Use overlap joins to find relative enrichment
We are now ready to compute overlaps between RNA-seq genes (our DE set and
bootstrap sets) and the ATAC-seq peaks. In plyranges, overlaps are defined as
joins between two GRanges objects: a left and a right GRanges object.
In an overlap join, a match is any range on the left GRanges that is
overlapped by the right GRanges. One powerful aspect of the overlap joins
is that the result maintains all (metadata) columns from each of the left and
right ranges which makes downstream summaries easy to compute.
To combine the DE genes with the DA peaks, we perform a left overlap join. This
returns to us the all_genes ranges (potentially with duplication), but with
the metadata columns from those overlapping DA peaks. For any gene that has no
overlaps, the DA peak columns will have NA’s.
genes_olap_peaks <- all_genes %>%
join_overlap_left(da_peaks)
genes_olap_peaks
## GRanges object with 30449 ranges and 8 metadata columns:
## seqnames ranges strand | gene_id de_log2FC
## <Rle> <IRanges> <Rle> | <character> <numeric>
## [1] chr1 196641878-196661877 + | ENSG00000000971.15 4.98711
## [2] chr6 46119993-46139992 + | ENSG00000001561.6 1.92722
## [3] chr4 17567192-17587191 + | ENSG00000002549.12 2.93373
## [4] chr4 17567192-17587191 + | ENSG00000002549.12 2.93373
## [5] chr4 17567192-17587191 + | ENSG00000002549.12 2.93373
## ... ... ... ... . ... ...
## [30445] chr12 47823132-47843131 - | ENSG00000061273.17 -0.330255
## [30446] chr12 47823132-47843131 - | ENSG00000061273.17 -0.330255
## [30447] chr12 47823132-47843131 - | ENSG00000061273.17 -0.330255
## [30448] chr12 47823132-47843131 - | ENSG00000061273.17 -0.330255
## [30449] chr12 47823132-47843131 - | ENSG00000061273.17 -0.330255
## de_padj resample origin peak_id da_log2FC da_padj
## <numeric> <integer> <factor> <character> <numeric> <numeric>
## [1] 6.85285e-14 0 de ATAC_peak_21236 -0.546582 1.15274e-04
## [2] 1.58739e-05 0 de ATAC_peak_231183 1.453297 9.73225e-17
## [3] 1.00655e-11 0 de ATAC_peak_193578 0.222371 3.00939e-11
## [4] 1.00655e-11 0 de ATAC_peak_193579 -0.281615 7.99889e-05
## [5] 1.00655e-11 0 de ATAC_peak_193580 0.673705 7.60043e-15
## ... ... ... ... ... ... ...
## [30445] 9.17085e-05 10 not_de ATAC_peak_61944 -0.864385 2.15479e-10
## [30446] 9.17085e-05 10 not_de ATAC_peak_61945 -1.263839 2.83702e-21
## [30447] 9.17085e-05 10 not_de ATAC_peak_61946 -0.854639 2.02495e-16
## [30448] 9.17085e-05 10 not_de ATAC_peak_61947 -0.421568 5.22070e-04
## [30449] 9.17085e-05 10 not_de ATAC_peak_61948 -0.538309 1.46133e-08
## -------
## seqinfo: 25 sequences (1 circular) from hg38 genome
Now we can ask, how many DA peaks are near DE genes relative to “other” non-DE
genes? A gene may appear more than once in genes_olap_peaks, because
multiple peaks may overlap a single gene, or because we have re-sampled the
same gene more than once, or a combination of these two cases.
For each gene (that is the combination of chromosome, the start, end, and
strand), and the “origin” (DE vs not-DE) we can compute the distinct number of
peaks for each gene and the maximum peak based on LFC. This is achieved via
reduce_ranges_directed, which allows an aggregation to result in a GRanges
object via merging neighboring genomic regions. The use of the directed suffix
indicates we’re maintaining strand information. In this case, we are simply
merging ranges (genes) via the groups we mentioned above. We also have to
account for the number of resamples we have performed when counting if there
are any peaks, to ensure we do not double count the same peak:
gene_peak_max_lfc <- genes_olap_peaks %>%
group_by(gene_id, origin) %>%
reduce_ranges_directed(
peak_count = sum(!is.na(da_padj)) / n_distinct(resample),
peak_max_lfc = max(abs(da_log2FC))
)
We can then filter genes if they have any peaks and compare the peak fold
changes between non-DE and DE genes using a boxplot:
library(ggplot2)
gene_peak_max_lfc %>%
filter(peak_count > 0) %>%
as.data.frame() %>%
ggplot(aes(origin, peak_max_lfc)) +
geom_boxplot()
In general, the DE genes have larger maximum DA fold changes relative to the
non-DE genes.
