1 of 13

Persist-seq: a set of reproducible single cell RNA-seq analysis pipelines for understanding early persister cells in cancer.

Pablo Moreno, Anil S Thanki, Anca Farcas, Martin Miller, Ultan McDermott

2 of 13

2

PERSIST-SEQ consortium: understanding persister cells with an end-to-end Single Cell RNA-seq pipeline.

“Persisters”

Colon, lung and breast cancer
5 million single cells / 5 years
16 European consortium partners
New targets for drug combinations

Resistance

Industry partners

* Consortium co-leads

Academic partners

SME partners

* Unified sequencing

*

* Standardised analysis

*

scRNA-seq pipeline, based on battle tested EBI scRNA-seq pipeline

Processed so far 14 pilots and 16 real experiments

3 of 13

Persist-seq End-to-End scRNA-seq data analysis operational and automated

3

SC Discoveries

Count matrices

Count matrix available

Interactive view of results

Download and decrypt

Main results object available in Data Library

Run standard pipeline

- Matrix to clusters, dimreds and packaged results.

- Battle tested (>350 datasets).

- galaxy-workflow-executor package

- Using Shared Data Library in Galaxy

Cellxgene
UCSC CellBrowser
TODO:

Jupyter notebooks

4 of 13

What are persister cells and what is the context for this work

Drug-tolerant persister cells (DTPs) contribute to therapy failure and cancer relapse ("the deadly survivors")

4

The pathways involved in DTP survival are poorly understood but the mechanisms are primarily non-genetic

Single cell RNA sequencing has become the best way to characterise mechanism(s) which allow DTPs to survive

5 of 13

Persister datasets sequenced/in the pipeline

5

Cancer cell lines

3D organoids

Mouse xenografts

Cell line (CDX)
Patient-derived (PDX)

Patients

PC9
HCC827
II18
NCIH3255

HUB-07-B2-051 (L858R)
Tempus AZ-574812 (L858R)

PC9-CDX

Co-culture

CAFs

GEMM

EGFRm/osimertinib

Model

dose clinically equivalent

(2 weeks treatment, dosed 2x week)

DTPs

washout

Because we want combination targets for treatment in patients, we want to see the phenomena at multiple different levels, to avoid pursuing things that show up only on a single biological model.
By biological model here we mean cell lines, organoids, etc all the way to patient samples.
Here we show data from 7 different models that are all EGFR mutants and are treated with Osimertinib (and EGFR kinase inhibitor usually given to patients with mutations on EGFR)

The experimental design, shown in pink, is usually that you start with the model cells, also referred to as parental, which are sequenced before treatment, then these are subject to clinically relevant doses of the drug (osimertinib) for 2 weeks and then surviving cells sequenced (once the reversible mechanism is confirmed). In some cases, we also sequence after drug wash outs.

6 of 13

Persist-seq main pipeline includes best practices and produces an object that can be explored for detailed analysis

Ingest count matrix

Filtering & QC

Normalization, scaling, HVGs

PCA, k-nn graph for UMAP

tSNE & clustering

Marker genes for clusters and metadata fields

Object merging final AnnData

Optional batch correction

Mark cell cycle phase

Pseudo-bulk

pipeline

7 of 13

Persist-seq pseudo-bulk RNA-seq for DE calling fights p-value over-inflation and enables complex modelling

Ingest main pipeline results

Aggregate cells per groups definitions - Decoupler

EdgeR filtering + DE

Volcano for each contrast

GSEA with multiple collections, per contrast -

Fgsea

OncoEnrichR for Top DE, per contrast

Sanitisation for DE

8 of 13

Automatic execution

8

parameters

workflow

inputs.yaml

allowed_errors.yaml

Credentials

Upload datasets
Upload workflow
Run & wait for result.
Allow errors on specific steps.
Retrieve results.

