Integration stage (`01_integration`)

This stage imports multiple SingleCellExperiment objects and performs integration using one or several methods.

⚙️ Config file: config/integration/01_integration.yaml

📋 HTML report target (in config/pipeline.yaml): DRAKE_TARGETS: ["report_integration"]

🪜 Structure

Reading in data from:
- drake caches of multiple single-sample pipelines
- Rds files with SingleCellExperiment objects
Subsetting to common column and row data, and metadata
Normalization of each sample for inter-batch sequencing depth (batchelor::multiBatchNorm(), theory)
Combining of HVGs by either (theory):
- A HVG metric (gene variance or CV2), then apply top HVGs selection as in the single-sample pipeline. All samples must have the same HVG metric used before.
- Set operations (without duplicates): intersection, union, or take them all
Perform integration using one or more methods:
- Rescaling (batchelor::rescaleBatches(), theory): scale counts so that the average count within each batch is the same for each gene
- Regression (batchelor::regressBatches(), theory): fit a linear model to each gene regress out uninteresting factors of variation, returning a matrix of residuals
- Fast mutual nearest neighbors (batchelor::fastMNN(), theory)
- Harmony

The steps below are performed for each integration method, plus separately for HVGs with removed cell cycle-related genes (if this removal was requested).

Calculation of PCA and selection of number of principal components (PCs) which will be used downstream. This is similar to what is done in the single-sample pipeline (02_norm_clustering stage), but, in this case, the selection of PCs can be configured for each integration method.
Dimensionality reduction, which is same as in the single-sample pipeline, but as PCA, it can be configured for each integration method
Expression plots of selected markers
Mutual nearest neighbors (MNN) clustering to assess integration results
Integration diagnostics:
- Assignment of cells of each sample to MNN clusters (tables, plots)
- Rand indices: used to evaluate biological heterogeneity preservation by summarizing the agreement between clusterings

Config for this stage is stored in the config/integration/01_integration.yaml file. Directory with this file is read from SCDRAKE_INTEGRATION_CONFIG_DIR environment variable upon scdrake load or attach, and saved as scdrake_integration_config_dir option. This option is used as the default argument value in several scdrake functions.

Integration sources

INTEGRATION_SOURCES:
  - pbmc1k:
      path: "path/to/.drake or path/to/sce.Rds"
      path_type: "drake_cache"
      description: "10x Genomics PBMC 1k dataset"
      hvg_rm_cc_genes: True
      hvg_cc_genes_var_expl_threshold: 5
    pbmc3k:
      path: "path/to/.drake or path/to/sce.Rds"
      path_type: "drake_cache"
      description: "10x Genomics PBMC 3k dataset"
      hvg_rm_cc_genes: False
      hvg_cc_genes_var_expl_threshold: null

Type: list of named lists

This parameter specifies datasets to be integrated. Those are read in from single-sample pipeline results (the sce_final_norm_clustering targets) taken from corresponding drake caches or from SCE objects saved in Rds format. The latter should correspond to sce_final_norm_clustering targets, but one can modify their e.g. colData() as needed.

Each name of a nested list specifies the name of a single-sample, e.g. pbmc1k. Parameters for each single-sample are:

path: "path/to/.drake or path/to/sce.Rds" (character scalar): a path to drake cache directory or SCE object Rds file (depends on path_type).
path_type: "drake_cache" (character scalar: "drake_cache" | "sce"): a type of input data - either a drake cache directory or SCE object Rds file.
description: "10x Genomics PBMC 1k dataset" (character scalar): a description of the single-sample, will appear in reports.
hvg_rm_cc_genes: True (logical scalar): if True, cell cycle-related genes will be removed prior to selection of highly variable genes (HVGs). This is the same method of cell cycle correction as in the 02_norm_clustering stage of the single-sample pipeline (see vignette("stage_norm_clustering") for more details).
hvg_cc_genes_var_expl_threshold: 5 (numeric scalar): a threshold on cell cycle variance explained. Genes with var. expl. greater than this threshold will be marked as CC-related.

Currently, there are some limitations of which datasets can be integrated:

All single-samples must be normalized by {scran} method.
If the HVG selection in integrated data is based on combined HVG metrics (see the relevant parameters below), all single-samples had to use the same HVG metric (HVG_METRIC of "gene_var" or "gene_cv2").

