timothy-barry / import-frangieh-2021

Frangieh (2021) Data Documentation

Gene Katsevich; May 4, 2022

Overview

This repository contains code to import and process the Frangieh 2021 data. Frangieh et al developed the perturb-CITE-seq protocol, a new single cell CRISPR screen assay providing both gene and protein expression readouts. They applied perturb-CITE-seq to study 248 genes and 20 proteins involved in cancer immunotherapy resistance in a large screen with over 200,000 patient-derived melanoma cells in total. These cells represent three experimental conditions (control, treated with IFN-gamma, and co-cultured with tumor-infiltrating lymphocytes).

The frangieh-2021 directory structure is as follows:

├── processed
│   └── perturb-cite-seq
│       ├── gene
│       │   ├── gene_expression_matrix.odm
│       │   ├── gene_expression_metadata.rds 
│       │   ├── gene_expression_metadata_control.rds
│       │   ├── gene_expression_metadata_ifn-gamma.rds
│       │   └── gene_expression_metadata_co-culture.rds
│       ├── gRNA
│       |   ├── ...
|       └── protein
│           ├── ...
├── raw
│   ├── ...

The contents of the raw directory are suppressed, as they are unimportant. The processed directory contains the single subdirectory perturb-cite-seq, which in turn contains subdirectories for the three measured modalities: gene, gRNA, protein. Each of these directories has five files each: one ODM file and four metadata files. The ODM file contains the expression information for all cells and features present in the raw data deposited in the Single Cell Portal. There is one metadata file corresponding to this full set of cells and features. The three additional metadata files contain cells from each of the three condnitions, and which contain exactly one gRNA.

Experimental design

248 genes with putative roles in immunotherapy resistance were targeted by three gRNAs each, giving 744 targeting guide RNAs. There were additionally 74 negative control gRNAs, for a total of 818 gRNAs. The following table shows the breakdown of the cells into the three experimental conditions:

processed_dir <- sprintf("%s/processed", 
                         .get_config_path("LOCAL_FRANGIEH_2021_DATA_DIR"))
processed_gene_dir <- sprintf("%s/perturb-cite-seq/gene", processed_dir)
gene_odm_fp <- sprintf("%s/gene_expression_matrix.odm", processed_gene_dir)
gene_metadata_fp <- sprintf("%s/gene_expression_metadata.rds", processed_gene_dir)
gene_odm <- ondisc::read_odm(gene_odm_fp, gene_metadata_fp)
gene_odm |> ondisc::get_cell_covariates() |> dplyr::pull(condition) |> table()

## 
## Co-culture    Control       IFNγ 
##      73114      57627      87590

Interestingly, this study contains not only negative control gRNAs (i.e., negative control treatments) but also negative control proteins (i.e. negative control outcomes). Let’s take a look:

processed_protein_dir <- sprintf("%s/perturb-cite-seq/protein", processed_dir)
protein_odm_fp <- sprintf("%s/protein_expression_matrix.odm", processed_protein_dir)
protein_metadata_fp <- sprintf("%s/protein_expression_metadata.rds", processed_protein_dir)
protein_odm <- ondisc::read_odm(protein_odm_fp, protein_metadata_fp)

Note that there are 24 features. Let’s take a closer look:

protein_odm |> ondisc::get_feature_ids()

##  [1] "CD117"       "CD119"       "CD140a"      "CD140b"      "CD172a"     
##  [6] "CD184"       "CD202b"      "CD274"       "CD29"        "CD309"      
## [11] "CD44"        "CD47"        "CD49f"       "CD58"        "CD59"       
## [16] "CD61"        "HLA_A"       "Rat_IgG2a"   "Mouse_IgG1"  "Mouse_IgG2a"
## [21] "Mouse_IgG2b" "HLA_E"       "CD9"         "CD279"

There are 20 proteins of interest, along with the following four controls: Rat_IgG2a, Mouse_IgG1, Mouse_IgG2a Mouse_IgG2b. My understanding is that these are four extra antibodies targeting proteins that are not actually present in the population of (human) cells under investigation. So whatever fluctuations of these four protein expression are detected must be technical rather than biological. Theoretically, we may pair these protein expressions with any perturbations (not necessarily just negative control perturbations) to assess calibration. Note that the expression of each protein was normalized not against sequencing depth but against the expression of “its corresponding IgG control.” This implies that each of the 20 proteins of interest has associated with it one of the four controls, but I’m not sure what this association is.

Guide RNA assignments

Unlike the other datasets we are working with, this dataset does not come with raw gRNA expressions. Instead, it comes with binary gRNA assignments. Below is a histogram of the number of gRNAs per cell:

processed_gRNA_dir <- sprintf("%s/perturb-cite-seq/gRNA", processed_dir)
gRNA_odm_fp <- sprintf("%s/gRNA_assignments_ungrouped.odm", processed_gRNA_dir)
gRNA_metadata_fp <- sprintf("%s/gRNA_assignments_ungrouped_metadata.rds", processed_gRNA_dir)
gRNA_odm <- ondisc::read_odm(gRNA_odm_fp, gRNA_metadata_fp)
gRNA_odm |> 
  ondisc::get_cell_covariates() |>
  dplyr::filter(n_nonzero <= 5) |>
  ggplot2::ggplot(ggplot2::aes(x = n_nonzero, y = stat(count) / sum(count))) +
  ggplot2::geom_histogram(binwidth = 1, colour = "black") +
  ggplot2::scale_x_continuous(breaks = 0:5) +
  ggplot2::labs(x = "Number of gRNAs", y = "Frequency") +
  ggplot2::theme_bw() +
  ggplot2::theme(panel.grid.major.x = ggplot2::element_blank(),
                 panel.grid.minor.x = ggplot2::element_blank())

