HOME  ›   pipelines

Customized Secondary Analysis with cellranger reanalyze

The cellranger reanalyze command reruns secondary analysis performed on the feature-barcode matrix (dimensionality reduction, clustering, and visualization) using different parameter settings.

Command line interface

These are the most common command line arguments (run cellranger reanalyze --help for a full list):

ArgumentDescription
--id=IDA unique run ID string: e.g. AGG123_reanalysis
--matrix=H5Path to a filtered_feature_bc_matrix.h5 or raw_feature_bc_matrix.h5 or sample_filtered_feature_bc_matrix.h5 from a completed pipestance (from either cellranger count, cellranger multi, or cellranger aggr). Use the raw_feature_bc_matrix.h5 when specifying a value for --force-cells that exceeds the original cell count.
--params=CSVOptional. Path to a CSV file containing a list of valid parameters and the values to use for them (see Parameters).
--agg=CSVOptional. Path to a CSV file that was used for cellranger aggr. This allows you to retain any metadata associated with the samples for display in Loupe Browser. This argument is required if you want to enable Chemistry Batch Correction in your reanalysis.
--barcodes=listOptional. Path to a file containing a list of barcodes (one barcode per line) to use for reanalysis. The first line of file must be a header called Barcode. Header is case sensitive. All barcodes must be present in the matrix.
--genes=listOptional. Path to a file containing a list of gene IDs (one gene ID per line) to use for reanalysis (corresponding to the gene_id field of the reference GTF). The first line of the genes file must be a header called Gene. Header is case sensitive. All gene IDs included in the list must be present in the matrix. Only gene features are used in the secondary analysis. Feature Barcode features are ignored. An example gene.txt supplied to this flag:
Gene
ENSG00000243485
ENSG00000237613
ENSG00000186092
ENSG00000238009
ENSG00000239945
ENSG00000239906
ENSG00000241599
ENSG00000236601
ENSG00000284733
ENSG00000235146
ENSG00000284662

--exclude-genes=listOptional. Path to a file containing a list of gene IDs (one gene ID per line) to exclude for reanalysis (corresponding to the gene_id field of the reference GTF). All gene IDs must be present in the matrix. The first line of the genes file must be a header called Gene. The exclusion can be applied with or without the --genes list. When a --genes list is supplied, --exclude-genes searches within the supplied list for exclusion. Note that only gene features are used in secondary analysis. Feature Barcode features are ignored.
--force-cells=NUMOptional. Force pipeline to use this number of cells, bypassing the cell detection algorithm. Use this if the number of cells estimated by Cell Ranger is not consistent with the barcode rank plot. If specifying a value that exceeds the original cell count, you must use the raw_feature_bc_matrices_h5.h5. Starting with Cell Ranger 6.1, it is no longer possible to run --force-cells of an aggr output in reanalyze with more cells than were originally called.

After specifying these input arguments, run cellranger reanalyze. In this example, we're reanalyzing the results of an aggregation named AGG123 from cellranger count outputs:

$cd /home/jdoe/runs$ ls -1 AGG123/outs/*.h5 # verify the input file exists
AGG123/outs/filtered_feature_bc_matrix.h5
AGG123/outs/filtered_molecules.h5
AGG123/outs/raw_feature_bc_matrix.h5
AGG123/outs/raw_molecules.h5
\$ cellranger reanalyze --id=AGG123_reanalysis \
--matrix=AGG123/outs/filtered_feature_bc_matrix.h5 \
--params=AGG123_reanalysis.csv


The pipeline will begin to run, creating a new folder named with the reanalysis ID you specified (e.g. /home/jdoe/runs/AGG123_reanalysis) for its output. If this folder already exists, cellranger will assume it is an existing pipestance and attempt to resume running it.

