API Walkthrough

Creating and connecting

Creating .loom files

To create a loom file from data, you need to supply a main matrix (numpy ndarray or scipy sparse matrix) and two dictionaries of row and column attributes (with attribute names as keys, and numpy ndarrays as values). If the main matrix is N×M, then the row attributes must have N elements, and the column attributes must have M elements.

For example, the following creates a loom file with a 100x100 main matrix, one row attribute and one column attribute:

import numpy as np
import loompy
filename = "test.loom"
matrix = np.arange(10000).reshape(100,100)
row_attrs = { "SomeRowAttr": np.arange(100) }
col_attrs = { "SomeColAttr": np.arange(100) }
loompy.create(filename, matrix, row_attrs, col_attrs)
loompy.create accepts numpy dense matrices (np.ndarray) as well as scipy sparse matrices (scipy.coo_matrix, scipy.csc_matrix,
or scipy.csr_matrix). For example:
import numpy as np
import loompy
import scipy.sparse as sparse
filename = "test.loom"
matrix = sparse.coo_matrix((100, 100))
row_attrs = { "SomeRowAttr": np.arange(100) }
col_attrs = { "SomeColAttr": np.arange(100) }
loompy.create(filename, matrix, row_attrs, col_attrs)

You can also create a file by combining existing loom files. The files will be concatenated along the column axis, and therefore must have the same number of rows. If the rows are potentially not in the same order, you can supply a key argument; the row attribute corresponding to the key will be used to sort the files. For example, the following code will combine files and use the “Accession” row attribute as the key:

loompy.combine(files, output_filename, key="Accession")

You can import a 10X Genomics cellranger output folder:

loompy.create_from_cellranger(folder, output_filename)

Connecting to .loom files

In order to work with a loom file, you must first connect to it. This does not load the data or attributes, so is very quick regardless of the size of the file. It’s more like connecting to a database than reading a file. Loom supports Python context management, so normally you should use a with statement to take care of the connection:

with loompy.connect("filename.loom") as ds:
  # do something with ds

The connection will be automatically closed at the end of the with block.

Sometimes, especially in interactive use in a Jupyter notebook, you may want to just open the file and keep the connection around:

ds = loompy.connect("filename.loom")

In that case, you should close the file when you are done:


In most cases, forgetting to close the file will do no harm, but may (for example) prevent concurrent processes from accessing the file, and will leak file handles.

In the rest of the documentation below, ds is assumed to be an instance of LoomConnection obtained by connecting to a .loom file.

Manipulate data

Shape, indexing and slicing

The shape property returns the row and column count as a tuple:

>>> ds.shape
(100, 2345)

The data stored in the main matrix can be retrieved by indexing and slicing. The following are supported:

  • Indices: anything that can be converted to a Python long
  • Slices (i.e. : or 0:10)
  • Lists of the rows/columns you want (i.e. [0, 34, 576])
  • Mask arrays (i.e. numpy array of bool indicating the rows/columns you want)

Lists and mask arrays are supported along one dimension at a time only. Since the main matrix is two-dimensional, two arguments are always needed. Examples:

ds[:, :]          # Return the entire matrix
ds[0:10, 0:10]    # Return the 10x10 submatrix starting at row and column zero
ds[99, :]         # Return the 100th row
ds[:, 99]         # Return the 100th column
ds[[0,3,5], :]    # Return rows with index 0, 3 and 5
ds[:, bool_array] # Return columns where bool_array elements are True

Note that performance will be poor if you select many individual rows (columns) out of a large matrix. For example, in a dataset with shape (27998, 160796), loading ten randomly chosen individual full columns took 914 ms, whereas loading 1000 columns took 1 minute and 6 seconds, and loadingh 5000 columns took 13 minutes. This slowdown is caused by a performance bug in h5py.

If the whole dataset fits in RAM, loading it in full and then selecting the row/columns you want will be faster. If it doesn’t, consider using the scan() method (see below), which in this example took 1 minute and 12 seconds regardless of how many columns were selected. As a rule of thumb, scan() will be faster whenever you are loading more than about 1% of the rows or columns (randomly selected).

Sparse data

On disk, every layer is stored chunked and block-compressed, for efficient storage and access along both axes.

The main matrix and additional layers can be assigned from dense or sparse matrices (any scipy.sparse format).

