Computational analyses identify CDS1-CDS2 as a common and cancer-associated SLI
To query for SLIs based on differential gene expression between cancer and normal tissue, we set up a bioinformatic pipeline using public datasets on gene dependencies and RNA expression (Fig. 1a). This pipeline identifies anchor-target gene pairs, in which the anchor gene shows a relatively reduced expression level in cancer compared to normal tissue, and in which disruption of the target gene results in cancer lethality.
The DepMap includes additional types of data that we incorporated in our pipeline to score SLIs, including promoter methylation data, protein expression data, damaging mutation data and cancer type data33,34,37,38. RNA expression data provided good power to detect previously established synthetic lethal interactions, including BRCA1-PARP1 and WRN-MLH1 (Fig. 1b, Extended Data Fig. 1a)2,3,5,6,14,15,18,19,24,30,39. The inclusion of all cancer types simultaneously resulted in a high resolution to detect established synthetic lethal interactions compared to analyzing specific cancer types separately (Extended Data Fig. 1a). Therefore, we determined the SLI scores for all gene pairs with RNA expression data for all cancer types simultaneously.
The CDS1-CDS2 and NAA10-NAA11 gene pairs received high SLI scores (Fig. 1c, Extended Data Fig. 1b, c and Supplementary Table 1). However, the NAA10-NAA11 SLI failed to show cancer specificity and was therefore not pursued. In contrast, the CDS1-CDS2 SLI demonstrated significant cancer specificity. The high SLI score for CDS1-CDS2 is consistent with other analyses and screens scoring CDS1-CDS2 as a candidate synthetic lethal pair, but to our knowledge this has not yet been pursued18,30,32,40. CDS1 and CDS2 are enzymes that are conserved in plants and yeast. They serve to convert phosphatidic acid into cytidine diphosphate diacylglycerol, conceivably representing the bottleneck in phosphatidylinositol (PI) synthesis41–43. PI constitutes an essential component of cellular membranes and is also used as a critical kinase substrate regulating cell proliferation and survival44,45. Aberrations in PI signaling components act as common cancer drivers and are clinically targeted with PI3K inhibitors46.
Our analysis confirmed FAM50A-B, DDX3X-DDX3Y and EIF1AX-EIF1AY as cancer-specific SLIs, in agreement with previous reports (Extended Data Fig. 1c)18,30. Similarly, we identified PARP-associated synthetic lethality for breast cancer cell lines (Fig. 1d, left panel). For several other SLIs, cancer specificity was previously inferred from genomic data5,6,14,18,19, but did not score here as such, for example WRN-MLH1. This is conceivably due to the relatively low frequency of genomic alterations in MLH1 coupled to our focus on transcriptomic rather than genomic data.
In contrast, we noted that CDS2 loss was associated with lethality in a large number of cancer types (Fig. 1d, right panel). Several of these concerned common cancers, including lung, blood, brain and skin cancer (Extended Data Fig. 1d). From here onwards, lethality as a function of CDS2 loss will be referred to as ΔCDS2 lethality. Patients with CDS1-low cancers showed significantly worse survival compared to CDS1-high cancer patients (Fig. 1e). For these reasons we focus here on the CDS1-CDS2 SLI.
To confirm and quantify cancer-specific loss of CDS1 expression, we compared pan-cancer DepMap expression data and pan-tissue GTEx expression data (Fig. 1f). The data was calibrated on housekeeping genes, while control analyses on reference genes were performed to determine reliability. In line with the TCGA data (Fig. 1c), we observed common absence, or low expression, of CDS1 in cancer cell lines compared to healthy tissue (64% average reduction, p-value < 0.0001). As expected, given the SLI, low or no CDS1 expression strongly correlated with ΔCDS2 lethality in cancer cell lines (90% average reduction, p-value < 0.0001). We also observed that lung, blood, brain and skin cancers have reduced CDS1 levels compared to their lineages of origin (Extended Data Fig. 1e), suggesting that these cancers may suppress transcription of CDS1 during cancer development. For lung cancer, differential CDS1 levels were confirmed using a cohort comprising patient-matched proteomics data from healthy and tumor tissue (Fig. 1g)47. Thus, our in silico analyses predict that the CDS1-CDS2 gene pair constitutes a human cancer-associated SLI, specifically in common CDS1-low cancers.
CDS1 and CDS2 constitute a synthetic lethal gene pair across cancer types
For wet lab validation of the computational predictions, we used a panel of cancer cell lines, which are also in the DepMap (Fig. 2a). Their CDS1 RNA levels were confirmed by qPCR analysis (Fig. 2b). Low throughput CRISPR perturbations were used to quantify the lethality inferred from genome-wide CRISPR knockout screens. By including fluorescent tracker cells serving as an internal control for each experimental condition, we quantified lethality over extended timeframes, showing minimal variation (Fig. 2c).
