• Reproducing results in the "Harmonization and Annotation of Single-cell Transcriptomics data with Deep Generative Models" paper
• Demonstration of how to use scVI and scANVI for the harmonization and annotation problem

Contact

chenlingantelope [at] berkeley [dot] edu

Datasets

Analysis	Associated Script	Datasets	Technology	Number of Cells	Ref.
Figure 2: Benchmark	PBMC8KCITE.py	PBMC-8K; PBMC-CITE	10x	8,381; 7,667	10x DatasetsStoeckius, Marlon, et al. 2017
Supplementary Figure 2: UMAP Visualization	PBMC8KCITE.py	PBMC-8K; PBMC-CITE	10x	8,381; 7,667	10x Datasets; Stoeckius, Marlon, et al. 2017
Figure 2: Benchmark	MarrowTM.py Tech1.pretty.ipynb	MarrowTM-10x; MarrowTM-ss2	10x; SmartSeq2	4,112;5,351	Quake, Stephen R., et al. 2018
Supplementary Figure 1: Robustness Analysis for Hyperparameter Choice	Robustness_study.ipynb	MarrowTM-10x; MarrowTM-ss2	10x; SmartSeq2	4,112;5,351	Quake, Stephen R., et al. 2018
Supplementary Figure 3: UMAP Visualization	MarrowTM.py	MarrowTM-10x; MarrowTM-ss2	10x; SmartSeq2	4,112;5,351
Figure 2: Benchmark	Pancreas.py	Pancreas-InDrop; Pancreas-CEL-Seq2	inDrop; CEL-Seq2	8,569; 2,449	Baron, Maayan, et al. 2016; Muraro, Mauro J., et al. 2016
Supplementary Figure 4: UMAP Visualization	Pancreas.py	Pancreas-InDrop; Pancreas-CEL-Seq2	inDrop; CEL-Seq2	8,569; 2,449	Baron, Maayan, et al. 2016; Muraro, Mauro J., et al. 2016
Figure 2: Benchmark	DentateGyrus.py	DentateGyrus-10x; DentateGyrus-C1	10x; Fluidigm C1	5,454; 2,303	Hochgerner, Hannah, et al. 2018
Supplementary Figure 5: UMAP Visualization	DentateGyrus.py	DentateGyrus-10x; DentateGyrus-C1	10x; Fluidigm C1	5,454; 2,303	Hochgerner, Hannah, et al. 2018
Figure 3: Robustness Analysis by subsampling cells Supplementary Figure 10	NoOverlapSCANVI.py PopRemoveSCANVI.py SCANVI_posterior-NoOverlap.ipynb SCANVI_posterior_poprm.ipynb	PBMC-8K; PBMC-CITE	10x	8,381; 7,667	10x Datasets; Stoeckius, Marlon, et al. 2017
Figure 4: Continuous Trajectory Supplementary Supplementary Figure 6: UMAP	continuous.ipynb	HEMATO-Tusi; HEMATO-Paul	inDrop; MARS-seq	4,016 ; 2,730	Tusi, Betsabeh Khoramian, et al. 2018; Paul, Franziska, et al. 2015
Figure 5: External Validation by Experimentally Derived Labels, Supplementary Figure 11	harmonization-CitePure-SCANVI.ipynb	PBMC-68K; PBMC-Sorted; PBMC-CITE	10x	68,579; 94,655; 7,667	Zheng, Grace XY, et al. 2017; Stoeckius, Marlon, et al. 2017
Figure 6: Semi-Supervised Annotation of T Cell Subtypes, Supplementary Figure 12	SCANVI-mild-annot-Clustering.ipynb	PBMC-Sorted T cell Subtypes	10x	42919	Zheng, Grace XY, et al. 2017; Stoeckius, Marlon, et al. 2017
Hierarchical Semi-Supervised Annotation	Hierarchical.ipynb	CORTEX	10x	160,796	Zeisel, Amit, et al. "Molecular architecture of the mouse nervous system." bioRxiv (2018): 294918.
Supplementary Figure 7: Scalability Analysis	scanorama.ipynb	SCANORAMA	Mixed	105,476	Hie, Brian L., Bryan Bryson, and Bonnie Berger. "Panoramic stitching of heterogeneous single-cell transcriptomic data." bioRxiv (2018): 371179.
Supplementary Figure 13: Differential Expression	DE-final.ipynb	PBMC-8K; PBMC-68K	10x	8,381; 68,579	10x Datasets; Zheng, Grace XY, et al. 2017

• Supplemtary Figure 2,3,4,5,8,9 are generated using scripts in Additional_Scripts/ using output from the analysis python scripts including scanvi_acc.R, KNNcurves.py and BE_curves.py.
• Boxplots for Figure 3 are generated using poprm_boxplot.R in Additional_Scripts/
• The Additional_Scripts also contains code for running Seurat directly from commandline runSeurat.R and SeuratPCA.R.
• All .gmt files in Additional_Scripts/ are gene signatures.

Installation

• Clone the github repository, install the dependencies and call functions from the modules scVI
• Install time (< 10 min)

Requirements

• Pytorch V0.4.1
• Python 3
• scikit-learn V0.19.1

Instructions

• To reproduce results from the paper, look up the relevant datasets, python notebooks (located in notebooks/), or python scripts (located in the root directory).
• Download the relevant datasets except for the ones already wrapped for the scVI package (PBMC-8K, PBMC-CITE, PBMC-68K, PBMC-Sorted, MarrowTM-10x, MarrowTM-ss2 can be loaded directly with the dataloader functions)
• Annotation files generated by us when the original study did not provide annotation (cite.seurat.labels) can be found in the scvi-data repository
• Run the analysis and results should match those of the paper.
• This repository contains functions written uniquely to produce some of the analysis in this paper. For more up-to-date package refer to main scVI repository