Next we examine how thresholds on the DA LFC modify the enrichment we observe
of DA peaks near DE or non-DE genes. First, we want to know how the number of
peaks within DE genes and non-DE genes change as we change threshold values on
the peak LFC. As an example, we could compute this by arbitrarily chosen LFC
thresholds of 1 or 2 as follows:
origin_peak_lfc <- genes_olap_peaks %>%
group_by(origin) %>%
summarize(
peak_count = sum(!is.na(da_padj)) / n_distinct(resample),
lfc1_peak_count =sum(abs(da_log2FC) > 1, na.rm=TRUE)/ n_distinct(resample),
lfc2_peak_count = sum(abs(da_log2FC) > 2, na.rm=TRUE)/ n_distinct(resample)
)
origin_peak_lfc
## DataFrame with 2 rows and 4 columns
## origin peak_count lfc1_peak_count lfc2_peak_count
## <factor> <numeric> <numeric> <numeric>
## 1 not_de 2625.7 406.8 35.7
## 2 de 3726.0 1181.0 246.0
Here we see that DE genes tend to have more DA peaks near them, and that the
number of DA peaks decreases as we increase the DA LFC threshold (as expected).
We now show how to compute the ratio of peak counts from DE compared to non-DE
genes, so we can see how this ratio changes for various DA LFC thresholds.
For all variables except for the origin column we divide the first row’s
values by the second row, which will be the enrichment of peaks in DE genes
compared to other genes. This requires us to reshape the summary table from
long form back to wide form using the tidyr package. First we pivot the
results of the peak_count columns into name-value pairs, then pivot again to
place values into the origin column. Then we create a new column with the
relative enrichment:
origin_peak_lfc %>%
as.data.frame() %>%
tidyr::pivot_longer(cols = -origin) %>%
tidyr::pivot_wider(names_from = origin, values_from = value) %>%
mutate(enrichment = de / not_de)
## # A tibble: 3 × 4
## name not_de de enrichment
## <chr> <dbl> <dbl> <dbl>
## 1 peak_count 2626. 3726 1.42
## 2 lfc1_peak_count 407. 1181 2.90
## 3 lfc2_peak_count 35.7 246 6.89
The above table shows that relative enrichment increases for a larger LFC
threshold.
Due to the one-to-many mappings of genes to peaks, it is unknown if we have the
same number of DE genes participating or less, as we increase the threshold on
the DA LFC. We can examine the number of genes with overlapping DA peaks at
various thresholds by grouping and aggregating twice. First, the number of
peaks that meet the thresholds are computed within each gene, origin, and
resample group. Second, within the origin column, we compute the total number
of peaks that meet the DA LFC threshold and the number of genes that have more
than zero peaks (again averaging over the number of resamples).
genes_olap_peaks %>%
group_by(gene_id, origin, resample) %>%
reduce_ranges_directed(
lfc1 = sum(abs(da_log2FC) > 1, na.rm=TRUE),
lfc2 = sum(abs(da_log2FC) > 2, na.rm=TRUE)
) %>%
group_by(origin) %>%
summarize(
lfc1_gene_count = sum(lfc1 > 0) / n_distinct(resample),
lfc1_peak_count = sum(lfc1) / n_distinct(resample),
lfc2_gene_count = sum(lfc2 > 0) / n_distinct(resample),
lfc2_peak_count = sum(lfc2) / n_distinct(resample)
)
## DataFrame with 2 rows and 5 columns
## origin lfc1_gene_count lfc1_peak_count lfc2_gene_count lfc2_peak_count
## <factor> <numeric> <numeric> <numeric> <numeric>
## 1 not_de 300 406.8 33.4 35.7
## 2 de 564 1181.0 162.0 246.0
To do this for many thresholds is cumbersome and would create a lot of
duplicate code. Instead we create a single function called
count_above_threshold that accepts a variable and a vector of thresholds, and
computes the sum of the absolute value of the variable for each element in the
thresholds vector.
count_if_above_threshold <- function(var, thresholds) {
lapply(thresholds, function(.) sum(abs(var) > ., na.rm = TRUE))
}
The above function will compute the counts for any arbitrary threshold, so we
can apply it over possible LFC thresholds of interest. We choose a grid of one
hundred thresholds based on the range of absolute LFC values in the da_peaks
GRanges object:
thresholds <- da_peaks %>%
mutate(abs_lfc = abs(da_log2FC)) %>%
with(
seq(min(abs_lfc), max(abs_lfc), length.out = 100)
)
The peak counts for each threshold are computed as a new list-column called
value. First, the GRanges object has been grouped by the gene, origin, and
the number of resamples columns. Then we aggregate over those columns, so each
row will contain the peak counts for all of the thresholds for a gene, origin,
and resample. We also maintain another list-column that contains the threshold
values.