Note that only about 60% of cells received exactly one gRNA. About 10% of cells had no detected gRNAs, and about 30% of cells had more than one gRNA detected. It is unclear whether the association analysis in the paper was restricted to cells with exactly one gRNA, but we decided it’s safer to do so. Below is the number of cells in each experimental condition with exactly one gRNA:

dplyr::left_join(
  gene_odm |>
    ondisc::get_cell_covariates() |>
    tibble::rownames_to_column(var = "cell_barcode") |>
    dplyr::select(cell_barcode, condition),
  gRNA_odm |>
    ondisc::get_cell_covariates() |>
    tibble::rownames_to_column(var = "cell_barcode") |>
    dplyr::select(cell_barcode, n_nonzero),
  by = "cell_barcode"
) |>
  dplyr::filter(n_nonzero == 1) |>
  dplyr::count(condition)

##    condition     n
## 1 Co-culture 46427
## 2    Control 30486
## 3       IFNγ 50053

QC’d datasets

The QC’d datasets retain all of the original features, but only those cells that (a) have a given experimental condition and (b) have exactly one gRNA. Let’s take the control dataset as an example.

gRNA_metadata_control_fp <- sprintf("%s/gRNA_assignments_ungrouped_metadata_control.rds", processed_gRNA_dir)
gRNA_control_odm <- ondisc::read_odm(gRNA_odm_fp, gRNA_metadata_control_fp)
gRNA_control_odm

## A covariate_ondisc_matrix with the following components:
##  An ondisc_matrix with 818 features and 30486 cells.
##  A cell covariate matrix with columns n_nonzero, n_umis.
##  A feature covariate matrix with columns mean_expression, coef_of_variation, n_nonzero, target, target_type.

gene_metadata_control_fp <- sprintf("%s/gene_expression_metadata_control.rds", processed_gene_dir)
gene_control_odm <- ondisc::read_odm(gene_odm_fp, gene_metadata_control_fp)
gene_control_odm

## A covariate_ondisc_matrix with the following components:
##  An ondisc_matrix with 23712 features and 30486 cells.
##  A cell covariate matrix with columns n_nonzero, n_umis, condition, cluster_x, cluster_y.
##  A feature covariate matrix with columns mean_expression, coef_of_variation, n_nonzero.

protein_metadata_control_fp <- sprintf("%s/protein_expression_metadata_control.rds", processed_protein_dir)
protein_control_odm <- ondisc::read_odm(protein_odm_fp, protein_metadata_control_fp)
protein_control_odm

## A covariate_ondisc_matrix with the following components:
##  An ondisc_matrix with 24 features and 30486 cells.
##  A cell covariate matrix with columns n_nonzero, n_umis, condition.
##  A feature covariate matrix with columns mean_expression, coef_of_variation, n_nonzero.

Note that all three of these datasets have 30486 cells (we computed this number at the end of the previous section).

The gRNA data have 818 features, which is the total number of gRNAs used in the experiment.

The gene expression data have 23712 features, which is the total number of genes measured in the experiment. Frangieh et al did additional feature QC: “Genes detected in fewer than 200 cells were…removed from further analysis.” Also, Frangieh et al did not analyze all genes surviving the above QC. They restricted their attention to the “1,000 highly variable genes were selected using the Scanpy v.1.4.4 implementation of highly variably gene selection…In addition, all features from each condition’s ten most significant Jackstraw PCA programs were included.” Luckily, neither the feature QC step nor the highly variable gene selection has been done on the available data. We’ll apply our own feature QC.

The protein expression data have 24 features. This includes the 20 proteins of interest as well as four controls, as discussed above.

Absent information

At least two kinds of information are absent from the data: batch information and cell cycle information.

• Batch. With this number of cells, it’s almost certain that multiple sequencing batches were used. This seems to be suggested in the methods section: “15,000 cells loaded onto each of eight channels per condition using the 10X Chromium system…” I am guessing that “channel” here means sequencing batch or lane or something like that. However, batch effects are not accounted for in the analysis, or discussed at all in the paper. Unfortunately, batch information is also absent from the data published in the Single Cell Portal.
• Cell cycle. The authors used cell cycle as a covariate for their MIMOSCA analysis. However, cell cycle information is apparently not available in the published data. According to the paper, cell cycle state was “assigned with Scanpy’s implementation of scoring cell cycle genes.” So we might need to rerun this portion ourselves to obtain the cell cycle information.

Note: data size

The raw gene expression data, a CSV file of normalized gene expressions, is 25GB. There should be no reason for lab members to download this file onto their local machine, since it is already processed. Therefore, I discourage using hpcc pull FRANGIEH_2021 to get the processed data onto a local machine (ideally, at some point we should extend the hpcc pull functionality to allow pulling of only the processed data). If you’d like to get the processed data onto your machine, I recommend using rsync directly:

rsync -rltvP $REMOTE_FRANGIEH_2021_DATA_DIR/processed/ $LOCAL_FRANGIEH_2021_DATA_DIR/processed/

Note that even the processed data are about 15GB in size.