Discovery of novel tumor suppressors from CRISPR screens reveals lipid-sensitive 1 subtype of AML 2 3

CRISPR knockout screens in hundreds of cancer cell lines have revealed a substantial number of context-specific essential genes that, when associated with a biomarker such as lineage or oncogenic mutation, offer candidate tumor-specific vulnerabilities for targeted therapies or novel drug development. Data-driven analysis of knockout fitness screens also yields many other functionally coherent modules that show emergent essentiality or, in some cases, the opposite phenotype of faster proliferation. We develop a systematic approach to classify these suppressors of proliferation, which are highly enriched for tumor suppressor genes, and define a network of 103 genes in 22 discrete modules. One surprising module contains several elements of the glycerolipid biosynthesis pathway and operates exclusively in a subset of AML lines, which we call Fatty Acid Synthesis/Tumor Suppressor (FASTS) cells. Genetic and biochemical validation indicates that these cells operate at the limit of their carrying capacity for saturated fatty acids. Overexpression of saturated acyltransferase GPAT4 or its regulator CHP1 confers a survival advantage in an age-matched cohort of AML patients, indicating the in vitro phenotype reflects a clinically relevant subtype, and suggesting a previously unrecognized risk in clinical trials for fatty acid synthesis pathway inhibitors.

Although known TSG act as PS genes in only a subset of cell lines, we observed patterns of co-201 occurrence among functionally related genes. PTEN co-occurs with mTOR regulators NF2 50 (P 202 < 2x10 -6 , Fisher's exact test) and the TSC1/TSC2 complex (P-values both < 2x10 -13 ) 51 , as well 203 as Programmed Cell Death 10 (PDCD10) 52 , a proposed tumor suppressor 8,53 (Figure 2a). The 204 p53 regulatory cluster (TP53, CDKN1A, CHECK2, TP53BP1) also exhibited a strong co-205 occurrence pattern that was independent of the mTOR regulatory cluster (Figure 2a). does not appear to be a strong tissue specificity signature for most clusters (Figure 2c), the 234 fatty acid metabolism cluster shows a strong enrichment for AML cell lines (P = 1.1x10 -5 ). AML, 235 like most cancers, typically relies on increased glucose consumption for energy and diversion of 236 glycolytic intermediates for the generation of biomass required for cell proliferation. Membrane 237 biomass is generated by phospholipid biosynthesis that uses fatty acids as building blocks, with To learn whether additional elements of lipid metabolism were associated with the FAS cluster, 250 we examined the differential correlation of shuffled Z-scores in AML cells. We and others have 251 shown that genes with correlated gene knockout fitness profiles in CRISPR screens are likely to 252 be involved in the same biological pathway or process ("co-functional") 18-21 , analogous to 253 correlated genetic interaction profiles in yeast 25,27,65 . Strikingly, all gene pairs within the fully 254 connected clique in the FAS cluster (containing genes FASN, ACACA, GPAT4, CHP1, and GPI, 255

Cas12a-mediated Genetic Interaction Screens Confirm Rewired Lipid Metabolism 298 299
We sought to confirm whether gene knockout confers improved cell fitness, and to gather some 300 insight into why some AML cells show the FASTS phenotype and others do not. We designed a 301 CRISPR screen that measures the genetic interactions between eight selected "query genes" 302 and ~100 other genes ("array genes"). The query genes include FASN and ACACA, from the 303 cluster of proliferation-suppressor genes, as well as lipid homeostasis transcription factor 304 SREBF1, anticorrelated with the FAS cluster in the differential network analysis, and 305 uncharacterized gene c12orf49, previously implicated in lipid metabolism by coessentiality 21 and 306 a recent genetic interaction study 60 . Additional query genes include control tumor suppressor 307 genes TP53 and PTEN and control context-dependent essential genes GPX4 and PSTK 308 (Figure 4a). The array genes include two to three genes each from several metabolic pathways, 309 including various branches of lipid biosynthesis, glycolysis and glutaminolysis, oxphos, 310 peroxisomal and mitochondrial fatty acid oxidation. We include the query genes in the array 311 gene set (Figure 4a) to test for screen artifacts and further add control essential and 312 nonessential genes to measure overall screen efficacy (Supplementary Table 3-4). 313

