BiocFrame - Bioconductor-like data frames¶
BiocFrame class is a Bioconductor-friendly data frame class. Its primary advantage lies in not making assumptions about the types of the columns - as long as an object has a length (__len__) and supports slicing methods (__getitem__), it can be used inside a BiocFrame.
This flexibility allows us to accept arbitrarily complex objects as columns, which is often the case in Bioconductor objects. Also check out Bioconductor’s S4Vectors package, which implements the DFrame class on which BiocFrame was based.
Advantages of BiocFrame¶
One of the core principles guiding the implementation of the BiocFrame class is “what you put is what you get”. Unlike Pandas DataFrame, BiocFrame makes no assumptions about the types of the columns provided as input. Some key differences to highlight the advantages of using BiocFrame are especially in terms of modifications to column types and handling nested dataframes.
Inadvertent modification of types¶
As an example, Pandas DataFrame modifies the types of the input data. These assumptions may cause issues when interoperating between R and python.
import pandas as pd
import numpy as np
from array import array
df = pd.DataFrame({
"numpy_vec": np.zeros(10),
"list_vec": [1]* 10,
"native_array_vec": array('d', [3.14] * 10) # less used but native python arrays
})
print("type of numpy_vector column:", type(df["numpy_vec"]), df["numpy_vec"].dtype)
print("type of list_vector column:", type(df["list_vec"]), df["list_vec"].dtype)
print("type of native_array_vector column:", type(df["native_array_vec"]), df["native_array_vec"].dtype)
print(df)
type of numpy_vector column: <class 'pandas.core.series.Series'> float64
type of list_vector column: <class 'pandas.core.series.Series'> int64
type of native_array_vector column: <class 'pandas.core.series.Series'> float64
numpy_vec list_vec native_array_vec
0 0.0 1 3.14
1 0.0 1 3.14
2 0.0 1 3.14
3 0.0 1 3.14
4 0.0 1 3.14
5 0.0 1 3.14
6 0.0 1 3.14
7 0.0 1 3.14
8 0.0 1 3.14
9 0.0 1 3.14
With BiocFrame, no assumptions are made, and the input data is not cast into (un)expected types:
from biocframe import BiocFrame
import numpy as np
from array import array
bframe_types = BiocFrame({
"numpy_vec": np.zeros(10),
"list_vec": [1]* 10,
"native_array_vec": array('d', [3.14] * 10)
})
print("type of numpy_vector column:", type(bframe_types["numpy_vec"]))
print("type of list_vector column:", type(bframe_types["list_vec"]))
print("type of native_array_vector column:", type(bframe_types["native_array_vec"]))
print(bframe_types)
type of numpy_vector column: <class 'numpy.ndarray'>
type of list_vector column: <class 'list'>
type of native_array_vector column: <class 'array.array'>
BiocFrame with 10 rows and 3 columns
numpy_vec list_vec native_array_vec
<ndarray[float64]> <list> <array>
[0] 0.0 1 3.14
[1] 0.0 1 3.14
[2] 0.0 1 3.14
[3] 0.0 1 3.14
[4] 0.0 1 3.14
[5] 0.0 1 3.14
[6] 0.0 1 3.14
[7] 0.0 1 3.14
[8] 0.0 1 3.14
[9] 0.0 1 3.14
Handling complex nested frames¶
Pandas DataFrame does not support nested structures; therefore, running the snippet below will result in an error:
df = pd.DataFrame({
"ensembl": ["ENS00001", "ENS00002", "ENS00002"],
"symbol": ["MAP1A", "BIN1", "ESR1"],
"ranges": pd.DataFrame({
"chr": ["chr1", "chr2", "chr3"],
"start": [1000, 1100, 5000],
"end": [1100, 4000, 5500]
}),
})
print(df)
However, it is handled seamlessly with BiocFrame:
bframe_nested = BiocFrame({
"ensembl": ["ENS00001", "ENS00002", "ENS00002"],
"symbol": ["MAP1A", "BIN1", "ESR1"],
"ranges": BiocFrame({
"chr": ["chr1", "chr2", "chr3"],
"start": [1000, 1100, 5000],
"end": [1100, 4000, 5500]
}),
})
print(bframe_nested)
BiocFrame with 3 rows and 3 columns
ensembl symbol ranges
<list> <list> <BiocFrame>
[0] ENS00001 MAP1A chr1:1000:1100
[1] ENS00002 BIN1 chr2:1100:4000
[2] ENS00002 ESR1 chr3:5000:5500
Construction¶
Creating a BiocFrame object is straightforward; just provide the data as a dictionary.