Upon perturbation of CDS2 with one of two separate sgRNAs, we were able to confirm ΔCDS2 lethality in CDS1-low or CDS1-negative cell lines across several cancer types (Fig. 2d and Extended Data Fig. 2a, b). Furthermore, as predicted, a CDS1-high cell line was ΔCDS2 non-lethal. Compared to the DepMap data, we observed remarkably strong lethality. Furthermore, ΔCDS2 lethality was still evident in cell lines with low levels of CDS1 RNA (Fig. 2d; green).
To validate the CDS1 dependency for the ΔCDS2 lethality, we either ectopically expressed CDS1 in CDS1-negative cancer cell lines or perturbed CDS1 in CDS1-proficient cancer cells. When CDS1 was introduced into two CDS1-negative cell lines, the ΔCDS2 lethality was largely negated. Conversely, when CDS1 was perturbed in a CDS1-high cell line it exhibited increased ΔCDS2 lethality (Fig. 2e). To validate these observations, we admixed control GFP-expressing cancer cell lines with CDS1-restored cell lines in an additional panel of four human melanoma models. The results confirmed strong synthetic lethality upon perturbation of CDS2 in this panel (Fig. 2f and Extended Data Fig. 2c). Together, these functional experiments confirm that CDS1 and CDS2 constitute a synthetic lethal pair across a panel of cancer cell lines in vitro.
ΔCDS2 lethality is associated with apoptosis
We suspected ΔCDS2 lethality may result in apoptotic cell death in vitro. To investigate this, we collected tumor samples and measured cleaved caspase-3 by quantitative western blotting as a measure of apoptosis (Fig. 2g). For quantification the cleaved caspase-3 signal is compared to total protein signal in the same capillary. As a positive control, we analyzed cleaved caspase-3 levels upon induction of apoptosis by TPCA-1 + TNF (BLM) or staurosporine (SK-MEL-2) (Extended Data Fig. 2d). We observed a significant increase in cleaved caspase-3 in ΔCDS2 cancer cells, which indicates ΔCDS2 lethality is associated with tumor cell apoptosis.
ΔCDS2 lethality in vivo
Next, we investigated whether the SLI between CDS1 and CDS2 observed in silico and in vitro can be recapitulated in vivo. For this purpose, we again admixed control GFP-expressing cancer cell lines with CDS1-restored cell lines. Tumor cells were inoculated into immunodeficient NOD-Scid IL2Rgnull mice and analyzed by flow cytometry of the tumors 17 days later. We observed a striking inability of CDS2-perturbed cells to contribute to tumor formation in vivo (Fig. 2h and Extended Data Fig. 2e, f; note that the tumor growth curves are derived from cell mixes including rescued cells). Together, these results demonstrate that CDS1 and CDS2 form a robust synthetic lethal pair in cancer, in silico, in vitro and in vivo.
No common escape mechanism for Δ CDS2 lethality
The computational and in vitro and in vivo functional validation data above demonstrate the broad cancer range and reproducibility of the CDS1-2 SLI, prompting us to further challenge its robustness. Specifically, the rate-limiting role of the CDS enzymes in PI synthesis led us to investigate whether any cells can rewire their signaling network such that they can escape from this SLI. To investigate this in an unbiased, genome-wide fashion, we performed CRISPR knockout rescue screens in a panel of four CDS1-negative human cancer cell lines and, as a control, one CDS1-high cancer cell line (Fig. 3a and Supplementary Table 2). In parallel, cells from the screens were used to track ΔCDS2 lethality during the screen. These analyses confirmed ΔCDS2 lethality in four CDS1-negative cancer cell lines and extended our data on the lack of ΔCDS2 lethality in CDS1-high cancer cell lines to an additional cancer cell line (Fig. 3b). In addition, colony formation assays were performed to visualize the lethal effect during the screen (Extended Data Fig. 3a; note that for K562 a different readout was used because it is a suspension cell line).
Analysis of the dropout of essential genes confirmed the high quality of the screens (Fig. 3c). Potential escape mechanisms were determined using ΔCDS2 lethality quantified with tracker cells. However, the results of the screens (Fig. 3d and Extended Data Fig. 3b) indicated no common escape mechanism of ΔCDS2 lethality. For example, the screen performed in SK-MEL-2 cells yielded no significant enrichment even after an additional 14 days (32 days in total; Extended Data Fig. 3c), while the other screens yielded only some cell line-specific rescue (Extended Data Fig. 3d-f). These findings suggest that no common escape to the combined loss of CDS1 and CDS2 is possible, which is in agreement with the idea that CDS1 and CDS2 together serve as a bottleneck for PI synthesis.
