Inhibition of Aurora Kinase Induces Endogenous Retroelements to Induce a Type I/III IFN Response via RIG-I

IFNs are cytokine mediators of the innate antiviral immune response. Type I IFNs (most commonly α or β), can be produced by most cells in response to viral infection. Type III IFNs (λ IFNs), are produced by surface-exposed epithelial cells and engage distinct receptors. However, inducers, downstream signaling and target gene expression patterns of type I and type III IFNs are largely identical. Major inducers of IFN expression include the pattern recognition receptor pathways that recognize signatures of viral infection such as inappropriately modified or localized DNA (e.g., cGAS/STING) or RNA (e.g., RIG-I/MDA5/MAVS; ref. 1). The role of IFNs is to orchestrate a suite of intracellular defenses against viral replication, increase class I MHC expression and antigen presentation, and stimulate production of cytokines such as CXCL10 that stimulate immune cell infiltration.

While immune checkpoint blockade (ICB) agents, such as those targeting PD-L1 or CTLA4 pathways, have demonstrated clinical efficacy in some cancers, current ICB treatments do not have activity in all cancers, sometimes due to an unfavorable tumor immune microenvironment (2). Triggering a local innate immune response in tumor cells could be one strategy to stimulate tumor immunogenicity. Indeed, the efficacy of many clinically approved small-molecule drugs is thought to be at least partially due to activation of local tumor-intrinsic immunity. For example, DNA methyltransferase inhibitors (DNMTi) have been shown to induce expression of suppressed endogenous retroelements (ERV), resulting in induction of IFN (3, 4), and expression of neoantigens (5, 6), both of which can stimulate immune responsiveness. Radiation and DNA-damaging chemotherapy have also been shown to promote tumor immunity via cGAS/STING-mediated induction of IFN (reviewed in ref. 7), which is sometimes accompanied by derepression of ERVs (8, 9). Thus, desilencing ERVs and triggering the activation of type I/III IFNs in tumor cells could represent a promising strategy to augment existing immune therapies.

Because IFN signal transduction is well characterized, several screens have been described taking advantage of well-defined STAT-binding DNA sequences driving artificial promoter-reporters to screen for activators of type I/III IFNs (10–16). Only one of these screens (13) tested responses longer than 24 hours. However, drugs that act through epigenetic or indirect mechanisms (such as DNMTis), could take longer than 24 hours to induce IFN, and might require interactions with chromatin of target genes in their native genomic context.

To identify novel activators of the innate immune response, we created a reporter cell line designed to capture inducers of the type I/III IFN pathway in the endogenous genomic context. We used it to screen a library of 1,443 well-annotated compounds, and discovered that Aurora kinase inhibitors (AURKi) were strong inducers of IFN. This activation was dependent on the innate RNA-sensing pathway components MAVS and RIG-I, and independent of STING in lines with cGAS deficiency, but dependent on STING in cells with wildtype (WT) STING signaling. Induction of ERVs was found to follow AURKi treatment. We also observed induction of IFN after AURKi administration in tumors grown in mice. The ability of AURKi to slow tumor growth depended partly on an intact immune compartment, suggesting that AURKi treatment augments antitumor immunity.

All cell lines were short tandem repeat–authenticated cell lines purchased directly from ATCC, expanded, and stock vials frozen down within two passages of receipt. Cells were grown in 10% FBS/RPMI, renewed from fresh stocks after 1 month in culture, and were confirmed negative for Mycoplasma monthly. Reagents for cell culture and lipid-based transfection were from Thermo Fisher Scientific. Reagents for nucleofection were from Lonza. Nucleic acid purification kits were from Qiagen. RNA sequencing (RNA-seq) library prep kits were from New England Biolabs. Oligonucleotides, Cas9 protein, guide cRNAs, and trans-activating CRISPR RNA (tracrRNAs) were from IDT, single-guide RNAs were from Synthego. Selective antibiotics were from Invivogen. IFNα (universal type I IFN) and IFNβ was from PBL Assay Science, IFNγ was from Sigma, and IFNλ2 was from R&D Systems. The drug library used for the screen (the “FDA-approved drug library” catalog L-1300) contained 1,443 different inhibitors and was from Selleckchem. Individual drugs were either from Selleckchem or MedChemExpress. Nocodazole was from Sigma.

