Journal: Molecular Systems Biology

Article Title: A method for benchmarking genetic screens reveals a predominant mitochondrial bias

doi: 10.15252/msb.202010013

Figure Lengend Snippet: FLEX inputs a CRISPR screening dataset and functional reference standards to compute gene‐level performance and module‐level (e.g., protein complex) performance summaries (see Appendix Fig for details). Precision‐recall (PR) performance of gene–gene co‐essentiality scores using the CORUM complex standard to define true positives (TP). This is a traditional PR curve with the following modifications: (i) the absolute number of TP instead of fractional recall (0‐1) on the x‐axis (simply a scaling of the axis) and (ii) use of a log‐scale on the x‐axis (highlights high precision part of the curve). Pearson correlation coefficients (PCC) are computed between CERES score profiles across the 563 19Q2 DepMap screens for all possible gene pairs. Contribution diversity of CORUM complexes to PR performance (B). Functional composition of different complexes (x‐axis, as a fraction) to the set of TP pairs predicted at different precision levels (y‐axis) are plotted. Only the minimum number of complexes to cover the set of TP pairs (for a certain precision) are considered (see Materials and Methods for details). Complexes with a fraction smaller than 0.01 (1%) at any precision are collectively shown in light gray. The background (bg) contribution diversity represents the functional contribution of complexes across the entire CORUM standard. Highlighted complexes are defined in (D). Size and individual CORUM complex PR performance. Area under the PR curve (AUPRC) was computed per complex on a fractional precision‐recall (0‐1) scale. Dot size corresponds to the mean within‐complex CERES profile PCC, adjusted by the standard error. Protein complexes with at least 30 members (genes) are defined as large, otherwise small. Complexes with an AUPRC of at least 0.4 are defined as high AUPRC, otherwise low. All sub‐complexes mapping to the ETC I or 55S mitochondrial ribosome are shown in the respective color. PR performance of gene–gene co‐essentiality scores (see (B)). Black line shows complete data, colored lines show the performance after sets of complexes (defined in (C)) were removed from the data and standard. The inset barchart shows the percentage of TP lost at a precision of 0.5 after either set of complexes is excluded. Module PR (mPR) curve summarizes performance at a functional module level (here, CORUM protein complexes). This is a modified version of a precision‐recall curve (B) with the number of unique complexes (x‐axis) covered and plotted (instead of unique gene pairs) at each precision cutoff (y‐axis) (see Materials and Methods for details). Comparison of two methods measuring co‐essentiality in the DepMap using PR and mPR plots. The method proposed by Wainberg and colleagues is compared with the standard PCC‐based method (top). The well‐balanced coverage of complexes is shown after their ETC‐related complex exclusion (dotted lines, top) as well as in the mPR curve (bottom). The approach from Wainberg et al bases gene‐pair similarity scores on FDR corrected P ‐values (1 ‐ fdr) resulting in a ‘late start’ of the PR curve (many values at top are the same, 1.0).

Article Snippet: First, we compared dependency data from 149 cell lines in the 19Q2 DepMap that had been screened both at the Broad Institute (hereafter referred to as Broad DepMap) and the Sanger Institute (Sanger DepMap).

Techniques: CRISPR, Functional Assay, Modification, Comparison

Journal: Molecular Systems Biology

Article Title: A method for benchmarking genetic screens reveals a predominant mitochondrial bias

doi: 10.15252/msb.202010013

Figure Lengend Snippet: Co‐essentiality networks from the DepMap dataset are calculated using different similarity measures: cosine similarity, dot product similarity, Pearson correlation, and Spearman correlation. Gene‐level performance comparison of different similarity metrics across CORUM, Pathway, and GO‐BP standards. Contribution diversity of CORUM complexes for different similarity measures.

Article Snippet: First, we compared dependency data from 149 cell lines in the 19Q2 DepMap that had been screened both at the Broad Institute (hereafter referred to as Broad DepMap) and the Sanger Institute (Sanger DepMap).

Techniques: Comparison

Journal: Molecular Systems Biology

Article Title: A method for benchmarking genetic screens reveals a predominant mitochondrial bias

doi: 10.15252/msb.202010013

Figure Lengend Snippet: Compared are different alternative methods to infer functional relationships from DepMap 18Q3 release CERES score co‐essentiality profiles and CORUM 3.0 complex relations as a standard. The methods are PCC of co‐essentiality profiles as baseline data treatment (pink) and approaches published in Wainberg and colleagues (Wainberg et al, ) (violet), Boyle and colleagues (Boyle et al, ) (orange) and Kim and colleagues (Kim et al, ) (green). Gene‐level performance to capture co‐complex membership on the full data. Gene‐level performance to capture co‐complex membership when ETC‐related complexes are removed from data and standard. Module‐level performance to capture number of complexes with TP co‐complex membership on the full data. Module‐level performance to capture number of complexes with TP co‐complex membership when ETC‐related complexes are removed from data and standard.

Article Snippet: First, we compared dependency data from 149 cell lines in the 19Q2 DepMap that had been screened both at the Broad Institute (hereafter referred to as Broad DepMap) and the Sanger Institute (Sanger DepMap).