from biocframe import BiocFrame
obj = {
"ensembl": ["ENS00001", "ENS00002", "ENS00003"],
"symbol": ["MAP1A", "BIN1", "ESR1"],
}
bframe = BiocFrame(obj)
print(bframe)
BiocFrame with 3 rows and 2 columns
ensembl symbol
<list> <list>
[0] ENS00001 MAP1A
[1] ENS00002 BIN1
[2] ENS00003 ESR1
obj = {
"ensembl": ["ENS00001", "ENS00002", "ENS00002"],
"symbol": ["MAP1A", "BIN1", "ESR1"],
"ranges": BiocFrame({
"chr": ["chr1", "chr2", "chr3"],
"start": [1000, 1100, 5000],
"end": [1100, 4000, 5500]
}),
}
bframe2 = BiocFrame(obj, row_names=["row1", "row2", "row3"])
print(bframe2)
BiocFrame with 3 rows and 3 columns
ensembl symbol ranges
<list> <list> <BiocFrame>
row1 ENS00001 MAP1A chr1:1000:1100
row2 ENS00002 BIN1 chr2:1100:4000
row3 ENS00002 ESR1 chr3:5000:5500
The row_names parameter is analogous to index in the pandas world and should not contain missing strings. Additionally, you may provide:
column_data: ABiocFrameobject containing metadata about the columns. This must have the same number of rows as the numbers of columns.metadata: Additional metadata about the object, usually a dictionary.column_names: If different from the keys in thedata. If not provided, this is automatically extracted from the keys in thedata.
With other DataFrame libraries¶
Pandas¶
BiocFrame is intended for accurate representation of Bioconductor objects for interoperability with R, many users may prefer working with pandas DataFrame objects for their actual analyses. This conversion is easily achieved:
from biocframe import BiocFrame
bframe3 = BiocFrame(
{
"foo": ["A", "B", "C", "D", "E"],
"bar": [True, False, True, False, True]
}
)
df = bframe3.to_pandas()
print(type(df))
print(df)
<class 'pandas.core.frame.DataFrame'>
foo bar
0 A True
1 B False
2 C True
3 D False
4 E True
Converting back to a BiocFrame is similarly straightforward:
out = BiocFrame.from_pandas(df)
print(out)
BiocFrame with 5 rows and 2 columns
foo bar
<list> <list>
0 A True
1 B False
2 C True
3 D False
4 E True
Polars¶
Similarly, you can easily go back and forth between BiocFrame and a polars DataFrame:
from biocframe import BiocFrame
bframe3 = BiocFrame(
{
"foo": ["A", "B", "C", "D", "E"],
"bar": [True, False, True, False, True]
}
)
pl = bframe3.to_polars()
print(pl)
shape: (5, 2)
┌─────┬───────┐
│ foo ┆ bar │
│ --- ┆ --- │
│ str ┆ bool │
╞═════╪═══════╡
│ A ┆ true │
│ B ┆ false │
│ C ┆ true │
│ D ┆ false │
│ E ┆ true │
└─────┴───────┘
Extracting data¶
BiocPy classes follow a functional paradigm for accessing or setting properties, with further details discussed in functional paradigm section.
