An interferon-related gene signature for DNA damage resistance is a predictive marker for chemotherapy and radiation for breast cancer

The parental SCC61 human squamous cell carcinoma cancer cell line was selected in vivo for resistance to ionizing radiation, resulting in the Nu61 subline that differentially expresses the 49 genes in the IRDS (7). To determine if IRDS genes are more broadly associated with resistance to RT, we analyzed 34 cancer cell lines from the NCI60 panel (8). Thirty-six IRDS genes were among the top 25% of all genes ranked by their correlation with the surviving fraction after 2 Gy of radiation (SF2), a result that is statistically significant based on the enrichment score calculated from gene set enrichment analysis (Fig. 1B). Of these 36 genes, 32 are a subset of the 40 genes up-regulated in the IRDS. Of the IRDS genes, STAT1 showed the highest correlation to the SF2 and ranked in the top 1% of all genes considered (Fig. 1 B and C).

Similar to RT, IRDS(+) Nu61 xenografts are more resistant to doxorubicin chemotherapy compared with IRDS(−) SCC61 tumors (Fig. 2A). Knockdown of STAT1 using stable shRNA led to decreased expression of other IRDS genes (Fig. 2B) and a re-sensitization of Nu61 tumors to doxorubicin in vivo (Fig. 2C) and to RT (9). This re-sensitization was observed over a dose range of doxorubicin (Fig. 2D). Conversely, not only does constitutive expression of STAT1 in parental SCC61 confer resistance to DNA damaging agents as previously shown (7), resistance to doxorubicin can also be transferred to the SKBR3 human breast cancer cell line (Fig. 2E). To test whether other IRDS genes merely act as markers for STAT1 activity or can themselves mediate resistance, we also targeted ISG15 and IFIT1 by shRNA. ISG15 is a ubiquitin-like protein involved in posttranslational modification (10), and IFIT1 has been shown to regulate translation initiation (11). To the best of our knowledge, neither of these genes have been previously implicated in DNA damage resistance, and both of these genes appear to be regulated by STAT1 based on decreased expression resulting from shRNA targeting of STAT1 (Fig. 2 B and F) and induced expression after 5 h of IFN treatment (7). Although knockdown of ISG15 and IFIT1 (≈80% for ISG15 by protein and 90% for IFIT1 by quantitative RT-PCR) had no or only marginal effects on STAT1 levels (Fig. 2F), decreased expression of either gene re-sensitized Nu61 to doxorubicin (Fig. 2G). IRDS expression was not associated with enhanced metastatic ability as determined by lung metastasis assay in mice (Fig. 2H). In total, these results suggest that IRDS genes primarily regulate experimental resistance to DNA damage but not metastasis.

Given that IRDS genes can associate with resistance across cell lines of different cancer types, we explored whether the expression profile of IRDS genes found in different primary human tumors might resemble either Nu61 or SCC61. Individual patient samples from DNA microarray databases from a variety of different cancers (breast, head and neck, prostate, lung, glioma) were directly compared with the IRDS gene profiles from Nu61 and SCC61 by using rank correlation analysis. For each primary cancer type, the IRDS genes were able to divide the tumors into two groups that show similarity to either the Nu61 or the SCC61 profile, which define the IRDS(+) and IRDS(−) cell line states, respectively [supporting information (SI) Fig. S1A]. Motivated by these results, and to avoid the possibility of imposing idiosyncrasies of the cell line data in the class assignment process, we also performed unsupervised clustering using the IRDS genes to divide patient samples from each tumor type into two groups (Fig. 3A). The groups with positive correlation to the Nu61 centroid (r² = 0.45 to 0.71) are defined to be IRDS(+), which represents 37%, 48%, 29%, 46%, and 50% of head and neck, lung, prostate, breast, and high-grade glioma cases, respectively. The groups correlated with the SCC61 centroid (r² = 0.41 to 0.65) are defined as IRDS(−). As expected, class assignments made from the clustering approach are highly similar to results from the rank analysis (r² = 0.56 to 0.90). Particularly high correlation is observed among the 78 patients with breast cancer, with a correlation between the IRDS(+) centroid and the Nu61 centroid of 0.71 and correlation between rank correlation results and clustering results of 0.90 (Fig. 3B and Fig. S1B). For all cancer types, many of the highly correlated genes are known to be involved in the IFN pathway (Fig. 3B). In total, these data demonstrate that IRDS genes can segregate a variety of human cancers into groups resembling either Nu61 or SCC61 in their IRDS profiles, representing IRDS(+) and IRDS(−) states, respectively.

The clustering method used to define IRDS status for the 78 patients with breast cancer (Fig. 3A) is not clinically practical for the classification of new samples. Instead, we used this data set to train a classifier to predict IRDS status using the top scoring pairs (TSP) method (12) (see SI Text). This classifier, denoted the TSP IRDS, uses simple non-parametric decision rules by measuring pair-wise relative expression between only seven gene pairs. Each pair contains an IRDS gene and a gene used for comparison. The sum of the results from each pair-wise comparison defines an eight-point ordinal scale from zero to seven, which we denote the TSP IRDS score (Fig. S2A). Notably, STAT1, IFIT1, and ISG15, which all affect experimental resistance to DNA damage (Fig. 2 C–E and G), were selected to be among the seven gene pairs in the classifier.

