Many si/shRNAs can kill cancer cells by targeting multiple survival genes through an off-target mechanism

Over 80% of multiple-tested siRNAs and shRNAs targeting CD95 or CD95 ligand (CD95L) induce a form of cell death characterized by simultaneous activation of multiple cell death pathways preferentially killing transformed and cancer stem cells. We now show these si/shRNAs kill cancer cells through canonical RNAi by targeting the 3’UTR of critical survival genes in a unique form of off-target effect we call DISE (death induced by survival gene elimination). Drosha and Dicer-deficient cells, devoid of most miRNAs, are hypersensitive to DISE, suggesting cellular miRNAs protect cells from this form of cell death. By testing 4666 shRNAs derived from the CD95 and CD95L mRNA sequences and an unrelated control gene, Venus, we have identified many toxic sequences - most of them located in the open reading frame of CD95L. We propose that specific toxic RNAi-active sequences present in the genome can kill cancer cells.


5
To test whether CD95L or CD95 proteins could protect cancer cells from death, we 137 introduced silent mutations into the targeted sites of three very toxic shRNAs: shL1 and shL3 138 (both targeting CD95L) and shR6 (targeting CD95). We first introduced eight silent mutations 139 into the sites targeted by either shL1 or shL3 ( Figure 1B) and expressed these proteins in NB7 140 cells ( Figure 1C). Both mutant constructs were highly resistant to knockdown by their cognate 141 shRNA but still sensitive to knockdown by the other targeting shRNA (Figure 1C). 142 Overexpression of these shRNA-resistant versions of the CD95L ORF did not protect the cells 143 from shL1 or shL3, respectively ( Figure 1D). Interestingly, expression of full length CD95L 144 slowed down the growth of the NB7 cells right after infection with the lentivirus despite the 145 absence of caspase-8 (data not shown). Infection with shRNAs was therefore performed 9 days 146 after introducing CD95L when the cells had recovered and expressed significant CD95L protein 147 levels ( Figure 1C). We then mutated the CD95 mRNA in the targeted site of shR6 ( Figure 1E). 148 Neither expression of wild-type (wt) nor mutated (MUT) CD95 in MCF-7 cells ( Figure 1F) 149 reduced the toxicity when cells were infected with the pLKO-shR6 or another toxic lentiviral 150 shRNA, pLKO-shR7 ( Figure 1G). These data suggested that neither exogenously added 151 recombinant CD95L or exogenously expressed CD95L or CD95 protein can protect cells from 152 toxic shRNAs derived from these genes. 153 To determine whether we could prevent cancer cells from dying by this form of cell death by 154 deleting the endogenous targeted sites, we used CRISPR/Cas9 gene-editing to excise sites 155 targeted by different shRNAs and siRNAs in both alleles of the CD95 and CD95L genes. We 156 first deleted a 41 nt piece of the CD95L gene in 293T cells, that contained the target site for shL3 157 (Figure 2A, 2C). While internal primers could not detect CD95L mRNA in three tested clones, 158 primers outside of the deleted area did detect CD95L mRNA ( Figure 2D, and data not shown). 159 Three clones with this shL3 Δ41 deletion were pooled and tested for toxicity by shL3 expressed 160 from a Tet-inducible plasmid (pTIP-shL3). Compared to a pool of control cells transfected only 161 with the Cas9 plasmid, the 293T shL3 Δ41 cells were equally sensitive to the toxic shRNA 162 ( Figure 2G). This was also observed when the clones were tested individually (data not shown). 163 To exclude the possibility that shL3 was inducing cell death due to a unique activity of shL3 164 and/or 293T cells, we deleted the same 41 nt in CD95L in the ovarian cancer cell line HeyA8; 165 We also generated HeyA8 clones in which we either removed a 64 nt region containing the target 166 site for the siRNA siL3 in the CD95L coding sequence or a 227 nt region containing the target 167 site for shR6 in CD95 (Figure 2A, 2B and Figure 2 -figure supplement 1). In all cases, 6 homozygous deletions were generated ( Figure 2E). To confirm the deletion of the shR6 target 169 site, we infected HeyA8 cells treated with the Cas9 plasmid only and HeyA8 with a homozygous 170 deletion of the shR6 site with shR6 and, as positive controls, with shR2 (targeting the CD95 171 ORF) and shR6' (targeting the CD95 3'UTR). Five days after infection, CD95 mRNA was 172 quantified by real time PCR using a primer located outside the 227bp deletion ( Figure 2F). The 173 mutated CD95 mRNA was still detectable in the shR6 Δ227 cells. While shR2 and shR6' (both 174 targeting outside the deleted region) caused knockdown of CD95 mRNA in both the Cas9 175 expressing control and the shR6 Δ227 cells, shR6 could only reduce mRNA expression in the 176 Cas9 control cells. These data document that HeyA8 CD95 shR6 Δ227 cells no longer harbor the 177 sequence targeted by shR6. 