You can load the main matrix or any layer as sparse:

ds.layers["exons"].sparse()  # Returns a scipy.sparse.coo_matrix()
ds.layers["unspliced"].sparse(rows, cols)  # Returns only the indicated rows and columns (ndarrays of integers)

Global attributes

Global attributes are available at ds.attrs and can be accessed by name or as a dictionary. You create new attributes by assignment, and delete them using the del statement:

>>> ds.attrs.title
"The title of the dataset"

>>> ds.attrs.title = "New title"
>>> ds.attrs["title"]
"New title"

>>> del ds.attrs.title

You can list the attributes and loop over them as you would with a dictionary:

>>> ds.attrs.keys()
["title", "description"]

>>> for key, value in ds.attrs.items():
>>>   print(f"{key} = {value}")
title = New title
description = Fancy dataset

Global attributes can be scalars, or multidimensional arrays of any shape, and the elements can be integers, floats or strings. See below for the exact types allowed.

Row and column attributes

Row and column attributes are accessed at ds.ra and ds.ca, respectively, and support the same interface as global attributes. For example:

ds.ra.keys()       # Return list of row attribute names
ds.ca.keys()       # Return list of column attribute names
ds.ra.Gene = ...   # Create or replace the Gene attribute
a = ds.ra.Gene     # Assign the array of gene names (assuming the attribute exists)
del ds.ra.Gene     # Delete the Gene row attribute

Attributes can be any of the following:

  • One-dimensional arrays of integers, floats or strings. The number of elements in the array must match the corresponding matrix dimension.
  • Multidimensional arrays of any of the same element types. The length along the first dimension of a row attribute must equal the number of rows in the main matrix (and vice versa for column attributes). Remaining dimensions can be any size.

For example, if the main matrix has M columns, the result of a dimensionality reduction (for example, a PCA) to 20 dimensions could be stored as a column attribute with shape (M, 20).

You can assign attributes using almost any array or list-like type, but attributes will always return numpy array (np.ndarray).

Using attributes as masks for indexing the main matrix results in a very compact and readable syntax for selecting subarrays:

>>> ds[ds.ra.Gene == "Actb", :]
array([[  2.,   9.,   9., ...,   0.,  14.,   0.]], dtype=float32)

>>> ds[(ds.ra.Gene == "Actb") | (ds.ra.Gene == "Gapdh"), :]
array([[  2.,   9.,   9., ...,   0.,  14.,   0.],
       [  0.,   1.,   4., ...,   0.,  14.,   3.]], dtype=float32)

>>> ds[:, ds.ca.CellID == "AAACATACATTCTC-1"]
array([[ 0.],
       [ 0.],
       [ 0.],
       [ 0.],
       [ 0.],
       [ 0.]], dtype=float32)

Note that numpy logical functions overload the bitwise, not the boolean operators. Use | for ‘or’, & for ‘and’ and ~ for ‘not’. You also must place parentheses around the comparison expressions to ensure proper operator precedence. For example:

(a == b) & (a > c) | ~(c <= b)

Adding columns

You can add columns to an existing loom file. It’s not possible to add rows or to delete any part of the matrix.

ds.add_columns(submatrix, col_attrs)

You need to provide a submatrix corresponding to the new columns, as well as a dictionary of column attributes with values for all the new columns.

You can also add the contents of another .loom file:

ds.add_loom(other_file, key="Gene")

The content of the other file is added as columns on the right of the current dataset. The rows must match for this to work. That is, the two files must have exactly the same number of rows. If key is given, the rows will be ordered based on the key attribute. Furthermore, the two datasets must have the same column attributes (but of course can have different values for those attributes at each column). Missing attributes can be given default values using the fill_values argument. If the files contain any global attribute with conflicting values, you can automatically convert such attributes into column attributes by passing convert_attrs=True to the method.

There is also a convenient function to create or append to a loom file:

loompy.create_append(filename, layers, row_attrs, col_attrs)

This will create the file if it doesn’t exist, and append to it if it does. This function is commonly used when combining several loom files while performing a selection on the columns:

for f in input_files:
  with loompy.connect(f) as ds:
    cells = # select the desired columns in ds
    for (_, _, view) in ds.scan(items=cells):
      loompy.create_append(outout_file, view.layers, view.ra, view.ca)

The code above loops over a number of input files, then scans across each file to select a desired subset of the columns (cells) and writes them to the output file. Since it uses scan(), it will never load entire datasets in RAM and will work no matter how big the input datasets are.


