Bulk RNA-seq yield many molecular readouts that are hard to interpret by themselves. One way of summarizing this information is by inferring transcription factor (TF) activities from prior knowledge.

In this notebook we showcase how to use decoupleR for transcription factor activity inference with a bulk RNA-seq data-set where the transcription factor FOXA2 was knocked out in pancreatic cancer cell lines.

The data consists of 3 Wild Type (WT) samples and 3 Knock Outs (KO). They are freely available in GEO.

1 Loading packages

First, we need to load the relevant packages:

## We load the required packages
library(decoupleR)
library(dplyr)
library(tibble)
library(tidyr)
library(ggplot2)
library(pheatmap)
library(ggrepel)

2 Loading the data-set

Here we used an already processed bulk RNA-seq data-set. We provide the normalized log-transformed counts, the experimental design meta-data and the Differential Expressed Genes (DEGs) obtained using limma. We can open the data like this:

inputs_dir <- system.file("extdata", package = "decoupleR")
data <- readRDS(file.path(inputs_dir, "bk_data.rds"))

From data we can extract the mentioned information. Here we see the normalized log-transformed counts:

# Remove NAs and set row names
counts <- data$counts %>%
  dplyr::mutate_if(~ any(is.na(.x)), ~ if_else(is.na(.x),0,.x)) %>% 
  column_to_rownames(var = "gene") %>% 
  as.matrix()
head(counts)
#>          PANC1.WT.Rep1 PANC1.WT.Rep2 PANC1.WT.Rep3 PANC1.FOXA2KO.Rep1 PANC1.FOXA2KO.Rep2 PANC1.FOXA2KO.Rep3
#> NOC2L        10.052588     11.949123     12.057774          12.312291          12.139918          11.494205
#> PLEKHN1       7.535115      8.125993      8.714880           8.048196           8.290154           8.621239
#> PERM1         6.281242      6.424582      6.589668           6.293285           6.486136           6.775344
#> ISG15        10.938252     11.469081     11.425415          11.549986          11.371464          11.178157
#> AGRN          6.956335      7.196108      7.522550           7.061549           7.485534           7.071555
#> C1orf159      9.546224      9.788721      9.794589           9.850830           9.988069           9.965357

The design meta-data:

design <- data$design
design
#> # A tibble: 6 Ă— 2
#>   sample             condition    
#>   <chr>              <chr>        
#> 1 PANC1.WT.Rep1      PANC1.WT     
#> 2 PANC1.WT.Rep2      PANC1.WT     
#> 3 PANC1.WT.Rep3      PANC1.WT     
#> 4 PANC1.FOXA2KO.Rep1 PANC1.FOXA2KO
#> 5 PANC1.FOXA2KO.Rep2 PANC1.FOXA2KO
#> 6 PANC1.FOXA2KO.Rep3 PANC1.FOXA2KO

And the results of limma, of which we are interested in extracting the obtained t-value and p-value from the contrast:

# Extract t-values per gene
deg <- data$limma_ttop %>%
    select(ID, logFC, t, P.Value) %>% 
    filter(!is.na(t)) %>% 
    column_to_rownames(var = "ID") %>%
    as.matrix()
head(deg)
#>             logFC          t      P.Value
#> RHBDL2  -1.823940 -12.810588 3.030276e-06
#> PLEKHH2 -1.568830 -10.794453 9.932046e-06
#> HEG1    -1.725806  -9.788112 1.939734e-05
#> CLU     -1.786200  -9.761618 1.975813e-05
#> FHL1     2.087082   8.950191 3.552199e-05
#> RBP4    -1.728960  -8.529074 4.904579e-05

3 CollecTRI network

CollecTRI is a comprehensive resource containing a curated collection of TFs and their transcriptional targets compiled from 12 different resources. This collection provides an increased coverage of transcription factors and a superior performance in identifying perturbed TFs compared to our previous DoRothEA network and other literature based GRNs. Similar to DoRothEA, interactions are weighted by their mode of regulation (activation or inhibition).

For this example we will use the human version (mouse and rat are also available). We can use decoupleR to retrieve it from OmniPath. The argument split_complexes keeps complexes or splits them into subunits, by default we recommend to keep complexes together.

net <- get_collectri(organism='human', split_complexes=FALSE)
net
#> # A tibble: 45,398 Ă— 3
#>    source target   mor
#>    <chr>  <chr>  <dbl>
#>  1 MYC    TERT       1
#>  2 SPI1   BGLAP      1
#>  3 SMAD3  ADAM2      1
#>  4 SMAD3  JUN        1
#>  5 SMAD4  ADAM2      1
#>  6 SMAD4  JUN        1
#>  7 STAT5A IL2        1
#>  8 STAT5B IL2        1
#>  9 RELA   FAS        1
#> 10 WT1    NR0B1      1
#> # ℹ 45,388 more rows

4 Activity inference with Weighted Mean

To infer activities we will run the Weighted Mean method (wmean). It infers regulator activities by first multiplying each target feature by its associated weight which then are summed to an enrichment score wmean. Furthermore, permutations of random target features can be performed to obtain a null distribution that can be used to compute a z-score norm_wmean, or a corrected estimate corr_wmean by multiplying wmean by the minus log10 of the obtained empirical p-value.

In this example we use wmean but we could have used any other. To see what methods are available use show_methods().

To run decoupleR methods, we need an input matrix (mat), an input prior knowledge network/resource (net), and the name of the columns of net that we want to use.