178 Now having HeyA8 cells lacking one of three RNAi-targeted sites in either CD95 or CD95L, 179 we could test the role of the CD95 and CD95L gene products in protecting HeyA8 cells from the 180 death induced by either shRNA (shL3 and shR6, two different vectors: pLKO or the Tet 181 inducible pTIP) or the siRNA siL3. In all cases, the shRNA or siRNA that targeted the deleted 182 region was still fully toxic to the target-site deleted cells (Figure 2H and 2I). We saw efficient 183 growth reduction and cell death in siL3 site deleted cells transfected with as little as 1 nM siL3 184 ( Figure 2I, and data not shown). These data firmly establish that cells were not dying due to the 185 knockdown of either CD95 or CD95L. 186 187

Involvement of canonical RNAi 188
shRNAs and early generation naked siRNAs showed general toxicity when introduced in large 189 amounts, presumably by eliciting an interferon (IFN) response (Marques & Williams, 2005) or 190 by saturating the RISC (Grimm et al., 2006). However, both chemically modified siRNAs at very 191 low concentrations and lentiviral shRNAs at an MOI<1 were still toxic (data not shown). We 192 therefore decided to test whether the observed toxicity involved canonical RNAi and activity of 193 the RISC. To test shRNAs or siRNAs targeting CD95L, we introduced the Venus-CD95L sensor 194 (inset in Figure 3A, right panel) into HeyA8 CD95 protein k.o. cells we had generated in the 195 process of deleting the shR6 site (Figure 2 -figure supplement 1, clone # 2 was used for the 196 following studies; see figure legend for strategy and characterization of the clones). While 197 double-stranded (ds)-siL3 effectively silenced Venus expression and induced toxicity, neither the 198 sense nor the antisense single-stranded (ss)RNAs significantly decreased Venus expression or 199 induced toxicity ( Figure 3A). In addition, no activity was found when ds-siL3, synthesized as 200 9 harvested in all cases at either the 50-hour time point (before the onset of cell death) or at the 264 100-hour time point (during cell death) ( Figure 4A). To achieve high stringency, the data were 265 then analyzed in two ways: first, using a conventional alignment-based analysis to identify genes 266 for which the mRNA changed more than 1.5-fold (and an adjusted p-value of less than 0.05) and 267 second, by a read-based method, in which we first identified all reads that changed >1.5-fold and 268 then subjected each read to a BLAST search to identify the gene it was derived from. Only 269 RNAs that were detected by both methods were considered (Supplementary File 1). The 270 combination of the analyses resulted in one mRNA that was upregulated and 11 mRNAs that 271 were downregulated ( Figure 4B). Using an arrayed qPCR approach, most of these detected 272 mRNA changes were validated for both cell lines (Figure 4 -figure supplement 1A). 273 Interestingly, for nine of the eleven genes, published data suggest they are either highly 274  (Supplementary File 2). Considering these two screens 280 only identified 6.6% of human genes to be critical for cell survival, we found a significant 281 enrichment (54.5%, p-value = 3 x 10 -6 according to binomial distribution) of these survival genes 282 among the genes downregulated during the cell death induced by either shL3 or shR6. All six 283 survival genes are either highly amplified or mutated in human cancers (Figure 4 -figure  284 supplement 2A). In addition to these six genes, GNB1 and HIST1H1C were reported to be 285 required fitness genes in a recent high-resolution CRISPR-based screen (Hart et al., 2015). A 286 kinetic analysis showed most of the deregulated mRNAs were downregulated early with a 287 significant effect already at 14 hours, more than two days before the onset of cell death (Figure 4 288 -figure supplement 1C and data not shown). This suggested the cells were dying because of the 289 silencing of multiple critical survival genes, providing an explanation for why multiple cell death 290 pathways were activated. We therefore call this type of cell death DISE (for Death Induced by 291

