Batch integration embed

Removing batch effects while preserving biological variation (embedding output)

Description

This is a sub-task of the overall batch integration task. Batch (or data) integration integrates datasets across batches that arise from various biological and technical sources. Methods that integrate batches typically have three different types of output: a corrected feature matrix, a joint embedding across batches, and/or an integrated cell-cell similarity graph (e.g., a kNN graph). This sub-task focuses on all methods that can output joint embeddings, and includes methods that canonically output corrected feature matrices with subsequent postprocessing to generate a joint embedding. Other sub-tasks for batch integration can be found for:

• graphs, and
• corrected features

This sub-task was taken from a benchmarking study of data integration methods.

Summary

viewof color_by_rank = Inputs.toggle({label: "Color by rank"})
viewof scale_column = Inputs.toggle({label: "Rescale per column"})

funkyheatmap(
    funky_heatmap_args.data,
    funky_heatmap_args.columns,
    funky_heatmap_args.column_info,
    funky_heatmap_args.column_groups,
    funky_heatmap_args.palettes,
    funky_heatmap_args.expand,
    funky_heatmap_args.col_annot_offset,
    funky_heatmap_args.add_abc,
    scale_column,
    {
        fontSize: 14,
        rowHeight: 26,
        rootStyle: 'max-width: none',
        colorByRank: color_by_rank
    }
);

Figure 1: Overview of the results per method. This figures shows the mean of the scaled scores (group Overall), the mean scores per dataset (group Dataset) and the mean scores per metric (group Metric).

funkyheatmap = (await require('d3@7').then(d3 => {
  window.d3 = d3;
  return import('https://unpkg.com/funkyheatmap-js@0.1.7');
})).default;

Metrics

• ARI¹: ARI (Adjusted Rand Index) compares the overlap of two clusterings. It considers both correct clustering overlaps while also counting correct disagreements between two clustering.

• Cell Cycle Score¹: The cell-cycle conservation score evaluates how well the cell-cycle effect can be captured before and after integration.

• Graph connectivity¹: The graph connectivity metric assesses whether the kNN graph representation, G, of the integrated data connects all cells with the same cell identity label.

• Isolated label F1¹: Isolated cell labels are identified as the labels present in the least number of batches in the integration task. The score evaluates how well these isolated labels separate from other cell identities based on clustering.

• Isolated label Silhouette¹: This score evaluates the compactness for the label(s) that is(are) shared by fewest batches. It indicates how well rare cell types can be preserved after integration.

• kBET²: kBET determines whether the label composition of a k nearest neighborhood of a cell is similar to the expected (global) label composition. The test is repeated for a random subset of cells, and the results are summarized as a rejection rate over all tested neighborhoods.

• NMI¹: NMI compares the overlap of two clusterings. We used NMI to compare the cell-type labels with Louvain clusters computed on the integrated dataset.

• PC Regression¹: This compares the explained variance by batch before and after integration. It returns a score between 0 and 1 (scaled=True) with 0 if the variance contribution hasn’t changed. The larger the score, the more different the variance contributions are before and after integration.

• Silhouette¹: The absolute silhouette with is computed on cell identity labels, measuring their compactness.

• Batch ASW¹: The absolute silhouette width is computed over batch labels per cell. As 0 then indicates that batches are well mixed and any deviation from 0 indicates a batch effect, we use the 1-abs(ASW) to map the score to the scale [0;1].

Results

Results table of the scores per method, dataset and metric (after scaling). Use the filters to make a custom subselection of methods and datasets. The “Overall mean” dataset is the mean value across all datasets.

Details

Methods

• Random Integration by Batch¹²: Feature values, embedding coordinates, and graph connectivity are all randomly permuted within each batch label. Links: Docs.

• Random Embedding by Celltype¹²: Cells are embedded as a one-hot encoding of celltype labels. Links: Docs.

• Random Embedding by Celltype (with jitter)¹²: Cells are embedded as a one-hot encoding of celltype labels, with a small amount of random noise added to the embedding. Links: Docs.

• Random Graph by Celltype¹²: Cells are embedded as a one-hot encoding of celltype labels. A graph is then built on this embedding. Links: Docs.

• Random Integration by Celltype¹²: Feature values, embedding coordinates, and graph connectivity are all randomly permuted within each celltype label. Links: Docs.

• Combat (full/scaled)⁵: ComBat uses an Empirical Bayes (EB) approach to correct for batch effects. It estimates batch-specific parameters by pooling information across genes in each batch and shrinks the estimates towards the overall mean of the batch effect estimates across all genes. These parameters are then used to adjust the data for batch effects, leading to more accurate and reproducible results. Links: Docs.

