned to a particular cell type, overcoming the identifiability issues affecting reference-free methods. This flexible framework enables BayesDeBulk to perform the deconvolution by integrating proteomic and transcriptomic data measured for the same set of samples. Improved performance of BayesDeBulk over state-of-the-art deconvolution algorithms such as Cibersort and xCell is shown on different synthetic and real data examples.
Availability Software available at http://www.bayesdebulk.com/
Contact For any information, please contact francesca.petralia{at}mssm.edu
Competing Interest Statement
The authors have declared no competing interest.
Footnotes
The proposed algorithm was compared with additional existing deconvolution tools.