Few lymphocytes detected by cell type deconvolution tools in lung adenocarcinoma

Question

[350769816]

Lung adenocarcinoma(LUAD) is a kind of solid tumor which is usually thought to originate from AT2 cells in aveolar epithelium. Recently I tried BayesPrism, SCDC and MuSiC to deconvolute LUAD bulk RNA seq data with single cell from same patients as reference. Figure below is the deconvolution result of BayesPrism. .

Few lymphocytes can be detected. Then I also tried deconvolution on TCGA LUAD and a singapore LUAD datasets. They are all same.
However, I found an article "Precise reconstruction of the TME using bulk RNA-seq and a machine learning algorithm trained on artificial transcriptomes". The result of the machine learning tool called "Kassandra" of the article showed a much higher lymphocyte fraction (lymphocyte fraction even came to ~50% in that article).
Another article "Spatial Organization and Molecular Correlation of Tumor-Infiltrating Lymphocytes Using Deep Learning on Pathology Images" explore lymphocyte fraction based on H&E images. They found the TIL fraction level of TCGA LUAD is lower than 10% but definitely higher than my results. But this result is based on a different system.
I'm really confused. Appreciate anyone sharing your experience.

Answer 1

[350769816]

This is the answer from author of BayesPrism, may be also same for MuSiC, SCDC and Cibersortx.
One possible cause of the lower lymphocyte fraction can be attributed to the fact that BayesPrism infers the reads from each cell type% rather than cell count%. As lymphocyte expresses way lower amount of reads (you may observe this in your scrna-seq reference), the inferred fraction can be lower than H&E.

[ATpoint]

lymphocyte expresses way lower amount of reads

Is this the case, do you see this in your data? Just curious, can you show above plot but with log10(total reads per cell) in y-axis? Thx.

[350769816]

Same dataset but cells are divided into cell subtypes. RNA count of T cell and other lymphocytes is much lower than malignant cells. But this is result of all genes. When I performed deconvolution, only protein coding genes were included.