Properties can be directly accessed from the object:
print("shape:", bframe.shape)
print("column names (functional style):", bframe.get_column_names())
print("column names (as property):", bframe.column_names) # same as above
shape: (3, 2)
column names (functional style): ['ensembl', 'symbol']
column names (as property): ['ensembl', 'symbol']
We can fetch individual columns:
print("functional style:", bframe.get_column("ensembl"))
print("w/ accessor", bframe["ensembl"])
functional style: ['ENS00001', 'ENS00002', 'ENS00003']
w/ accessor ['ENS00001', 'ENS00002', 'ENS00003']
And we can get individual rows as a dictionary:
bframe.get_row(2)
{'ensembl': 'ENS00003', 'symbol': 'ESR1'}
sliced_with_bools = bframe[1:2, [True, False, False]]
print("Subset using booleans: \n", sliced_with_bools)
sliced_with_names = bframe[[0,2], ["symbol", "ensembl"]]
print("\nSubset using column names: \n", sliced_with_names)
# Short-hand to get a single column:
print("\nShort-hand to get a single column: \n", bframe["ensembl"])
Subset using booleans:
BiocFrame with 1 row and 1 column
ensembl
<list>
[0] ENS00002
Subset using column names:
BiocFrame with 2 rows and 2 columns
symbol ensembl
<list> <list>
[0] MAP1A ENS00001
[1] ESR1 ENS00003
Short-hand to get a single column:
['ENS00001', 'ENS00002', 'ENS00003']
Setting data¶
Preferred approach¶
For setting properties, we encourage a functional style of programming to avoid mutating the object directly. This helps prevent inadvertent modifications of BiocFrame instances within larger data structures.
modified = bframe.set_column_names(["column1", "column2"])
print(modified)
BiocFrame with 3 rows and 2 columns
column1 column2
<list> <list>
[0] ENS00001 MAP1A
[1] ENS00002 BIN1
[2] ENS00003 ESR1
Now let’s check the column names of the original object,
# Original is unchanged:
print(bframe.get_column_names())
['ensembl', 'symbol']
To add new columns, or replace existing ones:
modified = bframe.set_column("symbol", ["A", "B", "C"])
print(modified)
modified = bframe.set_column("new_col_name", range(2, 5))
print(modified)
BiocFrame with 3 rows and 2 columns
ensembl symbol
<list> <list>
[0] ENS00001 A
[1] ENS00002 B
[2] ENS00003 C
BiocFrame with 3 rows and 3 columns
ensembl symbol new_col_name
<list> <list> <range>
[0] ENS00001 MAP1A 2
[1] ENS00002 BIN1 3
[2] ENS00003 ESR1 4
Change the row or column names:
modified = bframe.\
set_column_names(["FOO", "BAR"]).\
set_row_names(['alpha', 'bravo', 'charlie'])
print(modified)
BiocFrame with 3 rows and 2 columns
FOO BAR
<list> <list>
alpha ENS00001 MAP1A
bravo ENS00002 BIN1
charlie ENS00003 ESR1
We also support Bioconductor’s metadata concepts, either along the columns or for the entire object:
modified = bframe.\
set_metadata({ "author": "Jayaram Kancherla" }).\
set_column_data(BiocFrame({"column_source": ["Ensembl", "HGNC" ]}))
print(modified)
BiocFrame with 3 rows and 2 columns
ensembl symbol
<list> <list>
[0] ENS00001 MAP1A
[1] ENS00002 BIN1
[2] ENS00003 ESR1
------
column_data(1): column_source
metadata(1): author
The not-preferred-way¶
Properties can also be set by direct assignment for in-place modification. We prefer not to do it this way as it can silently mutate BiocFrame instances inside other data structures.
