Community assessment of methods to deconvolve cellular composition from bulk gene expression.

Area of Science:

• Computational Biology
• Immunology
• Bioinformatics

Background:

• Deconvolution methods estimate immune cell infiltration from bulk tumor gene expression data.
• Accurate immune cell profiling is crucial for understanding tumor microenvironments and developing immunotherapies.

Purpose of the Study:

• To comprehensively evaluate existing and novel deconvolution methods.
• To identify limitations in current methods, particularly for specific immune cell states.
• To establish a benchmark for future deconvolution method development.

Main Methods:

• A community-wide DREAM Challenge was organized to assess deconvolution algorithms.
• In vitro and in silico transcriptional profiles of admixed cancer and healthy immune cells were used for evaluation.
• Six published and 22 community-contributed methods were analyzed.

Main Results:

• Several published methods accurately predicted most cell types but struggled with specific CD8+ T cell states.
• Community-contributed methods, including a deep learning approach, improved accuracy for challenging cell types.
• Deconvolution methods demonstrated good performance in predicting tumor-infiltrating immune cells, even when trained on healthy immune cells.

Conclusions:

• Deep learning represents a promising paradigm for improving immune cell deconvolution.
• The developed transcriptional profiles serve as a valuable resource for advancing deconvolution techniques.
• Further development is needed to enhance the sensitive identification of functional CD4+ T cell states.