To create the IFI27 reporter, we synthesized homology arms flanking the stop codon, either 814 bp upstream or 747 bp downstream of the stop codon, of the human IFI27 gene (Genscript). We introduced a silent C→T mutation located 28 bp upstream of the IFI27 STOP codon in the left homology arm, to destroy the PAM site following the guide sequence (bottom strand): CCTCGCAATGACAGCCGCAA. Between the homology regions, in place of the natural stop codon, a P2A self-cleaving peptide sequence (GSGATNFSLLKQAGDVEENPGP; ref. 17) was added, followed by the Clover EGFP gene with a C-terminal 3X SV40 nuclear localization signal, followed by a T2A self-cleaving peptide sequence (GSGEGRGSLLTCGDVEENPGP), followed by the nanoluciferase sequence, modified to contain the secretion leader peptide from the IL6 gene, and finally a new stop codon. Outside the right homology arm, we added the HSV-TK gene driven by a mini-TK promoter to enable negative selection against random integrants. The CRISPR RNA (crRNA) for the IFI27 guide sequence (CCTCGCAATGACAGCCGCAA) was annealed with tracrRNA to comprise the two-component guide RNA (gRNA), and transfected along with the donor plasmid into an HCT116 cell line with stable Cas9 expression (18). The correctly integrated reporter was verified by PCR of genomic DNA and by RT-PCR for the hybrid IFI27 transcript after IFNα treatment, then positive cells were enriched by 2 weeks 10 µg/mL ganciclovir selection. Cells were then treated 24 hours with IFNα and GFP+ cells isolated for single-cell cloning. The clone that had the highest induction of GFP/luciferase, and correct DNA sequence showing integration of the reporter into all copies of the IFI27 locus, was used for all subsequent studies.

The HCT116-IFI27 reporter line was plated in 384-well plates, at 2,500 cells per well in 50 µL/well Fluorobrite DMEM with 10% FBS. The next day, drug library was added to a 2.5 µmol/L final concentration and cells were cultured for 7 days, then half the volume of cell supernatants was transferred to fresh plates for luciferase assay, while the remaining plate contents were subjected to Cell Titer Glo viability assay (Promega). For experiments with individual drugs, cells were incubated with drugs for 5 days prior to assay unless otherwise noted.

For the assay of conditioned medium from various cell lines using the HCT116-IFI27 reporter, medium from drug-treated cells was diluted 50% and placed on the reporter cells for 24 hours, which is sufficient time to detect a direct IFN response, but not long enough for the residual drugs in the conditioned medium to themselves activate the reporter.

RNAi to Aurora kinase A and B was done using OTP smartpools from GE/Dharmacon, or using the individual siRNAs from these pools (Supplementary Table S1). The PPIB housekeeping gene was used as a negative control in some experiments, and the non-targeting control #2 was used in others. Transfections were done using RNAiMAX, with 10 nmol/L each siRNA or siRNA pool, and samples collected at 72 hours posttransfection.

CRISPR gene knockout (KO) was accomplished by lentiviral infection of guide expressing vectors in the HCT116-IFI27 reporter cells, which stably express Cas9. The plasmid used for stable expression of gRNAs was described previously (18). Guide sequences were taken from the published Avana library sequences (19) or designed using the online CRISPOR tool Supplementary Table S1). In cell lines that did not have stable Cas9 expression, CRISPR KO was accomplished by transfection of RNP complexes as described (20). KO validation was done by Western blotting or by ICE analysis of PCR amplicons of target sites (Synthego; https://ice.synthego.com/#/).

RNA was extracted using the Qiagen RNeasy kit with DNAse digestion. For RNA-seq, libraries were prepared using the NEBNext Ultra II Directional RNA Library Prep Kit for Illumina. RNA-seq analysis was executed and visualized using an in-house, web-based platform, in which sequencing quality control was performed using FastQC (v0.11.5). Transcript expression was then quantified using Salmon (ref. 21; v0.9.1) in pseudoalignment mode, without adapter trimming, producing transcript-per-million (TPM) estimates, using GRCh38 for human reference transcriptomes. Differential expression analysis was performed in R using the Sleuth package (ref. 22; v0.29.0), producing gene-level effect sizes and Q-values for each comparison. The list of genes ranked by q-value was then used to perform gene set enrichment analysis (GSEA; refs. 23, 24). Hierarchical clustering and principal component analysis were also performed on the gene-by-sample matrix using built-in R functions. For qPCR, cDNA was prepared using the Bio-Rad iScript kit and targets amplified using Taqman Gene Expression 2X master mix along with the manufacturer recommended, predeveloped Taqman Gene Expression assay primer/probe mix (Thermo Fisher Scientific) for each target.

For S10-H3 phosphorylation, HCT116-IFI27 cells were treated with 200 ng/mL nocodazole and Aurora kinase inhibitors, and lysates prepared 24 hours later. For activity of individual Aurora siRNAs on Aurora kinase expression and phosphorylation, cells were transfected with siRNAs 72 hours prior to harvest, and treated as indicated with 200 ng/mL nocodazole 24 hours prior to harvest. Cell lysates were prepared in RIPA with protease/phosphatase inhibitors (Pierce) followed by sonication. Lysates were run on NuPAGE 4%–12% gels (Thermo Fisher Scientific) and transferred to nitrocellulose membranes. After 1–2 hours blocking, blots were probed overnight with primary antibodies (Supplementary Table S1). Secondary antibodies conjugated to fluorescent IR800 or IR680 (Licor) or horseradish peroxidase (Bio-Rad) were used at the manufacturer recommended dilutions.

To evaluate the EGFP response to IFN, HCT116-IFI27 reporter cells were treated with IFNα for 24 hours. Cells were then harvested and analyzed on a Fortessa X-20 flow cytometer (BD).