• Combat (full/unscaled)⁵: ComBat uses an Empirical Bayes (EB) approach to correct for batch effects. It estimates batch-specific parameters by pooling information across genes in each batch and shrinks the estimates towards the overall mean of the batch effect estimates across all genes. These parameters are then used to adjust the data for batch effects, leading to more accurate and reproducible results. Links: Docs.

• Combat (hvg/scaled)⁵: ComBat uses an Empirical Bayes (EB) approach to correct for batch effects. It estimates batch-specific parameters by pooling information across genes in each batch and shrinks the estimates towards the overall mean of the batch effect estimates across all genes. These parameters are then used to adjust the data for batch effects, leading to more accurate and reproducible results. Links: Docs.

• Combat (hvg/unscaled)⁵: ComBat uses an Empirical Bayes (EB) approach to correct for batch effects. It estimates batch-specific parameters by pooling information across genes in each batch and shrinks the estimates towards the overall mean of the batch effect estimates across all genes. These parameters are then used to adjust the data for batch effects, leading to more accurate and reproducible results. Links: Docs.

• FastMNN embed (full/scaled)¹⁰: fastMNN performs a multi-sample PCA to reduce dimensionality, identifying MNN paris in the low-dimensional space, and then correcting the target batch towards the reference using locally weighted correction vectors. The corrected target batch is then merged with the reference. The process is repeated with the next target batch except for the PCA step. Links: Docs.

• FastMNN embed (full/unscaled)¹⁰: fastMNN performs a multi-sample PCA to reduce dimensionality, identifying MNN paris in the low-dimensional space, and then correcting the target batch towards the reference using locally weighted correction vectors. The corrected target batch is then merged with the reference. The process is repeated with the next target batch except for the PCA step. Links: Docs.

• FastMNN embed (hvg/scaled)¹⁰: fastMNN performs a multi-sample PCA to reduce dimensionality, identifying MNN paris in the low-dimensional space, and then correcting the target batch towards the reference using locally weighted correction vectors. The corrected target batch is then merged with the reference. The process is repeated with the next target batch except for the PCA step. Links: Docs.

• FastMNN embed (hvg/unscaled)¹⁰: fastMNN performs a multi-sample PCA to reduce dimensionality, identifying MNN paris in the low-dimensional space, and then correcting the target batch towards the reference using locally weighted correction vectors. The corrected target batch is then merged with the reference. The process is repeated with the next target batch except for the PCA step. Links: Docs.

• Harmony (full/scaled)³: Harmony is a method that uses PCA to group the cells into multi-dataset clusters, and then computes cluster-specific linear correction factors. Each cell is then corrected by its cell-specific linear factor using the cluster-weighted average. The method keeps iterating these four steps until cell clusters are stable. Links: Docs.

• Harmony (full/unscaled)³: Harmony is a method that uses PCA to group the cells into multi-dataset clusters, and then computes cluster-specific linear correction factors. Each cell is then corrected by its cell-specific linear factor using the cluster-weighted average. The method keeps iterating these four steps until cell clusters are stable. Links: Docs.

• Harmony (hvg/scaled)³: Harmony is a method that uses PCA to group the cells into multi-dataset clusters, and then computes cluster-specific linear correction factors. Each cell is then corrected by its cell-specific linear factor using the cluster-weighted average. The method keeps iterating these four steps until cell clusters are stable. Links: Docs.

• Harmony (hvg/unscaled)³: Harmony is a method that uses PCA to group the cells into multi-dataset clusters, and then computes cluster-specific linear correction factors. Each cell is then corrected by its cell-specific linear factor using the cluster-weighted average. The method keeps iterating these four steps until cell clusters are stable. Links: Docs.

• Liger (full/unscaled)¹¹: LIGER or linked inference of genomic experimental relationships uses iNMF deriving and implementing a novel coordinate descent algorithm to efficiently do the factorization. Joint clustering is performed and factor loadings are normalised. Links: Docs.

• Liger (hvg/unscaled)¹¹: LIGER or linked inference of genomic experimental relationships uses iNMF deriving and implementing a novel coordinate descent algorithm to efficiently do the factorization. Joint clustering is performed and factor loadings are normalised. Links: Docs.

• MNN (full/scaled)⁴: MNN first detects mutual nearest neighbours in two of the batches and infers a projection of the second onto the first batch. After that, additional batches are added iteratively. Links: Docs.

• MNN (full/unscaled)⁴: MNN first detects mutual nearest neighbours in two of the batches and infers a projection of the second onto the first batch. After that, additional batches are added iteratively. Links: Docs.

• MNN (hvg/scaled)⁴: MNN first detects mutual nearest neighbours in two of the batches and infers a projection of the second onto the first batch. After that, additional batches are added iteratively. Links: Docs.

• MNN (hvg/unscaled)⁴: MNN first detects mutual nearest neighbours in two of the batches and infers a projection of the second onto the first batch. After that, additional batches are added iteratively. Links: Docs.