# Run wmean
sample_acts <- run_wmean(mat=counts, net=net, .source='source', .target='target',
                  .mor='mor', times = 100, minsize = 5)
sample_acts
#> # A tibble: 10,512 Ă— 5
#>    statistic  source condition          score p_value
#>    <chr>      <chr>  <chr>              <dbl>   <dbl>
#>  1 corr_wmean ABL1   PANC1.FOXA2KO.Rep1 1.08     0.7 
#>  2 corr_wmean ABL1   PANC1.FOXA2KO.Rep2 0.399    0.88
#>  3 corr_wmean ABL1   PANC1.FOXA2KO.Rep3 0.415    0.88
#>  4 corr_wmean ABL1   PANC1.WT.Rep1      0        1   
#>  5 corr_wmean ABL1   PANC1.WT.Rep2      0.601    0.82
#>  6 corr_wmean ABL1   PANC1.WT.Rep3      1.33     0.64
#>  7 corr_wmean AHR    PANC1.FOXA2KO.Rep1 0.962    0.76
#>  8 corr_wmean AHR    PANC1.FOXA2KO.Rep2 4.27     0.3 
#>  9 corr_wmean AHR    PANC1.FOXA2KO.Rep3 4.52     0.28
#> 10 corr_wmean AHR    PANC1.WT.Rep1      0.960    0.76
#> # ℹ 10,502 more rows

5 Visualization

From the obtained results, we will select the norm_wmean activities and we will observe the most variable activities across samples in a heat-map:

n_tfs <- 25

# Transform to wide matrix
sample_acts_mat <- sample_acts %>%
  filter(statistic == 'norm_wmean') %>%
  pivot_wider(id_cols = 'condition', names_from = 'source',
              values_from = 'score') %>%
  column_to_rownames('condition') %>%
  as.matrix()

# Get top tfs with more variable means across clusters
tfs <- sample_acts %>%
  group_by(source) %>%
  summarise(std = sd(score)) %>%
  arrange(-abs(std)) %>%
  head(n_tfs) %>%
  pull(source)
sample_acts_mat <- sample_acts_mat[,tfs]

# Scale per sample
sample_acts_mat <- scale(sample_acts_mat)

# Choose color palette
palette_length = 100
my_color = colorRampPalette(c("Darkblue", "white","red"))(palette_length)

my_breaks <- c(seq(-3, 0, length.out=ceiling(palette_length/2) + 1),
               seq(0.05, 3, length.out=floor(palette_length/2)))

# Plot
pheatmap(sample_acts_mat, border_color = NA, color=my_color, breaks = my_breaks) 

We can also infer pathway activities from the t-values of the DEGs between KO and WT:

# Run wmean
contrast_acts <- run_wmean(mat=deg[, 't', drop=FALSE], net=net, .source='source', .target='target',
                  .mor='mor', times = 100, minsize = 5)
contrast_acts
#> # A tibble: 1,752 Ă— 5
#>    statistic  source condition    score p_value
#>    <chr>      <chr>  <chr>        <dbl>   <dbl>
#>  1 corr_wmean ABL1   t          0.133      0.54
#>  2 corr_wmean AHR    t          0.0162     0.74
#>  3 corr_wmean AIRE   t         -0.0469     0.24
#>  4 corr_wmean AP1    t          0.246      0.04
#>  5 corr_wmean APEX1  t          0.00919    0.9 
#>  6 corr_wmean AR     t          0.00740    0.32
#>  7 corr_wmean ARID1A t         -0.0442     0.46
#>  8 corr_wmean ARID3A t          1.04       0.1 
#>  9 corr_wmean ARID3B t          0.310      0.36
#> 10 corr_wmean ARID4A t         -0.0570     0.48
#> # ℹ 1,742 more rows

We select the norm_wmean activities and then we show the changes in activity between KO and WT:

# Filter norm_wmean
f_contrast_acts <- contrast_acts %>%
  filter(statistic == 'norm_wmean') %>%
  mutate(rnk = NA)

# Filter top TFs in both signs
msk <- f_contrast_acts$score > 0
f_contrast_acts[msk, 'rnk'] <- rank(-f_contrast_acts[msk, 'score'])
f_contrast_acts[!msk, 'rnk'] <- rank(-abs(f_contrast_acts[!msk, 'score']))
tfs <- f_contrast_acts %>%
  arrange(rnk) %>%
  head(n_tfs) %>%
  pull(source)
f_contrast_acts <- f_contrast_acts %>%
  filter(source %in% tfs)

# Plot
ggplot(f_contrast_acts, aes(x = reorder(source, score), y = score)) + 
    geom_bar(aes(fill = score), stat = "identity") +
    scale_fill_gradient2(low = "darkblue", high = "indianred", 
        mid = "whitesmoke", midpoint = 0) + 
    theme_minimal() +
    theme(axis.title = element_text(face = "bold", size = 12),
        axis.text.x = 
            element_text(angle = 45, hjust = 1, size =10, face= "bold"),
        axis.text.y = element_text(size =10, face= "bold"),
        panel.grid.major = element_blank(), 
        panel.grid.minor = element_blank()) +
    xlab("Pathways")

The TFs HLF and IRX1 are deactivated in KO when compared to WT, while FOSL1 and MYC seem to be activated.

We can further visualize the most differential target genes in each TF along their p-values to interpret the results. For example, let’s see the genes that are belong to FOXA2:

tf <- 'FOXA2'

df <- net %>%
  filter(source == tf) %>%
  arrange(target) %>%
  mutate(ID = target, color = "3") %>%
  column_to_rownames('target')

inter <- sort(intersect(rownames(deg),rownames(df)))
df <- df[inter, ]
df[,c('logfc', 't_value', 'p_value')] <- deg[inter, ]
df <- df %>%
  mutate(color = if_else(mor > 0 & t_value > 0, '1', color)) %>%
  mutate(color = if_else(mor > 0 & t_value < 0, '2&#