DUX4 is a common driver of immune evasion and immunotherapy failure in metastatic cancers

Cancer immune evasion contributes to checkpoint immunotherapy failure in many patients with metastatic cancers. The embryonic transcription factor DUX4 was recently characterized as a suppressor of interferon-γ signaling and antigen presentation that is aberrantly expressed in a small subset of primary tumors. Here, we report that DUX4 expression is a common feature of metastatic tumors, with ~10–50% of advanced bladder, breast, kidney, prostate, and skin cancers expressing DUX4. DUX4 expression is significantly associated with immune cell exclusion and decreased objective response to PD-L1 blockade in a large cohort of urothelial carcinoma patients. DUX4 expression is a significant predictor of survival even after accounting for tumor mutational burden and other molecular and clinical features in this cohort, with DUX4 expression associated with a median reduction in survival of over one year. Our data motivate future attempts to develop DUX4 as a biomarker and therapeutic target for checkpoint immunotherapy resistance.

(A) As in (Figure 2A), but the Gene Ontology (GO) enrichment network analysis corresponding to DUX4-upregulated genes, compared against the set of coding genes, is shown.(B) Interferon-gamma (IFN-γ) signature score (Ayers et al., 2017).The p-value was estimated via a Mann-Whitney U test.
(C) CD8 T cell score from (Danaher et al., 2017).The p-value was estimated via a Mann-Whitney U test.(D)As in (C), showing the other immune cell signatures available in (Danaher et al., 2017).(E) Chemokine signature score (Coppola et al., 2011).The p-value was estimated via a Mann-Whitney U test.(F) DUX4 expression (TPM) in inflamed, immune excluded, and immune desert tumors.The phenotypes are based on CD8 + T cell abundance and degree of tumor infiltration determined by anti-CD8 staining of tumor FFPE sections in the original study (Mariathasan et al., 2018)    The transparent ribbon corresponds to the 95% confidence interval.(I) Partial plot illustrating the marginal effect of DUX4 expression status.The points correspond to the RSF prediction for mortality for each patient when DUX4 expression status is fixed to the indicated value for the entire cohort.

Figure S1 .
Figure S1.The DUX4 transcript is likely polyadenylated.(A) As in (Figure 1A), but The Cancer Genome Atlas (TCGA) cancer cohorts without matched advanced metastatic counterparts analyzed in our study are shown.(B) A comparison of DUX4 expression values (TPM, transcripts per million) measured from sequencing libraries prepared via poly(A) capture or hybrid capture.(C) As in (B), but a heatmap where patient samples (columns) were stratified according to the indicated categories of DUX4 expression.

Figure S2 .
Figure S2.DUX4-positivity is correlated with an embryonic gene expression signature, downregulation of interferon-gamma signaling, and exclusion of diverse immune cell types.(A)As in (Figure2A), but the Gene Ontology (GO) enrichment network analysis corresponding to DUX4-upregulated genes, compared against the set of coding genes, is shown.(B) Interferon-gamma (IFN-γ) signature score(Ayers et al., 2017).The p-value was estimated via a Mann-Whitney U test.
. The pvalues were estimated via the Mann-Whitney U test.(G) DUX4 expression (TPM) in advanced urothelial carcinoma tumors.The percentage of tumor cells with positive PD-L1 staining are indicated on the x-axis.The p-values were estimated via the Mann-Whitney U test.(H) As in (G), but showing the percentage of tumor-infiltrating immune cells (lymphocytes, macrophages, and dendritic cells) with positive PD-L1 staining on the x-axis.

Figure S3 .
Figure S3.DUX4 expression status stratifies patients according to survival.(A) Kaplan-Meier (KM) estimates of overall survival (solid lines), 95% confidence intervals (transparent ribbons), and censored events (crosses) for ICI-treated advanced urothelial carcinoma patients stratified by DUX4 expression status.The p-value was estimated via a logrank test.

Figure S4 .
Figure S4.Cox Proportional Hazards regression models containing DUX4 expression status as a predictor have a better fit to the data.(A) Akaike information criterion (AIC) measurements for goodness of fit for the full (TMB, Clinical, DUX4 expression) vs. reduced Cox PH models, where lower values indicate better fit.The bootstrapped AIC mean and the 95% confidence interval are illustrated.Clinical (ECOG Performance Status and Platinum treatment history).(B) Harrell's concordance indices (C-index) for the full (TMB, Clinical, DUX4 expression) vs. reduced Cox PH models, where high values indicate better model performance.The bootstrapped C-index mean and the 95% confidence interval are illustrated.(C) Kaplan-Meier (KM) estimates of overall survival, 95% confidence interval (transparent ribbon), and censored events (crosses) for low-risk (solid gray line) and high-risk (solid orange line) patients in the training (left) and test (right) sets.Risk group assignments were based on risk scores estimated by the Cox PH model with only TMB as a predictor.p-values were estimated via a log-rank test.(D) As in (C), but the risk group assignments were based on risk scores estimated by the Cox PH model with TMB, ECOG Performance Status, and Platinum treatment history as predictors.(E) Time-dependent Brier scores for the full and reduced Cox PH models applied on the training (left) and test (right) sets.The Continuous Ranked Probability Scores (CRPS), defined as the integrated Brier score divided by time, are shown in parentheses.Reference refers to the Kaplan-Meier prediction model.A Brier score = 0.25 indicates random guessing (gray dashed line).

Figure
Figure S5.A Random Survival Forest model quantifies the effect of DUX4 status on overall survival probability in the context of immune checkpoint inhibition.(A) Error (1 -Harrell's concordance index) as a function of the number of trees in the Random Survival Forest (RSF) model.The training out-of-bag error (OOB error, solid gray line) and the test error (solid orange line) are shown.1500 trees were used in the final model.(B) Time-dependent Brier scores for the RSF model estimated from the training (solid turquoise line) or test (solid teal line) sets.OOB survival predictions were used to calculate the Brier score for the training set.The Continuous Ranked Probability Scores (CRPS), defined as the integrated Brier score divided by time, for both sets are shown.A Brier score = 0.25 indicates random guessing (gray dashed line).