Comprehensive evaluation of human brain gene expression deconvolution methods

摘要: Gene expression measurements, similarly to DNA methylation and proteomic are influenced by the cellular composition of sample analysed. Deconvolution bulk transcriptome data aims estimate a from its gene data, which in turn can be used correct for differences across samples. Although multitude deconvolution methods have been developed, it is unclear whether their performance consistent tissues with different complexities composition. For example, human brain unique transcriptomic diversity, complexity cellularity, yet comprehensive assessment accuracy on currently lacking. Here we carry out first comparative evaluation assess tissue-specificity our key observations comparison pancreas. We evaluate 22 approaches, covering all main classes: 3 partial methods, each applied 6 categories cell-type signature 2 enrichment complete methods. test cell type estimates using silico mixtures single-cell RNA-seq neuronal glial RNA, as well nearly 2,000 Our results bring several important insights into deconvolution: (a) find that has stronger impact than choice method. In contrast, only mildly influences pancreas highlighting importance tissue-specific benchmarking. (b) demonstrate biological factors influencing (e.g. region, vitro culturing), effects outcome technical RNA sequencing platform). (c) outperform data. (d) differential analyses tissue-specific, more pronounced To facilitate wider implementation correction composition, develop novel signature, MultiBrain, integrates single-cell, immuno-panned, single-nucleus datasets. achieves improved over existing reference signatures. autism cases controls MultiBrain identified changes replicable studies, highlighted genes dysregulated autism.