Improved cell composition deconvolution method of bulk gene expression profiles to quantify subsets of immune cells

Area of Science:

• Computational biology
• Immunology
• Bioinformatics

Background:

• Investigating immune cell roles in tumors requires accurate cell composition analysis.
• Existing deconvolution methods lack robust reference gene expression profiles, especially for key immune cells.
• Imbalanced reference datasets limit the precision of predicting immune cell fractions in complex tissues.

Purpose of the Study:

• To develop a novel computational deconvolution method for predicting immune cell composition.
• To create an improved reference gene expression profile dataset with more immune cell replicates.
• To enhance the accuracy of immune cell fraction estimation in tumor microenvironments.

Main Methods:

• Utilized ε-support vector regression (ε-SVR) with an L1-norm penalty loss function.
• Constructed a reference gene signature matrix using differentially expressed genes selected via ANOVA and condition number minimization from 148 microarray profiles.
• Validated the method against CIBERSORT using in silico and real human samples.

Main Results:

• The new method demonstrated superior performance compared to CIBERSORT in in silico breast tissue-immune cell mixtures.
• Outperformed CIBERSORT in benchmarks using simulated bulk tissues and 164 human peripheral blood mononuclear cell (PBMC) samples.
• Achieved performance comparable to state-of-the-art tools, suggesting improved accuracy in immune cell deconvolution.

Conclusions:

• A novel deconvolution method and reference dataset were developed using publicly available R and Python packages.
• The new reference profiles enable more robust prediction of immune cell fractions from mixed cell expression data.
• Source code is available for download, facilitating further research in immune cell deconvolution for cancer studies.