Selection of highly variable genes (HVGs)

This is done prior to integration, and is same for all methods specified in the INTEGRATION_METHODS (see below).

If a single-sample has the hvg_rm_cc_genes set to True, cell cycle-related genes will be removed prior to HVG combination and selection.

HVG_COMBINATION_INT: "hvg_metric"

Type: character scalar ("hvg_metric" | "intersection" | "union" | "all")

How to combine HVGs from single-samples:

"hvg_metric": combine HVG metrics (gene variance or CV2) and then apply HVG selection similar to procedure in the 02_norm_clustering stage of the single-sample pipeline. Note that in order to combine HVG metrics, all samples had to have HVG_METRIC set to either "gene_var" (scran::combineVar()) or "gene_cv2" (scran::combineCV2()).
Combine HVGs by set operations:
- "intersection"
- "union"
"all": use all common genes from all samples as HVGs (without duplicates).

HVG_SELECTION_INT: "top"
HVG_SELECTION_VALUE_INT: 3000

Type: character scalar, integer scalar

HVG selection strategy. Only relevant when HVG_COMBINATION_INT is "hvg_metric", i.e. this selection will be applied to combined HVG metrics.

These parameters are similar to HVG_SELECTION and HVG_SELECTION_VALUE in the 02_norm_clustering stage of the single-sample pipeline (see vignette("stage_norm_clustering") for more details).

Integration methods

INTEGRATION_METHODS:
  - uncorrected:
      pca_selection_method: "forced"
      pca_forced_pcs: 15
      tsne_perp: 20
      tsne_max_iter: 1000
    rescaling:
      pca_selection_method: "forced"
      pca_forced_pcs: 15
      tsne_perp: 20
      tsne_max_iter: 1000
      integration_params:
        log.base: 2
        pseudo.count: 1
    regression:
      pca_selection_method: "corrected"
      pca_forced_pcs: 15
      tsne_perp: 20
      tsne_max_iter: 1000
      integration_params:
        d: 50
    mnn:
      pca_selection_method: "corrected"
      pca_forced_pcs: 15
      tsne_perp: 20
      tsne_max_iter: 1000
      integration_params:
        k: 20
        prop.k: null
        cos.norm: True
        ndist: 3
        d: 50
        merge.order: null
        auto.merge: True
    harmony:
      pca_selection_method: null
      pca_forced_pcs: null
      tsne_perp: 20
      tsne_max_iter: 1000
      integration_params:
        dims.use: 50
        theta: null
        lambda: null
        sigma: 0.1
        nclust: null
        tau: 0
        block.size: 0.05
        max.iter.harmony: 10
        max.iter.cluster: 20
        epsilon.cluster: 0.00001
        epsilon.harmony: 0.0001

Type: list of named lists

Each named list is specifying parameters for one integration method. The ones specified in the integration_params list are used directly in an integration method, while the others are used prior/after integration. The "uncorrected" and at least one integration method must be set (by default, all methods are performed). Here are descriptions and possible parameters (some of them are used by multiple methods, and thus described only once) for each integration method:

uncorrected: correction for inter-batch sequencing depth with batchelor::multiBatchNorm(). This method is required, because it is used for differential expression analysis (cluster markers and contrasts).
- pca_selection_method: "forced" (character scalar: "forced" | "elbow" | "technical_noise"): see the PCA_SELECTION_METHOD parameter in vignette("stage_norm_clustering").
- pca_forced_pcs: 15 (integer scalar): force this number of selected PCs (if pca_selection_method is "forced").
- tsne_perp: 20 (numeric scalar): t-SNE perplexity.
- tsne_max_iter: 1000 (positive integer scalar): a maximum number of t-SNE iterations.
rescaling: scale counts so that the average count within each batch is the same for each gene (batchelor::rescaleBatches()).
- pca_selection_method: "forced"
- pca_forced_pcs: 15
- tsne_perp: 20
- tsne_max_iter: 1000
- integration_params: directly passed to batchelor::RescaleParam() used for batchelor::rescaleBatches(). Default values from this function are used by default.
  - log.base: 2 (numeric scalar): a base of the log-transformation.
  - pseudo.count: 1 (numeric scalar): a pseudo-count used for the log-transformation.
regression: fit a linear model to each gene regress out uninteresting factors of variation, returning a matrix of residuals (batchelor::regressBatches()).
- pca_selection_method: "corrected": in this integration method, batchelor::multiBatchPCA() is used, and all PCs (50 by default) from this method will be used if "corrected" is specified. Otherwise any other PCA selection method can be used.
- pca_forced_pcs: 15
- tsne_perp: 20
- tsne_max_iter: 1000
- integration_params: directly passed to batchelor::RegressParam() used for batchelor::regressBatches(). Default values from this function are used by default.
  - d: 50 (integer scalar): a number of PCs to compute in batchelor::multiBatchPCA().
mnn: fast mutual nearest neighbors (batchelor::fastMNN()).
- pca_selection_method: "corrected"
- pca_forced_pcs: 15
- tsne_perp: 20
- tsne_max_iter: 1000
- integration_params: directly passed to batchelor::FastMnnParam() used for batchelor::fastMNN(). Default values from this function are used by default.
  - k: 20 (integer scalar): a number of nearest neighbors to consider when identifying MNNs.
  - prop.k: null (numeric scalar): a proportion of cells in each dataset to use for mutual nearest neighbor searching. If set, the number of nearest neighbors used for the MNN search in each batch is redefined as max(k, prop.k*N) where N is the number of cells in that batch.
  - cos.norm: True (logical scalar): if True, cosine normalization is performed on the input data prior to PCA.
  - ndist: 3 (numeric scalar): a threshold beyond which neighbors are to be ignored when computing correction vectors. Each threshold is defined as a multiple of the number of median distances.
  - d: 50
  - merge.order: null (integer vector, list of lists, or null): a merge order of single-samples. Alternatively, a list of lists representing a tree structure specifying a hierarchical merge order (not tested, see batchelor::fastMNN() for more details).
  - auto.merge: True (logical scalar): if True, automatically identify the “best” merge order.
harmony: projects cells into a shared embedding in which cells group by cell type rather than dataset-specific conditions (harmony::RunHarmony(), Nature Methods). Default values from this function are used by default. Note that Harmony does not compute an integrated expression matrix, but only reduced dimensions corrected for batch effect that are subsequently used for UMAP and t-SNE, and clustering.
- pca_selection_method: null; pca_forced_pcs: null: not used, see dims.use below
- tsne_perp: 20
- tsne_max_iter: 1000
- integration_params: directly passed to harmony::RunHarmony(). Default values from this function are used by default, except dims.use.
  - dims.use: 50 (integer scalar): number of first PCs to use for Harmony integration. Selection of PCs is not applied for Harmony - PCA with dims.use PCs is computed prior to the integration.
  - theta: null: diversity clustering penalty parameter. Specify for each variable in group.by.vars ("batch"). Default theta = 2, theta = 0 does not encourage any diversity. Larger values of theta result in more diverse clusters.
  - lambda: null: didge regression penalty parameter. Specify for each variable in group.by.vars ("batch"). Default lambda = 1. Lambda must be strictly positive. Smaller values result in more aggressive correction.
  - sigma: 0.1: width of soft kmeans clusters. Default sigma = 0.1. Sigma scales the distance from a cell to cluster centroids. Larger values of sigma result in cells assigned to more clusters. Smaller values of sigma make soft k-means cluster approach hard clustering.
  - nclust: null: number of clusters in model. nclust = 1 equivalent to simple linear regression.
  - tau: 0: protection against overclustering small datasets with large ones. tau is the expected number of cells per cluster.
  - block.size: 0.05: what proportion of cells to update during clustering. Between 0 to 1, default 0.05. Larger values may be faster but less accurate.
  - max.iter.harmony: 10: maximum number of rounds to run Harmony. One round of Harmony involves one clustering and one correction step.
  - max.iter.cluster: 20: maximum number of rounds to run clustering at each round of Harmony.
  - epsilon.cluster: 0.00001: convergence tolerance for clustering round of Harmony. Set to -inf to never stop early.
  - epsilon.harmony: 0.0001: convergence tolerance for Harmony. Set to -inf to never stop early.

Integration diagnostics

Fast MNN clustering

Used to obtain basic integration diagnostics. See scran::buildSNNGraph() and bluster::makeSNNGraph() for more details.

