Influence of parameters on scRNA-seq dimension reduction methods.

This repository contains the code used in Tuning parameters of dimensionality reduction methods forsingle-cell RNA-seq analysis

The methods selected for DR are:

• scran
• seurat
• ZinbWAVE
• DCA
• scVI

Data used

The data that we used are the Zhengmix4eq, Zhengmix4uneq, and Zhengmix8eq from Duo et al.. sc_10x, sc_10x_5cl, sc_celseq2, and sc_celseq2_5cl from Tian et al. And a mixture of the 'FACS' (smart-SEQ2) data from TabulaMuris containing all the Brain_Myeloid, Large_Intestine, Skin and Spleen cells.

Finally we also created Zhengmix5eq and Zhengmix8uneq with the cell_mixer software with the following settings:

Rsript cell_mixer.R \
--data_path=$DATA_PATH \
--name=Zhengmix5eq \
--format=SingleCellExperiment \
--seed=1234 \
--qc_count_mad_lower=3 \
--qc_feature_count_mad_lower=3 \
--qc_mito_mad_upper=3 \
--naive_cytotoxic=1000 \
--regulatory_t=1000 \
--cd4_t_helper=1000 \
--memory_t=1000 \
--naive_t=1000

Rsript cell_mixer.R \
--data_path=$DATA_PATH \
--name=Zhengmix8uneq \
--format=SingleCellExperiment \
--seed=1234 \
--qc_count_mad_lower=3 \
--qc_feature_count_mad_lower=3 \
--qc_mito_mad_upper=3 \
--b_cells=500 \
--naive_cytotoxic=250 \
--cd14_monocytes=1000 \
--regulatory_t=1500 \
--cd4_t_helper==500 \
--cd56_nk=250 \
--memory_t=1000 \
--naive_t==1500

Note that you can create Zhengmix4eq, Zhengmix4uneq and Zhengmix8eq (with the mitochondrial preprocessing) with the following commands:

Rsript cell_mixer.R \
--data_path=$DATA_PATH \
--name=Zhengmix4eq \
--format=SingleCellExperiment \
--seed=1234 \
--qc_count_mad_lower=3 \
--qc_feature_count_mad_lower=3 \
--qc_mito_mad_upper=3 \
--b_cells=1000 \
--naive_cytotoxic=1000 \
--cd14_monocytes=1000 \
--regulatory_t=1000

Rsript cell_mixer.R \
--data_path=$DATA_PATH \
--name=Zhengmix4uneq \
--format=SingleCellExperiment \
--seed=1234 \
--qc_count_mad_lower=3 \
--qc_feature_count_mad_lower=3 \
--qc_mito_mad_upper=3 \
--b_cells=1000 \
--naive_cytotoxic=500 \
--cd14_monocytes=2000 \
--regulatory_t=3000

Rsript cell_mixer.R \
--data_path=$DATA_PATH \
--name=Zhengmix8eq \
--format=SingleCellExperiment \
--seed=1234 \
--qc_count_mad_lower=3 \
--qc_feature_count_mad_lower=3 \
--qc_mito_mad_upper=3 \
--b_cells=500 \
--naive_cytotoxic=400 \
--cd14_monocytes=600 \
--regulatory_t=500 \
--cd4_t_helper=400 \
--cd56_nk=600 \
--memory_t=500 \
--naive_t=500

The cell lines and TabulaMuris dataset are created in the Generate Cell Lines and Generate Tabula Muris R Jupyter notebooks.

Launching the scripts

All the scripts can be run locally (examples in the run.sh), however if you want to do the full benchmark you will nedd a distribution system as it would take multiple years.

For the R methods, since they don't need GPUs we launched them on google cloud default machines using dsub which is a similar to qsub. We provide generate_dsub_conf.py which will generate the task files for the different methods and store both the metrics and the embedding on GCS. Be warned that launching them will cost you >10k$. These script will generate one CSV and one Loompy file per configuration.

For the python methods we used another distribution infrastructure. Since the number of parameter combinations was very high, the launch scripts evaluate their own grid of parameters and takes ~3 days on a Tesla P100. These scripts do not save the embedding (by default) as there are simply too many of them and will write the metrics of multiple runs on a single CSV. Note that the scripts for scVI and DCA can be interrupted, and will restart where they left off so that it can be launched on a shared GPU cluster with preemption.

Data availability

The evaluation for all the runs as well as 5000 embeddings can be found in this Google Drive folder

The files named df_scran.csv, df_seurat.csv, df_zinbwave.csv, df_dca.csv, and df_scvi.csv contain one row per configuration that we ran successfully.

The files named DATASET.METHOD.h5ad are encoded with anndata v0.7.0 (be careful as they are not readable with previous versions) and contain 100 embeddings each. The embeddings are in the obsm attribute of the object. All the embeddings can be listed with the obsm_keys() method. The name of the embedding contains the parameters used to generate that embedding and are written like that method=zinbwave.dims=10.epsilon=1000.features=300.gene_covariate=0.