Survival gene Elimination). 292
To confirm some of the downregulated genes were also critical survival genes for HeyA8 293 cells, we transfected HeyA8 cells with siRNA SmartPools targeting each of the eleven genes. 294 Individual knockdown of seven of the targeted genes resulted in reduced cell growth when 295 compared to cells transfected with a pool of scrambled siRNAs ( Figure 4C). To mimic the 296 effect of the CD95 and CD95L-derived shRNAs, we treated HeyA8 cells with a combination of 297 siRNA pools targeting these seven genes. Remarkably, 1 nM of this siRNA mixture (35.7 pM of 298 each individual siRNA) was sufficient to effectively reduce growth of the cells (Figure 4 -299 figure supplement 2B) and also cause substantial cell death (Figure 4 -figure supplement 2C), 300 suggesting it is possible to kill cancer cells with very small amounts of siRNAs targeting a 301 network of these survival genes. 302 To test the generality of this phenomenon, we inducibly expressed another CD95L derived 303 shRNA, shL1, in 293T cells using the pTIP vector, and transfected HeyA8 cells with 25 nM 304 siL3. We subjected the cells to RNA-Seq analysis 100 hours and 48 hours after addition of Dox 305 or after transfection, respectively. To determine whether survival genes were downregulated in 306 all cases of sh/siRNA induced cell death, we used a list of 1883 survival genes and 423 genes not 307 required for survival (nonsurvival genes) recently identified in a CRISPR lethality screen 308 (Supplementary File 2). We subjected the four ranked RNA-Seq data sets to a gene set 309 enrichment analysis using the two gene sets ( Figure 4D). In all cases, survival genes were 310 significantly enriched towards the top of the ranked lists (most downregulated). In contrast, 311 nonsurvival genes were not enriched. One interesting feature of DISE that emerged was the 312 substantial loss of histones. Of the 16 genes that were significantly downregulated in cells treated 313 with any of the four sh/siRNAs, 12 were histones ( Figure 4E). While it might be expected that 314 dying cells would downregulate highly expressed genes such as histones, we believe that losing 315 histones is a specific aspect of DISE because a detailed analysis revealed the downregulated 316 histones were not the most highly expressed genes in these cells (Figure 4 -figure supplement  317 3). In addition, almost as many genes with similarly high expression were found to be 318 upregulated in cells after DISE induction.  when NB7 cells were infected with either of these two pools individually (see Figure 6 -456 figure supplement 1B). The shRNAs of these two toxic pools were highly enriched in the 457 underrepresented shRNAs in the two pooled experiments (CD95L and CD95). Their toxicity was 458 also evident when all shRNAs in each pool (2362 shRNAs in the CD95L and 3004 shRNAs in 459 the CD95 pool) were ranked according to the highest fold downregulation (Figure 6C). The 460 three subpools in each experiment are shown separately. Thus, again this analysis identified the 461 ORF of CD95L and the 3'UTR of CD95 as the subpool in each analysis with the highest 462 enrichment of underrepresented shRNAs ( Figure 6C). 463 This analysis allowed us to describe the toxicity landscape of CD95L and CD95 ORFs and 464 their 3'UTRs ( Figure 6D). All shRNAs significantly underrepresented at least five-fold (red dots 465 in Figure 6 -figure supplement 2B) are shown along the CD95L pool (Figure 6D, left) and the 466 CD95 pool (Figure 6D, right) sequences. For both CD95L and CD95, toxic shRNAs localized 467 into distinct clusters. The highest density of toxic sequences was found in the stretch of RNA 468 that codes for the intracellular domain of CD95L (underlined in green in Figure 6D). 469 470

Predicting shRNA toxicity -the toxicity index (TI) and GC content 471
