1 of 12

Downstream Effects of Pipeline Modifications on scRNA-Seq Data

Presentation by: Taylor Gaito

Mentored by: Uma Chandran, PhD.

Special thanks to: Alex Chang and William Schwarzmann

2 of 12

Single Cell RNA Sequencing (scRNA-Seq) measures the RNA of individual cells

Source: Yijyechern. “RNA-Seq.” RNA-Seq, en.wikipedia.org/wiki/RNA-Seq.

DNA

mRNA

Protein

3 of 12

scRNA-Seq preserves the heterogeneity of cells, unlike prior methods of bulk analysis

Source: 10x Genomics Website

4 of 12

Companies like 10x provide platforms for the complex sample preparation

Source: 10x Genomics Website

This is the 10x Genomics specific pipeline. Use microfluidics to isolate individual cells, and pair them with a gel bead:

contain millions of oligo-primers help with cDNA synthesis but also provide markers for beads and UMIs) and an oil droplet so that they separate (each Gel bead contains the same 10x Barcode but different UMIs)
Droplets formed during partitioning are referred to as GEMs

As the gel dissolves, primers are released from the gel beads and hybridize to the polyA tail mRNAs released from the lysed cell; then using a reverse transcription enzyme they create cDNA
the cDNA is purified from oil and then amplified using PCR (polymerase chain reaction)
the library is prepared through a series of quality control (10x barcode tells which cell the transcript is from and the UMI tells how many copies of the transcript was present in the cell)
the library is put through a sequencer (Illumina sequencer in the case of 10x) which gives the nucleotide sequence of fragments

5 of 12

Cell Ranger provides quality control and analysis of data after sequencing

Source: 10x Genomics Website

Focus on normalization algorithm and clustering visualization (resolution)

The 10x platform employs the Cell Ranger pipeline, which performs barcode and UMI multiplexing, alignment to a reference, and quality control to ultimately identify the different cell types within the population using known genetic markers; however if you didn’t know what cell types to expect, the problem of determining them becomes exponentially more difficult
Before clustering, however, the data must be filtered - many factors can influence the single cell experiment including not enough cells, clumped or dead cells, or multiple cells within a suspension - all of which manifest in the html output of Cell Ranger, which is easily readable
Cell Ranger is just one stepping stone in the process of analysis, and if the outputs of CellRanger provide adequate depth and purity, then the outgoing matrix can be used in a pipeline for further data filtering and analysis

6 of 12

Data analysis required to differentiate cells is lengthy and multi-stepped

Source: Satija Lab

B cells

One common downstream analysis package, or set of tools for data analysis of Cell Ranger Outputs, is Seurat which is used within the R programming language
Seurat uses a variety of quality control and normalization algorithms along with principal component analysis

Within Seurat alone, you’re able to alter the t-SNE dimensions to change the shape of clusters, resolution which alters the number of clusters, thresholds and numerous others
UMap is specifically the algorithm to produce the t-SNE chart; Seurat ultimately provides a visualization of the cell clusters and a differential list from which scientists are able to generate violin plots of what genes are expressed in what clusters

Due to the multistep and lengthy nature of the process, the results are very dependent on the algorithms, filters, platform used, complexity of tissues, and number of cells

EX Ms4a1 is highly expressed in B cells, so cluster 3 (which is green) is labelled B cells

7 of 12

Our project focuses on varying the pipeline and noting its downstream effects

Seurat

LogNormalize

(Seurat)

Scran

Scater

Seurat

Variable Features

Resolution

Dimensions

Normalization

Visualization

Cell Ranger

8 of 12

Our goal was to see how changing parameters can subtly affect the clusters of cell types

PBMCs

Satija Lab list of biomarkers
10x Chromium 1000x PBMC

Source: Miltenyi Biotec

9 of 12

Changing the resolution changed the appearance, proximity, and shape of clusters

Scater

Scran

Raw Seurat (0.5 Resolution)

