Fast, sensitive and accurate integration of single-cell data with Harmony
Check out the manuscript in Nature Methods:
For Python users, check out the harmonypy package by Kamil Slowikowski.
Harmony has been tested on R versions >= 3.4. Please consult the DESCRIPTION file for more details on required R packages. Harmony has been tested on Linux, OS X, and Windows platforms.
To run Harmony, open R and install harmony from CRAN:
install.packages("harmony")If you'd like the latest development version, install from this github directly:
devtools::install_github("immunogenomics/harmony", build_vignettes=TRUE)Harmony is designed to be user-friendly and supports some SingleCellExperiment and Seurat R analysis pipelines. Alternatively, it can be used in standalone mode.
Check out this vignette for a quick start tutorial which demonstrates the usage of the tool in standalone mode.
At minimum the following parameters need to be specified to achieve an integration.
library(harmony)
my_harmony_embeddings <- RunHarmony(my_pca_embeddings, meta_data, "dataset")By default, the harmony API works on Seurats PCA cell embeddings and corrects them. You can run Harmony within your Seurat workflow with RunHarmony(). Prior RunHarmony() the PCA cell embeddings need to be precomputed through Seurat's API. For downstream analyses, use the harmony embeddings instead of pca.
For example, the following snippet run Harmony and then calculates UMAP of the corrected input embeddings:
seuratObj <- RunHarmony(seuratObj, "dataset")
seuratObj <- RunUMAP(seuratObj, reduction = "harmony")For a more detailed overview of the RunHarmony() Seurat interface check, the Seurat vignette
Harmony can integrate over multiple covariates. To do this, specify a vector covariates to integrate.
my_harmony_embeddings <- RunHarmony(
my_pca_embeddings, meta_data, c("dataset", "donor", "batch_id")
)Do the same with your Seurat object:
seuratObject <- RunHarmony(seuratObject, c("dataset", "donor", "batch_id"))The examples above all return integrated PCA embeddings. We created a detailed walkthrough that explores the internal data structures and mechanics of the Harmony algorithm.
R distributions can be bundled with different scientific computing libraries. This can drastically impact harmony's performance. Rstudio comes by default with BLAS. In contrast, conda distributions of R are bundled with OPENBLAS. Overall, our benchmarks show that harmony+OPENBLAS is substantially faster compared harmony+BLAS. Therefore users with large datasets will benefit using OPENBLAS.
One caveat is that OPENBLAS uses OPENMP to parallelize operations. By default, OPENBLAS will utilize all cores for these operations. While in theory this accelerates runtimes, in practice harmony is not optimized for multi-threaded performance and the unoptimized parallelization granularity may result in significantly slower run times and inefficient resource utilization (wasted CPU cycles). Therefore, by default harmony turns off multi-threading. However, very large datasets >1M may benefit from parallelization. This behavior can be controlled by the ncores parameter which expects a number threads which harmony will use for its math operation. Users are advised to increase gradually ncores and assess potential performance benefits.
Code to reproduce Harmony results from the Korsunsky et al 2019 manuscript will be made available on github.com/immunogenomics/harmony2019.