Our data suggest toxic shRNAs derived from either CD95L or CD95 kill cancer cells by 472 targeting a network of genes critical for survival through canonical RNAi. Therefore, we 473 wondered how many 8mer seed sequences derived from these toxic shRNAs would have 474 corresponding seed matches in the 3'UTR of critical survival genes in the human genome. Would 475 it be possible to predict with some certainty in an in silico analysis what shRNAs would be toxic 476 to cells? To calculate such a hypothetical toxicity index, we used the ranked CRISPR data set 477 (Wang et al., 2015) with 1883 survival genes (SGs) and 423 nonSGs. Based on our RNA-Seq 478 analyses, we hypothesized the survival genes contained more putative seed matches for toxic 479 shRNAs in their 3'UTRs than the nonsurvival genes ( Figure 7A, left) and that the number of 480 seed matches in the 3'UTRs of survival genes divided by the number of seed matches in the 481 3'UTR of nonsurvival genes would, to some extent, predict toxicity of an si/shRNA ( Figure 7A, 482 right). 483 To establish a Toxicity Index (TI) for each shRNA, we first gathered 3'UTR sequences for 484 1846 of the survival genes and 416 of the nonsurvival genes. We then generated a list containing 485 a normalized ratio of occurrences of every possible 8mer seed match in the 3'UTRs of the 486 survival and non-survival gene groups. This resulted in a ratio for each of the 65,536 possible 487 8mer combinations (Supplementary File 4), the TI. We then assigned to each of the 4666 488 shRNAs in our screen its TI, and ranked each pool within the two experiments of our screen 489 according to the highest TI (red stippled lines in Figure 7B). We then further separated the 490 shRNAs into two groups: those that were toxic just after infection and those toxic after addition 491 of Dox (Figure 7B, Supplementary File 5). In each ranked list, we could now assess whether the 492 experimentally determined toxicity of shRNAs correlated with the in silico predicted TI. 493 Remarkably, the highest enrichment of toxic shRNAs was found amongst those with higher TI 494 for the subpool of shRNAs targeting the CD95L ORF followed by shRNAs in the subpool 495 targeting the CD95 3'UTR. To confirm the significance of this finding, we repeated the analysis 496 10,000 times by randomly assigning 8mers and their associated TIs to the two shRNA pools and 497 again sorted the data from highest to lowest TI. The reported p-values were calculated based on 498 these permutated datasets using Mann-Whitney U tests. 499 We noticed that survival genes tend to be more highly expressed than nonsurvival genes 500 (data not shown). To address the question whether toxic si/shRNAs only target survival genes or 501 all genes that are highly expressed, we recalculated the TI based on a set of 850 highly expressed 502 and expression matched survival and nonsurvival genes (Figure 7 -figure supplement 1A). This 503 alternative TI tracked slightly less well with the toxic shRNAs we identified, but the enrichment 504 of toxic shRNAs towards the top of the list ranked according to the new TI was still statistically 505 significant (Figure 7 -figure supplement 1B). This analysis demonstrates survival genes contain 506 more seed matches for toxic shRNAs in their 3'UTR than nonsurvival genes regardless of the 507 expression level. This suggests, to a certain extent, it is possible to predict the experimental 508 toxicity of shRNAs based on the in silico calculated TI. 509 Our data suggest DISE results from a sequence-specific off-target activity that depends on 510 the presence of certain seed matches in the 3'UTR of survival genes. Thus, DISE inducing RISC 511 associated small RNAs behave in manner similar to miRNAs. This raised the question whether 512 these seed matches have special properties. While we did not find a sequence motif that was 513 present in all toxic si/shRNAs, we did find that sequence composition, specifically GC content, 514 which has been reported to affect the specificity of shRNAs (Gu et al., 2014;Ui-Tei et al., 2004), 515 correlated with the toxicity of shRNAs. When the GC content of the 6mer seed sequences of all 516 underrepresented shRNAs detected in the shRNA screen across the CD95L ORF was plotted we 517 found a significant correlation between the GC content and higher toxicity (indicated by 518 underrepresentation) (Figure 7C and 7D). This correlation was even more pronounced when 519 plotting GC content versus the 6mer toxicity index (Supplementary File 4) (Figure 7E). 520 While not an absolute requirement, higher GC content made shRNAs more toxic, consistent 521 with reports demonstrating that shRNAs with high GC content in the seed region showed 522 decreased on-target and increased off-target activity (Gu et al., 2014;Ui-Tei et al., 2004). In 523 summary, our data suggest that si-and/or shRNAs with certain seed sequences are toxic to 524 cancer cells by targeting critical survival genes through an RNAi mechanism independent of 525 both Drosha and Dicer. Furthermore, the data suggest high miRNA content, presumably through 526 competing for occupancy in the RISC, might render cells less sensitive to DISE. 527 528 18

