Cell Ranger 3.1 (latest), printed on 08/25/2019
Cell Ranger uses an aligner called STAR, which peforms splicing-aware alignment of reads to the genome. Cell Ranger then uses the transcript annotation GTF to bucket the reads into exonic, intronic, and intergenic, and by whether the reads align (confidently) to the genome. A read is exonic if at least 50% of it intersects an exon, intronic if it is non-exonic and intersects an intron, and intergenic otherwise.
For reads that align to a single exonic locus but also align to 1 or more non-exonic loci, the exonic locus is prioritized and the read is considered to be confidently mapped to the exonic locus with MAPQ 255.
Cell Ranger further aligns exonic reads to annotated transcripts, looking for compatibility. A read that is compatible with the exons of an annotated transcript, and aligned to the same strand, is considered mapped to the transcriptome. If the read is compatible with a single gene annotation, it is considered uniquely (confidently) mapped to the transcriptome. These confidently mapped reads are the only ones considered for UMI counting.
Before counting UMIs, Cell Ranger attempts to correct for sequencing errors in the UMI sequences. Reads that were confidently mapped to the transcriptome are placed into groups that share the same barcode, UMI, and gene annotation. If two groups of reads have the same barcode and gene, but their UMIs differ by a single base (i.e., are Hamming distance 1 apart), then one of the UMIs was likely introduced by a substitution error in sequencing. In this case, the UMI of the less-supported read group is corrected to the UMI with higher support.
Cell Ranger again groups the reads by barcode, UMI (possibly corrected), and gene annotation. If two or more groups of reads have the same barcode and UMI, but different gene annotations, the gene annotation with the most supporting reads is kept for UMI counting, and the other read groups are discarded. In case of a tie for maximal read support, all read groups are discarded, as the gene cannot be confidently assigned.
After these two filtering steps, each observed barcode, UMI, gene combination is recorded as a UMI count in the unfiltered feature-barcode matrix. The number of reads supporting each counted UMI is also recorded in the molecule info file.
Cell Ranger 3.0 introduces an improved cell-calling algorithm that is better able to identify populations of low RNA content cells, especially when low RNA content cells are mixed into a population of high RNA content cells. For example, tumor samples often contain large tumor cells mixed with smaller tumor infiltrating lymphocytes (TIL) and researchers may be particularly interested in the TIL population. The new algorithm is based on the EmptyDrops method (Lun et al., 2018).
The algorithm has two key steps:
In the first step, the original Cell Ranger cell calling algorithm is used to identify the primary mode of high RNA content cells, using a cutoff based on the total UMI count for each barcode. Cell Ranger takes as input the expected number of recovered cells, N (see --expect-cells). Let m be the 99th percentile of the top N barcodes by total UMI counts. All barcodes whose total UMI counts exceed m/10 are called as cells in the first pass.
In the second step, a set of barcodes with low UMI counts that likely represent ‘empty’ GEM partitions is selected. A model of the RNA profile of selected barcodes is created. This model, called the background model, is a multinomial distribution over genes. It uses Simple Good-Turing smoothing to provide a non-zero model estimate for genes that were not observed in the representative empty GEM set. Finally, the RNA profile of each barcode not called as a cell in the first step is compared to the background model. Barcodes whose RNA profile strongly disagrees with the background model are added to the set of positive cell calls. This second step identifies cells that are clearly distinguishable from the profile of empty GEMs, even though they may have much lower RNA content than the largest cells in the experiment.
Below is an example of a challenging cell-calling scenario where 300 high RNA content 293T cells are mixed with 2000 low RNA content PBMC cells. On the left is the cell calling result with the cell calling algorithm prior to Cell Ranger 3.0 and on the right is the current Cell Ranger 3.0 result. You can see that low RNA content cells are successfully identified by the new algorithm.
The plot shows the count of filtered UMIs mapped to each barcode. Barcodes can be determined to be cell-associated based on their UMI count or by their RNA profiles. Therefore some regions of the graph can contain both cell-associated and background-associated barcodes. The color of the graph represents the local density of barcodes that are cell-associated.
