Computing on X using online (incremental) algorithms
Source:vignettes/census_compute_over_X.Rmd
census_compute_over_X.RmdThis tutorial showcases computing per-gene and per-cell statistics for a user-defined query using out-of-core operations, where incremental (online) algorithms are needed.
NOTE: when query results are small enough to fit in memory, it’s usually easier to use get_seurat() or get_single_cell_experiment().
Contents
- Incremental mean calculation
- Counting cells grouped by dataset and gene
Incremental mean calculation
Many statistics, such as marginal means, are easy to calculate incrementally. Let’s begin with a query on the X$raw sparse matrix of unnormalized read counts, which will return results in shards incrementally accumulate the read count for each gene, then divide by the cell count to get the mean reads per cell for each gene.
First define a query - in this case a slice over the obs axis for cells with a specific tissue & sex value, and all genes on the var axis. The query$X() method returns an iterator of results, each as an Arrow Table. Each table will contain the sparse X data and obs/var coordinates, using standard SOMA names:
-
soma_data- the X values (float32) -
soma_dim_0- the obs coordinate (int64) -
soma_dim_1- the var coordinate (int64)
Important: the X matrices are joined to var/obs axis DataFrames by an integer join “id” (aka soma_joinid). They are NOT positionally indexed, and any given cell or gene may have a soma_joinid of any value (e.g., a large integer). In other words, for any given X value, the soma_dim_0 corresponds to the soma_joinid in the obs dataframe, and the soma_dim_1 coordinate corresponds to the soma_joinid in the var dataframe.
For convenience, the query class includes a utility to simplify operations on query slices. query$indexer is an indexer used to wrap the output of query$X(), converting from soma_joinids to positional indexing in the query results. Positions are [0, N), where N are the number of results on the query for any given axis.
Key points:
- it is expensive to query and read the results - so rather than make multiple passes over the data, read it once and perform multiple computations.
- by default, data in the census is indexed by
soma_joinidand not positionally.
library("tiledbsoma")
library("cellxgene.census")
census <- open_soma()
query <- census$get("census_data")$get("mus_musculus")$axis_query(
measurement_name = "RNA",
obs_query = SOMAAxisQuery$new(value_filter = "tissue=='brain' && sex=='male'")
)
genes_df <- query$var(column_names = c("feature_id", "feature_name"))$concat()
genes_df <- as.data.frame(genes_df)
n_genes <- nrow(genes_df)
# accumulator vector (for each gene) for the total count over all cells in X("raw")
raw_sum_by_gene <- numeric(n_genes)
names(raw_sum_by_gene) <- genes_df$feature_id
# iterate through in-memory shards of query results
tables <- query$X("raw")$tables()
while (!tables$read_complete()) {
table_part <- tables$read_next()
# table_part is an Arrow table with the columns mentioned above. The result
# order is not guaranteed!
# table_part$soma_dim_1 is the var/gene soma_joinid. But note that these are
# arbitrary int64 id's, and moreover each table_part may exhibit only a subset
# of the values we'll see over all query results. query$indexer helps us map
# any given soma_dim_1 values onto positions in query$var() (genes_df), that is
# the union of all values we'll see.
gene_indexes <- query$indexer$by_var(table_part$soma_dim_1)$as_vector()
stopifnot(sum(gene_indexes >= n_genes) == 0)
# sum(table_part) group by gene, yielding a numeric vector with the gene_index
# in its names
sum_part <- tapply(as.vector(table_part$soma_data), gene_indexes, sum)
# update the accumulator vector
which_genes <- as.integer(names(sum_part)) + 1 # nb: gene_indexes is zero-based
stopifnot(sum(which_genes > n_genes) == 0)