genes_peak_all_thresholds <- genes_olap_peaks %>%
group_by(gene_id, origin, resample) %>%
reduce_ranges_directed(
value = count_if_above_threshold(da_log2FC, thresholds),
threshold = list(thresholds)
)
genes_peak_all_thresholds
## GRanges object with 9042 ranges and 5 metadata columns:
## seqnames ranges strand | gene_id origin
## <Rle> <IRanges> <Rle> | <character> <factor>
## [1] chr1 196641878-196661877 + | ENSG00000000971.15 de
## [2] chr6 46119993-46139992 + | ENSG00000001561.6 de
## [3] chr4 17567192-17587191 + | ENSG00000002549.12 de
## [4] chr7 150790403-150810402 + | ENSG00000002933.8 de
## [5] chr4 15768275-15788274 + | ENSG00000004468.12 de
## ... ... ... ... . ... ...
## [9038] chr2 197695369-197715368 + | ENSG00000247626.4 not_de
## [9039] chr5 90519584-90539583 - | ENSG00000176018.12 not_de
## [9040] chr12 6679572-6699571 - | ENSG00000126746.17 not_de
## [9041] chr19 974271-994270 + | ENSG00000065268.10 not_de
## [9042] chr12 47823132-47843131 - | ENSG00000061273.17 not_de
## resample value threshold
## <integer> <IntegerList> <NumericList>
## [1] 0 1,1,1,... 0.0658243,0.1184840,0.1711436,...
## [2] 0 1,1,1,... 0.0658243,0.1184840,0.1711436,...
## [3] 0 6,6,6,... 0.0658243,0.1184840,0.1711436,...
## [4] 0 4,4,4,... 0.0658243,0.1184840,0.1711436,...
## [5] 0 11,11,11,... 0.0658243,0.1184840,0.1711436,...
## ... ... ... ...
## [9038] 10 6,5,5,... 0.0658243,0.1184840,0.1711436,...
## [9039] 10 2,2,1,... 0.0658243,0.1184840,0.1711436,...
## [9040] 10 2,2,2,... 0.0658243,0.1184840,0.1711436,...
## [9041] 10 3,3,3,... 0.0658243,0.1184840,0.1711436,...
## [9042] 10 7,7,7,... 0.0658243,0.1184840,0.1711436,...
## -------
## seqinfo: 25 sequences (1 circular) from hg38 genome
Now we can expand these list-columns into a long GRanges object using the
expand_ranges() function. This function will unlist the value and
threshold columns and lengthen the resulting GRanges object. To compute
the peak and gene counts for each threshold, we apply the same summarization as
before:
origin_peak_all_thresholds <- genes_peak_all_thresholds %>%
expand_ranges() %>%
group_by(origin, threshold) %>%
summarize(
gene_count = sum(value > 0) / n_distinct(resample),
peak_count = sum(value) / n_distinct(resample)
)
origin_peak_all_thresholds
## DataFrame with 200 rows and 4 columns
## origin threshold gene_count peak_count
## <factor> <numeric> <numeric> <numeric>
## 1 not_de 0.0658243 778.1 2625.3
## 2 not_de 0.1184840 768.0 2548.1
## 3 not_de 0.1711436 753.8 2391.3
## 4 not_de 0.2238033 739.3 2183.3
## 5 not_de 0.2764629 714.6 1961.2
## ... ... ... ... ...
## 196 de 5.06849 2 2
## 197 de 5.12115 0 0
## 198 de 5.17381 0 0
## 199 de 5.22647 0 0
## 200 de 5.27913 0 0
Again we can compute the relative enrichment in LFCs in the same manner as
before, by pivoting the results to long form then back to wide form to compute
the enrichment. We visualize the peak enrichment changes of DE genes relative
to other genes as a line chart:
origin_threshold_counts <- origin_peak_all_thresholds %>%
as.data.frame() %>%
tidyr::pivot_longer(cols = -c(origin, threshold),
names_to = c("type", "var"),
names_sep = "_",
values_to = "count") %>%
select(-var)
origin_threshold_counts %>%
filter(type == "peak") %>%
tidyr::pivot_wider(names_from = origin, values_from = count) %>%
mutate(enrichment = de / not_de) %>%
ggplot(aes(x = threshold, y = enrichment)) +
geom_line() +
labs(x = "logFC threshold", y = "Relative Enrichment")
## Warning: Removed 4 rows containing missing values or values outside the scale range
## (`geom_line()`).
We computed the sum of DA peaks near the DE genes, for increasing LFC
thresholds on the accessibility change. As we increased the threshold, the
number of total peaks went down (likewise the mean number of DA peaks per
gene). It is also likely the number of DE genes with a DA peak nearby with such
a large change went down. We can investigate this with a plot that summarizes
many of the aspects underlying the enrichment plot above.
origin_threshold_counts %>%
ggplot(aes(x = threshold,
y = count + 1,
color = origin,
linetype = type)) +
geom_line() +
scale_y_log10()
