How to import data into immuneML

The first step of any immuneML analysis is to import the dataset that will be used. There exist three types of datasets in immuneML:

  • repertoire datasets should be used when making predictions per repertoire, such as predicting a disease state. When importing a repertoire dataset, you should create a metadata file.

  • sequence datasets should be used when predicting values for single immune receptor chains, such as antigen specificity.

  • receptor datasets are the paired variant of sequence datasets, and should be used to make a prediction for each receptor chain pair.

A broad range of different import formats can be specified, including AIRR, MiXCR, VDJdb, ImmunoSEQ (Adaptive Biotechnologies), 10xGenomics, OLGA and IGoR. For the complete list of supported data formats, and extensive documentation see Datasets. If you are using a custom format, or your preferred format is not yet supported, any type of tabular file can also be imported using Generic import. When possible, using format-specific importers is preferred over Generic import, as they require less options to be set and might take care of automatic reformatting of certain fields.

Alternatively to importing data from files, it is also possible to generate datasets containing random immune receptor sequences on the fly, see How to generate a random sequence, receptor or repertoire dataset.

What should the metadata file look like?

The metadata file is a simple .csv file describing metadata fields for a repertoire dataset where each row corresponds to one repertoire. Metadata files are only used for repertoire datasets, for receptor and sequence datasets the metadata information should be defined as additional columns in the same file that contains the sequences.

In case of repertoire datasets, each repertoire is represented by one file in the given format (e.g., AIRR/MiXCR/Adaptive). For all repertoires in one dataset, a single metadata file should be defined containing the following columns:

../_images/metadata.png

The columns filename and subject_id are mandatory. Other columns may be defined by the user. There are no restrictions as to what type of information these columns should represent, but typically they will represent diseases, HLA, age or sex. These columns can be used as a prediction target (also known as labels) when training ML models. When writing a YAML specification, the labels are defined by using the same name as the user-defined column(s) in the metadata file.

For an example of a metadata file, see here the metadata file used to replicate the analysis by Emerson et al. 2017.

YAML specification for importing data from files

Data import must be defined as a part of the YAML specification. First, we choose a name which will be used to refer to the dataset in the subsequent analyses:

definitions:
  datasets:
    my_dataset: # this is the name of the dataset we will use in the YAML specification
      ... # here, format and input parameters will be specified

The name is defined by the user. It can consist of letters, numbers and underscores.

Under the dataset name key, the format of the data must be specified, as well as additional parameters under a key named params. Under format, any of the formats listed under Datasets may be filled in. Under params, the parameter path is always required when importing data from files. All the files must be stored in a single folder, and this folder must set through the parameter path.

Here is an incomplete example specification using AIRR format:

definitions:
  datasets:
    my_dataset: # this is the name of the dataset we will use in the YAML specification
      format: AIRR
      params:
        path: path/to/data/
        ... # other import parameters will be specified here

Specifying params for repertoire dataset import

By default, it is assumed that a repertoire dataset should be imported. In this case, the path to the metadata_file must be specified. The metadata file is a .csv file which contains one repertoire (filename) per row, and the metadata labels for that repertoire. These metadata labels can be used as a prediction target when training ML models. For more details on structuring the metadata file, see What should the metadata file look like?. Note that only the repertoire files that are present in the metadata file will be imported.

Other parameters that are specific to the format may be specified under params as well, and are explained in more detail for each format under Datasets.

A complete specification for importing a repertoire dataset from AIRR format with default parameters may look like this:

definitions:
  datasets:
    my_dataset: # this is the name of the dataset we will use in the YAML specification
      format: AIRR
      params:
        # required parameters
        path: path/to/data/
        metadata_file: path/to/metadata.csv
        # is_repertoire is by default True, and may be omitted
        is_repertoire: True
        # Other parameters specific to AIRR data may be specified here

Specifying params for receptor or sequence dataset import

If you want to import a sequence or receptor dataset, set the parameter is_repertoire to false, and set paired to either false (sequence dataset) or true (receptor dataset). For sequence and receptor dataset, metadata labels must be specified directly as columns in the input files. These metadata labels can be used as a prediction target when training ML models. For example, a column ‘binding’ can be added, which may have values ‘true’ and ‘false’. The metadata labels are specified through parameter metadata_column_mapping, which is a mapping from the names of the columns in the file to the names that will be used internally in immuneML (for example: when specifying labels in the TrainMLModel instruction). It is recommended that the immuneML-internal names contain only lowercase letters, numbers and underscores.

A complete specification for importing a sequence dataset from AIRR format with default parameters may look like this:

definitions:
  datasets:
    my_dataset: # this is the name of the dataset we will use in the YAML specification
      format: AIRR
      params:
        # required parameters
        path: path/to/data/
        is_repertoire: false
        paired: false # must be true for receptor dataset and false for sequence datasets
        metadata_column_mapping: # metadata column mapping AIRR: immuneML
          binding: binding # the names could just be the same
          Epitope.gene: epitope_gene # if the column name contains undesired characters, it may be renamed for internal use
        # Other parameters specific to AIRR data may be specified here

For receptor datasets, the additional parameter receptor_chains needs to be set, which determines the type of chain pair that should be imported. The resulting specification may look like this:

definitions:
  datasets:
    my_dataset: # this is the name of the dataset we will use in the YAML specification
      format: AIRR
      params:
        # required parameters
        path: path/to/data/
        is_repertoire: false
        paired: true # must be true for receptor dataset and False for sequence datasets
        receptor_chains: TRA_TRB # choose from TRA_TRB, TRG_TRD, IGH_IGL and IGH_IGK
        metadata_column_mapping: # metadata column mapping AIRR: immuneML
          binding: binding # the names could just be the same
          Epitope.gene: epitope_gene # if the column name contains undesired characters, it may be renamed for internal use
        # Other parameters specific to AIRR data may be specified here

Importing previously generated immuneML datasets

When you import a dataset into immuneML for the first time, it is converted to an optimized binary format, which speeds up the analysis. The main resulting file has an .iml_dataset extension, and may be accompanied by several other .yaml and .npy files. When running immuneML locally, you can by default find these immuneML dataset files in the folder ‘datasets’, which is located in the main output folder of your analysis.

Some instructions (Simulation, DatasetExport, Subsampling) also explicitly export binarized immuneML datasets when selecting ‘ImmuneML’ as the export format.

These .iml_dataset files can later be imported easily and with few parameters, and importing from .iml_dataset is also faster than importing from other data formats. A YAML specification for ImmuneML data import is shown below. Important note: ImmuneML files might not be compatible between different immuneML (sub)versions.

definitions:
  datasets:
    my_dataset:
      format: ImmuneML
      params:
        path: path/to/dataset.iml_dataset
        # specifying a metadata_file is optional, it will update the dataset using this new metadata.
        metadata_file: path/to/metadata.csv