Discussion 529
Most current uses of RNAi are aimed toward highly specific silencing with little OTE. In fact, 530 OTEs represent one of the largest impediments to the use of RNAi in phenotypic screening 531 applications. We now demonstrate DISE is a unique form of OTE that results in the simultaneous 532 activation of multiple cell death pathways in cancer cells. The discovery that DISE involves loss 533 of multiple survival genes now provides an explanation for the unique properties we described That means that the number of shRNAs that are toxic due to a possible OTE or general 552 toxicity would be expected to be very small. In contrast, we found that >80% of the shRNAs 553 and siRNAs that were designed to target either CD95 or CD95L exhibited toxicity in multiple 554 cell lines. Consistent with our data analysis a parallel genome-scale loss of function screen 555 confirmed that the majority of the tested shRNAs derived from either CD95L and CD95 were 556 toxic to a majority of the tested 216 cell lines when used as a pooled library (Cowley et al., 557 2014). These also included a number of hematopoietic cell lines suggesting that the DISE 558 effect is not limited to solid cancers. Interestingly, in this study the authors did not consider 559 the data on most of the CD95L and CD95 targeting shRNAs to be significant as they received 560 a low consistency score. A high consistency score predicts the observed phenotype (cell 561 death or growth reduction in this case) is caused by knocking down the targeted gene (Shao et 562 al., 2013). However, we have demonstrated here that the toxicity of an shRNA is solely 563 dependent on its seed and the transcriptome of the treated cells. Therefore, the results of every 564 shRNA should be considered individually as far as the DISE inducing effect is concerned. 565 2) High concentrations of siRNAs can saturate the RISC, preventing the access of crucial 566 endogenous miRNAs (Khan et al., 2009). We have demonstrated that, in general, 5 nM of 567 CD95L-derived siRNAs are sufficient to kill cancer cells. We have even seen very efficient 568 cell death with as little as 1 nM of siRNA (see Figure 2I and lies in the analysis of the chimeras we generated between siL3 and a non-toxic scrambled 586 oligonucleotide (see Figure 3H). This analysis demonstrated that the seed match positions of 587 siL3 are critical for its toxicity. In fact, just replacing one nucleotide in a critical position in 588 the center of the seed match almost completely abolished toxicity of the siRNA. Our data provide strong evidence that the toxicity observed is a sequence-specific event 592 caused by seed matches present in the targets of the toxic si/shRNAs rather than by a toxic motif 593 enriched in all toxic si/shRNAs (i.e. the UGGC motif described before (Fedorov et al., 2006)). 594 We did find a correlation between the toxicity of shRNAs (both predicted by the TI and 595 experimentally determined in the shRNA screen) and the GC content in their seed region. While 596 this correlation was significant, it was not a requirement as some of the most toxic si-and 597 shRNAs had a low 8mer seed GC content (shL3, 25%; shR6, 25%; siL3, 37.5%). Our data 598 suggests that survival genes may contain different types of seed matches (based on base 599 composition or sequence) when compared to nonsurvival genes. Such a distinction has indeed 600 been described before (Stark, Brennecke, Bushati, Russell, & Cohen, 2005). In a study in 601 Drosophila, it was determined that survival genes are depleted of seed matches targeted by 602 highly expressed miRNAs. These authors concluded that evolution must have selected against 603 the presence of seed matches for highly expressed miRNAs in the 3'UTR of survival genes. It is 604 therefore not surprising that a gene ontology (GO) analysis of all miRNA targets (the "targets") 605 in this study described these genes as being involved in development and differentiation (Stark et 606 al., 2005). In contrast, genes not targeted by miRNAs (the "antitargets") grouped in GO clusters 607 that were consistent with cell survival (Stark et al., 2005). A similar phenomenon was also 608 shown in mammalian cells; genes with fewer miRNA target sites, as predicted by Targetscan, 609 contained distinct enriched GO terms from those enriched in genes with many predicted target 610 sites. The genes with fewer sites were enriched in GO terms like ribosomal subunits and 611 respiratory chain, whereas target-heavy genes were more enriched in regulatory-related GO 612 terms (Zare, Khodursky, & Sartorelli, 2014). It is possible the DISE inducing si/shRNAs carry 613 seed sequences that preferentially target seed matches present in the 3'UTRs of the "anti-614 targets". However, as our data on the miR-30 based shRNAs suggest, DISE-inducing shRNAs 615 must be expressed at a certain level to be toxic. 616

DISE is caused by loading of the guide strand of toxic si/shRNAs into the RISC 618
Part of our data was generated using a widely used first generation stem loop shRNA platform, 619 the TRC library. The TRC shRNAs have recently been found to be prone to cause OTE. Gu et al. 620 showed that the loop design of this system results in imprecise Dicer cleavage and, consequently, 621 the production of different mature small-RNA species that increase passenger loading, one major 622 source of OTE (Gu et al., 2012). More recently it was reported that most guide RNAs derived 623 21 from the TRC hairpin were shifted by 4 nt 3' of the expected 5' start site (Watanabe,Cuellar,624 & Haley, 2016). While we did see a shift in processing of these stem loop shRNAs, we did not 625 see such a high level of imprecision in the cleavage of our toxic shRNAs. In fact, 99.4% of the 626 shR6 guide RNAs started at the same nucleotide position (Figure 5 -figure supplement 1A). 627 The majority of the processing of both our pTIP and pLKO-based shRNAs was shifted by one 628 nucleotide (Figure 5 -figure supplement 1A). This shift was consistent with the defined seed 629 matches that were detected in the Sylamer analyses. In general, one major seed match was 630 detected with one other minor species (this was less obvious for shL1, Figure 5  suggest that DISE is not limited to one platform and requires sequence specific targeting. This 643 conclusion is also consistent with a previous report that suggested that sequence-dependent off-644 target transcript regulation is independent of the delivery method (Jackson et al., 2006). The 645 authors found the same enrichment of 6mers and 7mers in 3'UTRs of targeted mRNAs for 646 siRNAs and shRNAs (Jackson et al., 2006). 647 648

The role of Dicer in DISE 649
We previously reported that Dicer Exo5-/-HCT116 cells (with deleted Exon 5) were at least as 650 sensitive to induction of DISE (by either shL3 or shR6) than wt cells suggesting that Dicer 651 deficient cells could be killed by DISE (Hadji et al., 2014). It is has been reported that these 652 in a way that they usually cannot escape from. We have not found a way to block cancer cells 665 from dying by DISE. We provide strong evidence to suggest this is due to the simultaneous 666 targeting of multiple survival genes that result in the activation of multiple cell death pathways. 667 It will be difficult to prove cells are dying due to the preferential targeting of survival genes. It It is therefore possible that when cells are subjected to genotoxic or oncogenic stress that they 682 generate numerous small RNAs that can be taken up by the RISC and in combination execute 683 DISE. Hence, our analysis of CD95/CD95L will likely be applicable to other genes. 684

A model for why DISE preferentially kills cancer cells 686
We interpret the hypersensitivity of both Drosha -/and Dicer -/cells to DISE in the following 23 way: Most of the small RNAs in the cells that are loaded into the RISC are miRNAs. Using 688 AGO pull-down experiments we determined 98.4% of AGO associated RNAs in HCT116 cells 689 to be miRNAs (99.3% in HeyA8 cells, data not shown). It was recently reported that Drosha -/-690 cells showed a reduction of miRNA content from roughly 70-80% to 5-6%, and Dicer -/cells 691 showed a reduction down to 14-21% (Y. K. Kim et al., 2016). Since neither Drosha -/nor Dicer -/-692 cells express reduced AGO2 protein levels (see subset Figure 3E), it is reasonable to assume that 693 their RISC can take up many more of the toxic DISE inducing RNAs than the RISC in wt cells 694 explaining the super toxicity of both DISE inducing si/shRNAs and CD95L mRNAs in these 695