Galaxy-workflow-executor (CLI)

bioblend

Kubernetes on OpenStack

Kubernetes on other clouds

Dedicated instance on

AZ Slurm Cluster

results

https://github.com/ebi-gene-expression-group/galaxy-workflow-executor

pip install galaxy-workflow-executor

conda install galaxy-workflow-executor

(Bioconda channels set beforehand)

9 of 13

9

Osimertinib-generated persisters are transcriptionally distinct from untreated cells

HCC827 DMSO (2,411 cells)

HCC827 DTP (2,014 cells)

II18 DMSO (4,787 cells)

II18 DTP (3,771 cells)

NCIH3255 DMSO (3,470 cells)

NCIH3255 DTP (2,196 cells)

PC9 DMSO (12,241 cells)

PC9 DTP (10,181 cells)

UMAP1

UMAP2

(n=25 neighbours)

HUB 07-B2-051 DMSO (5,584 cells)

HUB 07-B2-051 OSI 160nM (6,631 cells)

HUB 07-B2-051 WASHOUT (7,240 cells)

TEMPUS AZ-574812 DMSO (5,794 cells)

TEMPUS AZ-574812 OSI 160nM (3,801 cells)

TEMPUS AZ-574812 WASHOUT (5,182 cells)

DMSO

DTP

WASHOUT

HUB

TEMPUS

DMSO

WASHOUT

DTP

Vehicle (7,271 cells)

2 days treatment (2,840 cells)

7 days treatment (2,552 cells)

10 days treatment (1,335 cells)

14 days treatment (8,597 cells)

Vehicle

Day 2

Day 7

Day 10

Day 14

Cell lines

Organoids

In vivo: PC9-CDX

PC9 DTP

PC9 DMSO

II18 DMSO

II18 DTP

HCC827 DTP

HCC827 DMSO

NCIH3255 DTP

NCIH3255 DMSO

DUSP6

Evidence of target engagement

First, explain with the first example that this is a UMAP plot, where each point is a cell, and where cells that are transcriptionally more similar tend to be closer on the map. Then points are coloured based on metadata (cell line and persister/parental state in this case).

Here in cell line DMSO and DTP cells are clustered separately showing that DTPs (or persisters) are distinct transcriptionally from their parental cells (growing on DMSO).

In organoids, again we see separate parental and persister cells, and how the washout (no drug treatment after the two weeks of treatment) cells revert partly to a behaviour similar to parental cells

PC9 CDX: In vivo vehicles are the parental cells, day 2, 7, 10 and 14 are persister cells at those points of treatment.

DUSP is confirmed to be down regulated in the presence of treatment (Osimertinib, this is a confirmation of something known)

10 of 13

Models activate different biological pathways: are all persisters created equal?

10

metabolic and signaliing pathways

Metabolic and signaling pathways relevant for cell growth are down regulated

Certain reversions, for instance Epithelial mesenchymal transitions is activated in-vitro in some cell lines, but partly deactivated in organoids and in-vivo.

Upregulation of inflammation related pathways (interferons) in many model

11 of 13

Persister cells consistently reduce cell cycling (G1 arrest); a cycling population remains

11

S

G2M

G1

Score every cell for relevant G2M and S genes

Cell population proportions plot

On the left we see the UMAP for cell lines again only, and in each case when moving from parentals to DTPs/persisters, you can see that a larger proportion of cells are in G1 arrest.

G1 arrest stop cycling

CLICK This can be seen better on a cell population proportions plot, where on the first plot you can see that for each pair of columns (each part for a cell line), on every transition from DMSO to DTP you can see a much larger blue (G1) bar.

CLICK the same is visible on the cell organoids (combined), and here you can see as well a reversion of that behaviour when going back to washout (drug free).

CLICK the same is visible in the in-vivo experiment (cell allograft in mice), where moving from vehicle to any of the treatment days you see an increase in G1 arrested cells.

Right G1 arrest cells Continue increasing - I wouldn’t say this this is not significant probably.

12 of 13

Summary

scRNA-seq pipelines operational, all datasets analyzed and shared (~1 million cells so far)

Persister cells arrested in G1, high heterogeneity.

Programmatic access and UIs (Data Libraries) help to cater for consortium needs.

Infrastructure is modular and re-deployable.

Two large workflows, new Galaxy tools and tool updates contributed.

12

13 of 13

Acknowledgement

13