decoupleR 2.6.0
decoupleR is an R package distributed as part of the Bioconductor project. To install the package, start R and enter:
install.packages("BiocManager")
BiocManager::install("decoupleR")
Alternatively, you can instead install the latest development version from GitHub with:
BiocManager::install("saezlab/decoupleR")
decoupleR (Badia-i-Mompel, Santiago, Braunger, Geiss, Dimitrov, Müller-Dott, Taus, Dugourd, Holland, Flores, and Saez-Rodriguez, 2022) contains different statistical methods to extract biological activities from omics data using prior knowledge. Some of them are:
In this vignette we showcase how to use it with some toy data.
decoupleR can be imported as:
library(decoupleR)
# Extra libraries
library(dplyr)
library(pheatmap)
decoupleR needs a matrix (mat
) of any molecular readouts (gene
expression, logFC, p-values, etc.) and a network
that relates target
features (genes, proteins, etc.) to “source” biological entities (pathways,
transcription factors, molecular processes, etc.). Some methods also require
the mode of regulation (MoR) for each interaction, defined as negative or
positive weights.
To get an example data-set, run:
data <- get_toy_data()
mat <- data$mat
head(mat,5)[,1:5]
#> S01 S02 S03 S04 S05
#> G01 9.3709584 9.3888607 9.8951935 8.7844590 8.431446
#> G02 8.5646982 8.2787888 8.4304691 8.8509076 8.655648
#> G03 8.3631284 8.1333213 8.2572694 10.4142076 8.321925
#> G04 8.6328626 8.6359504 9.7631631 8.0361226 8.783839
#> G05 0.4042683 0.2842529 0.4600974 0.2059986 1.575728
network <- data$network
network
#> # A tibble: 10 × 3
#> source target mor
#> <chr> <chr> <dbl>
#> 1 T1 G01 1
#> 2 T1 G02 1
#> 3 T1 G03 0.7
#> 4 T2 G06 1
#> 5 T2 G07 0.5
#> 6 T2 G08 1
#> 7 T3 G06 -0.5
#> 8 T3 G07 -3
#> 9 T3 G08 -1
#> 10 T3 G11 1
This example consists of two small populations of samples (S, cols) with different gene expression patterns (G, rows):
pheatmap(mat, cluster_rows = F, cluster_cols = F)
Here we can see that some genes seem to be more expressed in one group of samples than in the other and vice-versa. Ideally, we would like to capture these differences in gene programs into interpretable biological entities. In this example we will do it by summarizing gene expression into transcription factor activities.
The toy data also contains a simple net consisting of 3 transcription factors (Ts) with specific regulation to target genes (either positive or negative). This network can be visualized like a graph. Green edges are positive regulation (activation), red edges are negative regulation (inactivation):
According to this network, the first population of samples should show high activity for T1 and T3, while the second one only for T2.
decoupleR contains several methods. To check how many are available, run:
show_methods()
#> # A tibble: 12 × 2
#> Function Name
#> <chr> <chr>
#> 1 run_aucell AUCell
#> 2 run_consensus Consensus score between methods
#> 3 run_fgsea Fast Gene Set Enrichment Analysis (FGSEA)
#> 4 run_gsva Gene Set Variation Analysis (GSVA)
#> 5 run_mdt Multivariate Decision Trees (MDT)
#> 6 run_mlm Multivariate Linear Model (MLM)
#> 7 run_ora Over Representation Analysis (ORA)
#> 8 run_udt Univariate Decision Tree (UDT)
#> 9 run_ulm Univariate Linear Model (ULM)
#> 10 run_viper Virtual Inference of Protein-activity by Enriched Regulon anal…
#> 11 run_wmean Weighted Mean (WMEAN)
#> 12 run_wsum Weighted Sum (WSUM)
Each method models biological activities in a different manner, sometimes
returning more than one estimate or providing significance of the estimation.
To know what each method returns, please check their documentation like this
?run_mlm
.