0.5 Seurat Resolution

(10 Clusters)

1.0 Seurat Resolution

(12 Clusters)

Changing the normalization changed the appearance, proximity, and shape of clusters

look very similar so that it is likely that the major cell types are picked up all methods

Loupe which is provided by Cell Ranger is actually the most different

The bottom three are increasing Seurat resolution; we are going to focus on the resolutions of these two Seurat plots (.5 and 1.0)
we don’t know exactly what the resolution parameter does other than increase or decrease number of subclusters within a cluster. The number of clusters is determined by the dimensions chosen

Cluster 4, and 8 are being subclustered into 6, 8 and 9 when Seurat resolution is increased

What is considered a CD8 region changes to NK regions

Overlay of known markers shows cluster 4 as CD8 but overlay of known markers in Seurat 1.0 show it as NK
So, question is why do 2 clusters go to 3 and is the 3rd cluster Natural Killer cells or CD8+ cells since the identity changes between the 2 resolutions?

When you label, the clusters look the same because using biomarkers to define the clusters condenses the multiple clusters into one

although using known markers is extremely useful and a great way to get started identifying the subtypes, but further analysis will need to be done to understand the subclusters within the B cell and monocyte region which are present in both Seurat 0.5 and Seurat 1.0

So, we turned to the violin plots of known markers to understand whether these regions are CD8 T or NK cells

10 of 12

Biomarkers can consolidate potentially unique subclusters

Seurat 0.5

Seurat 1.0

4

8

6

8

9

CD8+ T

NK

CD8+T

Seurat 1.0

Violin Plots

So as I mentioned previously, violin plots are one way to check for the promise of clusters; here is an example of one using the B cells which differentiate, not in resolution but rather consolidation when biomarkers are added

Both .5 and 1.0 are identical, having 2 clusters but upon the addition of a biomarker they consolidate
As you can see, despite being unique in gene expression, clusters 3 and 7 end up consolidating - however they might be different subtypes
Additionally, the reverse of this can happen as well - where some of the same cells may be expressing different genes and they might inaccurately display 2 clusters

new methods of testing the validity of clusters are being developed, here is an example of an output of the statistical testing of the DE genes of a pbmc that seurat found - really only the S1 gene is truly correct in both
Another method is using ARI (The Hubert-Arabje Adjusted Index) to compare partitions (basically within potential clusters) the closer you are to one the better the agreement between them

11 of 12

Conclusions and Future Developments

Sampling and data pipelines for scRNA-seq are complex
It is difficult to mix and match pipeline components
Clustering results are dependent on inputs given by the user
Results aren’t definitive and need new methods of fidelity assessment

12 of 12

References and Acknowledgements

Unending thanks Dr. Uma Chandran, Alex Chang, Will Schwarzmann, and the rest of the lab team who were so thorough, caring, knowledgeable and kind as I learned from and worked with them. Special thanks to Dr. Boone, the Hillman Academy, University of Pittsburgh, and Hillman Cancer Center for this inspiring and unforgettable opportunity and Solomon Livshits for coordinating. Thank you to the CoSBBI gang who made this summer unforgettable!

Zheng, Grace et al. “Massively Parallel Digital Transcriptional Profiling of Single Cells.” Nature News, Nature Publishing Group, 16 Jan. 2017,
Zhang, Jesse M., et al. “Valid Post-Clustering Differential Analysis for Single-Cell RNA-Seq.” Bio Rxiv, 26 June 2019, doi:10.1101/463265.
Daniszewski, Maciej, et al.. “Single Cell RNA Sequencing of Stem Cell-Derived Retinal Ganglion Cells.” Scientific Data, 13 Feb. 2018, doi:10.1101/191395.
Tian, Luyi, et al.. “Benchmarking Single Cell RNA-Sequencing Analysis Pipelines Using Mixture Control Experiments.” Nature Methods, vol. 16, no. 6, 27 May 2019, pp. 479–487., doi:10.1038/s41592-019-0425-8.