For evaluation of γ-H2AX and micronuclei, HCT116-IFI27 or HT29 cells were treated for 5 days with the indicated amounts of alisertib or barasertib, then fixed with 4% paraformaldehyde and stained with anti-phospho-H2AX Ser139 (Cell Signaling Technology) and anti-rabbit Alexa 568 secondary antibody, followed by Hoechst 33342 (both from Life Technologies). Cells were imaged using an Opera Phenix instrument using 100 ms exposure time and 20% power, and quantified using built-in Harmony software. The sum of means per nucleus were plotted for γ-H2AX and the total area of micronuclei per well were determined using the Harmony micronuclei detection function using the Hoechst channel.

HCT116-IFI27 reporter cells were plated at 2.3 × 10⁵/well of 6-well plates. The next day, drugs were added (alisertib 1 µmol/L, barasertib 100 nmol/L, or decitabine 100 nmol/L) or DMSO. Medium was changed and drugs refreshed at day 3. At day 5, cells were harvested and genomic DNA extracted with a Qiagen DNeasy kit. The Illumina Infinium MethylationEPIC array BeadChip (850K) and subsequent bioinformatics analysis was carried out by Epigenomic Services at Diagenode (catalog no. G02090000).

All animal studies were reviewed and approved by each institution's Institutional Animal Care and Use Committee prior to conduct. Care and use of animals was in accordance with the regulations of the Association for Assessment and Accreditation of Laboratory Animal Care. The HCT116 xenograft study was performed by WuXi. Athymic nude mice (Charles River Laboratories), ages 6–8 weeks, were implanted subcutaneously in the flank with 2 × 10⁶ HCT116 cells suspended 1:1 in matrigel. When tumors reached 100 mm³, mice were randomized to groups, and 10 were treated by intraperitoneal injection with 0.5 mg/kg decitabine, daily for 5 days, with a 2-day break, then daily for another 5 days. The remaining 20 mice were untreated for 5 days, then 10 mice were treated orally daily with vehicle alone (10% 2-hydroxypropyl-β-cyclodextrin and 1% sodium bicarbonate) while the other 10 were treated vehicle containing alisertib at 30 mg/kg, for 7 days. Tumors were measured every 2 days and mice weighed twice weekly. 1–2 hours after the final dose, mice were sacrificed, tumors flash frozen, and plasma collected for pharmacokinetic analysis.

For experiments with CT26 cells, BALB/cJ female mice, either WT or NSG (NOD.Cg-Prkdc^scidIl2rg^tm1Wjl/SzJ) strains, 8–10 weeks of age were used (Jackson Labs). CT26 cells were inoculated subcutaneously in the flank at 10⁶ cells per injection. Once tumor sizes reached an average of about 50 mm³, mice were randomized to the different groups, and dosing initiated with indicated doses of alisertib or decitabine, or vehicle only, by gavage daily. Tumors were measured with calipers and mice weighed twice weekly. At 14–16 days of treatment, tumors were harvested and flash-frozen or formalin-fixed for paraffin embedding. Tumor samples were sent to HistoBridge for embedding, sectioning, and CD8 staining and quantification.

In vitro experiments were performed at least three times and a representative experiment is shown. Data are presented as the means ± SDs. Two-tailed Student t tests using equal variance were performed in Excel, or for more complex datasets, two-way ANOVA in GraphPad Prism to assess significance between the comparisons indicated in each figure. Asterisks are used to indicate statistical significance [n.s., or no callout, statistically nonsignificant (P > 0.05); *, P < 0.05; **, P < 0.01; ***, P < 0.001].

RNA-seq data will be made available upon request to the corresponding author.

To create an endogenous type I IFN reporter, we chose the colorectal cancer cell line HCT116, which is highly efficient for CRISPR-based gene editing, and surveyed IFNα induction of canonical type I IFN target genes (IFIT1, IFIT2, IFIT3, IFI27, IRF7, OASL, RSAD2; Fig. 1A). We found that the IFI27 gene was highly induced, with low baseline expression and high transcript levels after IFN treatment. RNA-seq confirmed IFI27 is the second most highly induced IFNα response gene in HCT116 (Supplementary Table S2). We targeted the 3′ end of the IFI27 coding region and inserted a nuclear-localized Clover GFP and a secreted nanoluciferase gene immediately before the stop codon (Fig. 1B). The reporter genes are separated from the IFI27 gene by 2A sequences, allowing generation of intact proteins from a single transcript that retains all regulatory sequences intact.

Single-cell clones were verified for induction GFP and luciferase with IFNα treatment, then one clone, with genomic integration into all copies of the IFI27 locus and similar morphology as the parent cell line, was selected for the subsequent experiments in this study. Induction of GFP could be detected starting at around 6 hours and was observed in most cells by 24 hours (Supplementary Fig. S1A–S1C). The IFNα response could also be measured by luciferase assays of cell supernatants and was completely inhibited by the JAK inhibitor ruxolitinib (Supplementary Fig. S1D). Induction of the reporter could also be seen in response to IFNβ and IFNλ2, a type III IFN (Supplementary Fig. S1D). Type II IFN could activate the reporter but was less effective, consistent with its reduced ability to induce the IFI27 gene (Supplementary Fig. S1E).