Nonetheless:
testframe = BiocFrame({ "A": [1,2,3], "B": [4,5,6] })
testframe.column_names = ["column1", "column2" ]
print(testframe)
BiocFrame with 3 rows and 2 columns
column1 column2
<list> <list>
[0] 1 4
[1] 2 5
[2] 3 6
/home/runner/work/BiocFrame/BiocFrame/.tox/docs/lib/python3.11/site-packages/biocframe/BiocFrame.py:457: UserWarning: Setting property 'column_names' is an in-place operation, use 'set_column_names' instead
warn(
Similarly, we could set or replace columns directly:
testframe["column2"] = ["A", "B", "C"]
testframe[1:3, ["column1","column2"]] = BiocFrame({"x":[4, 5], "y":["E", "F"]})
/home/runner/work/BiocFrame/BiocFrame/.tox/docs/lib/python3.11/site-packages/biocframe/BiocFrame.py:808: UserWarning: This method performs an in-place operation, use 'set_column' instead
warn(
/home/runner/work/BiocFrame/BiocFrame/.tox/docs/lib/python3.11/site-packages/biocframe/BiocFrame.py:802: UserWarning: This method performs an in-place operation, use 'set_slice' instead
warn(
Iterate over rows¶
You can iterate over the rows of a BiocFrame object. name is None if the object does not contain any row_names. To iterate over the first two rows:
for name, row in bframe[:2,]:
print(name, row)
None {'ensembl': 'ENS00001', 'symbol': 'MAP1A'}
None {'ensembl': 'ENS00002', 'symbol': 'BIN1'}
Combining objects¶
BiocFrame implements methods for the various combine generics from BiocUtils. For example, to combine by row:
import biocutils
bframe1 = BiocFrame({
"odd": [1, 3, 5, 7, 9],
"even": [0, 2, 4, 6, 8],
})
bframe2 = BiocFrame({
"odd": [11, 33, 55, 77, 99],
"even": [0, 22, 44, 66, 88],
})
combined = biocutils.combine_rows(bframe1, bframe2)
print(combined)
BiocFrame with 10 rows and 2 columns
odd even
<list> <list>
[0] 1 0
[1] 3 2
[2] 5 4
[3] 7 6
[4] 9 8
[5] 11 0
[6] 33 22
[7] 55 44
[8] 77 66
[9] 99 88
Similarly, to combine by column:
bframe3 = BiocFrame({
"foo": ["A", "B", "C", "D", "E"],
"bar": [True, False, True, False, True]
})
combined = biocutils.combine_columns(bframe1, bframe3)
print(combined)
BiocFrame with 5 rows and 4 columns
odd even foo bar
<list> <list> <list> <list>
[0] 1 0 A True
[1] 3 2 B False
[2] 5 4 C True
[3] 7 6 D False
[4] 9 8 E True
Relaxed combine operation¶
By default, the combine methods assume that the number and identity of columns (for combine_rows()) or rows (for combine_columns()) are the same across objects. In situations where this is not the case, such as having different columns across objects, we can use relaxed_combine_rows() instead:
from biocframe import relaxed_combine_rows
modified2 = bframe2.set_column("foo", ["A", "B", "C", "D", "E"])
combined = biocutils.relaxed_combine_rows(bframe1, modified2)
print(combined)
BiocFrame with 10 rows and 3 columns
odd even foo
<list> <list> <list>
[0] 1 0 None
[1] 3 2 None
[2] 5 4 None
[3] 7 6 None
[4] 9 8 None
[5] 11 0 A
[6] 33 22 B
[7] 55 44 C
[8] 77 66 D
[9] 99 88 E
Sql-like join operation¶
Similarly, if the rows are different, we can use BiocFrame’s merge function. This function uses the row_names as the index to perform this operation; you can specify an alternative set of keys through the by parameter.
from biocframe import merge
modified1 = bframe1.set_row_names(["A", "B", "C", "D", "E"])
modified3 = bframe3.set_row_names(["C", "D", "E", "F", "G"])
combined = merge([modified1, modified3], by=None, join="outer")
print(combined)
BiocFrame with 7 rows and 4 columns
odd even foo bar
<list> <list> <list> <list>
A 1 0 None None
B 3 2 None None
C 5 4 A True
D 7 6 B False
E 9 8 C True
F None None D False
G None None E True
Empty Frames¶
We can create empty BiocFrame objects that only specify the number of rows. This is beneficial in scenarios where BiocFrame objects are incorporated into larger data structures but do not contain any data themselves.
empty = BiocFrame(number_of_rows=100)
print(empty)
BiocFrame with 100 rows and 0 columns
Most operations detailed in this document can be performed on an empty BiocFrame object.
print("Column names:", empty.column_names)
subset_empty = empty[1:10,:]
print("\nSubsetting an empty BiocFrame: \n", subset_empty)
Column names: []
Subsetting an empty BiocFrame:
BiocFrame with 9 rows and 0 columns
Further reading¶
Explore more details the reference documentation.
Additionally, take a look at Bioconductor’s S4Vectors package, which implements the
DFrameclass upon whichBiocFramewas built.