Output files

A successful run should conclude with a message similar to this:

2018-10-09 11:05:58 [runtime] (run:local)       ID.AGG123_reanalysis.SC_RNA_REANALYZER_CS.SC_RNA_ANALYZER.SUMMARIZE_ANALYSIS.fork0.join
2018-10-09 11:06:01 [runtime] (join_complete)   ID.AGG123_reanalysis.SC_RNA_REANALYZER_CS.SC_RNA_ANALYZER.SUMMARIZE_ANALYSIS

Outputs:
- Secondary analysis output CSV:          /home/jdoe/runs/AGG123_reanalysis/outs/analysis_csv
- Secondary analysis web summary:         /home/jdoe/runs/AGG123_reanalysis/outs/web_summary.html
- Copy of the input parameter CSV:        /home/jdoe/runs/AGG123_reanalysis/outs/params_csv.csv
- Copy of the input aggregation CSV:      /home/jdoe/runs/AGG123_reanalysis/outs/aggregation_csv.csv
- Loupe Browser file:                /home/jdoe/runs/AGG123_reanalysis/outs/cloupe.cloupe
- Filtered feature-barcode matrices MEX:  /home/jdoe/runs/AGG123_reanalysis/outs/filtered_feature_bc_matrix
- Filtered feature-barcode matrices HDF5:  /home/jdoe/runs/AGG123_reanalysis/outs/filtered_feature_bc_matrix.h5

Pipestance completed successfully!


Refer to the Analysis page for an explanation of the secondary analyses outputs.

Parameters

The CSV file passed to --params should have 0 or more rows, one for every parameter that you want to customize. There is no header row. If a parameter is not specified in your CSV, its default value will be used. See Common Use Cases for some examples.

Here is a detailed description of each parameter. For parameters that subset the data, a default value of null indicates that no subsetting happens by default.