Because DNMTis have been shown to activate IFN signaling (3, 4), we tested the IFI27 reporter for DNMTi responsiveness. We found that the reporter was induced by DNMTis 5-azacytidine (5-Aza) and decitabine (Fig. 1C). Consistent with published reports and the mechanism of action of these drugs, which require cell division to incorporate into DNA, the induction of the reporter in response to DNMTis was slow relative to IFN treatment itself, barely detectable at 2 days but easily detectable by 5 days.

We treated our reporter cell line with a library of 1,443 well-annotated bioactive compounds and, to ensure sufficient time for epigenetic changes to elicit a phenotype, assayed after 7 days of treatment. The assay yielded a small number of hits that were not correlated with effects on viability (Fig. 1D). The strongest scoring hits fell into a few categories (Fig. 1E). Most of these drugs (inhibitors of DNMT, MEK, CDK, DNA synthesis, and topoisomerase) have been previously shown to induce type I/III IFN signaling (3, 4, 25–27), indicating that the screen worked properly. In addition, we saw reporter activation in response to inhibitors alisertib (modestly Aurora A selective), barasertib (Aurora B selective), and tozasertib (pan-Aurora). Kinetics of the reporter activation showed that, like with DNMTis, signal after AURKis or MEKis was low at 48 hours, and increased with time (Fig. 1F). The induction of IFN signaling by Aurora kinase inhibition had not, to our knowledge, been reported previously.

To determine whether AURKis were activating IFN signaling more generally, we performed Western blot analysis to confirm STAT1 phosphorylation in HCT116-IFI27 cells after alisertib treatment (Fig. 2A). This phosphorylation was similar in magnitude and kinetics to that induced by the positive control decitabine. Both compounds acted more slowly than direct treatment with IFN itself, as expected. To further confirm the activation of IFN signaling, we performed qPCR of canonical type I/III IFN target genes IFI27, IFIT1, IFIT3, and CXCL10, as well as IFNβ, and IFNλ1. We observed robust activation of all these target genes after alisertib treatment (Fig. 2B).

We tested Aurora isoform-selective compounds in the reporter line to assess the respective roles for Aurora kinase A or B in IFN signal induction. We found that Aurora B selective compounds were highly potent, whereas Aurora A selective compounds were less potent (Fig. 2C). Alisertib, which preferentially inhibits Aurora A, also inhibits Aurora B at higher doses (28), and activated the reporter to a similar extent as Aurora B selective compounds. The activity of higher doses of alisertib against Aurora B was confirmed by Western blotting showing inhibition of H3-S10 phosphorylation, an Aurora B target, by alisertib, but not by the highly Aurora A selective compound MK8745 (Fig. 2D). The effect of Aurora B inhibition on the reporter was blocked by cotreatment with the pan-JAK inhibitor ruxolitinib, confirming that IFN induction was responsible for reporter activation, not direct action of Aurora inhibition on the IFI27 promoter (Fig. 2E).

As an orthogonal approach to targeting Aurora A versus Aurora B with small-molecule inhibitors, we also performed siRNA-mediated knockdowns. Transfected pooled siRNAs showed that activation of the reporter is associated with knockdown of Aurora B and not Aurora A (Fig. 2F and G), despite the induction of the enlarged cell phenotype typical of loss of either Aurora's function (Fig. 2G). As with Aurora small-molecule inhibitors, activation of the reporter by siRNA to Aurora B was inhibited by ruxolitinib (Supplementary Fig. S2). To better visualize Aurora kinases and activation-specific phosphorylation at the protein level, we transfected HCT116-IFI27 cells with individual siRNAs against Aurora A or B, in the presence of nocodazole, which elevates AURK expression (29), and performed Western blot analysis. While both Aurora A and B were present, only Aurora B was detectably phosphorylated (Fig. 2H). Interestingly, knockdown of Aurora B also led to reduction of Aurora A. Whether this reduction is due to a dependence of expression of Aurora A on Aurora B, or cross-reactivity of the individual Aurora B siRNAs to the Aurora A sequence is not clear, although the sequence identity for each Aurora B siRNA for Aurora A is quite low (Supplementary Table S1). Reporter assays showed that the ability of individual Aurora siRNAs to induce IFN signaling corresponded to reduction in Aurora B protein (Fig. 2I). We conclude that the activation of IFN by AURKi requires Aurora B.

To better characterize gene expression changes occurring in response to AURKis and selected other hits from the screen, we performed RNA-seq analysis of the reporter line after 5 days drug exposure (Supplementary Table S3). Gene expression patterns clustered into three groups (Fig. 3A). Inhibition of DNMTs by decitabine and 5-Aza represented Group 1; Aurora kinase inhibitors alisertib and tozasertib formed a second group that included other hits tested, foretinib, axitinib, and irinotecan; and MEK inhibitors binimetinib and selumetinib formed a third group. These results suggest that the IFN induction in response to AURKis may not be a direct effect of inhibiting Aurora kinase, but rather the state of the cells resulting from growth arrest or senescence. Cell morphology was concordant with the drug treatment categories identified within the RNA-seq analysis; cells that clustered with the AURKis in Group 2 were much bigger, with enlarged and abnormal looking nuclei compared with either DNMTi- (Group 1) or MEKi-treated (Group 3) cells (Supplementary Fig. S3A).