Loom supports multiple layers. There is always a single main matrix, but optionally one or more additional layers having the same number of rows and columns. Layers are accessed using the layers property on the LoomConnection object.

Layers support the same pythonic API as attributes:

ds.layers.keys()            # Return list of layers
ds.layers["unspliced"]      # Return the layer named "unspliced"
ds.layers["spliced"] = ...  # Create or replace the "spliced" layer
a = ds.layers["spliced"][:, 10] # Assign the 10th column of layer "spliced" to the variable a
del ds.layers["spliced"]     # Delete the "spliced" layer

The main matrix is availabe as a layer named “” (the empty string). It cannot be deleted but otherwise supports the same operations as any other layer.

As a convenience, layers are also available directly on the connection object. The above expressions are equivalent to the following:

ds["unspliced"]      # Return the layer named "unspliced"
ds["spliced"] = ...  # Create or replace the "spliced" layer
a = ds["spliced"][:, 10] # Assign the 10th column of layer "spliced" to the variable a
del ds["spliced"]     # Delete the "spliced" layer


Loom supports sparse graphs with either the rows or the columns as nodes. For example, a sparse graph of cells (stored in the columns) could represent a K nearest-neighbors graph of the cells. In that case, the cells are the nodes (so there are M nodes in the graph if there are M columns in the main matrix), which are connected by an arbitrary number of edges. The graph could be considered directed or undirected, and can have float-valued weights on the edges. Loom even supports multigraphs (permitting multiple edges between pairs of nodes). Graphs are stored as arrays of edges and the associated edge weights.

Row and column graphs are accessed at ds.row_graphs and ds.col_graphs, respectively, and support the same interface as attributes. For example:

ds.row_graphs.keys()      # Return list of row graphs
ds.col_graphs.KNN = ...   # Create or replace the column-oriented graph KNN
a = ds.col_graphs.KNN     # Assign the KNN column graph to variable a
del ds.col_graphs.KNN     # Delete the KNN graph

Graphs are returned as scipy.sparse.coo_matrix, and can be created/assigned from any scipy sparse format as well as from a numpy dense matrix or ndarray. In each case, the matrix represents the adjacency matrix of the graph.


Loompy views are in-memory views of a slice through the underlying loom file. Views can be created explicitly by slicing:

ds.view[:, 10:20]

This will create a view, fully loaded in memory, containing all the rows of the underlying loom file, but only columns 10 through 19 (zero-based). You can use fancy indexing including slices, arrays of integers (to pick out specific rows/columns) and boolean arrays.

The power of the view is that it slices through everything: the main matrix, every layer, every attribute, and every graph. This hides a lot of messy and error-prone code, and makes it easy to extract relevant subsets of a loom file.

The most common use of a view is in scanning through a file (see scan() below).



You can map one or more functions across all rows (all columns), while avoiding loading the entire dataset into memory:

ds.map([np.mean, np.std], axis=1)

The functions will receive an array (of floats or integers) as their only argument, and should return a single float or integer value. Internally, map() uses scan() to loop across the file.

Note that you must always provide a list of functions, even if it has only one element, and that the result is a list of vectors, one per function that was supplied.


Permute the order of the rows or columns:

ordering = np.random.permutation(np.arange(ds.shape[1]))
ds.permute(ordering, axis=1)

This permutes the order of rows or columns in the file, without loading the entire file in RAM. The ordering argument should be a numpy array of ds.shape[axis] elements, in the desired order.


For very large loom files, it’s very useful to scan across the file (along either rows or columns) in batches, to avoid loading the entire file in memory. This can be achieved using the scan() method:

for (ix, selection, view) in ds.scan(axis=1):
  # do something with each view

Inside the loop, you get access to the current view into the file. It has all the attributes, graphs and data of the original loom file, but only for the columns included in selection (or rows, if axis=0).

In essence, you get a succession of slices through the loom file, corresponding to bands of columns (rows). The ix variable tells you the starting column of the band, whereas the selection gives you the list of columns contained in the current view.

You can also scan across a selected subset of the columns or rows. For example:

cells = # List of columns you want to see
for (ix, selection, view) in ds.scan(items=cells, axis=1):
  # do something with each view

This works exactly the same, except that each selection and view now include only the columns you asked for.