In some cases the set of barcodes called as cells may not match the desired set of barcodes based on visual inspection. This can be remedied by either re-running count or reanalyze with the --force-cells option, or by selecting the desired barcodes from the raw feature-barcode matrix in downstream analysis. Custom barcode selection can also be done by specifying --barcodes to reanalyze.
When using Cell Ranger in Feature Barcoding Only Analysis mode, only step #1 of the cell calling algorithm is used. The cells called by this step of the algorithm are returned directly.
When multiple genomes are present in the reference (for example, Human and Mouse, or any other pair), Cell Ranger runs a special multigenome analysis to detect barcodes associated with partitions where cells from two different genomes were present. Among all barcodes called as cell-associated (See "Calling Cell Barcodes"), they are initially classified as Human or Mouse by which genome has more total UMI counts for that barcode. Barcodes with total UMI counts that exceed the 10th percentile of the distributions for both Human and Mouse are called as observed multiplets. Because Cell Ranger can only observe the (Human, Mouse) multiplets, it computes an inferred multiplet rate by estimating the total number of multiplets (including (Human, Human) and (Mouse, Mouse)). This is done by estimating via maximum likelihood the total number of multiplet GEMs from the observed multiplets and the inferred ratio of Human to Mouse cells. If this ratio is 1:1, the inferred multiplet rate is approximately twice the observed (Human, Mouse) multiplets.
In order to reduce the gene expression matrix to its most important features, Cell Ranger uses Principal Components Analysis (PCA) to change the dimensionality of the dataset from (cells x genes) to (cells x M) where M is a user-selectable number of principal components (via num_principal_comps). The pipeline uses a python implementation of IRLBA algorithm, (Baglama & Reichel, 2005), which we modified to reduce memory consumption. The reanalyze pipeline allows the user to further reduce the data by randomly subsampling the cells and/or selecting genes by their dispersion across the dataset. Note that if the data contains Feature Barcode data, only the gene expression data will be used for PCA and subsequent analysis.
For visualizing data in 2-d space, Cell Ranger passes the PCA-reduced data into t-SNE (t-Stochastic Neighbor Embedding), a nonlinear dimensionality reduction method. (Van der Maaten, 2014) The C++ reference implementation by Van der Maaten was modified to take a PRNG seed for determinism and to expose various parameters which can be changed in
reanalyze. We also decreased its runtime by fixing the number of output dimensions at compile time to 2 or 3.
Cell Ranger also supports visualization with UMAP (Uniform Manifold Approximation and Projection), which estimates a topology of the high dimensional data and uses this information to estimate a low dimensional embedding that preserves relationships present in the data. (Leland McInnes et al, 2018) The pipeline uses the python implementation of this algorithm by Leland McInnes. The
reanalyze pipeline allows the user to customize the parameters for the UMAP, including
metric etc. Below shows the t-SNE (left) and UMAP (right) visualizations of our public dataset 5k PBMCs.
Cell Ranger uses two different methods for clustering cells by expression similarity, both of which operate in the PCA space.
The graph-based clustering algorithm consists of building a sparse nearest-neighbor graph (where cells are linked if they among the k nearest Euclidean neighbors of one another), followed by Louvain Modularity Optimization (LMO; Blondel, Guillaume, Lambiotte, & Lefebvre, 2008), an algorithm which seeks to find highly-connected "modules" in the graph. The value of k, the number of nearest neighbors, is set to scale logarithmically with the number of cells. An additional cluster-merging step is done: Perform hierarchical clustering on the cluster-medoids in PCA space and merge pairs of sibling clusters if there are no genes differentially expressed between them (with B-H adjusted p-value below 0.05). The hierarchical clustering and merging is repeated until there are no more cluster-pairs to merge.
The use of LMO to cluster cells was inspired by a similar method in the R package Seurat
Cell Ranger also performs traditional K-means clustering across a range of K values, where K is the preset number of clusters. In the web summary prior to 1.3.0, the default selected value of K is that which yields the best Davies-Bouldin Index, a rough measure of clustering quality.
In order to identify genes whose expression is specific to each cluster, Cell Ranger tests, for each gene and each cluster, whether the in-cluster mean differs from the out-of-cluster mean.