Many of the top genes induced by alisertib treatment were IFN pathway–related genes (Fig. 3B; Supplementary Table S3). GSEA of AURKi-treated cells revealed strong enrichment for the MSigDB hallmark IFNα response pathway (Fig. 3C); it is the most strongly enriched among all the hallmark enriched gene sets in both AURKi treatments (Fig. 3D; Supplementary Fig. S3B). This pathway is also enriched by decitabine (Fig. 3D), but less strongly than by AURKis. GSEA of the other drugs in Group 2, as well as Group 3, also showed strong enrichment of the hallmark IFNα pathway (Supplementary Fig. S3C–S3G).

To establish that IFN induction in response to alisertib is not limited to the reporter cell line, we performed RNA-seq on other colorectal cancer cell lines (LoVo, Ls174T, HT29, SW480, SW1417, and Ls123) treated with alisertib or decitabine (Fig. 3E; Supplementary Table S4). We observed that enrichment of the hallmark IFNα gene set was highly significant in most cell lines, and did not clearly track with microsatellite instability or Ras mutation status; although WT p53 was associated with all but HT-29. Within this panel, the IFN response is more frequently observed with AURKi than with DNMTi (Fig. 3E). In addition, we observed a modest (3-fold) increase in PD-L1 gene expression in HT29 which is consistent with a recent report on Aurora A kinase inhibition (30). However, we did not observe a significant change in PD-L1 expression in six of the seven cell lines tested (Supplementary Tables S3 and S4). We confirmed induction of IFN expression by qPCR of IFNβ, IFNλ1, IFNλ2, and IFIT3 (Fig. 3F). We also found that the mouse syngeneic colorectal cancer cell line CT26 exhibited an IFN response to AURKi treatment (Fig. 3E and G).

We next interrogated the signaling pathways required for this response. RNA expression data indicated that Aurora inhibition caused increases in IFNλ1, IFNλ2, and IFNλ3 at the transcript level, as well as IFNβ (Fig. 2B; Supplementary Table S3). We were unable to detect IFNs in the conditioned medium of drug-treated cells using commercial ELISA kits. However, we were able to detect active IFN in conditioned medium from drug-treated parental HCT116 cells using the HCT116-IFI27 reporter. To determine the contributions of type I versus type III IFN receptors to this signal, we generated HCT116-IFI27 reporter lines with CRISPR KO of either IFNAR1 or IFNLR1. Validated KO lines (Supplementary Fig. S4A) were then used to assay conditioned medium from drug-treated cells, to show that the activation of the reporter is dependent on IFNLR1 and independent of IFNAR1 (Fig. 4A). However, direct assay of AURKi induced IFI27-luc activity of IFNAR1 or IFNLR1 KO reporter cell lines themselves indicated KO of either IFNAR1 or IFNLR1 reduces reporter signal, suggesting both contribute (Fig. 4B). Because these cells do induce IFNβ at the transcript level, we speculate that the quantity of type I IFN in the conditioned medium may be too low to activate the reporter. We can conclude that active IFNs are made by cells in response to AURKi and they are responsible for IFI27 reporter activation.

Previous studies have shown the STING pathway is required for inducing the IFN response to radiation (31), chemotherapeutic drugs (32), topoisomerase inhibitors (33), DNA mismatch repair deficiency (34), and senescence (35). Other studies have shown that the MAVS pathway mediates the IFN response to decitabine (3, 6), as well as irradiation (8). We observed morphologic changes, including micronuclei formation and increased γ-H2AX staining, in HCT116-IFI27 (Supplementary Fig. S4B–S4F) and HT-29 (Supplementary Fig. S4G–S4K), in response to AURKi, suggesting possible involvement of DNA damage induced STING pathway activation. To explore which pathways may be required for AURKi to induce IFN, we made CRISPR KOs of JAK1, RELA, IRF3, MAVS, TMEM173, and TLR3 in the HCT116-IFI27 reporter cell line, verified knockdown of the target proteins (Supplementary Fig. S4L–S4Q) and measured the response to alisertib, as well as decitabine or IFNα for most lines. The reporter response to IFNα was dependent on JAK1, as expected (Supplementary Fig. S4R). The response to alisertib treatment was dependent on IRF3, JAK1, RELA, and MAVS, but independent of STING (Fig. 4B and C). In contrast, the response to decitabine was only partially dependent on JAK1, RELA, or MAVS, but also independent of STING (Supplementary Fig. S4S). This dependence on MAVS but not STING was recapitulated when using MAVS or STING KO single-cell clones rather than KO pools (Supplementary Fig. S4T and S4U). Barasertib, DNMTis, foretinib, and axitinib (other drugs from Group 2 in the RNA-seq study; Supplementary Fig. S4T–S4V) but not IFN (Supplementary Fig. S4W), were also dependent on MAVS and independent of STING. RNA-seq analysis examining the transcriptional response of MAVS KO clones to alisertib (Fig. 4D; Supplementary Table S5) revealed very few alisertib-mediated gene expression changes inhibited by MAVS KO (Fig. 4D, boxed). These were largely IFN target genes, and most of the genes from the MSigDB hallmark IFNα geneset that are induced by alisertib are no longer induced in the MAVS KO clones (Fig. 4E), suggesting that MAVS is required specifically for the IFN response, but not the overall response, to Aurora kinase inhibition.