cells. 696
We previously showed expression of either shL3 and shR6 induced DISE in immortalized 697 normal ovarian fibroblasts much more efficiently than in matching nonimmortalized cells (  Glutamine, and 1% penicillin/ streptomycin. All cell lines were authenticated using STR 748 profiling and tested monthly for mycoplasm using PlasmoTest (Invitrogen). 749 42 All lentiviruses were generated in 293T cells using pCMV-dR8.9 and pMD.G packaging 750 plasmids. Retroviruses were generated in Phoenix AMPHO cells using the VSVg packaging 751 plasmid. 752 Figure 1C and D 753

MCF-7 cells overexpressing CD95 cDNAs used in
The HeyA8 cells used in Figure 3D  were identified using the CRISPR gRNA algorithm found at http://crispr.mit.edu/; only gRNAs 818 with scores over 50 were used. These 6 gene blocks were sub-cloned into the pSC-B-amp/kan 819 plasmid using the StrataClone Blunt PCR Cloning kit (Agilent Technologies #240207). 820 The target sites of siL3, shL3, and shR6 were homozygously deleted from target cells by co- performed to confirm that the proper deletion had occurred. Three clones were pooled for each 842 si/shRNA target site deletion except for HeyA8 ΔshR6 for which only clone #11 showed 843 homozygous deletion of the shR6 site; clones #1 and 2 were not complete shR6 deletion mutants, 844 but frame-shift mutations did occur in each allele (as in clone #11) making them CD95 845 knockout clones as depicted in Figure 2 -figure supplement 1A (Figures 2I and 4D, Figure 1 -figure  892 supplement 1A, Figure 5 -supplement 2) or synthesized by IDT (Figure 3A) as sense and 893 antisense RNA (or DNA for Figure 3B, Figure 5 -supplement 1B,) oligos and annealed. The were used in Figure 3B. Blunt siL3 and siScr RNA oligos without the deoxynucleotide 902 overhangs as well as siL2 and siL3 RNA oligos with Cy5-labelled 5' or 3' ends (IDT) were used 903 in Figure 3C. DsiRNA used in Figure 1 -figure supplement 1   To perform arrayed real-time PCR (Figure 4 -figure supplement 1) After treatment/infection, cells were seeded at 500 to 4,000 per well in a 96-well plate at least in 999 triplicate. Images were captured at indicated time points using the IncuCyte ZOOM live cell 1000 imaging system (Essen BioScience) with a 10x objective lens. Percent confluence, red object 1001 count, and the green object integrated intensity were calculated using the IncuCyte ZOOM 1002 software (version 2015A). 1003

RNA-Seq analysis 1005
The following describes the culture conditions used to produce samples for RNA-Seq in Figure  1006 4. HeyA8 ΔshR6 clone #11 cells were infected with pLKO-shScr or pLKO-shR6. A pool of three 1007 293T ΔshL3 clones was infected with either pTIP-shScr or pTIP-shL3. After selection with 1008 puromycin for 2 days, the pTIP-infected 293T cells were plated with Dox in duplicate at 500,000 1009 cells per T175 flask. The pLKO-infected HeyA8 cells were plated at 500,000 cells per flask. 1010 Total RNA was harvested 50 hours and 100 hours after plating. In addition, 293T cells were 1011 infected with either pLKO-shScr or pLKO-shL1 and RNA was isolated (100 hrs after plating) as 1012 To identify differentially abundant RNAs in cells expressing either shL3 or shR6, using a 1036 method unbiased by genome annotation, we also analyzed the raw 100 bp reads for differential 1037 abundance. First, the second end in each paired end read was reverse complemented, so that both 1038 reads were on the same strand. Reads were then sorted and counted using the core UNIX utilities 1039 sort and uniq. Reads with fewer than 128 counts across all 16 samples were discarded. A table  1040 with all of the remaining reads was then compiled, summing counts from each sequence file 1041 corresponding to the same sample. This table contained a little over 100,000 reads. The R 1042 package edgeR (http://bioinformatics.oxfordjournals.org/content/26/1/139) was used to identify 1043 differentially abundant reads, and then these reads were mapped to the human genome using blat 1044 (http://genome.cshlp.org/content/12/4/656.abstract) to determine chromosomal