Techniques: Functional Assay

Journal: Molecular Systems Biology

Article Title: A method for benchmarking genetic screens reveals a predominant mitochondrial bias

doi: 10.15252/msb.202010013

Figure Lengend Snippet: Protein complex‐level differences in fitness effects between the Broad and Sanger DepMap screens. The 149 cell lines and 16,464 genes common to both datasets are compared. For each CORUM complex, the median differential CERES score (x‐axis) and a paired Wilcoxon rank sum P ‐value with BH‐correction are shown. Mitochondrial ribosome (yellow) and ETC I (orange) sub‐complexes are highlighted. Dot size is proportional to complex size. Protein stability of CORUM complexes. Protein half‐life data were taken from B cells, hepatocytes, and monocytes, and summarized on the CORUM complex level. Half‐life data were z‐transformed, and the minimum z‐score set to 0 to emphasize large z‐scores. Complexes for which at least 5 members contributed data across the three cell lines are shown. Scheme of time‐resolved genome‐wide CRISPR/Cas9 screens in HAP1 cells. Temporal fitness profile similarity was estimated by computing the pairwise PCC between genes with 32 unique measurements across time. The dropout speed was derived from profiles interpolated from the 32 measurements after correcting for maximal dropout effects (see Materials and Methods). Precision‐recall (PR) curve showing HAP1 temporal fitness profile similarity performance using CORUM complexes as a pairwise functional standard. Black line shows complete data, red line performance after ETC I, V, and mitochondrial ribosome (ETC‐related complexes) are removed from the data and standard. Contribution diversity of HAP1 temporal fitness profile similarity PR performance using the CORUM complex standard. Shown are the fraction of TP pairs for CORUM complexes (distributions across the x‐axis) at different precision cutoffs (down the y‐axis). The minimum number of complexes to cover the complete set of TPs is shown (see Materials and Methods). Complexes with a fraction smaller than 0.01 (1%) at any precision are collectively shown in light gray. The background (bg) functional diversity represents the distribution of categories across the entire reference standard (i.e., the expected distribution in a random selection of gene pairs). Module‐level performance of HAP1 temporal fitness profile similarity shows CORUM complex size and AUPRC. Dot size corresponds to the mean within‐complex similarity, adjusted by the standard error. All sub‐complexes mapping to the ETC‐related complexes are shown in the respective color. Comparison of module‐level performance between Broad DepMap co‐essentiality and temporal fitness. AUPRC measures the performance of each dataset in reconstructing CORUM complex co‐memberships. Dot size is proportional to complex size. Dropout speed for ETC‐related and other selected essential complexes. Dropout speed is a normalized estimate of the derivative of an LFC profile (across time) for each guide (see Materials and Methods). A positive dropout speed indicates faster relative dropout, while a negative dropout speed indicates slower dropout (see left panel for hypothetical LFC profile examples and their corresponding dropout speeds). The average dropout speed across all genes in each of the indicated complexes is plotted as a function of screen sampling time (right). tSNE embedding groups CORUM complexes with similar dropout speed (see Materials and Methods). The six selected complexes on the right are indicated in the tSNE plot (large colored dots) and sub‐complexes are labeled with matching colors (bottom).

Article Snippet: First, we compared dependency data from 149 cell lines in the 19Q2 DepMap that had been screened both at the Broad Institute (hereafter referred to as Broad DepMap) and the Sanger Institute (Sanger DepMap).

Techniques: Transformation Assay, Genome Wide, CRISPR, Derivative Assay, Functional Assay, Selection, Comparison, Sampling, Labeling

Journal: Molecular Systems Biology

Article Title: A method for benchmarking genetic screens reveals a predominant mitochondrial bias

doi: 10.15252/msb.202010013

Figure Lengend Snippet: Broad DepMap genome‐wide CRISPR/Cas9 screens ranked by the median CERES score across the ETC‐related complexes. The middle red line indicates the median, the vertical lines the 25 and 75% quantiles of a given screen. Gray lines represent the same metrics for all genes in the genome. Pairwise comparison of Broad and Sanger DepMap screens based on their median CERES score of ETC‐related complexes. Highlighted are 149 cell lines common to both datasets. To rank those cell lines, Sanger data from those 149 screens were added to the 563 Broad DepMap screens and all screens were ranked. Green lines indicate a higher ranking of the Broad screen (assay length 21 days) and brown a higher ranking for the corresponding Sanger screen (assay length 14 days). Rank of HAP1 time course genome‐wide screens in the Broad DepMap screens based on the adjusted median ETC‐related LFC. HAP1 screens were performed with the TKOv3 library, and LFC values were adjusted by centering non‐essential genes around 0 and core essential genes around −1 (see Materials and Methods). HAP1 screens sampled at T3 are shown as circles indicating that they have not been used for computing the Spearman’s rank correlation coefficient and the associated statistical significance (see Materials and Methods for details). Wild‐type protein abundance of two protein complexes is schematically displayed over the course of a CRISPR screen. The measured phenotype (e.g., gRNA abundance as a proxy for cell fitness) depends on the presence of a sufficient amount of protein to fulfill a cellular function. Stability of proteins, the rate of cell doublings that redistribute residual protein, protein levels required for normal function, or more stable epistatic protein complexes determine the penetrance of cellular fitness phenotypes throughout the course of a CRISPR experiment.

Article Snippet: First, we compared dependency data from 149 cell lines in the 19Q2 DepMap that had been screened both at the Broad Institute (hereafter referred to as Broad DepMap) and the Sanger Institute (Sanger DepMap).

Techniques: Genome Wide, CRISPR, Comparison, Quantitative Proteomics, Cell Function Assay