The dependence of Aurora kinase inhibition on MAVS instead of the STING pathway to induce IFN is somewhat surprising, because Aurora kinase inhibitors have been documented to cause DNA damage (36), which has been reported to activate IFN via the cGAS/STING pathway (31–33, 37). While we demonstrated expression of STING in HCT116 when we validated the STING KO, we found that HCT116 cells failed to activate IFN signaling in response to STING agonist (Supplementary Fig. S4X and S4Y). Literature investigation revealed that HCT116 cells, as well as several other colorectal cancer lines, are reported to have deficiency in cGAS expression (38), and cannot activate STING. Our panel of colorectal cancer lines includes some lines with reported intact cGAS/STING signaling (HT-29, LoVo, CT26) and some with reported cGAS deficiency (HCT116, Ls174T, and SW1417; refs. 38, 39).

To determine whether the response to alisertib was MAVS dependent in colorectal cancer cells, we made CRISPR KOs of MAVS or STING in HT-29, CT26, Ls174T, and SW1417 cell lines (Supplementary Table S6) and assayed for IFN pathway induction in response to AURKis. We found that the cell lines with cGAS deficiency (SW1417, Ls174T; ref. 38) were dependent on MAVS for the IFN response (Fig. 4F and G), whereas cell lines with functioning cGAS/STING signaling (HT-29, CT26) exhibited STING dependency, but had reduced or no MAVS dependency (Fig. 4H and I).

MAVS is activated by binding to the cytosolic double-stranded RNA (dsRNA) sensor proteins RIG-I or MDA5, or the endosomal dsRNA sensor TLR3 (40). The IFN response to doxorubicin (41) and decitabine has been shown to depend on TLR3 (3) or by MDA5 and RIG-I (4), suggesting cellular context likely helps specify relative roles of the RNA-sensing pathways. Because we did not observe dependency on TLR3, we performed further CRISPR KO analysis to investigate the contributions of RIG-I and MDA5. We found the ability of alisertib to induce IFN signaling relied on RIG-I, but was independent of MDA-5 (Fig. 4J). Together, these data indicate that Aurora kinase inhibition induces IFN signaling by triggering the MAVS/RIG-I innate immune RNA-sensing pathway, which act through IRF3 and NFkB to induce type III (and to a lesser extent type I) IFN, which in turn activates the cognate IFN receptors to execute the IFN gene expression response via JAK1.

We next considered the source of AURKi-induced dsRNA that activates RIG-I/MAVS. One source might be ERVs, which have been shown to mediate the IFN response to decitabine treatment (3, 4), SETDB1 loss (42), and senescence (43). To determine whether ERV upregulation was associated with the induction of IFN by AURK inhibition, we used TEtranscripts (44) to quantify transposable element (TE) subfamily expression from our RNA-seq data (from Fig. 3C). As expected, DNMTis caused a strong upregulation of TEs, led by the HERV9-int subfamily (log₂FC = 4.7 and 4.8 for 5-Aza and decitabine, respectively, see Supplementary Table S7). All Group 2 compound treatments, which includes the AURKis alisertib and tozasertib, broadly resulted in an upregulation of TEs, exemplified by the MER65C subfamily (log₂FC > 1.9 in all Group 2 conditions). Interestingly, the three compound groups induced distinct subfamilies of TEs (Fig. 5A) where Group 2 compounds tended to cluster together, suggesting a conserved mechanism of TE upregulation shared across these compounds distinct from that of DNMTis. Next, we investigated whether upregulation of TEs by AURKis might be associated with loss of genomic DNA methylation, like DNMTis (3). We observed that decitabine treatment resulted in a profound loss of DNA methylation (n = 566,352 of 723,483 peaks with Padj < 0.05), while both alisertib and barasertib resulted in zero significantly (Padj < 0.05) methylated peaks (Supplementary Fig. S5A and S5B), consistent with a distinct mechanism. Importantly, alisertib resulted in similar TE upregulation in both MAVS KO cells and control clones even though the IFN signature was not triggered in the absence of MAVS (Fig. 5B). Thus, AURKis and other Group 2 drugs likely activate IFN signaling via derepression of retroelements in a manner distinct from loss of DNA methylation. Histone methylation, such as H3K9 and/or H3K27 is another established mechanism for ERV inhibition (45). We looked at global H3K9me3 and H3K27me3 levels after Aurora inhibition or siRNA, but were not able to detect reductions in the levels of these marks by Western that are consistent with IFN induction (Supplementary Fig. S5C and S5D).