In order to find differentially expressed genes between groups of cells, Cell Ranger uses the quick and simple method sSeq (Yu, Huber, & Vitek, 2013), which employs a negative binomial exact test. When the counts become large, Cell Ranger switches to the fast asymptotic beta test used in edgeR (Robinson & Smyth, 2007). For each cluster, the algorithm is run on that cluster versus all other cells, yielding a list of genes that are differentially expressed in that cluster relative to the rest of the sample.
Cell Ranger's implementation differs slightly from that in the paper: in the sSeq paper, the authors recommend using DESeq's geometric mean-based definition of library size. Cell Ranger instead computes relative library size as the total UMI counts for each cell divided by the median UMI counts per cell. As with sSeq, normalization is implicit in that the per-cell library-size parameter is incorporated as a factor in the exact-test probability calculations.
To correct the batch effects between chemistries, the Cell Ranger uses an algorithm based on mutual nearest neighbors (MNN; Haghverdi et al, 2018) to identify similar cell subpopulations between batches. The mutual nearest neighbor is defined as a pair of cells from two different batches that is contained in each other’s set of nearest neighbors. The MNNs are detected for every pair of user-defined batches to balance among batches, as proposed in (Park et al, 2018).
The cell subpopulation matches between batches will then be used to merge multiple batches together (Hie et al, 2018). The difference in expression values between cells in a MNN pair provides an estimate of the batch effect. A correction vector for each cell is obtained as a weighted average of the estimated batch effects, where a Gaussian kernel function up-weights matching vectors belonging to nearby points (Haghverdi et al, 2018).
The batch effect score is defined to quantitatively measure the batch effect before and after correction. For every cell, we calculate how many of its k nearest-neighbors belong to the same batch and normalize it by the expected number of same batch cells when there isn't batch effect. The batch effect score is calculated as the average of above metric in randomly sampled 10% of the total number of cells. If there isn't batch effect, we would expect that each cells’ nearest neighbors would be evenly shared across all batches and the batch effect score is close to 1.
In the example below, the PBMC mixture was profiled separately by Single Cell 3' v2 (brown) and Single Cell 3' v3 (blue). On the left is the t-SNE plot after aggregating two libraries without the batch correction, and on the right is the t-SNE plot with the batch correction. The batch effect score decreased from 1.81 to 1.31 with the chemistry batch correction.
Baglama, J. & Reichel, L. Augmented Implicitly Restarted Lanczos Bidiagonalization Methods. SIAM Journal on Scientific Computing 27, 19–42 (2005).
Blondel, V. D., Guillaume, J.-L., Lambiotte, R. & Lefebvre, E. Fast unfolding of communities in large networks. Journal of Statistical Mechanics: Theory and Experiment 2008, (2008).
Haghverdi, L., Lun, A. T., Morgan, M. D., & Marioni, J. C. Batch effects in single-cell RNA-sequencing data are corrected by matching mutual nearest neighbors. Nature biotechnology, 36(5), 421 (2018).
Hie, B. L., Bryson, B., & Berger, B. Panoramic stitching of heterogeneous single-cell transcriptomic data. bioRxiv 2018.
Lun, A., Riesenfeld, S., Andrews, T., Dao, T.P., Gomes. T., Participants in the 1st Human Cell Atlas Jamboree, Marioni, J. Distinguishing cells from empty droplets in droplet-based single-cell RNA sequencing data. bioRxiv 2018
McInnes, L, Healy, J, UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction. arXiv 2018
Park, J. E., Polanski, K., Meyer, K., & Teichmann, S. A. Fast Batch Alignment of Single Cell Transcriptomes Unifies Multiple Mouse Cell Atlases into an Integrated Landscape. bioRxiv 2018.
Robinson, M. D. & Smyth, G. K. Small-sample estimation of negative binomial dispersion, with applications to SAGE data. Biostatistics 9, 321–332 (2007). Link to edgeR source.
Van der Maaten, L.J.P. Accelerating t-SNE using Tree-Based Algorithms. Journal of Machine Learning Research 15, 3221-3245 (2014).
Yu, D., Huber, W. & Vitek, O. Shrinkage estimation of dispersion in Negative Binomial models for RNA-seq experiments with small sample size. Bioinformatics 29, 1275–1282 (2013).