Figure 3. Pleiotropy underlies the fitness effect of gene knockout across human cancer cell lines.

(A) The Cancer Dependency Map (DepMap) fitness screen collection. Individual human cancer cell lines were screened for genetic dependencies for cell growth by comparing cell counts before and after infection with a genome-scale CRISPR-Cas9 gene knockout library.

(B) The DepMap data were preprocessed to a set of 2,921 high-variance fitness genes screened across 675 cell lines. From this data alone, Webster learned a dictionary matrix of 220 fitness effects reflecting inferred biological functions and approximated each gene effect in terms of four functional effects.

(C) Webster approximated the fitness effect of SHOC2 knockout as a mixture of four functional effects (activated KRAS, activated NRAS, EGFR signaling, and FGFR signaling), each of which were strongest in cell lines harboring corresponding genomic alterations (KRAS mutation, NRAS mutation, activated EGFR, and FGFR expression, respectively). This decomposition reflects the pleiotropic interactions underlying SHOC2’s overall function downstream of these signaling pathways.

(D) Joint UMAP embedding of fitness effects for genes and functions inferred from DepMap data. Each of the 220 functions (triangles) and 2,921 genes (circles) are co-embedded in a 2D layout. Selected functions of interest are labeled on the map. Gene effect data from Cancer Dependency Map 19Q4v3 release (https://doi.org/10.6084/m9.figshare.11384241.v3). Functional effects inferred with Webster (this study).

(E) For each function from (C), genes in the joint embeddings are colored according to their loadings on each function, including an inset focusing in on the immediate neighborhood of the function of interest. To the right, the top 10 ranked genes are shown with a heatmap of their respective gene loadings. Gene effect data from Cancer Dependency Map 19Q4v3 release (https://doi.org/10.6084/m9.figshare.11384241.v3). Functional effects inferred with Webster (this study).

(F) Top: relationships learned from fitness data for SHOC2 and RAF1, which also acts downstream of activated RAS proteins. Each arrow corresponds to an inferred gene-to-function loading. Bottom: Illustration of SHOC2/RAF1’s shared biological role in activated RAS signaling.

(G) A network of SHOC2-related functions. Each node is a function inferred by Webster from fitness data. Pleiotropic genes with loadings on two functions are plotted as edges in the network, with the number of pleiotropic genes connecting functions represented by the line thickness. The four bolded functions are those used in the SHOC2 gene effect approximation, while the three additional functions shared at least two pleiotropic genes with one of these four functions.