• No Integration¹²: Cells are embedded by PCA on the unintegrated data. A graph is built on this PCA embedding. Links: Docs.

• No Integration by Batch¹²: Cells are embedded by computing PCA independently on each batch. Links: Docs.

• Random Integration¹²: Feature values, embedding coordinates, and graph connectivity are all randomly permuted. Links: Docs.

• SCALEX (full)⁶: SCALEX is a method for integrating heterogeneous single-cell data online using a VAE framework. Its generalised encoder disentangles batch-related components from batch-invariant biological components, which are then projected into a common cell-embedding space. Links: Docs.

• SCALEX (hvg)⁶: SCALEX is a method for integrating heterogeneous single-cell data online using a VAE framework. Its generalised encoder disentangles batch-related components from batch-invariant biological components, which are then projected into a common cell-embedding space. Links: Docs.

• Scanorama (full/scaled)⁹: Scanorama is an extension of the MNN method. Other then MNN, it finds mutual nearest neighbours over all batches and embeds observations into a joint hyperplane. Links: Docs.

• Scanorama (full/unscaled)⁹: Scanorama is an extension of the MNN method. Other then MNN, it finds mutual nearest neighbours over all batches and embeds observations into a joint hyperplane. Links: Docs.

• Scanorama (hvg/scaled)⁹: Scanorama is an extension of the MNN method. Other then MNN, it finds mutual nearest neighbours over all batches and embeds observations into a joint hyperplane. Links: Docs.

• Scanorama (hvg/unscaled)⁹: Scanorama is an extension of the MNN method. Other then MNN, it finds mutual nearest neighbours over all batches and embeds observations into a joint hyperplane. Links: Docs.

• Scanorama gene output (full/scaled)⁹: Scanorama is an extension of the MNN method. Other then MNN, it finds mutual nearest neighbours over all batches and embeds observations into a joint hyperplane. Links: Docs.

• Scanorama gene output (full/unscaled)⁹: Scanorama is an extension of the MNN method. Other then MNN, it finds mutual nearest neighbours over all batches and embeds observations into a joint hyperplane. Links: Docs.

• Scanorama gene output (hvg/scaled)⁹: Scanorama is an extension of the MNN method. Other then MNN, it finds mutual nearest neighbours over all batches and embeds observations into a joint hyperplane. Links: Docs.

• Scanorama gene output (hvg/unscaled)⁹: Scanorama is an extension of the MNN method. Other then MNN, it finds mutual nearest neighbours over all batches and embeds observations into a joint hyperplane. Links: Docs.

• scANVI (full/unscaled)⁷: ScanVI is an extension of scVI but instead using a Bayesian semi-supervised approach for more principled cell annotation. Links: Docs.

• scANVI (hvg/unscaled)⁷: ScanVI is an extension of scVI but instead using a Bayesian semi-supervised approach for more principled cell annotation. Links: Docs.

• scVI (full/unscaled)⁸: scVI combines a variational autoencoder with a hierarchical Bayesian model. Links: Docs.

• scVI (hvg/unscaled)⁸: scVI combines a variational autoencoder with a hierarchical Bayesian model. Links: Docs.

Baseline methods

• Random Integration by Batch: Feature values, embedding coordinates, and graph connectivity are all randomly permuted within each batch label.

• Random Embedding by Celltype: Cells are embedded as a one-hot encoding of celltype labels.

• Random Embedding by Celltype (with jitter): Cells are embedded as a one-hot encoding of celltype labels, with a small amount of random noise added to the embedding.

• Random Graph by Celltype: Cells are embedded as a one-hot encoding of celltype labels. A graph is then built on this embedding.

• Random Integration by Celltype: Feature values, embedding coordinates, and graph connectivity are all randomly permuted within each celltype label.

• No Integration: Cells are embedded by PCA on the unintegrated data. A graph is built on this PCA embedding.

• No Integration by Batch: Cells are embedded by computing PCA independently on each batch.

• Random Integration: Feature values, embedding coordinates, and graph connectivity are all randomly permuted.

Datasets

• Immune (by batch)¹: Human immune cells from peripheral blood and bone marrow taken from 5 datasets comprising 10 batches across technologies (10X, Smart-seq2).

• Lung (Viera Braga et al.)¹: Human lung scRNA-seq data from 3 datasets with 32,472 cells. From Vieira Braga et al. Technologies: 10X and Drop-seq.

• Pancreas (by batch)¹: Human pancreatic islet scRNA-seq data from 6 datasets across technologies (CEL-seq, CEL-seq2, Smart-seq2, inDrop, Fluidigm C1, and SMARTER-seq).

Download raw data

Task info Method info Metric info Dataset info Results Quality control

Quality control results

✓ All checks succeeded!

Visualization of raw results