Read anndata dataframes with pure R

read_ad_df.R

library(rhdf5)
library(tidyverse)

read_ad_df <- function (file, name) {
    x_attrs <- h5readAttributes(file, name)
    
    ## check requested entry is a dataframe
    ## TODO: do we need to check encoding-version?
    stopifnot(x_attrs[['encoding-type']] == "dataframe")
    
    ## rownames and columns in order
    idx_cols <- unlist(x_attrs[c("_index", "column-order")], use.names=FALSE)
    
    ## load the factor levels
    x_levels <- h5read(file, str_c(name, "/__categories"))
    
    ## load dataframe
    h5read(file, name)[idx_cols] %>% as_tibble() %>%
        ## replace categorical columns with proper factors
        mutate(across(any_of(names(x_levels)), ~ factor(x_levels[[cur_column()]][.x+1L])))
}

This can be used on h5ad files from anndata to load what were originially Pandas DataFrame objects. For example, for single-cell data the equivalent of colData() on a SingleCellExperiment object would be found in /obs in the HDF5 object:

df_colData <- read_ad_df("scrna.h5ad", "/obs")

and rowData is found at /var

df_rowData <- read_ad_df("scrna.h5ad", "/var")

---

Note on Motivation

There is an anndata R package, so why not use that? Because anndata is a wrapper for calling the Python anndata package, rather than a pure R package. For a well-defined specification, it seems unnecessary to load an entire Python subprocess for what can be done directly in R.