Given that IRDS genes can mediate experimental resistance to DNA damaging agents but is insufficient for metastasis, this provides a biological basis for testing the IRDS as a therapy-predictive marker for ADCT. To investigate this clinically, the following statistical expectations for a therapy-predictive marker were tested (see Introduction): (i) there is an interaction between the IRDS and whether patients receive ADCT (i.e., the effects of IRDS status on patient outcome depends on use of ADCT), (ii) the accuracy of outcome prediction for treated patients is improved when prognostic markers are combined with the IRDS by an amount approximately equal to the absolute benefit of treatment, and (iii) the IRDS can integrate into commonly used prognostic classifiers to identify patients with poor prognosis who are rendered low risk with adjuvant therapy.

To statistically examine the IRDS as a therapy-predictive marker, a data set of 295 patients with early-stage breast cancer (NKI295) was analyzed (6). To test for an interaction between the IRDS and ADCT, a multivariable Cox proportional-hazards model for metastatic risk was used. This analysis reveals a hazard ratio of 1.2 (i.e., a 1.2-fold increased risk of metastasis for each incremental increase in the TSP IRDS score from 0 to 7) specifically when an interaction with chemotherapy is considered (P = 0.05; Table S1). These results suggest that an association of the IRDS with clinical outcome depends on the use of ADCT.

The TSP IRDS shows moderate association with some standard clinical prognostic markers and other genomic classifiers (Table S2 and Fig. S2B). If the TSP IRDS is a therapy-predictive marker, it should add unique information when combined with prognostic markers and improve prediction accuracy among treated patients. How to test this hypothesis is not straightforward. The use of Cox models has known limitations that include relying on restrictive assumptions and concerns about the correct modeling of interactions and non-linear effects. Furthermore, for judging the predictive value of a new tumor marker, P values for calculated hazard ratios derived from Cox regression are mathematically unrelated to prediction (13). Our analysis method of choice is a multivariable random survival forest (RSF) analysis (see SI Text). RSF is a non-parametric ensemble partitioning tree method for survival data that automatically estimates non-linear effects for variables and multi-way interactions between variables, and is capable of imputing missing data (14). To properly measure the effect of the TSP IRDS and other variables on prediction accuracy, we calculate an importance score, which is a metric of how much the error rate of a model is improved by addition of each variable (more influential factors have higher scores). For comparison, results using Cox regression modeling are also shown throughout the article.

Among patients who received ADCT (see Table S3 for patient characteristics), a model that combines the TSP IRDS with standard clinicopathological factors reveals that the TSP IRDS has an importance score of ≈0.05 (Fig. 4A, green bar plot and blue crosses), meaning that the addition of the TSP IRDS to standard clinicopathological factors decreases the prediction error by 5%. This value for the importance score is within the expected range for a therapy-predictive marker of ADCT for early-stage breast cancer (i.e., approximately equal to the absolute benefit for ADCT). As expected, even after adjusting for other covariates and interactions, increasing TSP IRDS leads to a sharp increase in metastasis risk (Fig. 4B, Left). In contrast, among patients who do not receive ADCT, the TSP IRDS contributes significantly less to prediction (Fig. 4A, orange) and demonstrates no relationship to metastatic risk (Fig. 4B, Right). Standard clinicopathological factors contribute to prediction accuracy and/or associate with metastasis risk in an expected way regardless of treatment, confirming their primary role as prognostic markers (Fig. 4A and Fig. S3A). Similar results for the TSP IRDS are seen specifically among patients receiving ADCT when the TPS IRDS is combined with genomics-based markers (Fig. 4 C and D). The NKI 70, wound, and molecular subtypes are best defined as prognostic markers given their association with risk regardless of treatment (see Fig. S3B). In total, these results suggest that the TSP IRDS is a unique genomic classifier that increases the accuracy of outcome prediction not as a prognostic marker but as a therapy-predictive marker.

As a therapy-predictive marker, the IRDS is expected to identify patients who have outcomes better than predicted by prognostic markers as a result of successful treatment. To test this, we used the 2005 St. Gallen consensus criteria, the NKI 70 gene signature, or AOL to identify poor prognosis patients. IRDS status was assigned using a conservative TSP IRDS cut-off of <2 for IRDS(−) and ≥2 for IRDS(+) (see partial plots for metastasis risk as a function of TSP IRDS in Fig. 4 B and D and SI Text).

