decoupleR: ensemble of computational methods to infer biological activities from omics data

1Omics datasets, such as transcriptomics or phospho-proteomics, provide unbiased high-dimensional molecular profiles. However, their big dimensionality, combined with the highly connected nature of the molecules that are measured, makes it difficult to interpret them in a mechanistically relevant manner. Leveraging prior knowledge, we can use computational methods to infer which biological activities are relevant. For example, the activity of transcription factors (TFs) and kinases can be inferred robustly from downstream transcripts and phosphosite targets, respectively. Over the past decade, a plethora of methods that infer biological activity has emerged, each with its own assumptions and biases.

2Although comparisons and collections of these methods exist (;;; Supplementary Table S1), they do not incorporate recent methodological developments, such as modeling activities based on weighted mode of regulation (Supplementary Table S2). Here, we present decoupleR, an R and Python package containing a collection of methods adapted for biological activity estimation in bulk, single-cell and spatial omics data.

3Currently, decoupleR contains 11 different methods (Fig. 1A), these include popular methods such as AUCell, fast GSEA, GSVA, over-representation analysis, univariate linear model (ULM) adapted from, VIPER and others (Supplementary Table S1). The inputs of decoupleR are: (i) a matrix containing molecular feature values, either for single samples or from population comparisons, like normalized gene expression counts per sample or log fold changes and (ii) a prior knowledge resource such as a collection of gene sets. The user can then choose any method alone or many simultaneously. decoupleR also provides a consensus score obtained by computing a mean z-score across methods (Supplementary Note). Additionally, decoupleR offers easy to use wrappers to query the meta-database OmniPath, making it easy to flexibly access processed resources such as cell-type marker databases, gene regulatory networks or pathway footprints, and estimate biological activities from them.

4We used decoupleR to evaluate the performance of individual methods by recovering perturbed regulators—TFs and kinases—from two independent collections of transcriptomics and phospho-proteomics datasets (Supplementary Note), respectively, upon single-gene perturbation experiments. As resources, we used the gene regulatory network DoRothEA and a kinase substrate network, respectively.

5We built a benchmarking pipeline with decoupleR (Supplementary Note), which evaluates the performance of regulator activity scores from different methods, mainly focused on the sensitivity of methods. Furthermore, to evaluate the robustness of the methods to noise, we added or deleted a percentage of edges from the prior knowledge resources.

6Methods return different distributions of activities (Supplementary Fig. S1) but display general similarities (Supplementary Fig. S2), with a median Spearman correlation of activities between methods of 0.52, and 0.65 for transcriptomics and phospho-proteomics, respectively (Fig. 1B). There was also a moderate agreement between methods in the top 5% ranked regulators (median Jaccard indexes of 0.23 and 0.21, respectively; Supplementary Fig. S2).

7Despite these similarities, methods showed different performances at predicting perturbed regulators (Supplementary Fig. S3). Some of them performed consistently better than the others (Supplementary Table S3; Fig. 1C), the top three being: consensus, multivariate linear model and ULM. Moreover, methods that leverage weights perform better when those are taken into account (P-value <2.2e-16; one-sided Wilcoxon signed-rank test; Supplementary Fig. S4).

8Deleting edges in the resource had a greater effect than adding them across methods (Supplementary Fig. S5); with a median Spearman correlation of activities to the original ones of 0.84 and 0.77 for the addition and deletion, respectively (P-value <2.2e-16; one-sided Wilcoxon signed-rank test). Additionally, adding or deleting edges decreased predictability, and deleting edges had a worse effect than adding (adjusted P-values <2.2e-16 for normal-addition, <2.2e-16 for normal-deletion and <2.2e-16 for deletion-addition; F = 131; Tukey’s HSD posthoc test) (Supplementary Fig. S6).

9Finally, we evaluated decoupleR’s speed and found that methods run relatively fast in the R version, and orders of magnitude faster in the Python one [median across methods of 1.44 and 0.44 ms per sample and regulator in R and Python, respectively, with an Intel(R) Core(TM) i7-8550U CPU @ 1.80 GHz; Supplementary Fig. S7], enabling their use with larger datasets such as single-cell or spatial omics.

10In summary, decoupleR combines a variety of methods to infer biological activities into one efficient, robust, and user-friendly tool in the two most used programming languages for omics data analysis. With a common syntax for different methods, types of omics datasets, and knowledge sources available via OmniPath, it facilitates the exploration of different approaches and can be integrated in many workflows.

11We observed that the majority of methods return adequate estimates of regulator activities, but that their aggregation into a consensus score and linear models perform better than other methods. We welcome the addition of further methods by the community.