314
We used the enCas12a CRISPR endonuclease system to carry out multiplex gene knockouts 35 .
construction of specific guide pairs through pooled oligonucleotide synthesis (Figure 4b). The 317 library robustly measures single knockout fitness by pairing three Cas12a crRNA per target 318 gene each with five crRNA targeting nonessential genes 8,43 (n=15 constructs for single knockout 319 fitness), and efficiently assays double knockout fitness by measuring all guides targeting query-320 array gene pairs (n=9) (Figure 4c & Supplementary Table 4). Using this efficient design and 321 the endogenous multiplexing capability of enCas12a, we were able to synthesize a library 322 targeting 800 gene pairs with a single 12k oligonucleotide array.  growth effect, resulting in a suppressor interaction with a negative Z-score (Figure 4g-h). anticorrelated in blood (r= -0.41) but highly correlated ex-blood (r=0.59) due to their co-384 with CHP1 in most tissues but strongly correlated in AML, with underlying covariation largely 386 driven by the PS phenotype in FASTS cells (Figure 3e). This pathway sign reversal is 387 confirmed in the single knockout fitness observed in our screens: SCD is strongly essential in 388 MOLM-13 but not in NOMO-1 (Figure 4e). 389 390 Collectively these observations make a strong prediction about the metabolic processing of 391 specific lipid species. Faster proliferation upon knockout of genes related to saturated fatty acid 392 processing, coupled with increased dependency on fatty acid desaturase (Figure 5a), suggests 393 that these cells are at or near their carrying capacity for saturated fatty acids. To test this 394 prediction, we exposed three To explore whether the FASTS phenotype has clinical relevance, we compared our results with 408 patient survival information from public databases. Using genetic characterization data from 409 CCLE 71 , we did not find any lesion which segregated FASTS cells from other CD33+ AML cells 410 (Figure 6a), so no mutation is nominated to drive a FASTS phenotype in vivo. Instead, we 411 explored whether variation in gene expression was associated with patient outcomes. We 412 included genes in the core FASTS module as well as genes with strong genetic interactions with 413 ACACA/FASN in our screen (Figure 6a). To select an appropriate cohort for genomic analysis, 414 we first considered patient age. Although AML is present across every decade of life, patients 415 from whom FASTS cell lines were derived are all under 30 years of age (sources of other AML 416 cells ranged from 7 to 68 years; Figure 6b). With this in mind, we explored data from the 14 data, we calculated hazard ratios using univariate Cox proportional-hazards modeling with 419 continuous mRNA expression values for our genes of interest as independent variables. We 420 observed that both CHP1 and GPAT4 show significant, negative hazard ratios (HR), consistent 421 with a tumor suppressor signature (Figure 6d), and that no other gene from our set shows a 422 negative HR. Indeed, tumors in the top quartile of gene expression showed significantly 423 improved survival for both CHP1 (P-value 0.007, Figure 6e) and GPAT4 (P-value 0.035, Figure  424  phenotype, we lack the statistical power to discover associations in an unbiased way. However, 464 by narrowing our search to strong hits from the differential network analyses, we found a 465 significant survival advantage in a roughly age-matched cohort for GPAT4 and CHP1 466 overexpression. This finding is consistent with a wholly novel tumor suppressor signature for our 467 PS gene module. 468

469
The combination of genetic, biochemical, and clinical support for the discovery of a novel tumor 470 suppressor module has several implications. First, it provides a clinical signature that warrants 471 further research as a prognostic marker as well as a potential therapeutic target --and a high-472 risk group for fatty acid synthesis inhibitors. Second, it demonstrates the power of differential 473 network analysis, and in particular differential genetic interaction networks, to dissect the 474 rewiring of molecular pathways from modular phenotypes. And finally, it suggests that there still 475 may be much to learn from data-driven analyses of large-scale screen data, beyond the low-