Single-cell gene set enrichment analysis and transfer learning for functional annotation of scRNA-seq data

Abstract

Although an essential step, cell functional annotation often proves particularly challenging from single-cell transcriptional data. Several methods have been developed to accomplish this task. However, in most cases, these rely on techniques initially developed for bulk RNA sequencing or simply make use of marker genes identified from cell clustering followed by supervised annotation. To overcome these limitations and automatize the process, we have developed two novel methods, the single-cell gene set enrichment analysis (scGSEA) and the single-cell mapper (scMAP). scGSEA combines latent data representations and gene set enrichment scores to detect coordinated gene activity at single-cell resolution. scMAP uses transfer learning techniques to re-purpose and contextualize new cells into a reference cell atlas. Using both simulated and real datasets, we show that scGSEA effectively recapitulates recurrent patterns of pathways’ activity shared by cells from different experimental conditions. At the same time, we show that scMAP can reliably map and contextualize new single-cell profiles on a breast cancer atlas we recently released. Both tools are provided in an effective and straightforward workflow providing a framework to determine cell function and significantly improve annotation and interpretation of scRNA-seq data.

Single cell lineage tracing reveals clonal dynamics of anti-EGFR therapy resistance in triple negative breast cancer

“…After reads alignment, cells with fewer than 5,000 UMI were discarded. Next, putative cell doublets and cells expressing a high fraction of mitochondrial reads were removed using the filterCell function from the gficf version 2 R package (27, 80, 81) available at https://github.com/gambalab/gficf. Specifically, the filterCell function employs loess regression to fit the relationship between the total UMI count in a cell (in log scale) and the ratio between the total UMIs falling in mitochondrial genes over the total UMI count.…”

“…scRNA libraries were sequenced with NovaSeq 6000 machine using an SP 100 cycles flow cell. Raw reads pre-processing was performed as described in (81). Only high depth cells with at least 5,000 UMI were retained and used to test the scATRAL tool.…”

“…Putative cell doublet and cells expressing an high fraction of mitochondrial genes were removed with filterCell function of gficf R package (27,75,76). Finally, only highly covered cells with at least 5000 UMI were retained and used for downstream analyses.…”

Predicting drug response from single-cell expression profiles of tumours

“…To predict drug response at the single-cell level, we first used gf-icf normalization (https://github.com/gambalab/gficf) (27) to extract the top relevant genes from each cell. We then used these gene sets as input for Gene Set Enrichment Analysis (GSEA) (29) against each GDPS ranked-list to predict the sensitivity of a cell to a specific drug.…”

Interpreting single-cell and spatial omics data using deep networks training dynamics

“…to cell types and states can be inferred based on established marker genes [7,8], clustering of gene expression profiles [9][10][11], combinations thereof [12], or in a supervised manner using complementary pre-annotated datasets [1]. Annotations related to localization across spatial or temporal processes can be reconstructed based on prior knowledge of the topology of a biological signal [13,14], principles of tissue organization [15,16], trajectory reconstruction [17,18], or annotation transfer between datasets [19][20][21].…”