multideconv - an integrative pipeline for efficiently combining first and second generation cell type deconvolution results

The number of computational methods for cell type deconvolution from bulk RNA-seq data has been increasing in the last years, but their high feature complexity and variability of results across methods and signatures limit their utility and effectiveness for patient stratification. Applying multiple combinations of deconvolution methods and signatures often results in hundreds of redundant or contradictory cell type features describing the composition of complex tumour samples. Benchmarking efforts are inherently limited by the lack of bias-free ground truth, often yielding inconsistent results or no consensus.

To address these limitations, we present multideconv , an R package that reduces dimensionality and eliminates redundancy in deconvolution results, through unsupervised filtering and iterative correlation analyses. Built on top of existing frameworks, multideconv harmonizes outputs across methods to identify robust cell type proportion estimates and mitigate signature-driven heterogeneity. We benchmarked multideconv against two existing methods that provide similar functions and found it to yield more accurate estimations of cell type proportions based on virtual bulk reconstruction from single-cell expression datasets. We also increase computational efficiency by providing a meta-cell aggregation of the single-cell datasets, showing it preserves the samples’ complexity.

Despite our focus on tumour samples in the context of immuno-oncology, the tool is flexible and can be adapted to infer mixed sample composition from bulk RNAseq datasets.

The multideconv R package and tutorials are available at https://github.com/VeraPancaldiLab/multideconv . The code to reproduce the analysis and figures is available on github at https://github.com/VeraPancaldiLab/multideconv_paper .

Contact: marcelo.hurtado@inserm.fr or vera.pancaldi@inserm.fr