Usage

decoupler contains different statistical methods to infer enrichment scores from omics data using prior knowledge. In this notebook we showcase how to use it with some toy data.

Note

In Jupyter notebooks and lab, you can see the documentation for a python function by hitting SHIFT + TAB, hit it twice to expand the view, or by typing ?name_of_function.

Loading packages

decoupler can be imported as:

import decoupler as dc

# Only needed for visualization:
import matplotlib.pyplot as plt
import seaborn as sns

Note

The first time decoupler is imported in an enviroment can be slow. This is due to numba compiling its source code. After the first usage it will be stored in cache, making the import time normal again.

Loading toy data

decoupler needs a matrix (mat) of molecular readouts (gene expression, logFC, p-values, etc.) and a network (net) that relates target features (genes, proteins, etc.) to “source” biological entities (pathways, transcription factors, molecular processes, etc.)

To load the example data-set, simply run:

mat, net = dc.get_toy_data()

This example consists of two small populations of samples (S, rows) with different gene expression patterns (G, columns):

fig, ax = plt.subplots(1,1, figsize=(3,3), tight_layout=True, sharey=True)
sns.heatmap(mat, cmap='viridis', ax=ax)
plt.show()

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 5 transcription factors (Ts) with specific regulation to target genes (Gs), either positive or negative:

net

	source	target	weight
0	T1	G01	1.0
1	T1	G02	1.0
2	T1	G03	0.7
3	T2	G04	1.0
4	T2	G06	-0.5
5	T2	G07	-3.0
6	T2	G08	-1.0
7	T3	G06	1.0
8	T3	G07	0.5
9	T3	G08	1.0
10	T4	G05	1.9
11	T4	G10	-1.5
12	T4	G11	-2.0
13	T4	G09	3.1
14	T5	G09	0.7
15	T5	G10	1.1
16	T5	G11	0.1

This network can be visualized like a graph. Green edges are positive regulation (activation), red edges are negative regulation (inactivation):

dc.plot_network(
    net,
    node_size=50,
    figsize=(4, 4),
    s_cmap='white',
    t_cmap='white',
    c_pos_w='darkgreen',
    c_neg_w='darkred'
)

According to this network, the first population of samples should show high activity for T1 and T2, while the second one for T3 and T4. T5 should have no activity in all samples.

Methods

decoupler contains several methods. To check how many are available, run:

dc.show_methods()

	Function	Name
0	run_aucell	AUCell.
1	run_consensus	Consensus score from top methods.
2	run_gsea	Gene Set Enrichment Analysis (GSEA).
3	run_gsva	Gene Set Variation Analysis (GSVA).
4	run_mdt	Multivariate Decision Tree (MDT).
5	run_mlm	Multivariate Linear Model (MLM).
6	run_ora	Over Representation Analysis (ORA).
7	run_udt	Univariate Decision Tree (UDT).
8	run_ulm	Univariate Linear Model (ULM).
9	run_viper	Virtual Inference of Protein-activity by Enric...
10	run_wmean	Weighted mean (WMEAN).
11	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 inside decoupler have these shared arguments:

• mat : input matrix of molecular readouts.

• net : input prior knowledge information relating molecular features to biological entities.

• source,target and weight : column names where to extract the information from net.

  • source refers to the biological entities.

  • target refers to the molecular features.

  • weight refers to the “strength” of the interaction (if available, else 1s will be used). Only available for methods that can model interaction weights.

• min_n : 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 mat. For this example data-set we will have to keep it to 0 though.

• verbose : Whether to show progress.

• use_raw : When the input is an AnnData object, whether to use the data stored in it’s .raw atribute or not (True by default).

Running methods

Individual methods

As an example, let’s first run the Gene Set Enrichment Analysis method (gsea), one of the most well-known statistics:

# For this toy data, we need to set min_n to 0 (otherwise it is better to keep at 5):
acts, norm_acts, pvals = dc.run_gsea(mat, net, min_n=0, times=100)
acts.head()

source	T1	T2	T3	T4	T5
S01	0.888889	0.711597	-0.555556	-0.50	-0.666667
S02	0.888889	0.699385	-0.555556	-0.50	-0.444444
S03	0.888889	0.551198	-0.444444	-0.75	-0.666667
S04	0.888889	0.607223	-0.666667	-0.50	-0.444444
S05	0.888889	0.732963	-0.888889	-0.50	-0.444444

In the case of gsea, it returns a simple estimate of activities (acts), a normalised estimate (norm_acts) and pvals data-frames after doing permutations.

Note

If mat is an AnnData instance, results will instead be saved in its .obsm attribute.

Let us plot the obtained results:

fig, axes = plt.subplots(1,3, figsize=(9,3), tight_layout=True, sharey=True)

axes[0].set_title('acts')
sns.heatmap(acts, cmap='coolwarm', vmin=-1, vmax=1, ax=axes[0])
axes[1].set_title('norm_acts')
sns.heatmap(norm_acts, cmap='coolwarm', vmin=-1, vmax=1, ax=axes[1])
axes[2].set_title('pvals')
sns.heatmap(pvals, cmap='viridis_r', ax=axes[2], vmax=1)

plt.show()

We can observe that for transcription factors T1 and T3, the obtained activities correctly distinguish the two sample populations. T2, 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 the Univariate Linear Model method ulm:

# For this toy data, we need to set min_n to 0 (otherwise it is better to keep at 5):
acts, pvals = dc.run_ulm(mat, net, min_n=0)
acts.head()

	T1	T2	T3	T4	T5
S01	4.020366	1.667426	-1.378734	-0.243605	-1.188727
S02	4.060297	1.392356	-1.313069	-0.387175	-1.102454
S03	4.189780	1.086979	-1.346471	-0.294281	-1.139018
S04	4.149536	1.395950	-1.332407	-0.273508	-1.197696
S05	4.052075	1.610522	-1.534533	-0.311155	-0.990895

In this case, ulm only returns infered activities and their associated p-value.

As before, let us plot the resulting activities:

fig, axes = plt.subplots(1,2, figsize=(6,3), tight_layout=True, sharey=True)

axes[0].set_title('acts')
sns.heatmap(acts, cmap='coolwarm', vmin=-1, vmax=1, ax=axes[0])
axes[1].set_title('pvals')
sns.heatmap(pvals, cmap='viridis_r', ax=axes[1], vmax=1)

plt.show()

Since ulm models weights when estimating biological activities, it correctly assigns T2 as inactive in the second population of samples.

Multiple methods

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, decouple 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. If we wanted to only obtain the consensus score, we could also have used run_consensus, check ?run_consensus for more information.

# For this toy data, we need to set min_n to 0 (otherwise it is better to keep at 5):
results = dc.decouple(mat, net, min_n=0, verbose=False)

decouple either returns a dictionary of activities and p-values, or stores them in the AnnData instance provided.

Let us see how the consensus score looks like:

# Extract from dictionary
acts = results['consensus_estimate']
pvals = results['consensus_pvals']

fig, axes = plt.subplots(1,2, figsize=(6,3), tight_layout=True, sharey=True)

axes[0].set_title('acts')
sns.heatmap(acts, cmap='coolwarm', vmin=-1, vmax=1, ax=axes[0])
axes[1].set_title('pvals')
sns.heatmap(pvals, cmap='viridis_r', ax=axes[1], vmax=1)

plt.show()

We can observe that the consensus score correctly predicts that T1 and T2 should be active in the first population of samples while T3 and T4 in the second one. T5 is inactive everywhere as expected.