location 1045 whenever possible. Homer (http://homer.salk.edu/homer/) was used to annotate chromosomal 1046 locations with overlapping genomic elements (such as genes). Raw read counts in each sequence 1047 file were normalized by the total number of unique reads in the file. 1048 To identify the most significant changes in expression of RNAs both methods of RNAs-Seq 1049 analyses (alignment and read based) were used to reach high stringency. All samples were 1050 prepared in duplicate and for each RNA the average of the two duplicates was used for further 1051 analysis. In the alignment-based analysis, only RNAs that had a base mean of >2000 reads and 1052 were significantly deregulated between the groups (adjusted p-value <0.05) were considered for 1053 further analysis. RNAs were scored as deregulated when they were more than 1.5 fold changed 1054 in the shL3 expressing cells at both time points and in the shR6 expressing cells at either time 1055 points (each compared to shScr expressing cells) (Supplementary File 1). This was done because 1056 we found that the pLKO driven expression of shR6 was a lot lower than the pTIP driven 1057 expression of shL3 (see the quantification of the two shRNAs in Figure 5 -figure supplement  1058   1A). This likely was a result of the reduced cellular responses in the shR6 expressing cells. In the 1059 read based analysis, reads were only considered if they had both normalized read numbers of >10 1060 across the samples in each treatment, as well as less than 2 fold variation between duplicates and 1061 >1.5 fold change between treatment groups at both time points and both cell lines 1062 (Supplementary File 1). After filtering, reads were mapped to the genome and associated with 1063 genes based on chromosomal localization. Finally, All RNAs were counted that showed 1064 deregulation in the same direction with both methods. This resulted in the identification of 11 1065 RNAs that were down and 1 that was upregulated in cells exposed to the shRNAs shL3 and 1066 shR6. To determine the number of seed matches in the 3'UTR of downregulated genes, the 1067 53 1100

Construction of pTIP-shRNA libraries 1101
The pTIP-shRNA libraries were constructed by subcloning libraries of 143nt PCR inserts of the 1102

The toxicity index (TI) and GC content analysis 1191
The TI in Figure 7A is defined by the sum of the counts of a 6mer or 8mer seed match in the 1192 3'UTRs of critical survival genes divided by the seed match counts in the 3'UTRs of nonsurvival 1193 genes. We used the 1882 survival genes recently described in a CRISPR/Cas9 lethality screen by 56 <-0.1 and an adjusted p-value of <0.05. We chose as a control group to these top essential 1196 genes the bottom essential genes using inverse criteria (CRISPR score of >0.1 and adjusted p-1197 value of <0.05) and are referring to them as the "nonsurvival genes". Both counts were 1198 normalized for the numbers of genes in each gene set. 3'UTRs were retrieved as described above. 1199 For the survival genes 1846 and for the nonsurvival genes 416 3'UTRs were found. For each 1200 gene, only seed matches in the longest 3'UTR were counted. The TI was calculated for each of 1201 the 4096 possible 6mer combinations and each of the 65536 possible 8mer combinations 1202 (Supplementary File 4). These numbers were then assigned to the results of the shRNA screen 1203 (Supplementary File 5). An alternative TI was calculated in Figure 7 -figure supplement 1B  1204 and is based on the top 850 most highly expressed survival genes (all expressed >1000 average 1205 reads) and 850 expression matched genes not described to be critical for cancer cell survival 1206 were selected as controls. 1207 For the analyses in Figure 7C and D, the GC content % was calculated for every 6mer in 1208 the CD95L ORF shRNA pool. The GC content % was then plotted against the log(Fold down) 1209 for each shRNA in the CD95L ORF shRNA after infection (compared to the plasmid 1210 composition) in Figure 7C and after addition of Dox (compared to cells infected but not treated 1211 with Dox) in Figure 7D. In Figure 7E, the log(TI) and GC content % was extracted for every        0  40  80  120  0  40  80  120  0  40  80  120  160  0   40  80  120  0  40  80  120  0  40  80  120  0  40  80  120  160  0  40  80  120  0  40  80  120  0  40  80  120  0  40  80  120    LzCD95L (100ng/ml) /