Mesenchymal cancers depend on CDS2 for PI synthesis
To understand which cancer types show reduced CDS1 expression and, hence, ΔCDS2 lethality, we characterized CDS1-low cancers using publicly available data. First, we noted that CDS1-high cancer cells express high levels of the epithelial marker gene E-cadherin, whereas CDS1-low cancer cell lines instead express mesenchymal markers like ZEB1, ZEB2 and vimentin (Fig. 4a and Extended Data Fig. 4a)48,49. Mechanistically, this is in agreement with the notion that CDS1 expression is suppressed in mesenchymal cancers by the transcription factor ZEB1, previously reported to bind the CDS1 locus and to suppress CDS1 expression50,51. Besides this major regulatory mechanism, we also observed that a rare subset of the blood lineage cancer cell lines exhibit methylation of the CDS1 promoter (Extended Data Fig. 4b). Overall, these findings indicate that the suppression of CDS1 expression in mesenchymal-like cancers (by ZEB1) results in their strong dependency on CDS2.
Gene-set enrichment analysis indicated that CDS1-low cancers are enriched for the Hallmark EMT gene-set (p = 0.0001 DepMap, p = 0.009 TCGA, Extended Data Fig. 4c). Mesenchymal-transitioned cancers are common, highly metastatic and therapy-refractory48,52–55. These findings are in agreement with our observations that low CDS1 expression is common, more frequent in cancers compared to their healthy tissue of origin and associated with worse survival (Fig. 1d, e, Extended Data Fig. 1e). Together, these results suggest that suppression of CDS1 expression is an integral element of EMT in cancer.
Next, we zoomed in into the pathway in which CDS1 and CDS2 are involved, as has been defined by previous studies (Fig. 4b)41–43,56. We incorporated expression and dependency data from the DepMap of the enzymes involved. In line with the requirement for CDS1/2, the next enzyme in the pathway, cytidine diphosphate diacylglycerol synthase inositol-3-phosphatidyltransferase (CDIPT), is highly essential for survival. In addition, CDS2 and CDIPT are significantly co-dependent in DepMap cancer cell lines, which is indicative of a similar mechanism of dependency (Fig. 4c)27,57. Accordingly, when supplementing the cell culture media with exogenous PI, we observed a significant rescue of cell death, suggesting that ΔCDS2 lethality is due, at least in part, to insufficient availability of PI (Fig. 4d and Extended Data Fig. 4d).
CDS2 acts upstream of the PI3K signaling module regulating growth and survival. Unexpectedly, DepMap analysis revealed that ΔCDS2 lethality is not accompanied by lethality with either genetic loss of PIK3CA or pharmacologic PIK3CA inhibition (Alpelisib). Instead, we observed a strong anticorrelation between ΔCDS2 lethality and ΔPI3KCA lethality (Fig. 4c). This anticorrelation was independent of PI3K isoforms, as judged by pan-PI3K inhibitor responses. These results demonstrate an essential role of CDS for PI synthesis in mesenchymal cancers. Furthermore, they unexpectedly point to a contribution of CDS1 and CDS2 to survival signaling independent of the classical PI3K pathway.
Expression of either CDS1 or CDS2 is required for lipid homeostasis
To biochemically define the effects of CDS2 perturbation in CDS1-negative cancer cells, we performed multi-omic analyses in a panel of four CDS1-negative cancer cell lines and, as a control, one CDS1-high cancer cell line (Fig. 5a). Lipidomics was performed to quantify buildup or depletion of lipids, specifically those in the CDS pathway. In addition, proteomics was performed to identify deregulated cellular processes upon CDS2 loss.
Lipidomic analysis allowed for quantification of the major lipid classes in the cells, with phosphorylated PI (PIP) serving as the major signaling molecule44 (Extended Data Fig. 5a and Supplementary Table 3). Major changes as a function of CDS1 expression were detected upon CDS2 loss (Fig. 5b and Extended Data Fig. 5a, b; note that the stars indicate the number of CDS1-negative cancer cell lines with p-value < 0.01). For example, cholesterol esters and triglycerides were massively upregulated. Furthermore, we observed strong buildup of CDS2 substrates and depletion of the downstream product PI.
The proteomics analysis allowed for quantification of ~ 7000 proteins in each cell line (Supplementary Table 4). As expected, the levels of CDS2 protein had dropped 10-fold in all five cell lines upon CDS2 CRISPR perturbation (Fig. 5c). In addition, large groups of proteins were commonly and significantly down- or up-regulated in CDS1-low cancer cell lines, suggesting an orchestrated response (Fig. 5c and Extended Data Fig. 5c). GO-term enrichment analysis of these proteins revealed major induction of cholesterol import and production upon ΔCDS2 lethality (Extended Data Fig. 5d).
In line with the build-up of cholesterol esters and triglycerides, CDS2-perturbed cells formed large lipid droplets58, which were visible by both light and electron microscopy, staining positively for the lipid dyes BODIPY and Nile Red. Quantification of BODIPY-stained live cells revealed that on average 3% of the imaged cells comprised lipid (Fig. 5d). Combined light and electron microscopy of fixed cells showed these lipid droplets in more detail (Fig. 5e). Together, these findings again corroborate the CDS1/2 biochemical bottleneck to support lipid homeostasis, demonstrating that lethality upon CDS2 perturbation in CDS1-low cell lines is accompanied by major changes in lipid metabolism and expression of the cholesterol pathway (Fig. 5f).