To assess whether any well-characterized chromatin modifiers were systematically upregulated or downregulated after alisertib treatment, we reanalyzed our RNA-seq data with a focus on 429 well-characterized chromatin modifiers (46) in the five cell lines that showed an upregulation of IFN pathway genes (Supplementary Table S8). No genes scored as significant (q < 0.05) across all five cell lines, and only one (IDH1; downregulated upon treatment) scored as significantly differentially expressed across all cell lines with a more relaxed significance threshold (q < 0.1). Because loss of IDH1 would drive an increase of DNA and histone methylation by elevating D-2-Hydroxyglutarate levels and thus promote silencing (47), we reason that this is unlikely to be the driving mechanism behind ERV desilencing upon alisertib treatment.

Because AURKis have a mechanism for activating ERVs that is distinct from DNMTis, we tested combination treatments in the HCT116-IFI27 reporter. We observed strong synergy with decitabine (Fig. 5C; Supplementary Fig. S5E), whereas we observed either no or additive interaction between AURKis and DNA-damaging agents irinotecan (Fig. 5D; Supplementary Fig. S5F), etoposide, or 5-fluorouracil (Supplementary Fig. S5G–S5J). These data are consistent with AURKis acting on a similar pathway as DNA damaging agents but separate from DNMTis.

We next wanted to determine whether AURKi induced IFN in tumors. Alisertib or decitabine inhibited tumor growth of HCT116 cells grown as xenografts (Fig. 6A). Both IFNβ and IFNλ1 were significantly elevated in RNA prepared from tumors after alisertib administration, albeit to a lesser extent than after administration of decitabine (Fig. 6B). To address the consequences of IFN induction in the context of an intact immune system, we examined alisertib activity in a syngeneic model, CT26, which induces IFN signaling in response to alisertib treatment in vitro (Fig. 3E and G). Treatment of CT26 mice led to reduction of tumor growth (Fig. 6C) and increased expression of Ifnβ and IFN target gene expression (Fig. 6D).

To examine the dependency of alisertib on the adaptive immune system to affect tumor growth, we implanted CT26 cells into either WT Balb/c or NSG mice. We found that alisertib was ineffective at blocking tumor progression in the NSG mice, as opposed to WT mice (Fig. 6E). This result contrasts with results reported recently (30), but is consistent with another study (48). We could see a statistically significant induction of Ifnβ expression in either WT or NSG background after alisertib treatment, albeit at different timepoints (Fig. 6F). Levels of CD8⁺ T-cell infiltration of the tumors in WT hosts were found to be elevated in alisertib treated mice, consistent with alisertib and IFN induction, resulting in enhanced CD8⁺ T-cell recruitment (Fig. 6G).

CT26 cells induced IFN in vitro in a Sting-dependent manner (Fig. 4I). To determine whether Sting was required for the antitumor efficacy of alisertib and induction of IFN in vivo, we generated non-targeting control (NTC) control or Sting KO clones from CT26 cells (Supplementary Table S6), verifying Sting dependence of Ifnβ induction after alisertib treatment, and modest effects on cell viability (Supplementary Fig. S6A and S6B). When inoculated into WT mice, we confirmed that alisertib inhibited growth in NTC control tumors, but not in Sting KO tumors (Fig. 6H). Analysis of IFN pathway genes demonstrated elevated IFN signaling in tumor cells growing in vivo relative to in vitro. Moreover, a trend for increased IFN signaling in WT cells following alisertib treatment could be observed, which was not seen in the Sting KO cells (Fig. 6I). Taken together, the results suggest that Aurora kinase inhibition can activate IFN signaling in vivo and inhibit tumor growth in a manner at least partially dependent on the immune system and Sting signaling.

To find new activators of IFN signaling, we developed a screening platform that reports on the endogenous expression of a highly type I/III IFN activated gene, IFI27, and ran our screen for a sufficient time to capture epigenetic changes. Aurora kinase inhibitors were hits in our screen that had not previously reported to activate the IFN pathway. Indeed, one report (49) specifically tested whether AURKis could activate IFN, using a commercially available engineered cell line with an IFN response element driving a luciferase reporter, and observed no IFN induction response. This finding substantiates the utility of our CRISPR-based reporter, with its high degree of signal strength and structural authenticity that distinguish it from other IFN reporters.