raw_sum_by_gene[which_genes] <- raw_sum_by_gene[which_genes] + sum_part
}
# Divide each sum by cell count to get mean reads per cell (for each gene),
# implicitly averaging in all zero entries in X even though they weren't included
# in the sparse query results.
genes_df$raw_mean <- raw_sum_by_gene / query$n_obs
genes_df
#> feature_id feature_name raw_mean
#> 1 ENSMUSG00000051951 Xkr4 1.397121e+00
#> 2 ENSMUSG00000025900 Rp1 3.162902e-01
#> 3 ENSMUSG00000025902 Sox17 6.604085e+01
#> 4 ENSMUSG00000033845 Mrpl15 3.939172e+01
#> 5 ENSMUSG00000025903 Lypla1 1.986548e+01
#> 6 ENSMUSG00000033813 Tcea1 4.305924e+01
#> 7 ENSMUSG00000002459 Rgs20 3.496194e+00
#> 8 ENSMUSG00000033793 Atp6v1h 7.470932e+01
#> 9 ENSMUSG00000025905 Oprk1 4.568752e-01
#> 10 ENSMUSG00000033774 Npbwr1 1.241003e-04
#> 11 ENSMUSG00000025907 Rb1cc1 3.631679e+01
#> 12 ENSMUSG00000033740 St18 1.660110e+01
#> 13 ENSMUSG00000051285 Pcmtd1 5.410501e+01
#> 14 ENSMUSG00000025909 Sntg1 1.178725e+00
#> 15 ENSMUSG00000061024 Rrs1 2.098927e+01
#> 16 ENSMUSG00000025911 Adhfe1 1.266112e+01
#> 17 ENSMUSG00000079671 2610203C22Rik 9.474621e+00
#> 18 ENSMUSG00000025912 Mybl1 2.643129e-01
#> 19 ENSMUSG00000045210 Vcpip1 3.456668e+01
#> 20 ENSMUSG00000097893 1700034P13Rik 5.721023e-01
#> 21 ENSMUSG00000025915 Sgk3 2.012592e+01
#> 22 ENSMUSG00000098234 Snhg6 6.784314e+00
#> 23 ENSMUSG00000025916 Ppp1r42 2.585422e-01
#> 24 ENSMUSG00000025917 Cops5 7.909310e+01
#> 25 ENSMUSG00000056763 Cspp1 1.635604e+01
#> 26 ENSMUSG00000067851 Arfgef1 1.582897e+01
#> 27 ENSMUSG00000042501 Cpa6 1.880119e-02
#> 28 ENSMUSG00000048960 Prex2 2.283623e+01
#> 29 ENSMUSG00000057715 A830018L16Rik 9.992140e-01
#> 30 ENSMUSG00000016918 Sulf1 5.567469e+00
#> 31 ENSMUSG00000025938 Slco5a1 2.452015e-01
#> 32 ENSMUSG00000042414 Prdm14 6.142964e-03
#> 33 ENSMUSG00000005886 Ncoa2 1.707928e+01
#> [ reached 'max' / getOption("max.print") -- omitted 52384 rows ]Counting cells grouped by dataset and gene
The goal of this example is to count the number of cells with nonzero reads, grouped by gene and Census dataset_id. The result is a data frame with dataset, gene, and the number of cells with nonzero reads for that dataset and gene.
For this multi-factor aggregation, we’ll take advantage of dplyr routines instead of the lower-level vector indexer shown above. And for presentation purposes, we’ll limit the query to four genes, but this can be expanded to all genes easily.
library("dplyr")
query <- census$get("census_data")$get("mus_musculus")$axis_query(
measurement_name = "RNA",
obs_query = SOMAAxisQuery$new(value_filter = "tissue=='brain'"),
var_query = SOMAAxisQuery$new(value_filter = "feature_name %in% c('Malat1', 'Ptprd', 'Dlg2', 'Pcdh9')")
)
obs_tbl <- query$obs(column_names = c("soma_joinid", "dataset_id"))$concat()
obs_df <- data.frame(
# materialize soma_joinid as character to avoid overflowing R 32-bit integer
cell_id = as.character(obs_tbl$soma_joinid),
dataset_id = obs_tbl$dataset_id$as_vector()
)
var_tbl <- query$var(column_names = c("soma_joinid", "feature_name"))$concat()
var_df <- data.frame(
gene_id = as.character(var_tbl$soma_joinid),
feature_name = var_tbl$feature_name$as_vector()
)
# accumulator for # cells by dataset & gene
n_cells_grouped <- data.frame(
"dataset_id" = character(0),
"gene_id" = character(0),
"n_cells" = numeric(0)
)
# iterate through in-memory shards of query results
tables <- query$X("raw")$tables()
while (!tables$read_complete()) {
table_part <- tables$read_next()
# prepare a (dataset,gene,1) tuple for each entry in table_part
n_cells_part <- data.frame(
"cell_id" = as.character(table_part$soma_dim_0),
"gene_id" = as.character(table_part$soma_dim_1),
"n_cells" = 1
)
n_cells_part <- left_join(n_cells_part, obs_df, by = "cell_id")
stopifnot(sum(is.null(n_cells_part$dataset_id)) == 0)
# fold those into n_cells_grouped
n_cells_grouped <- n_cells_part %>%
select(-cell_id) %>%
bind_rows(n_cells_grouped) %>%
group_by(dataset_id, gene_id) %>%
summarise(n_cells = sum(n_cells)) %>%
ungroup()
}
# add gene names for display
n_cells_grouped <- left_join(n_cells_grouped, var_df, by = "gene_id")
stopifnot(sum(is.null(n_cells_grouped$feature_name)) == 0)
n_cells_grouped[c("dataset_id", "feature_name", "n_cells")]
#> # A tibble: 21 x 3
#> dataset_id feature_name n_cells
#> <chr> <chr> <dbl>
#> 1 3bbb6cf9-72b9-41be-b568-656de6eb18b5 Ptprd 79578
#> 2 3bbb6cf9-72b9-41be-b568-656de6eb18b5 Dlg2 79513
#> 3 3bbb6cf9-72b9-41be-b568-656de6eb18b5 Pcdh9 79476
#> 4 3bbb6cf9-72b9-41be-b568-656de6eb18b5 Malat1 79667
#> 5 58b01044-c5e5-4b0f-8a2d-6ebf951e01ff Ptprd 474
#> 6 58b01044-c5e5-4b0f-8a2d-6ebf951e01ff Dlg2 81
#> 7 58b01044-c5e5-4b0f-8a2d-6ebf951e01ff Pcdh9 125
#> 8 58b01044-c5e5-4b0f-8a2d-6ebf951e01ff Malat1 12622
#> 9 66ff82b4-9380-469c-bc4b-cfa08eacd325 Dlg2 856
#> 10 66ff82b4-9380-469c-bc4b-cfa08eacd325 Pcdh9 2910
#> # i 11 more rowsDon’t forget to close the census.
census$close()