In this tutorial we are going to use a GWAS dataset (accessible from this link) together with the whole ENCODE BroadPeak dataset to find which mutations (and their associated traits) are most represented in enhancer regions which are present in a limited set of cells.
As first thing let's download the data.
In [ ]:
%%bash
wget -q https://www.ebi.ac.uk/gwas/api/search/downloads/full -O tmp.tsv
cat tmp.tsv | \
awk 'BEGIN {FS="\t";OFS="\t"} {chrom=$12; gsub(chrom,"chr"chrom,$12)}{print $0}' | \
sed s/,//g > gwas.tsv
rm tmp.tsv
In [28]:
myBucket = "gs://fc-cad72548-2d6b-41ce-82aa-975cb7e8b764"
In order to run the query on HDFS, we have to put the file there. We will use the bucket for this Terra Notebook.
In [ ]:
!gsutil cp ./gwas.tsv $myBucket/
In [3]:
import gmql as gl
import matplotlib.pyplot as plt
import seaborn as sns
import pandas as pd
In [4]:
gl.set_master("yarn")
In [1]:
gmql_repository = "gs://geco_repository/"
gwas_path = myBucket + "/gwas.tsv"
The GWAS data comes from a single TSV file. Therefore we can import it using the load_from_file
function.
Notice that we have to specify a parser
to properly load our data. Therefore it is wise to take a look at the schema of the downloaded file.
We are mainly interested in the mutation position (11-th and 12-th column) and the associated trait (7-th).
In [5]:
gwas = gl.load_from_file(gwas_path,
parser=gl.parsers.RegionParser(
chrPos=11,
startPos=12,
stopPos=12,
otherPos=[(7, "trait", 'string')]))
In [6]:
gwas.head().regs
Out[6]:
We can also simply look at the schema
In [7]:
gwas.schema
Out[7]:
In [8]:
gwas_data = gwas.materialize().regs
We now plot the number of regions for each of the top 30 represented traits.
In [9]:
plt.figure(figsize=(20,5))
sns.countplot(data=gwas_data[gwas_data.trait.isin(
gwas_data.trait.value_counts().iloc[:30].index)], x='trait')
plt.xticks(rotation=90)
plt.title("Top represented GWAS traits", fontsize=20)
plt.show()
We now load the ENCODE BroadPeak dataset.
If the data come already in the GDM format, they can be loaded using the load_from_path
function. A GDM dataset is stored as a folder having the following structure:
/path/to/dataset/:
- sample1.gdm
- sample1.gdm.meta
- sample2.gdm
- sample2.gdm.meta
- ...
- schema.xml
The first dataset we load is the one from the GWAS study.
In [10]:
broad = gl.load_from_path(gmql_repository + "HG19_ENCODE_BROAD/")
In [11]:
broad.schema
Out[11]:
In [12]:
acetyl = broad[broad['experiment_target'] == 'H3K27ac-human']
We get the peak region of the Chip-Seq using the reg_project
function. The peak position (peak
) is given by the center of the region.
In [13]:
peaked = acetyl.reg_project(new_field_dict={
'peak': acetyl.right/2 + acetyl.left/2})
Once we have the peak, we extend the search region to $\pm 1500 bp$. We use again reg_project
In [14]:
enlarge = peaked.reg_project(new_field_dict={
'left': peaked.peak - 1500,
'right': peaked.peak + 1500})
We are interested in enhancers which are cell specific. Therefore it is important to group our data by cell line. In addition to this we merge the signals coming from different tracks for the same cell line. We can do both of these actions using the normal_cover
function.
In [15]:
enhancers_by_cell_line = enlarge.normal_cover(1, "ANY",
groupBy=['biosample_term_name'])
To select only the cell-specific enhancers we can now apply again normal_cover
and constraining the maximum number of overlaps between the regions to be a selected threshold.
In this case we select a threshold of 2.
In [16]:
max_overlapping = 2
cell_specific_enhancers = enhancers_by_cell_line.normal_cover(1, max_overlapping)
In [17]:
cell_specific_enhancers.schema
Out[17]:
In [18]:
cell_specific_enhancers_by_cell_line = enhancers_by_cell_line.join(
cell_specific_enhancers,
[gl.DLE(0)], 'left',
refName="en", expName="csen")
In [19]:
gwas.schema
Out[19]:
In [20]:
enhancer_gwas = cell_specific_enhancers_by_cell_line.map(
gwas, refName="csen", expName="gwas",
new_reg_fields={'traits': gl.BAG('trait')})
enhancer_gwas = enhancer_gwas.reg_project(
["count_csen_gwas", "traits"],
new_field_dict={'cell_line': enhancer_gwas['csen.en.biosample_term_name','string']})
In [21]:
enhancer_gwas = enhancer_gwas.materialize()
The traits
column of the resulting region is the list of traits associated with the cell specific enhancer. The data comes in the form of a string of trait names.
We convert the string to a list.
In [24]:
enhancer_gwas.regs['traits'] = enhancer_gwas.regs.traits\
.map(lambda x: x.split(",") if pd.notnull(x) else x)
The final part of the analysis regards the matching of cell lines and traits. We want to understand if a cell line (which is represented by its specific enhancers) has some particular mutation trait associated.
The analysis is performed in Pandas using the result region attributes traits
and cell_line
.
We build an association matrix between cell lines and traits by firstly converting the result to a list of (cell_line, trait)
, converting it to a Pandas DataFrame, and finally using the crosstab
Pandas function to extract the matrix.
In [25]:
cell_trait = pd.DataFrame.from_records([(k, v) for k, vs in enhancer_gwas.regs[enhancer_gwas.regs.count_csen_gwas > 0]\
.groupby("cell_line").traits.sum().to_dict().items() for v in vs],
columns=['cell_line', 'trait'])
In [26]:
cross = pd.crosstab(cell_trait.cell_line, cell_trait.trait)
We finally plot the result as an heatmap.
In [27]:
plt.figure(figsize=(50, 15))
sns.heatmap(cross[cross.sum(0).sort_values(ascending=False).iloc[:100].index], cmap='Reds', vmax=70, linewidths=1, annot=True, cbar=False)
plt.xticks(fontsize=20)
plt.yticks(fontsize=20)
plt.xlabel("Trait", fontsize=30)
plt.ylabel("Cell line", fontsize=30)
plt.show()