RNA-seq analyses of the screen hits we tested revealed three general groups. DNMTis and MEKis formed distinct groups, while AURKis grouped with the other drugs tested: foretinib, axitinib, and irinotecan (“Group 2”). Group 2 drugs all generated an enlarged cell phenotype, and induce cell cycle arrest in the G₂–M phase (50–53). MEKis arrest cells in G₁ (54, 55) whereas DNMTis, at the doses we used, are not expected to cause significant changes in cell cycle distribution (56). Thus, the three expression profile groups are associated with distinct effects on cell cycle progression. The G₂–M cell cycle arrest seen in Group 2 drugs are associated with defects in cell division that lead to DNA damage, aneuploidy and micronuclei (52, 57), which we observed (Supplementary Fig. S4C–S4F, S4H–S4K). Micronuclei can trigger cGAS-STING signaling (58); however, no direct connection between micronuclei and the RIG-I/MAVS pathway has yet been established. In contrast, DNA damage from irradiation has been shown to activate type I IFN signaling in a MAVS/RIG-I dependent manner; however, in this case, RIG-I was triggered by small noncoding RNA (sncRNAs) rather than ERVs (8). DNA demethylation cannot account for the induction of ERVs after AURKi treatment, and thus far we have not observed clear reductions in the global levels of ERV repression associated histone marks H3K9me3 or H3K27me3 that are consistent with induction of IFN. It is possible that such differences might become more obvious at the local chromatin level, and this question remains to be addressed. Currently, we speculate that the disordered state of cellular chromatin induced by treatment with Group 2 drugs, by interfering with DNA replication and cell division, results in failure to repress endogenous retroelements, and the compromised integrity of the nuclear membranes of G₂–M arrested cells allows escape of these dsRNAs into the cytoplasm where they are detected by RIG-I.

The unexpected dependence of AURKis on the MAVS pathway may represent a novel vulnerability of tumors deficient in STING signaling. Indeed, some tumors have been shown to repress the cGAS/STING pathway as a means of immune escape (reviewed in ref. 37), and such tumors have been documented at a significant frequency in colorectal cancer (38). Therefore, we were interested to determine whether Aurora kinase inhibitors exhibited dependency on MAVS in vivo. However, none of the syngeneic models we examined (CT26, B16, 4T1, MC38, LLC) exhibited IFN induction in response to AURKi, or MAVS dependence (Fig. 4I; Supplementary Fig. S6C–S6F). We were able to confirm that AURKis could induce IFN in vivo, and that tumor growth inhibition by alisertib depended on immune system involvement, consistent with the hypothesis that AURKi-mediated IFN induction and ERV expression created a more favorable environment for immune recognition. Other investigators have also observed dependence of alisertib on the immune system to inhibit growth in syngeneic tumor models. Alisertib has been previously shown to increase CD8⁺ infiltration into CT26, MC38, and 4T1 syngeneic model tumors (48, 59).

The role of STING in the tumor compartment has been somewhat controversial. Some groups find that only the production of cGAS, but not the STING response, by the tumor cells is important; tumor-produced extracellular cGAS activates STING in immune cells in the tumor microenvironment (39, 60). Others found that STING signaling within the tumor compartment itself is required for inhibition of tumor growth in response to chemotherapeutics or radiation (31, 32). Perhaps STING signaling mediates antitumor effects in both tumor and immune compartments. In our study, we observed that STING expression in the tumor compartment was necessary for alisertib to inhibit tumor growth. STING expression also supported a higher level of IFN pathway activation in the tumors. Relieving this activation by deleting STING did not have a significant effect on tumor growth by itself. However, Sting was necessary for alisertib to exert tumor growth inhibition in this context.

The work described here shows that there are multiple mechanisms by which damaged cancer cells can activate the antiviral response and IFN signaling, not just the DNA-sensing pathway. The idea that our IFN reporter screen could identify drug candidates that may successfully combine with checkpoint blockade was substantiated by the approval of one of the hits from our screen for use in combination with immune checkpoint blockade (axitinib; KEYNOTE-426; NCT02853331). Additional screening with our reporter system could yield novel compounds to foster an immunogenic tumor microenvironment.

L. Choy reports other from Calico Labs, LLC during the conduct of the study. G. Kolumam reports other from Calico Labs, LLC during the conduct of the study. A. Firestone reports other from Calico Life Sciences LLC during the conduct of the study. D. Stokoe reports other from Calico LLC during the conduct of the study. No disclosures were reported by the other authors.

L. Choy: Conceptualization, investigation, visualization, writing-original draft, writing-review and editing. S. Norrsi: Resources, data curation, software, visualization, methodology, writing-review and editing. X. Wu: Investigation, methodology. G. Kolumam: Formal analysis, supervision. A. Firestone: Formal analysis, supervision. J. Settleman: Conceptualization, formal analysis, supervision, writing-review and editing. D. Stokoe: Formal analysis, supervision, visualization, writing-review and editing.

The authors would like to thank Matt Pech for the HCT116 cell line expressing Cas9, Ben Haley for advice on CRISPR editing, Chirag Patel and Jonathan Paw for assistance with the in vivo studies and flow cytometry, Ismail Sergin for critical reading of the article and the γ-H2AX staining protocol, Timothy Hoffman for assistance with the γ-H2AX immunofluorescence staining quantification, Rob Keyser for help with dispensing the compound library, David Hendrickson for helpful discussions, and the Calico genomics lab for RNA-seq library preparation and sequencing. The authors would like to acknowledge that the top panel of Figure 4A was created with Biorender.com.