To have a unified framework, methods have these shared arguments:
mat
: input matrix of molecular readouts.network
: input prior knowledge information relating molecular features to
biological entities..source
,.target
and .mor
: column names where to extract the information
from network
.
.source
refers to the biological entities..target
refers to the molecular features..mor
refers to the “strength” of the interaction (if available, else 1s
will be used). Only available for methods that can model interaction weights.minsize
: Minimum of target features per biological entity (5 by default).
If less, sources are removed. This filtering prevents obtaining noisy activities
from biological entities with very few matching target features in matrix
. For
this example data-set we will have to keep it to 0 though.As an example, let’s first run the Gene Set Enrichment Analysis method (gsea
),
one of the most well-known statistics:
res_gsea <- run_fgsea(mat, network, .source='source', .target='target', nproc=1, minsize = 0)
#>
|
| | 0%
|
|======================================================================| 100%
#>
#>
|
| | 0%
|
|======================================================================| 100%
#>
#>
|
| | 0%
|
|======================================================================| 100%
#>
#>
|
| | 0%
|
|======================================================================| 100%
#>
#>
|
| | 0%
|
|======================================================================| 100%
#>
#>
|
| | 0%
|
|======================================================================| 100%
#>
#>
|
| | 0%
|
|======================================================================| 100%
#>
#>
|
| | 0%
|
|======================================================================| 100%
#>
#>
|
| | 0%
|
|======================================================================| 100%
#>
#>
|
| | 0%
|
|======================================================================| 100%
#>
#>
|
| | 0%
|
|======================================================================| 100%
#>
#>
|
| | 0%
|
|======================================================================| 100%
#>
#>
|
| | 0%
|
|=================================== | 50%
|
|======================================================================| 100%
#>
#>
|
| | 0%
|
|=================================== | 50%
|
|======================================================================| 100%
#>
#>
|
| | 0%
|
|=================================== | 50%
|
|======================================================================| 100%
#>
#>
|
| | 0%
|
|=================================== | 50%
|
|======================================================================| 100%
#>
#>
|
| | 0%
|
|=================================== | 50%
|
|======================================================================| 100%
#>
#>
|
| | 0%
|
|=================================== | 50%
|
|======================================================================| 100%
#>
#>
|
| | 0%
|
|=================================== | 50%
|
|======================================================================| 100%
#>
#>
|
| | 0%
|
|=================================== | 50%
|
|======================================================================| 100%
#>
#>
|
| | 0%
|
|=================================== | 50%
|
|======================================================================| 100%
#>
#>
|
| | 0%
|
|=================================== | 50%
|
|======================================================================| 100%
#>
#>
|
| | 0%
|
|=================================== | 50%
|
|======================================================================| 100%
#>
#>
|
| | 0%
|
|=================================== | 50%
|
|======================================================================| 100%
res_gsea
#> # A tibble: 144 × 5
#> statistic source condition score p_value
#> <chr> <chr> <chr> <dbl> <dbl>
#> 1 fgsea T1 S01 0.889 0.0755
#> 2 norm_fgsea T1 S01 1.24 0.0755
#> 3 fgsea T2 S01 -0.667 0.548
#> 4 norm_fgsea T2 S01 -1.11 0.548
#> 5 fgsea T3 S01 -0.75 0.01
#> 6 norm_fgsea T3 S01 Inf 0.01
#> 7 fgsea T1 S02 0.889 0.0764
#> 8 norm_fgsea T1 S02 1.29 0.0764
#> 9 fgsea T2 S02 0 0.977
#> 10 norm_fgsea T2 S02 0 0.977
#> # ℹ 134 more rows
Methods return a result data-frame containing:
statistic
: name of the statistic. Depending on the method, there can be more than one per method.source
: name of the biological entity.condition
: sample name.score
: inferred biological activity.p_value
: if available, significance of the inferred activity.In the case of gsea
, it returns a simple estimate of activities (fgsea
),
a normalized estimate (norm_fgsea
) and p-values after doing permutations.