In the absence of ADCT, IRDS status is not prognostic among patients with a poor prognosis according to NKI 70 (i.e., NKI 70(+)) or St. Gallen consensus criteria (Fig. 5 and Fig. S4). However, patients at high risk who are IRDS(−) and receive ADCT have an outcome better than their IRDS(+) counterparts and similar to those patients at low risk. Similarly, in the absence of ADCT, the estimated 10-year risk for metastasis is comparable to the predicted AOL risk regardless of IRDS status. In contrast, risk among patients with low TSP IRDS scores (either 0 or 1) who received ADCT is markedly lower than predicted by AOL across a wide range of AOL 10-year risk estimates. In total, these data demonstrate that the IRDS can be integrated with available prognostic tools to identify patients at risk for distant relapse who can be rendered low risk by ADCT.

Adjuvant RT reduces the risk of local-regional failure (LRF) following breast conservation therapy or mastectomy (1). As a mediator of DNA damage resistance, the IRDS should predict LRF after adjuvant RT. In a multivariable Cox model, the IRDS is independently associated with LRF (Table S4). Analysis of importance scores using RSF or a Cox model reveals that the IRDS significantly contributes to prediction accuracy for LRF (Fig. 5B). IRDS(+) patients who received adjuvant RT exhibit a high rate of LRF (Fig. 5C). Evaluation of patients not receiving adjuvant RT was not possible in this cohort because there are few events in the minority of patients who underwent mastectomy without RT (but see metaanalysis described later). However, a 30% to 40% LRF rate at 10 years, as seen with IRDS(+) patients, is within the expected range for patients who exhibit complete resistance to adjuvant RT (15).

To validate the properties of the IRDS, several independent breast cancer data sets were assembled (see Table S5). Cohort A is comprised of 292 patients from the Radcliffe, University of California San Francisco, and Stockholm data sets who all received ADCT and/or RT and was used to validate that the IRDS is a therapy-predictive marker for DNA damaging agents. For each of the three data sets in cohort A, a higher TSP IRDS is associated with a higher risk for distant failure and/or LRF, which is similar to results from the NKI295 (Table S6). Survival analysis of all patients from cohort A reveals that the IRDS(−) group has a markedly better recurrence-free survival compared with IRDS(+) patients (Fig. 6A). This improvement in recurrence-free survival is a result of both fewer distant relapses among the IRDS(−) patients treated with ADCT and lower LRF among IRDS(−) patients treated with adjuvant RT (Fig. 6 B and C). Analysis using cohort B, which consists of 277 patients who received only endocrine therapy as adjuvant systemic treatment (Fig. 6D), and cohort C, which is composed of 286 patients who did not receive adjuvant systemic treatment (Fig. 6E), confirms that the IRDS is neither a therapy-predictive marker for endocrine therapy nor prognostic for distant failure in the absence of ADCT. A similar lack of therapy-predictive effect with endocrine therapy or prognostic effect was noted for two additional cohorts (Fig. S5).

The NKI295, Radcliffe, University of California San Francisco, and Stockholm data sets were also used to validate the predictiveness of the IRDS by using each individual data set as a test set for a model trained on the other three. The patients from these four data sets differed in treatment regimens (e.g., cyclophosphamide/methotrexate/5-fluorouracil vs. an anthracycline regimen; Table S7) and patient characteristics. Nonetheless, for each of these data sets, the TSP IRDS improves prediction accuracy for metastasis-free survival and local-regional control for patients treated with ADCT or RT, respectively (Fig. S6). Importantly, as the TSP IRDS improves prediction in each test set, these results are unlikely a result of a confounding latent variable because it would have to be the same latent variable in all of the cohorts.

To provide best estimates of importance scores and their sampling error, and to test for effects of cohort heterogeneity, all data sets used in validation were combined with the NKI295, resulting in 1,573 patients. By using RSF we were able to perform a non-stratified metaanalysis whereby all 1,573 patients were analyzed simultaneously but the effects of treatment were extracted. Unlike with Cox regression, this non-stratified analysis is possible because of the ability of RSF to automatically model all possible interactions between variables. Bootstrap means and SEs for variable importance scores for metastasis-free survival confirm high values for the IRDS, specifically among patients treated with ADCT (Fig. 6F). Well established prognostic factors show importance scores of comparable magnitude regardless of treatment. Few or no cohort effects were seen, indicating that an adequate level of homogeneity across institutes was observed, and our analysis accounts for differences in these cohorts (Fig. S7A). Confirmation that the IRDS is a therapy-predictive marker is similarly seen with RT. Results from Cox regression, which are necessarily stratified by treatment and devoid of interaction effects between variables, are shown for comparison. Notable are the 3% to 8% gains in prediction accuracy for RSF over Cox regression (Fig. S7B). In total, these results suggest that the IRDS is a therapy-predictive marker that performs across patient populations that may differ in baseline characteristics and treatment.

See Tables S2 and S8 and SI Text for further details on all clinical and microarray data.

Derivation of Nu61 from the SCC61 cell line, analysis of the IRDS, and mouse xenografting have been described (7). For additional information, including Tables S9 and S10, please see SI Text.

See SI Text for full details.