INTEGRATION_SNN_K: 10

Type: integer scalar

A number of nearest neighbors to consider during graph construction.

INTEGRATION_SNN_TYPE: "rank"

Type: character scalar ("rank" | "number" | "jaccard")

A type of weighting scheme to use for shared neighbors.

INTEGRATION_SNN_CLUSTERING_METHOD: "walktrap"

Type: character scalar ("walktrap" | "louvain")

A type of MNN clustering algorithm. See igraph::cluster_walktrap() and igraph::cluster_louvain() for more details.

Input files

SELECTED_MARKERS_FILE: null

Type: character scalar or null

Similar to the SELECTED_MARKERS_FILE parameter in vignette("stage_norm_clustering").

INTEGRATION_REPORT_RMD_FILE: "Rmd/integration/01_integration.Rmd"

Type: character scalar

A path to RMarkdown file used for HTML report of this pipeline stage.

Output files

INTEGRATION_BASE_OUT_DIR: "01_integration"

Type: character scalar

A path to base output directory for this stage. It will be created under BASE_OUT_DIR specified in 00_main.yaml config.

INTEGRATION_SELECTED_MARKERS_OUT_DIR: "selected_markers"
INTEGRATION_REPORT_HTML_FILE: "01_integration.html"

Type: character scalar

Names of files and directories created under INTEGRATION_BASE_OUT_DIR. Subdirectories are not allowed.

HTML output parameters

INTEGRATION_KNITR_MESSAGE: False
INTEGRATION_KNITR_WARNING: False
INTEGRATION_KNITR_ECHO: False

Type: logical scalar

These are passed to knitr::opts_chunk() and used for rendering of stage’s HTML report.

Here you can find description of the most important targets for this stage. However, for a full overview, you have to inspect the source code of the get_integration_subplan() function.

HTML report target name: report_integration

`SingleCellExperiment` objects

sce_int_import: list of imported single-sample SCE objects, as defined in the INTEGRATION_SOURCES parameter. Because drake cannot watch for changes in a cache, this target is always rerun. Also, some constraints are checked (common normalization method and HVG metrics).

sce_int_raw_snn_clustering: sce_int_import with computed fast SNN clustering for each single-sample, used for integration diagnostics

sce_int_processed: sce_int_import subsetted to common colData, and features (rows) and their corresponding metadata (hvg_ids, hvg_metric_fit)

sce_int_multibatchnorm: sce_int_processed with SCE objects normalized for inter-batch sequencing depth (batchelor::multiBatchNorm())

sce_int_df: a tibble with integrated SCE objects. Each row is one method defined in the INTEGRATION_METHODS parameter and either with or without removed cell cycle-related genes from HVGs. See ?sce_int_df_fn for information about what is added or modified in metadata() of each integrated SCE object. The integrated assay is named "integrated", except for Harmony integration, which only computes a batch-corrected reduced dimensions (available in reducedDims(sce, "harmony")).

sce_int_pca_df: sce_int_df with computed PCA and selected number of PCs. Also includes PC selection statistics and plot

sce_int_clustering_df: sce_int_pca_df with computed MNN clustering used for integration diagnostics

sce_int_dimred_df: sce_int_pca_df with computed t-SNE and UMAP dimensionality reductions

Highly variable genes (HVGs) selection

hvg_int, hvg_int_with_cc: lists of HVGs and selection parameters. The latter is not NULL when any of the single-samples defined in the INTEGRATION_SOURCES parameter has hvg_rm_cc_genes set to True.

hvg_plots_int_df: a tibble with HVG diagnostic plots

Plots

sce_int_dimred_plots_df: a tibble with dimensionality reduction plots

Selected markers

selected_markers_int_plots_df: a tibble with plots of selected markers, similar to selected_markers_plots in the 02_norm_clustering stage of the single-sample pipeline

selected_markers_plots_files_out: make this target to export the plots

Document generated: 2024-09-14 13:58:45 UTC+0000

Integration sources

Selection of highly variable genes (HVGs)

Integration methods

Integration diagnostics

Fast MNN clustering

Input files

Output files

HTML output parameters

SingleCellExperiment objects

Highly variable genes (HVGs) selection

Plots

Selected markers

^{Document generated:
2024-09-14 13:58:45 UTC+0000}

`SingleCellExperiment` objects