Other methods can return different things, for example Univariate Linear Model
(ulm
):
res_ulm <- run_ulm(mat, network, .source='source', .target='target', .mor='mor', minsize = 0)
res_ulm
#> # A tibble: 72 × 5
#> statistic source condition score p_value
#> <chr> <chr> <chr> <dbl> <dbl>
#> 1 ulm T1 S01 4.21 0.00180
#> 2 ulm T1 S02 4.07 0.00224
#> 3 ulm T1 S03 3.85 0.00319
#> 4 ulm T1 S04 4.60 0.000979
#> 5 ulm T1 S05 3.90 0.00298
#> 6 ulm T1 S06 3.66 0.00442
#> 7 ulm T1 S07 4.31 0.00153
#> 8 ulm T1 S08 4.65 0.000902
#> 9 ulm T1 S09 4.49 0.00117
#> 10 ulm T1 S10 4.07 0.00225
#> # ℹ 62 more rows
In this case, ulm
returns just an estimate (ulm
) and its associated p-values.
Each method can return different statistics, we recommend to check their
documentation to know more about them.
Let us plot the obtained results, first for gsea
:
# Transform to matrix
mat_gsea <- res_gsea %>%
filter(statistic=='fgsea') %>%
pivot_wider_profile(id_cols = source, names_from = condition,
values_from = score) %>%
as.matrix()
pheatmap(mat_gsea, cluster_rows = F, cluster_cols = F, cellwidth = 15, cellheight = 40)
We can observe that for transcription factors T1 and T2, the obtained activities
correctly distinguish the two sample populations. T3, on the other hand, should
be down for the second population of samples since it is a repressor.
This mislabeling of activities happens because gsea
cannot model weights when
inferring biological activities.
When weights are available in the prior knowledge, we definitely recommend using
any of the methods that take them into account to get better estimates,
one example is ulm
:
# Transform to matrix
mat_ulm <- res_ulm %>%
filter(statistic=='ulm') %>%
pivot_wider_profile(id_cols = source, names_from = condition,
values_from = score) %>%
as.matrix()
pheatmap(mat_ulm, cluster_rows = F, cluster_cols = F, cellwidth = 15, cellheight = 40)
Since ulm
models weights when estimating biological activities, it correctly
assigns T3 as inactive in the second population of samples.
decoupleR also allows to run multiple methods at the same time.
Moreover, it computes a consensus score based on the obtained activities across
methods, called consensus
.
By default, deocuple
runs only the top performer methods in our benchmark (mlm
,
ulm
and wsum
), and estimates a consensus score across them. Specific
arguments to specific methods can be passed using the variable args
. For more
information check ?decouple
.
res_decouple <- decouple(mat,
network,
.source='source',
.target='target',
minsize = 0)
res_decouple
#> # A tibble: 432 × 6
#> run_id statistic source condition score p_value
#> <dbl> <chr> <chr> <chr> <dbl> <dbl>
#> 1 1 mlm T1 S01 3.52 0.00781
#> 2 1 mlm T2 S01 -1.13 0.290
#> 3 1 mlm T3 S01 -0.247 0.811
#> 4 1 mlm T1 S02 3.48 0.00831
#> 5 1 mlm T2 S02 -0.213 0.837
#> 6 1 mlm T3 S02 -0.353 0.733
#> 7 1 mlm T1 S03 3.15 0.0135
#> 8 1 mlm T2 S03 -0.638 0.541
#> 9 1 mlm T3 S03 0.0749 0.942
#> 10 1 mlm T1 S04 3.82 0.00512
#> # ℹ 422 more rows
Let us see the result for the consensus score in the previous decouple
run:
# Transform to matrix
mat_consensus <- res_decouple %>%
filter(statistic=='consensus') %>%
pivot_wider_profile(id_cols = source, names_from = condition,
values_from = score) %>%
as.matrix()
pheatmap(mat_consensus, cluster_rows = F, cluster_cols = F, cellwidth = 15, cellheight = 40)