openpipeline/src/annotate/random_forest_annotation/config.vsh.yaml

name: random_forest_annotation
namespace: annotate
description: Automated cell type annotation tool for scRNA-seq datasets on the basis of random forest.
authors:
  - __merge__: /src/authors/jakub_majercik.yaml
    roles: [ author ]

argument_groups:
  - name: Inputs
    description: Input dataset (query) arguments
    arguments:
      - name: "--input"
        type: file
        description: The input (query) data to be labeled. Should be a .h5mu file.
        direction: input
        required: true
        example: input.h5mu
      - name: "--modality"
        description: Which modality to process.
        type: string
        default: "rna"
        required: false
      - name: "--input_layer"
        type: string
        description: The layer in the input data to be used for cell type annotation if .X is not to be used. 
  - name: Reference
    description: Arguments related to the reference dataset.
    arguments:
      - name: "--reference"
        type: file
        description: "The reference data to train the CellTypist classifiers on. Only required if a pre-trained --model is not provided."
        example: reference.h5mu
        direction: input
        required: false
      - name: "--reference_layer"
        type: string
        description: The layer in the reference data to be used for cell type annotation if .X is not to be used. Data are expected to be processed in the same way as the --input query dataset.
        required: false
      - name: "--reference_obs_target"
        type: string
        description: Key in obs field of reference modality with cell-type information.
        required: true
  - name: Outputs
    description: Output arguments.
    arguments:
      - name: "--output"
        type: file
        description: Output h5mu file.
        direction: output
        example: output.h5mu
      - name: "--output_compression"
        type: string
        choices: ["gzip", "lzf"]
        required: false
        example: "gzip"
      - name: "--output_obs_predictions"
        type: string
        default: random_forest_pred
        required: false
        description: |
          In which `.obs` slots to store the predicted information.
      - name: "--output_obs_probability"
        type: string
        default: random_forest_probability
        required: false
        description: |
          In which `.obs` slots to store the probability of the predictions.
  - name: Model arguments
    description: Model arguments.
    arguments:
      - name: "--model"
        type: file
        description: "Pretrained model in pkl format. If not provided, the model will be trained on the reference data and --reference should be provided."
        required: false
        example: pretrained_model.pkl
      - name: "--n_estimators"
        type: integer
        required: false
        default: 100
        description: Number of trees in the random forest.
      - name: "--max_depth"
        type: integer
        required: false
        description: |
          Maximum depth of the trees in the random forest. 
          If not provided, the nodes are expanded until all leaves only contain a single sample.
      - name: "--criterion"
        type: string
        required: false
        choices: ["gini", "entropy", "log_loss"]
        default: "gini"
        description: The function to measure the quality of a split.
      - name: "--class_weight"
        type: string
        required: false
        default: "balanced_subsample"
        choices: ["balanced", "balanced_subsample", "uniform"]
        description: |
          Weights associated with classes.
          The `balanced` mode uses the values of y to automatically adjust weights inversely proportional to class frequencies in the input data.
          The `balanced_subsample` mode is the same as `balanced` except that weights are computed based on the bootstrap sample for every tree grown.
          The `uniform` mode gives all classes a weight  of one.
      - name: "--max_features"
        type: string
        default: "200"
        description: |
          The number of features to consider when looking for the best split. The value can either be a positive integer or one of `sqrt`, `log2` or `all`.
          If integer: consider max_features features at each split.
          If `sqrt`: max_features is the squareroot of all input features.
          If `log2`: max_features is the log2 of all input features.
          If `all`: max features equals all input features.

resources:
  - type: python_script
    path: script.py
  - path: /src/utils/setup_logger.py

test_resources:
  - type: python_script
    path: test.py
  - path: /resources_test/annotation_test_data/TS_Blood_filtered.h5mu
  - path: /resources_test/pbmc_1k_protein_v3/pbmc_1k_protein_v3_filtered_feature_bc_matrix.h5mu

engines:
  - type: docker
    image: python:3.12-slim
    setup:
      - type: apt
        packages:
          - libhdf5-dev
          - procps
      - type: python
        packages:
          - scikit-learn==1.4.2
      - type: python
        __merge__: [ /src/base/requirements/anndata_mudata.yaml, .]
    __merge__: [ /src/base/requirements/python_test_setup.yaml, .]
runners:
  - type: executable
  - type: nextflow