CRISPR: Dependencies

This page outlines raw data and the subsequent processing for use within the Cancer Dependency Map at Sanger. This data can be accessed through the Cell Model Passport and explored using DepMap Miner.

Raw Data

Corrected Fold Changes

Gene independent responses (e.g. copy number) to CRISPR-Cas9 were corrected using CRISPRcleanR.

16 additional screens generated at Sanger were added to the Project Score data. The screens were combined using ComBat, and correction vectors were estimated from eight models in both datasets to correct for differences in assay lengths. The Sanger day 14 screens were used as the baseline and the day 18 screens were batch corrected using the overlapping cell lines. Once the overlapping cell lines had been used to estimate the batch effects, only the day 14 screens were retained for further analysis. (Pacini, Cancer Cell 2024).

Independent datasets generated at Sanger and Broad showed batch effects due to differences in library design and assay length (Dempster et al. Nat.Commun 2019). To allow combined analysis of these two datasets they were batch corrected to account for these technical differences (Pacini et al. Nat Commun 2021). Batch correction was calculated using ComBat using a set of overlapping cell lines screened at both institutes. Where overlapping cell lines were present, the data from Sanger were preferentially selected for inclusion in the final combined datasets. Finally the first principal component is removed from the data that is associated with differences in media.

Defining Fitness Genes

Loss of fitness scores are generated from corrected FCs for both individual Sanger and Broad datasets and the combined CRISPR-Cas9 dataset using BAGEL2 (Kim et al. Genome Med 2021) to call significantly depleted genes. The gene-level Bayesian factors (BFs) are calculated as the average of the sgRNAs on a targeted-gene basis. As input to BAGEL2 we used reference sets of predefined essential and non-essential genes further curated to exclude high-confidence cancer driver genes. A statistical significance threshold for gene-level BFs is determined for each cell line. Each gene is assigned a scaled BF computed by subtracting the BF at the 5% FDR threshold (obtained from classifying reference essential/non-essential genes based on BF rankings) defined for each cell line from the original BF. For consistency of visualisation, all scaled BF values are multiplied by -1 resulting in significantly depleted values having a loss of fitness score < 0. Archived Project Score v1 data was processed using an implementation of BAGEL in R as described (Behan F, et al. Nature. 2019).

Gene Fitness Metrics

• Loss of Fitness Score: Scaled Bayesian Factors values are multiplied by -1. A score < 0 indicates a statistically significant effect on cell fitness.
• Corrected Fold Change: Copy-number-bias corrected gene depletion fold change, computed between average representation of targeting sgRNAs post-transfection at the end of the experiment versus plasmid library.
• Binary Gene Essentiality Score: Scaled Bayesian Factors >0 are assigned 1 as significantly depleted and 0 otherwise.

Core Fitness Genes

Fitness genes common to the majority of cell lines tested, or common within a cancer type, may be involved in cell essential processes - we refer to these as core fitness genes. In order to identify core fitness genes, we developed a statistical method, ADaM (Adaptive Daisy Model) which adaptively determines the minimum number of dependent cell models required for a gene to be classified as a core fitness gene. ADaM was implemented at both a cancer-type specific level and a pan-cancer specific level (code publicly available at https://github.com/DepMap-Analytics/ADAM2).

Processed Data (Organoids)

Descriptions of how the raw data was processed including algorithms and filtering. Processed datasets can be downloaded here, the active dataset can also be accessed using the DepMap web resources and API.

sgRNA Count Processing and Gene Knockout Fitness Effect Calling

Gene independent responses to CRISPR-Cas9 perturbations were corrected for library size, read-count distributions, and gene copy number bias using CRISPRcleanR. Batch effects between the two sgRNA libraries were corrected using ComBat on CRISPRcleanR normalised and copy number corrected gene log fold changes (LFCs). The resulting values were used as input for the BAGEL2 method (Kim et al. Genome Med 2021).

To improve comparability across organoids, LFCs were scaled using essential and non-essential genes as positive and negative controls, respectively. For each gene, the median LFC of non-essential genes was subtracted from the gene’s LFC, and the result was normalized by dividing it by the difference between the median LFCs of essential and non-essential genes. This scaling effectively adjusted gene-level LFCs based on a standard range defined by these control sets.

Loss of Fitness Scores were then generated from corrected fold changes using BAGEL2 (https://github.com/hart-lab/bagel) to identify significantly depleted genes. Gene level Bayesian Factors (BFs) were computed as the mean of sgRNAs targeting each gene. As input, we used reference sets of predefined essential and non-essential genes, further curated to exclude high-confidence cancer driver genes.

For each organoid, a statistical significance threshold for gene level BFs was determined by classifying reference essential/non-essential genes based on BF rankings at a 5% false discovery rate (FDR). Each gene was assigned a scaled BF, obtained by subtracting the BF at the 5% FDR threshold from the original BF. For visualization consistency, all scaled BF values were multiplied by −1, such that significantly depleted genes had loss-of-fitness scores < 0.

Gene Fitness Metrics

• Loss of Fitness Score: Scaled Bayesian Factors values are multiplied by -1. A score < 0 indicates a statistically significant effect on cell fitness.
• Scaled Corrected Fold Change: Copy-number-bias corrected gene depletion fold change, calculated from the average abundance of targeting sgRNAs at the end of the experiment relative to the plasmid library. Values were scaled across all organoids to improve cross-model comparability.
• Binary Gene Essentiality Score: Scaled Bayesian Factors >0 are assigned a value of 1 (significantly depleted), and 0 otherwise.

Core Fitness Genes

The Adaptive Daisy Model (ADaM) implemented in the CoRe R package was applied using the same curated reference essential gene set described above to estimate false positive rates and identify core fitness genes.

Common Essential Genes

Pan-cancer* common essential genes were identified using the Area Under the Curve (AUC) method implemented in the CoRe R package. This method quantifies gene essentiality across multiple CRISPR/Cas9 screens by integrating the cumulative distribution of fitness effects, enabling robust identification of genes that are consistently essential across diverse cancer types.

*Although the goal was to include all organoids, only three cancer types were ultimately analysed due to limited sample numbers (pancreatic, n = 4; gastric, n = 3), and therefore the resulting core fitness gene set does not completely represent a pan-cancer cohort.

Sanger Library & Data Generation (Organoids)

Organoids were screened using either the Yusa v1.1 library (Behan et al., Nature 2019) or the smaller derivative library MinLibCas9 (Gonçalves et al., Genome Biol 2021). Sixteen models were screened with both libraries. To enable data integration only sgRNAs shared between the two libraries were included. Applying this filter resulted in a final dataset for each organoid containing 16,562 genes targeted by at least two sgRNAs.

Publication reference: