Replication stress response defects are associated with response to immune checkpoint blockade in non-hypermutated cancers

Abstract

Treatment with immune checkpoint blockade (ICB) has resulted in durable responses for a subset of patients with cancer, with predictive biomarkers for ICB response originally identified largely in the context of hypermutated cancers. Although recent clinical data have demonstrated clinical responses to ICB in certain patients with non-hypermutated cancers, previously established ICB response biomarkers have failed to accurately identify which of these patients may benefit from ICB. Here, we demonstrate that a replication stress response (RSR) defect gene expression signature, but not other proposed biomarkers, is associated with ICB response in 12 independent non-hypermutated cohorts of patients with cancer across 7 tumor types, including those of the breast, prostate, kidney, and brain. Induction or suppression of RSR deficiencies was sufficient to modulate response to ICB in pre-clinical models of breast and renal cancers. Mechanistically, we find that despite robust activation of checkpoint kinase 1 (Chk1) signaling in RSR-deficient cancer cells, aberrant replication origin firing causes exhaustion of replication protein A, resulting in accumulation of immunostimulatory cytosolic DNA. We further found that deficient RSR coincided with increased intratumoral dendritic cells in both mouse cancer models and human tumors. Together, this work demonstrates the RSR defect gene signature can accurately identify patients who may benefit from ICB across numerous non-hypermutated tumor types, and pharmacological induction of RSR defects may further expand the benefits of ICB to more patients.

One sentence summary:

Endogenous and induced replication stress response defects promote tumor immunogenicity via generation of cytosolic single-stranded DNA.

INTRODUCTION

Immune checkpoint blockade (ICB) has provided robust, durable benefit to a subset of patients. Many initial ICB trials were focused on highly mutated cancer types, such as melanoma and lung cancer, largely predicated on the idea that mutation-derived neoantigens would allow for generation of tumor-specific T cells (1). Subsequent analysis of patient responses in these highly mutated cancer types confirmed that increased tumor mutation burden (TMB) corresponded with improved patient outcomes (2, 3). The extreme hypermutator phenotype caused by deficiencies in DNA mismatch repair (MMR), which can be detected by microsatellite instability (MSI), has received FDA approval as the first tissue-agnostic biomarker for ICB therapy (4), with a second pan-cancer approval of high TMB (>10 mutations/megabase) as an ICB response biomarker (5). Further clinical studies in hypermutated cancers identified additional predictive biomarkers, such as programmed death-ligand 1 (PD-L1) protein expression (6), and various gene expression signatures (7–9).

We recently demonstrated distinct tumor immunobiology between hypermutated and non-hypermutated tumor types (10), finding that relative neoantigen load/tumor mutation burden was only a relevant factor for immune infiltration in hypermutated tumor types. Consistent with this finding, clinical data have failed to detect predictive value for TMB to identify ICB responders in many non-hypermutated tumor types, such as breast cancer (11), ccRCC (12) and GBM (13), and accumulating evidence suggests TMB does not universally predict sensitivity to ICB (14–16). Moreover, PD-L1 protein expression has shown mixed results for predicting response to ICB in TNBC (17, 18), and the utility of other transcriptional biomarkers remains poorly explored. Thus, there remains a critical gap in knowledge as to how to identify which patients with non-hypermutated cancer may benefit from ICB.

Defects in DNA damage response (DDR) pathways including MMR, homologous recombination repair, and nucleotide excision repair have been associated with ICB sensitivity (19). During deoxyribonucleic acid (DNA) synthesis, the replication machinery must overcome stalling due to obstacles such as DNA lesions, ribonucleic acid (RNA) transcriptional machinery, or insufficient nucleotides, collectively known as replication stress. To overcome these obstacles and maintain genomic integrity, cells activate replication stress response (RSR), which slows DNA replication to allow for repair, thereby preventing replication fork collapse (20). In normal cells, excessive replication-associated DNA damage, such as oncogene-induced hyperproliferation, will induce senescence or apoptosis, preventing tumorigenesis (21). Deficiencies in RSR facilitate growth of pre-malignant cells by supporting escape of this oncogene-induced senescence (22).

In this study, we demonstrated how RSR deficiencies from aberrant replication origin firing lead to accumulation of immunostimulatory cytosolic single-stranded DNA. Using pre-clinical murine breast cancer models, we found that RSR defects predicted sensitivity to ICB. Sensitivity or resistance to ICB could be caused by induction or suppression of RSR defects, respectively. Application of a transcriptional RSR defect (RSRD) score (23) predicted ICB response in 12 independent patient non-hypermutated tumor cohorts across 7 different tumor types. These findings may help identify which patients will benefit from ICB and also provide the biological rationale for clinical approaches to pharmacologically induce RSR defects that may expand the benefits of ICB to a greater number of patients.

RESULTS

RSRD gene expression score predicts functional replication stress response defects across breast cancer cell lines and primary organoid cultures.

We recently developed a gene expression signature indicative of RSR defects (RSRD) using isogenic mammary epithelial model cell lines (23). We applied this RSRD gene signature to a panel of 20 human breast cancer cell lines and found robust prediction of RSR function, as measured by ability to complete S-phase following release from hydroxyurea (HU)-induced replication stress (Fig. 1A and fig. S1, A to C), with an enrichment in RSR defects in TNBC cell lines. To ensure this result was not a cell culture artifact, we further validated it in primary organoid cultures that also showed robust prediction of RSR function (Fig. 1B). Inability to recover from HU-induced replication stress was confirmed at the level of single replication forks by DNA fiber analysis (Fig. 1C). RSRD-high cells also exhibited increased basal replication stress (24), as quantified by both decreased replication fork speed (fig. S1D) and increased phosphorylation of ataxia-telangiectasia and Rad3 related (ATR) (fig. S1, E and F). Increased endogenous replication stress was supported by accumulation of single-stranded DNA (ssDNA) detected by native 5-chloro-2’-deoxyuridine (CldU) staining, which was exacerbated by HU treatment (Fig. 1D).

Fig. 1. RSRD gene expression signature predicts functional RSR defects in breast cancer cell lines and organoids.

A, RSR function assessed by fraction of cells capable of completing cell cycle after hydroxyurea (HU)-induced replication stress (57). See fig. S1, A to C. Two-tailed t-test.

B, RSR deficiency as assessed by cell cycle assay in 3 organoid models derived from PDXs (58), sorted by RSRD score. Dots represent biological replicates.

C, Quantification of fork restart following HU-induced replication stress in a panel of RSRD-low and -high cell lines (left) and representative images (right). Replicating DNA was labeled with CldU (green) before stalling replication forks with HU. After HU washout, restarted replication forks were labeled with IdU (red). Percentage of failed replication forks was determined for at least 100 CldU-labeled fibers per cell line. Scale bar indicates 10 μm. Two-tailed t-test.

D, Native CldU staining to detect ssDNA (green) in control (CTRL) cells or following 6-hour incubation with HU along with nuclear counterstain (blue). Representative cell lines shown to the right. N = 3 cell lines per condition. Two-way ANOVA.

All data are mean ± s.d. with each dot representing an individual cell line.

Replication stress response defects are caused by aberrant origin firing driving RPA exhaustion from and lead to accumulation of immunostimulatory cytosolic DNA.

We next sought to mechanistically understand the observed RSR defects. Our original model system was based on defective checkpoint activation (25), though we detected no loss of DNA damage checkpoint activation in RSRD high cell lines as measured by checkpoint kinase 1 (CHK1) phosphorylation or by global phosphorylation of ataxia-telangiectasia mutated (ATM)/ATR substrates (Fig. 2, A and B). Nucleotide deficiency has also been implicated in oncogene-induced replication stress (26); however, we found no phenotypic changes following nucleoside supplementation (fig. S2, A and B). We next analyzed DNA synthesis following replication stress and found that whereas RSR-intact cells exhibited a gradual recovery of DNA synthesis, RSR-defective cells rapidly initiated DNA synthesis but ultimately failed to complete S-phase (Fig. 2C). Based on this observation, we hypothesized that RSRD cells may exhibit aberrant DNA replication origin firing despite activation of checkpoint kinases. Our hypothesis was supported by ultraviolet (UV)-resistant DNA synthesis in RSRD-high cells (Fig. 2D). We found that aberrant origin firing in RSRD high cells led to exhaustion of replication protein A (RPA) pools (Fig. 2E and fig. S2C) and rapid accumulation of DNA double-strand breaks (fig. S2D).

Fig. 2. RSRD-high cell lines undergo RPA exhaustion from aberrant replication origin firing causing accumulation of cytosolic DNA.

A, CHK1 activation following treatment with HU. RSRD-low MCF-10A and RSRD-high MDA-MB-231 show similar amounts of CHK1 phosphorylation and CDC25a degradation by Western blot.

B, Quantification of phosphorylation of chromatin-bound ATM/ATR substrates by staining with anti-pSQ/TQ (59). Cells were incubated in 2 mM HU or vehicle control (water) for 6 hours. Dots represent individual cell lines; mean ± s.d.; inset P values determined by two-way ANOVA.

C, Re-initiation of DNA synthesis following HU-induced replication stress. Cells were incubated with 2 mM HU for 16 hours and moved to fresh growth media for indicated times. BrdU was added 15 minutes prior to fixation to identify S-phase cells. Data reported as percentage of cells that are BrdU⁺ relative to vehicle-treated control cells.

D, UV-resistant DNA synthesis. S-phase cells were labeled with BrdU prior to UV-irradiation (5 J/m²) or mock-treated. Post-UV DNA synthesis was detected by EdU incorporation. Values reported as relative EdU intensity in BrdU positive cells in UV-irradiated compared to mock-treated cells. N = 3 cell lines per condition. Two-tailed t-test.

E, Percentage of cells with RPA exhaustion in control (CTRL) cells or following 6-hour incubation with HU. See fig. S2C. N = 3 cell lines per condition. Two-way ANOVA.

F, Percentage of cells that are RSR defective measured by ability to complete DNA synthesis after HU-induced replication stress per Fig. 1A, either following suppression of origin firing with 10 μM roscovitine or mock DMSO 2 hours prior to HU treatment and maintained throughout the experiment. N = 3 cell lines per condition. Two-tailed t-test.

G, Quantification of fork restart following HU-induced replication stress in RSRD-high cells either in presence or absence of suppression of origin firing with 10 μM roscovitine. N = 3 cell lines per condition. Two-tailed t-test.

H, Immunostaining for cytosolic ssDNA (green) and nuclei (blue) in RSRD-low (i, MCF-10A; ii, ZR-75–1) and RSRD-high cell lines (iii, BT549; iv, HCC38). Scale bar indicates 10 μm.

I, Quantification of cytoplasmic DNA immunostaining in (H) as integrated ssDNA intensity normalized to DAPI. Two-tailed t-test.

J and K, Quantification of cytosolic ssDNA (J) or dsDNA (K) concentrations, reported normalized to cytosolic protein concentration. Two-tailed t-test.

All data are mean ± s.d. with each dot representing the average value for an individual cell line.

To validate that aberrant origin firing was causing RSR defects, we used the cyclin-dependent kinase (CDK) inhibitor roscovitine to normalize origin firing. Treatment with roscovitine restored the ability of three independent RSRD-high cell lines to recover from HU-induced replication stress, as quantified both by bulk ability to complete the cell cycle (Fig. 2F) and at the level of single replication forks by DNA fiber analysis (Fig. 2G). Furthermore, in both presence and absence of HU, roscovitine treatment prevented accumulation of ssDNA (fig. S2E) and of double strand breaks in S-phase cells (fig. S2F).

Because roscovitine can inhibit origin firing/cell cycle through multiple CDKs (27), to better understand the mechanism of action we used siRNA to suppress expression of CDK1, CDK2, CDK5, CDK7, CDK9, CCNA1, and CCNA2. Although the percentage of cells in S-phase was slightly reduced by some siRNAs (fig. S2G), analysis of DNA damage specifically in cells actively synthesizing DNA revealed the largest suppression of DNA damage by suppression of CCNA2 (fig. S2H). Consistent with this, expression of CCNA2 was positively correlated with RSRD score in patients with breast cancer (fig. S2I). Evaluation of RSR function in TNBC cell lines with stable suppression of CCNA2 revealed that CCNA2 suppression could phenocopy the restored RSR function observed with treatment with roscovitine (fig. S2, J to L).

Based on the depletion of RPA pools necessary to protect ssDNA in RSRD-high cells, we hypothesized that the unprotected ssDNA may accumulate within the cytoplasm (28). Immunostaining for cytosolic ssDNA revealed accumulation in RSRD-high cell lines (Fig. 2, H and I), which was validated by direct quantification of cytosolic ssDNA following cellular fractionation (Fig. 2J). Increased accumulation of cytoplasmic double-stranded DNA (dsDNA) was also observed, albeit at an order of magnitude lower in quantity (Fig. 2K). Capillary electrophoresis of the precipitated cytosolic DNA further suggested the primary species was ssDNA, as most of the precipitated cytosolic DNA was sensitive to digestion by nucleases specific to ssDNA, but not by nucleases specific to dsDNA (fig. S3A). Further analysis of the size of precipitated cytosolic DNA fragments revealed a log-normal distribution of lengths centered around approximately 100 bases in length, with an average of 90.7% falling under the size limitation for passive diffusion out of the nucleus(29) (fig. S3, B and C). Retention of larger DNA fragments above the passive diffusion size limit in the nucleus may explain the observed relatively conserved size distribution of cytosolic DNA fragments. Similarly, suppression of nucleases RBBP8 (CtIP), DNA2, EXO1, and MRE11A indicated that although EXO1 may generate a portion of this ssDNA, the majority is likely generated by replication fork collapse as opposed to exonuclease activity (fig. S3, D to F).

RSRD score predicts response to immune checkpoint blockade in pre-clinical murine breast cancer models.

Healthy, unperturbed mammalian cells have minimal accumulation of cytosolic DNA, and cytosolic nucleic acid content derived from various sources such as viral infection or endogenous DNA damage binds to certain receptors, activating immunostimulatory pathways to restore homeostasis (30). Based on this immunostimulatory role for cytosolic DNA, we hypothesized that ICB may show increased efficacy in RSRD-high tumors. To test this, we analyzed RSRD score in panel of murine breast tumor models and selected five models with varying RSRD scores for further study: 2208L, 4T1, HRM1, T11, and E0771 (Fig. 3A). Using this panel, we first validated that the RSRD signature predicted functional RSR defects in murine breast cancer models, finding that RSRD score was also functionally predictive in murine breast cancer models (Fig. 3B). Furthermore, we found that RSR defective murine breast cancer models displayed increased cytosolic DNA (Fig. 3, C and D), consistent with our observation in human RSR defective breast cancer (fig. 2, J and K).

Fig. 3. RSRD score predicts response to immune checkpoint blockade in pre-clinical breast cancer models.

A, RSRD scores across mouse breast cancer models, with models selected for further analysis indicated.

B, Percentage of cells that are RSR-defective by HU cell cycle assay described in Fig. 1A, plotted as a function of RSRD score for selected models.

C and D, Quantification of cytosolic single-stranded (C) and double-stranded (D) DNA as described in Fig. 2, J and K for selected models. Each dot represents a biological replicate.

E to I, Kaplan Meier plots for selected RSRD-low (E, 2208L; F, 4T1) and RSRD-high (G, HRM-1; H, T11; I, E0771) in vivo models following treatment with ICB (10 mg/kg anti-PD1 and 5 mg/kg anti-CTLA4, thrice weekly) compared to isotype control. Significance determined by log-rank test.

J, Hazard ratio with 95% confidence interval determined by log-rank test for ICB treatment for models in (E to I) plotted as function of RSRD score. R, P values determined by Pearson correlation.

After confirming the RSRD signature was functionally active in murine breast cancer models, we next tested whether RSRD may be predictive of response to ICB in this system. ICB treatment failed to improve survival in either RSRD-low model (Fig. 3, E and F), but did improved survival in all three RSRD-high models (Fig. 3, G to I). Overall, there was a strong correlation between a higher RSRD score and improved ICB response, as indicated by a lower hazard ratio upon ICB treatment (P = 0.0019, Fig. 3J).

Modulating replication stress response function can induce sensitivity and resistance to immune checkpoint blockade.

To test if the RSRD phenotype is functionally associated with response to ICB, we first sought to evaluate if restoring RSR function would mitigate response to ICB. Consistent with inhibition of origin firing with roscovitine restoring RSR function in human cells, treatment of murine models with roscovitine was sufficient to mitigate cytosolic ssDNA (Fig. 4, A and B). For in vivo testing, we pre-treated RSRD-high T11 tumors with roscovitine for 3 days to restore RSR function, and then repeated the ICB treatment with continual roscovitine administration. While roscovitine monotherapy slightly improved overall survival, presumably by slowing cell proliferation, it abrogated any benefit achieved by treatment with ICB (Fig. 4C). To confirm that these effects were specific to restoration of RSR function, and not caused by roscovitine inhibiting immune cell function, we generated cell lines with stable suppression of Ccna2 in RSRD-high T11 cells (Fig. 4D). Stable suppression of Ccna2 restored RSR function (Fig. 4E) and reduced accumulation of cytosolic ssDNA (Fig. 4F), but did not alter the percentage of cells synthesizing DNA (Fig. 4G). In vivo treatment of shCcna2 or shCTRL cells with ICB confirmed that restoring RSR function by suppression of Ccna2 could induce resistance to ICB (Fig. 4H). Although depletion of STING moderately suppressed ICB response, it failed to do so to the same degree observed with restoration of RSR function (fig. S4, A and B), suggesting multiple cytosolic DNA sensors may be activate in RSRD-high tumors.

Fig. 4. ICB response can be modulated by induction or suppression of RSR defects.

A, Immunostaining for ssDNA (green) and nuclei (blue) in RSRD-high T11 cells following treatment with 10 μM roscovitine (CDK inhibitor) for 36 hours. Scale bar indicates 10 μm.

B, RSRD-high T11 cells were treated with 10 μM roscovitine (CDK inhibitor) for 36 hours and then analyzed for cytoplasmic ssDNA as in 2J. Dots represent 3 biological replicates, values given as mean ± s.d. Two-tailed t-test.

C, Kaplan Meier plots for mice bearing RSRD-high/ICB-sensitive T11 tumor model treated with ICB in presence or absence of roscovitine. Roscovitine treatment started 3 days before initiation of ICB or IgG control antibodies. Treatment was ceased after 4 weeks. N = 10 mice per group. Significance determined by log-rank test.

D, Western blot confirming suppression of CCNA2 in T11 cells with stable expression of two shRNAs targeting Ccna2 (shCcna2-3, shCcna2-4) or non-targeting shRNA control (shCTRL).

E, Percentage of EdU positive T11 cells with stable expression of indicated shRNAs. One-way ANOVA and Holm-Sidak post-hoc test.

F, Percentage of RSR defective cells, determined as described in Fig. 1A, in T11 cells with stable expression of indicated shRNAs. One-way ANOVA and Holm-Sidak post-hoc test.

G, Quantification of cytosolic ssDNA, determined as described in Fig. 2J, in T11 cells with stable expression of indicated shRNAs. One-way ANOVA and Holm-Sidak post-hoc test.

H, Kaplan Meier plots for mice bearing RSRD-high T11 tumors with stable expression of indicated shRNAs treated with ICB or IgG control antibodies as described in Figure 3E, ceasing after 4 weeks. N = 8–10 per group. Log-rank test.

I, Staining for ssDNA after 12-hour treatment with varying concentrations of AZD7762 (CHKi), VE821 (ATRi), or DMSO vehicle control in RSRD-low 4T1 cells. Each dot represents and an individual cell.

J, Quantification of extracted cytosolic ssDNA relative to protein concentration in RSRD-low 4T1 cells following inhibition of CHK (100 nM AZD7762) for 24 hours and then analyzed for cytosolic ssDNA as in Fig. 2J. Shaded red area indicates mean ± std of RSRD-high cell lines. Dots represent 3 biological replicates, values given as mean ± std. Two-tailed t-test.

K, Kaplan Meier plots for mice bearing RSRD-low/ICB-resistant 4T1 tumors treated with either vehicle control, CHKi (25 mg/kg AZD7762, days 1/2), ICB (10 mg/kg anti-PD1 days 2/4/6), or combination of ICB and CHKi. Treatment ceased at day 19 (after three cycles) due to combination toxicity. N = 5 mice per arm, log-rank test.

L, Washout assay staining for ssDNA following 12-hour treatment with 100 nM LY2606368 (CHKi), 1 μM AZD6738 (ATRi) or DMSO vehicle control in RSRD-low 2208L cells. Each dot represents and an individual cell.

M, RSRD-low/ICB-resistant 2208L cells were treated with CHK1i (100 nM prexasertib) for 24 hours and then analyzed for cytoplasmic ssDNA as in Fig. 2J. Shaded red area indicates mean ± std of RSRD-high cell lines. Dots represent 3 biological replicates, values given as mean ± std. Two-tailed t-test.

N, Kaplan Meier plots for mice bearing RSRD-low/ICB-resistant 2208L tumors treated with either vehicle control, ICB, CHK1i (prexasertib), or combination of ICB and CHK1i. Treatment was ceased after 4 weeks. N = 10 mice per group. Significance determined by log-rank test.

After observing that restoration of RSR function could abrogate response to ICB, we next sought to test if induction of RSR defects could sensitize RSRD-low models to ICB, with the ultimate goal of identifying treatment strategies to expand the benefit of ICB to more patients. The top two clinically-relevant options for induction of RSR defects are inhibitors of ATR and CHK1, both critical regulators of RSR. Comparing the ATR inhibitor VE821 and CHK1/2 inhibitor AZD7762, we found that across all concentrations evaluated, CHK inhibition produced more ssDNA than ATR inhibition (Fig. 4I). Cytosolic fractionation validated that AZD7762 increased cytosolic ssDNA amounts compared to those observed in RSRD-high models (Fig. 4J). Combination of AZD7762 with ICB showed promising initial tumor growth responses, but ultimately the combination therapy proved too toxic. Despite stopping treatment after only 3 cycles, the combination still improved survival compared to either monotherapy alone (Fig. 4K).

To seek further improvement of this combination strategy, we next utilized the more CHK1-specific inhibitor prexasertib/LY2606368, which also induced more ssDNA than the ATR inhibitor AZD6738 at clinically achievable concentrations (fig. S4C). We recently found that Wee1 and ATR inhibitors exhibit persistent activity after drug removal, which could be leveraged to minimize toxicity observed in the context of combination with PARP inhibition (31). To evaluate if this concept was extendable to CHK inhibition, we analyzed ssDNA following LY2606368 washout for 48 hours and observed similar persistence (Fig. 4L). As observed with AZD7762, LY2606368 raised cytosolic ssDNA amounts to those observed in RSRD high cells (Fig. 4M). Based on the persistence of ssDNA for at least 48 hours following CHK1i washout, and to minimize combinatorial toxicity, CHK1i was administered only on days 1 and 4. Treatment of mice bearing highly ICB-resistant, RSRD-low 2208L tumors, with this dual ICB/CHK1i combinatorial strategy led to significantly improved survival compared to vehicle control (P = 1.1×10⁻⁵), CHKi alone (P = 2.0×10⁻⁵), and ICB alone (P = 5.0×10⁻⁶), whereas neither ICB (P = 0.56) nor CHKi (P = 0.06) alone significantly improved survival compared to vehicle control (Fig. 4N). No significant changes in mouse weights were observed using this therapeutic regimen (P = 0.99) (fig. S4D).

RSRD-high breast cancers have increased infiltration of dendritic cells and T cells in RSRD-high breast cancer.

To understand changes in the immune microenvironment between RSRD-high and RSRD-low tumors, we sought to understand how the RSRD score correlated with tumor cell expression of chemokines and cytokines. Altered chemokine/cytokine gene expression analysis in RSRD-high patient samples could be attributed to either modulation of expression in the tumor cells themselves, or cytokines secreted by RSRD-high tumor cells recruiting immune cells that in turn express additional chemokines/cytokines. To decouple alterations in chemokine/cytokine expression from tumor and immune cells, we looked for chemokines/cytokines correlated with RSRD score in both pure cell line cultures (Fig. 5A) and The Cancer Genome Atlas (TCGA) breast cancer samples (Fig. 5B). We found that RSRD score was significantly (FDR < 10%) positively correlated with expression of IL1A, IL1B, interferon alpha (IFNA1), GM-CSF (CSF2), IL11, and CXCL3 in both cell lines and primary breast cancer samples. Next, we used xCell (32) in the TCGA breast cancer cohort to interrogate how these increased cytokines/chemokines altered the immune cell microenvironment. After stratifying by subtype, numerous cell populations showed significant (FDR < 5%) Spearman correlation coefficients with RSRD score, including increases in multiple T-cell and dendritic cell populations (Fig. 5C).

Fig. 5. Immune microenvironment changes in RSRD-high breast cancer.

A and B, Relationship between RSRD score and expression of cytokines/chemokines in breast cancer cell lines (A) and samples from patients with breast cancer from TCGA (B). Highlighted cytokines indicate those significant in both cell lines and patients.

C, In silico immune populations from TCGA patients with breast cancer as determined by xCell. Dot color indicates magnitude of Spearman correlation coefficient. Relationships with an FDR<0.05 are indicated by black circles.

D, Representative images showing nuclei (blue) and dendritic cells (CD11c, red). Scale bar = 100 μm. Mice treated per Figure 3E–I for 10 days. IgG, IgG controls; ICB, immune checkpoint blockade.

E, Quantification of images from D. Dots represent tumors from independent mice (N = 5), values given as mean ± SEM. Significance determined by regression of average value for CD11c area per mouse with RSRD score and treatment status.

F, Representative multispectral images showing nuclei (blue), CD3 (all T cells, yellow), CD4 (CD4 T cells, green), CD8 (cytotoxic T cells, magenta), and FoxP3 (regulatory T cells, red). Scale bar indicates 100 μm. Mice treated per Fig. 3, E to I for 10 days.

G to J, Quantification of multispectral images from F, indicating total T cells (CD3⁺, G), cytotoxic T cells (CD3⁺CD8⁺, H), conventional CD4 T cells (CD3⁺CD4⁺FoxP3⁻, I), and regulatory T cells (CD3⁺CD4⁺FoxP3⁺, J). Dots represent tumors from independent mice (N = 5), values given as mean ± SEM. Significance determined by regression of average population value per mouse with RSRD score and treatment status.

We next sought to validate these computational predictions using our mouse pre-clinical models. For these experiments, tumors were immunostained following 10 days of treatment with either ICB or IgG controls. Quantification of dendritic cell infiltrates by staining for CD11c validated that RSRD high tumors were enriched in CD11c⁺ cells, which were further increased by treatment with ICB (Fig. 5, D and E). To analyze T-cell infiltrates, we developed a multispectral staining panel to probe CD3, CD4, CD8, and FoxP3 (Fig. 5F). Quantitative analysis of this panel indicated total T cells increased with RSRD score (P = 3×10⁻³), and that there was a significant positive interaction between RSRD score and ICB treatment (P = 0.04, Fig. 5G). Analysis of T cell subpopulations revealed a similar trend with cytotoxic CD8⁺ T cells (RSRD P = 5.5×10⁻³, RSRD*ICB P = 0.05, Fig. 5H) and conventional CD4⁺ T cells (RSRD P = 4.3×10⁻³, RSRD*ICB P = 0.03, Fig. 5I). Whereas T_regs were also increased with RSRD score as predicted (P = 3.6×10⁻⁴), there was no interaction effect observed with ICB treatment (Fig. 5J). Based on elevated amounts of dendritic cells present, we hypothesized that there may be increased immunological synapse activity leading to heightened T cell activation. Analyzing gene expression for components of the immune synapse, or supramolecular activation cluster (SMAC), indicated that RSRD-high tumors had increased expression of both central SMAC (c-SMAC) and peripheral SMAC (p-SMAC) components (Fig. S4E). Activation of c-SMAC was further supported by increased phosphorylation in RSRD-high tumors (fig. S4F) and may contribute to more active T cells in RSRD-high tumors. Consistent with these observed alterations in the immune microenvironment, we found that response to ICB in RSRD-high tumors was dependent on T cells because ICB treatment in nude mice lacking T cells no longer provided any survival benefit (fig. S4G), indicating that the observed response is not attributable to antibody-mediated cytotoxicity. To confirm this result, we re-challenged mice that had undergone complete tumor regression with ICB with tumor cells. Naïve immunocompetent mice implanted with RSRD-high E0771 and T11 tumors showed delayed tumor formation compared to nude mice, with a much longer delay in tumor formation in mice that had previously undergone a complete response to ICB (fig. S4, H and I).

High RSRD score predicts response to ICB in metastatic breast cancer patients.

In the TONIC breast cancer trial (11), a biopsy from a metastatic lesion was taken, and then patients were randomized to no treatment or four different low-dose induction treatment strategies for a two-week period to induce sensitivity to ICB. The four induction approaches utilized all induced DNA damage: (i) direct generation of DNA double-strand breaks with irradiation, (ii) generation of interstrand and intrastrand crosslinks with cyclophosphamide, (iii) inhibition of topoisomerase with the DNA intercalating agent doxorubicin, and (iv) generation of intrastand crosslinks with cisplatin. A second biopsy was taken prior to initiation of anti-PD1 therapy (Fig. 6A). Using RNA-seq data from biopsies collected immediately prior to anti-PD1 treatment (biopsy 2), we assessed the ability of RSRD score to predict response to anti-PD1 treatment and found that RSRD score predicted response, demonstrated by the receiver-operator characteristic (ROC) curve area under the curve (AUC) value of 0.78 (Fig. 6B). This predictive ability was moderately diminished by use of the pre-induction biopsies (fig. S5A). The original TONIC publication identified presence of stromal tumor-infiltrating lymphocytes (sTILs) and PD-L1 expression on immune cells by IHC as potential response biomarkers (11). Analyses of sTILs invasion and immune cell PD-L1 expression indicated that RSRD score is an independent biomarker of ICB response (fig. S5, B to D). Using in silico deconvolution of RNA-seq data, we also found a correlation between RSRD score and dendritic cells (Fig. 6C), matching prior in silico analysis of TCGA patient samples (Fig. 5C) and immunostaining of pre-clinical mouse models (Fig. 5, D and E).

Fig. 6. RSRD score is a dynamic biomarker to predict ICB response in metastatic breast cancer patients.

A, Schematic for metastatic breast cancer TONIC trial design (11).

B, ROC curves demonstrating ability of RSRD score to predict patients who achieved CR/PR/SD for at least 24 weeks following ICB treatment using RNA-seq from latest biopsy available.

C, Correlation of conventional/activated dendritic cells and RSRD score in the TONIC cohort. Spearman correlation coefficient.

D, ROC curves for alternative published ICB response biomarkers T Cell Inflamed (9, 33) gene expression signature, IMPRES (7), PD-L1 expression, and tumor mutation burden.

E, Response rate by induction treatment for patients in TONIC RNA-seq cohort.

F, Change in RSRD score following various induction strategies. Each pair of dots represents an individual patient (or single dots for patients lacking post-induction transcriptional data). Wilcoxon signed-rank test.

We recently demonstrated distinct tumor immunobiology between non-hypermutated tumor types, including breast cancer, and hypermutated tumors, such as melanoma and lung cancer, notably that relative neoantigen load/TMB was only a relevant factor in hypermutated types (10, 15). Consistent with this result, we found that TMB did not predict ICB response in the TONIC RNA-seq breast cancer cohort, consistent with prior results from the entire cohort (11) (Fig. 6D). We also compared the predictive ability of the RSRD score to PD-L1 expression and other published transcriptional biomarkers, including the T cell inflamed gene expression signature score, which was generated using multiple cancer types and has been applied across cancer types (9, 33) and the IMPRES gene expression score, which was generated from neuroblastoma data and validated in melanoma using the most independent cohorts published to date (N = 8) (7), but found that no other biomarker offered strong predictive ability (Fig. 6D).

When comparing patients within the RNA-seq TONIC cohort, we found that the various induction strategies induced varying response rates to anti-PD1 therapy (Fig. 6E), as was also described for the entire TONIC cohort in the original publication (11). The subset of samples with RNA-seq data generally mirrored response rates in the overall cohort, except for a slightly lower response rate for doxorubicin-induced (original = 35%, RNA-seq = 27%), and slightly higher response rate for cisplatin-induced (original = 23%, RNA-seq = 33.3%). To test if the RSRD gene expression signature could detect dynamic changes related to the induction treatments, we evaluated RSRD scores before and after various induction treatments (Fig. 6F). Patients who underwent no induction showed minimal change in RSRD score. Within the RNA-seq cohort, cisplatin-treated patients had the highest response rate, and we also detected a significant increase in RSRD score following induction with cisplatin (P = 0.01, Fig. 6F).

High RSRD score predicts ICB response in non-hypermutated tumor types.

We next sought to evaluate if prediction of ICB response by RSRD score could be generalized to other non-hypermutated cancer types treated with ICB. Analysis of an independent cohort of patients with cancer treated with ICB (3) confirmed the observation from the TONIC cohort that TMB does not offer prognostic value for ICB response in breast cancer (fig. S6, A and B). TMB also showed no prognostic value in ICB-treated patients with other non-hypermutated cancers, namely RCC and glioma (fig. S6, A to C). As was observed in breast cancer cell lines, RSRD-high ccRCC and GBM cells both exhibit increased cytosolic DNA, indicating that the RSRD signature is active in these other tumor types (fig. S7, A and B). To further evaluate the activity of the RSRD score in RCC, we next analyzed RSRD scores for two murine RCC tumor models, RENCA and RAG, compared to tumors derived from genetically engineered mouse models of kidney cancer driven by targeted deletion of either Vhl and Pbrm1 (34) or Vhl, Trp53, and Rb1 (35), as well as healthy kidney controls. RAG cells exhibited a high RSRD score, with a more intermediate score for RENCA cells (fig. S7C). As observed in murine breast cancer, high RSRD scores corresponded with functionally defective RSR (fig. S7D), increased accumulation of cytosolic ssDNA (fig. S7E), and increased sensitivity to treatment with ICB (fig. S7, F and G). Moreover, the combination of ICB with the CHK1 inhibitor prexasertib/LY2606368 in mice bearing RENCA tumors resulted in improved overall survival compared to either agent alone (fig. S7H), consistent with our observations in breast cancer models (Fig. 4, K and N).

Based on the ability of RSRD score to predict ICB response in breast cancer and observed activity in other pre-clinical models, we hypothesized that RSRD score may more broadly be predictive of clinical benefit in other non-hypermutated cancer types. To test if patients with RSRD-high ccRCC tumors exhibited increased response to ICB, we acquired samples from four cohorts of patients with ccRCC treated with ICB; a cohort treated with anti-PD1 from Ascierto et. al. (36); a cohort from NCT01358721 and an archival FFPE cohort treated with anti-PD1 from Miao et. al. (37); and patients treated with single agent anti-PD-L1 from the IMmotion150 trial (38). The RSRD score robustly predicted clinical responses to ICB across all 4 cohorts, as indicated by ROC curves with AUC values of 0.75 to 1.0 (Fig. 7A). Of other evaluated biomarkers, only the T cell inflamed gene expression signature score in the Miao archival cohort exceeded 0.7 (0.71), though it averaged an AUC of 0.48 in the other 3 cohorts analyzed. Furthermore, progression free survival (PFS) data from the Miao cohorts indicated that RSRD-high patients had improved outcomes by both univariate analysis (Fig. 7B) and multivariate analysis (Fig. 7C). Analysis of TCGA samples from patients with ccRCC who did not receive ICB demonstrated that high RSRD score was associated with worse prognosis, suggesting that improved survival in RSRD-high patients is ICB-specific (fig. S8, A and B). Previous studies have indicated that high amounts of endogenous retroviruses (39) and neoantigens (40) may promote ccRCC immunogenicity, but we did not find RSRD score to be associated with these factors (fig. S8, C to H).

Fig. 7. RSRD score predicts ICB response across multiple cancer types.

Prediction of response to ICB by published biomarkers T Cell Inflamed (9, 33) gene expression signature, IMPRES (7), PD-L1 expression, and mutation burden compared to RSRD score.

A, ROC curves demonstrating prediction of ICB response indicated biomarkers in cohorts of patients with ccRCC (36–38).

B, Progression free survival (PFS) for all in patients from Miao cohorts (37) based on RSRD score. Log-rank P-value.

C, Multivariate analysis of PFS in Miao cohorts (37) accounting for RSRD score, whether patients were treated in the first line, and PD-L1 expression, stratified by NCT1358721 vs archival cohort.

D, ROC curve showing ability of RSRD score to predict response in patients with GBM treated with anti-PD1 from the Zhao (41) cohort.

E, Overall survival analysis of anti-PD1-treated patients with GBM (13), considering treatment in either the neoadjuvant or adjuvant context as a clustering variable. All parameters were treated as continuous variables.

F and G, Kaplan-Meier curves showing overall survival from the Cloughesy (13) cohort treated either in the neoadjuvant (F) or adjuvant (G) context, and stratified by the RSRD score.

H, Overall survival of patients with metastatic prostate cancer treated with anti-CTLA4 stratified by RSRD score (53). Log-rank test.

I, Hazard ratio using overall survival data from (H) for indicated biomarkers. Error bars indicate 95% confidence interval. Log-rank test.

J to M, ROC curves showing ability of RSRD score and other indicated biomarkers to predict ICB response in low mutation subsets of patients with urothelial cancer treated with anti-PD-L1 from (J) Snyder (43) and (K) IMvigor210 (45), (L) patients with gastric cancer treated with anti-PD-L1 (54), and (M) patients with melanoma treated with anti-PD1 (55, 56).

N, Compilation of AUC values across all cohorts from Fig. 6 and 7 for evaluated biomarkers. P value assessed by one-sample t-test to determine significant difference from an AUC of 0.5 indicative of random assignment.

O, Correlation of conventional/activated dendritic cells and all patients treated with ICB. Spearman correlation coefficient.

P, Regression using a generalized linear mixed-effects model of RSRD score as a function of conventional dendritic cells, activated dendritic cells, or CD8 T-Cells. Patient cohorts were treated as a random effect. Regression coefficient with standard error.

Next, we sought to evaluate the predictive capability of the RSRD score in patients with GBM treated with ICB, which also exhibits the non-hypermutated phenotype of TMB/neoantigen-independent immunogenicity (10, 41). Analysis of RNA expression data from a patient cohort with GBM treated with anti-PD1(41) showed that the RSRD score alone robustly predicted clinical response (Fig. 7D). We confirmed this result in a second cohort of patients with GBM treated with anti-PD1 (13) in either the neoadjuvant or adjuvant setting. Although neoadjuvant treatment showed significantly better outcomes compared to adjuvant treatment (P = 0.01, fig. S8I), using a Cox proportional hazards model with treatment context as a clustering variable found that only a high RSRD score predicted improved overall survival (OS) (P <1×10⁻⁶, Fig. 7E). IMPRES score showed a trend towards improved prognosis, so we additionally performed multivariate analysis with both RSRD score and IMPRES score and found RSRD score still demonstrated robust prediction of outcomes following ICB treatment (fig. S8J). Despite small sample size in each arm, high RSRD score was still associated with improved OS regardless of whether anti-PD1 was administered in the neoadjuvant (HR = 0.23, P = 0.09, Fig. 7F) or adjuvant (HR = 0.31, P = 0.07, Fig. 7G) context. The RSRD score was not an independent prognostic biomarker in the TCGA cohort of patients with GBM who did not receive ICB (fig. S8K), suggesting the improved survival is specific to patients with GBM who received ICB.

We further validated the association between high RSRD score and improved response to ICB in a cohort of patients with metastatic prostate cancer treated with ipilimumab, a monoclonal antibody targeting cytotoxic T-lymphocyte-associated protein 4 (CTLA4). Consistent with observations in ccRCC and GBM, we found that only RSRD score significantly predicted improved overall survival following treatment with anti-CTLA4 (HR = 0.36, P = 0.04, Fig. 7H–I). RSRD score was not prognostic in patients with prostate cancer in TCGA who would have predominately not received ICB (fig. S8L)

Our previous study also indicated that some tumor types may have both hypermutated and non-hypermutated subtypes which display distinct immunobiology, most strongly observed when comparing hypermutated and non-hypermutated urothelial cancers (10, 42). Although ICB responses tend to be enriched in patients with hypermutated urothelial cancer, a subset of patients with non-hypermutated tumors also exhibit clinical benefit (43–45). We hypothesized that the RSRD score may identify which patients with non-hypermutated urothelial cancer may benefit from ICB. To test this, we evaluated the ability of the RSRD score to predict response to anti-PD-L1 in two cohorts of patients with urothelial cancer (43, 45) stratified into hypermutated and non-hypermutated subsets. As hypothesized, in urothelial cancer the RSRD score showed good predictive accuracy in patients with non-hypermutated tumors (Fig. 7, J and K), but not in patients with hypermutated tumors or the unstratified bulk population (fig. S9, A to D). This phenomenon was similarly observed when analyzing cohorts of patients with non-hypermutated gastric cancer (Fig. 7L and fig. S9, E and F), and was observable in low-frequency non-hypermutated melanoma after pooling two cohorts for sufficient sample number (Fig. 7M and fig. S9, G to I). Consistent with observations in ccRCC and GBM, the RSRD score also outperformed other biomarkers in these non-hypermutated subsets (Fig. 7, J to M). Response rates between patients with hypermutated tumors and patients with RSRD-high/low-TMB tumors were comparable (63.1% vs 54.1%), and enriched over three-fold relative to biomarker-negative patients (16.8%) (Fig. S9, B, D, F, and I). Together, upon compiling all 11 cohorts with objective response rates, only the RSRD score showed substantial predictive accuracy (mean AUC = 0.82, P = 6.5×10⁻⁷), with the second highest average AUC observed using PD-L1 expression (mean AUC = 0.55) (Fig. 7N). Confidence intervals for individual cohort AUC values are shown in fig. S9J, along with plots showing the relationship between positive predictive value and negative predictive value in fig. S9, K to M. As observed in breast cancer (Fig. 5, c and E, and Fig. 6C), the RSRD score was also positively correlated with dendritic cell infiltration when analyzing all patient cohorts (Fig. 7O). This positive relationship between RSRD score and dendritic cells was conserved when using a generalized linear mixed effects model, with a weaker relationship between RSRD score and CD8 T-cells observed (Fig. 7P).

DISCUSSION

Despite observations of robust clinical responses to ICB in a subset of patients with tumors lacking extensive mutation burdens, we and others have observed a failure of biomarkers established in the context of hypermutated tumors to identify which patients with non-hypermutated cancer types may receive clinical benefit from ICB. This study identifies a gene expression signature reflective of defective replication stress response that is functionally linked with increased DNA damage, RPA exhaustion, aberrant replication fork firing, and accumulation of immunostimulatory cytosolic DNA. This RSRD signature accurately predicts response to ICB in patients with non-hypermutated cancers across 12 independent cohorts, whereas other proposed biomarkers including (PD-L1) protein expression (6), and various gene expression signatures (7–9) showed minimal predictive value. It is noteworthy that the RSRD score predicts ICB outcomes in non-hypermutated, yet immunologically distinct cancer types—highly T-cell-infiltrated ccRCC tumors and in marginally T cell-infiltrated GBM tumors for example. Future studies further detailing the immune microenvironment features that sensitize to ICB in RSR-deficient tumors will be necessary to understand the biology of ICB response.

A critical limitation the current study is the predictive capacity of our RSRD signature on retrospective data, requiring prospective validation. Moreover, technical details, including how recent a biopsy is required as well as tumor purity, will need to be considered when performing any prospective studies. Any combination trials with agents to induce RSR defects, including ATR or CHK1 inhibition, will need to carefully consider scheduling and/or dose de-escalation to avoid unintentional immunosuppression by these agents. It is probable that DDR inhibitor dosages required for inducing an immunologic response and sensitizing to ICB are lower than necessary for monotherapy. Thus, a focus in trial design on the minimal effective DDR inhibitor dose to achieve an immunologic effect will be critical to avoid immunosuppressive side effects that could counteract the benefit from activation of inflammatory signaling.

Here, we found that regardless of concentration or inhibitor used, the inhibition of CHK1/2 induced a more robust phenotype than that observed with ATR inhibition in RSRD-low cells. Although CHK1 is primarily activated by ATR in response to replication stress, if ATR is inhibited CHK1 can also be activated via DNA-dependent protein kinase (DNA-PK)(46). The activity of this DNA-PK compensatory pathway could explain the enhanced induction of the RSRD phenotype following CHK1/2 inhibition compared to ATR inhibition. Moreover, the CHK1/2 inhibitor AZD7762 can simultaneously inhibit both checkpoint kinases (cell-free CHK1 IC₅₀ = cell-free CHK2 IC₅₀ = 5 nM) (47), and although prexasertib/LY2606368 exhibits an order of magnitude higher affinity for CHK1 (cell-free IC₅₀ = 0.9 nM) than CHK2 (cell-free IC₅₀ = 8.0 nM), it would likely still inhibit both kinases to a degree (48). Our comparison of ATRi and CHK1/2i was also performed in the context of RSRD-low cells that generally have lower endogenous replication stress. Previous work has demonstrated that ATR inhibitors more specifically target cells with high replication stress compared to CHK1 inhibitors (46), so it is possible that repeating the comparison in cells with higher replication stress would produce a more equivalent phenotype.

Prospective studies to evaluate the predictive ability of the RSRD gene expression score, genotype-to-phenotype relationships in RSRD-high tumors, and effects of pharmacological induction of RSR defects, are underway. For example, NCT04266912 is a correlative-rich early clinical phase study of avelumab (anti-PD-L1) and ATR inhibitor M6620/VX970. Additional trials in progress include combination of ATR inhibitor BAY1895344 and pembrolizumab in solid tumors (NCT04095273), ATR inhibitor M6620/VX970 and pembrolizumab in squamous non-small cell lung cancer (NCT04216316), CHK1 inhibitor LY2606368 and LY3300054 anti-PD-L1 in solid tumors (NCT03495323), and ATR inhibitor AZD6738 and durvalumab in biliary tract cancers (NCT04298008).

Mutations in select DDR genes have previously been used as biomarkers for response to various therapeutic modalities, such as PARP inhibition. However, gene mutation information alone does not always fully inform functional DDR defects. Detected mutations do not always have a known functional relevance, and even with known mutations, DDR function can be restored through secondary events (49). Furthermore, DDR defects may arise by pathway modulation outside of mutational events (50). Thus, relying on DDR gene mutations alone to serve as predictive biomarkers for treatment selection may exclude patients who have functional deficiencies in DDR pathways. Utilization of gene expression-based biomarkers circumvents these problems and allows for dynamic monitoring of DDR function, including following therapeutic interventions as observed here for patients with breast cancer from the TONIC trial.

In conclusion, this work demonstrates that defects in replication stress response arise from aberrant replication origin firing, exhausting RPA pools and causing accumulation of immunostimulatory cytosolic ssDNA. The RSRD gene expression signature can dynamically detect defects in replication stress response and can predict clinical benefit to ICB in patients with non-hypermutated tumor types, who currently lack accurate predictive biomarkers for ICB treatment stratification. Use of ICB in patients with RSRD-high tumors, as well as combination therapy approaches using DDR inhibitors to induce RSRD-high phenotype with ICB, might expand the number of patients who may benefit from these treatment modalities, with prospective validation in multiple early phase clinical trials in development and underway.

MATERIALS AND METHODS.

Study design

The initial objective of this study was to understand causes of DNA replication stress response defects in breast cancer, as well as to identify any therapeutic vulnerabilities induced in RSR-defective cancer cells. After identifying that RSR-defective pre-clinical murine breast cancer models are sensitive to immune checkpoint blockade, results were validated in cohorts of patients with breast cancer. Association of RSR defect score with ICB sensitivity was then generalized to multiple cancer types with similar immunobiology using pre-clinical models and clinical patient cohorts. All in vivo mouse experiments were approved by the Institutional Animal Care and Use Committee. After randomization, tumor measurements were performed by a blinded investigator. Microscopy was performed semi-blinded, coupled with automated image quantification. Multiple independent biological replicates were performed for each experiment, with sample size given in the corresponding figure legend along with details on statistical analysis. Sample sizes for mouse studies were chosen based on previous studies; sample sizes for human cohorts were based on data availability. Detailed methods are provided in Supplementary Materials, including descriptions of cell lines (table S1), antibodies (table S2), and shRNAs/siRNAs (table S3).

Statistical analysis

Statistical analysis was performed in Matlab, R, and GraphPad Prism. Unless otherwise noted, comparisons between two groups were performed using a two-sided t-test. For samples not normally distributed, a Wilcoxon rank-sum test (unpaired data) or Wilcoxon signed-rank test (paired data) was used. For comparing the effects of a perturbation on two groups, two-way analysis of variance (ANOVA) was used. No meaningful differences in variance were noted between groups being compared statistically. Univariate survival was assessed by log-rank test; multivariate survival was assessed using the Cox proportional hazards model. Correlation between two parameters were assessed using the Spearman correlation coefficient unless otherwise specified. Multiple comparisons were accounted for by the method of Benjamini and Hochberg.

Data and materials availability:

All data associated with this study are present in the paper or supplementary materials. The results here are in whole or part based upon data generated by the TCGA Research Network: http://cancergenome.nih.gov/. All TCGA data were downloaded using the TCGA data portal (https://portal.gdc.cancer.gov/) from the Pan-Cancer Atlas release (April 2018), except for phospho-proteomic data which was acquired from the corresponding manuscript’s supplementary information (51). Cell line gene expression data were acquired from CCLE (52). The TONIC breast cancer trial (11) data was acquired from EGAS0001003535, with associated patient outcome data acquired directly from the authors. All data from the Subudhi (53) cohort of metastatic prostate cancer treated with anti-CTLA4 were acquired from the paper’s supplementary materials. The Ascierto (36) anti-PD1 ccRCC cohort was downloaded from GEO (GSE67501). The Miao (37) anti-PD1 ccRCC data were downloaded from the publication’s supplemental information. For the ccRCC anti-PD-L1 IMmotion150 trial (38), data were downloaded from EGA (EGAS00001002928). The Zhao (41) GBM anti-PD1 cohort clinical data, including investigator-associated response, were downloaded from the publication’s supplemental data. Sequencing data (FASTQ files) were downloaded from SRA (PRJNA482620). The Cloughesy (13) GBM anti-PD1 clinical and transcriptional data were all downloaded from GEO (GSE121810). All data for urothelial cancer patients treated with anti-PD-L1 from Snyder et. al. (43) were downloaded from http://doi.org/10.5281/zenodo.546110. Data for the anti-PD-L1 IMvigor210 cohort(45) were downloaded from EGA (EGAS00001002556). Clinical data for metastatic gastric cancer treated with anti-PD-L1(54) were retrieved from the publication. RNA-seq data were downloaded from ENA (PRJEB25780). The Riaz (55) anti-PD1 melanoma cohort clinical data were downloaded from the publication’s supplemental information, and the RNA-seq data were downloaded from https://www.github.com/riazn/bms038_analysis. The Hugo (56) anti-PD1 melanoma cohort clinical data were downloaded from the publication’s supplemental information and the RNA-seq data were downloaded from GEO (GSE78220). Murine breast cancer data were acquired from GEO accessions GSE27101 and GSE69006. Newly generated murine RCC data are given in table S4 with additional data obtained from Harlander et al. supplemental data (35) and GSE83688 from Nargund et al. (34). Murine mouse breast cancer models T11 and 2208L are available from Dr. Jeffrey Rosen at Baylor College of Medicine following completion of a Material Transfer Agreement.