ParameterTypeDefault ValueRecommended RangeDescription
num_analysis_bcsintnullCannot be set higher than the available number of cells.Randomly subset data to N barcodes for all analysis. Reduce this parameter if you want to improve performance or simulate results from lower cell counts.
num_pca_bcsintnullCannot be set higher than the available number of cells.Randomly subset data to N barcodes when computing PCA projection (the most memory-intensive step). The PCA projection will still be applied to the full dataset, i.e. your final results will still reflect all the data. Try reducing this parameter if your analysis is running out of memory.
num_pca_genesintnullCannot be set higher than the number of genes in the reference transcriptome.Subset data to the top N genes (ranked by normalized dispersion) when computing PCA. Differential expression will still reflect all genes. Try reducing this parameter if your analysis is running out of memory.
num_principal_compsint1010-100, depending on the number of cell populations / clusters you expect to see.Compute N principal components for PCA. Setting this too high may cause spurious clusters to be called. The default value is 100 when the chemistry batch correction is enabled.
cbc_knnint105-20Specify the number of nearest neighbors used to identify mutual nearest neighbors. Setting this too high will increase runtime and may cause out of memory error. See Chemistry Batch Correction page for more details.
cbc_alphafloat0.10.05-0.5Specify the threshold of the percentage of matched cells between two batches, which is used to determine if the batch pair will be merged. See Chemistry Batch Correction page for more details.
cbc_sigmafloat15010-500Specify the bandwidth of the Gaussian smoothing kernel used to compute the correction vector for each cell. See Chemistry Batch Correction page for more details.
cbc_realign_panoramaboolfalse[true, false]Specify if two batches will be merged if they are already in the same panorama. Setting this to True will usually improve the performance, but will also increase runtime and memory usage. See Chemistry Batch Correction page for more details.
graphclust_neighborsint010-500, depending on desired granularityNumber of nearest-neighbors to use in the graph-based clustering. Lower values result in higher-granularity clustering. The actual number of neighbors used is the maximum of this value and that determined by neighbor_a and neighbor_b. Set this value to zero to use those values instead.
neighbor_afloat-230.0Determines how clustering granularity scales with cell count.The number of nearest neighbors, k, used in the graph-based clustering is computed as follows: k = neighbor_a + neighbor_b * log10(n_cells). The actual number of neighbors used is the maximum of this value and graphclust_neighbors.
neighbor_bfloat120.0Determines how clustering granularity scales with cell count.The number of nearest neighbors, k, used in the graph-based clustering is computed as follows: k = neighbor_a + neighbor_b * log10(n_cells). The actual number of neighbors used is the maximum of this value and graphclust_neighbors.
max_clustersint1010-50, depending on the number of cell populations / clusters you expect to see.Compute K-means clustering using K values of 2 to N. Setting this too high may cause spurious clusters to be called.
tsne_input_pcsintnullCannot be set higher than the num_principal_comps parameter.Subset to top N principal components for TSNE. Change this parameter if you want to see how the TSNE plot changes when using fewer PCs, independent of the clustering / differential expression. You may find that TSNE is faster and/or the output looks better when using fewer PCs.
tsne_perplexityint3030-50TSNE perplexity parameter (see the TSNE FAQ for more details). When analyzing 100k+ cells, increasing this parameter may improve TSNE results, but the algorithm will be slower.
tsne_thetafloat0.5Must be between 0 and 1.TSNE theta parameter (see the TSNE FAQ for more details). Higher values yield faster, more approximate results (and vice versa). The runtime and memory performance of TSNE will increase dramatically if you set this below 0.25.
tsne_max_dimsint2Must be 2 or 3.Maximum number of TSNE output dimensions. Set this to 3 to produce both 2D and 3D TSNE projections (note: runtime will increase significantly).
tsne_max_iterint10001000-10000Number of total TSNE iterations. Try increasing this if TSNE results do not look good on larger numbers of cells. Runtime increases linearly with number of iterations.
tsne_stop_lying_iterint250Cannot be set higher than tsne_max_iter.Iteration at which TSNE learning rate is reduced. Try increasing this if TSNE results do not look good on larger numbers of cells.
tsne_mom_switch_iterint250Cannot be set higher than tsne_max_iter.Iteration at which TSNE momentum is reduced. Try increasing this if TSNE results do not look good on larger numbers of cells. Cannot be set higher than tsne_max_iter.
umap_input_pcsintnullCannot be set higher than the num_principal_comps parameter.Subset to top N principal components for UMAP. Change this parameter if you want to see how the UMAP plot changes when using fewer PCs, independent of the clustering / differential expression. You may find that UMAP is faster and/or the output looks better when using fewer PCs.
umap_n_neighborsint30[5, 50]Determines the number of neighboring points used in local approximations of manifold structure. Larger values will usually result in more global structure at the loss of detailed local structure.
umap_max_dimsint2Must be 2 or 3.Maximum number of UMAP output dimensions. Set this to 3 to produce both 2D and 3D UMAP projections.
umap_min_distfloat0.3[0.001, 0.5]Controls how tightly the embedding is allowed to pack points together. Larger values make embedded points are more evenly distributed, while smaller values make the embedding more accurately with regard to the local structure.
umap_metricstringcorrelationlist of supported metricsDetermines how the distance is computed in the input space.
random_seedint0any integerRandom seed. Due to the randomized nature of the algorithms, changing this will produce slightly different results. If the TSNE or UMAP results don't look good, try running multiple times with different seeds and pick the TSNE or UMAP that looks best.

Common use cases

These examples illustrate what you should put in your --params CSV file in some common situations.

1. More principal components and clusters

For very large / diverse cell populations, the defaults may not capture the full variation between cells. In that case, try increasing the number of principal components and / or clusters. To run PCA with 50 components and k-means with up to 30 clusters, put this in your CSV:

num_principal_comps,50
max_clusters,30


2. Less memory usage

You can limit the memory usage of the analysis by computing the PCA projection on a subset of cells and genes. This is useful for large datasets (100k+ cells). If you have 100k cells, it's completely reasonable to use only 50% of them for PCA - the memory usage will be cut in half, but you'll still be well equipped to detect rare subpopulations. Limiting the number of genes will reduce memory even further. To compute the PCA projection using 50000 cells and 3000 genes, put this in your CSV:

num_pca_bcs,50000
num_pca_genes,3000


Note: Subsetting of cells is done randomly, to avoid bias. Subsetting of genes is done by binning genes by their mean expression across cells, then measuring the dispersion (a variance-like parameter) of each gene's expression normalized to the other genes in its bin.