Translation of the 27-gene immuno-oncology test (IO score) to predict outcomes in immune checkpoint inhibitor treated metastatic urothelial cancer patients

Immune checkpoint inhibitors targeting programmed cell death 1 or its ligand, programmed cell death 1 ligand 1, (PD-1 or PD-L1, respectively) are active in many different tumor types and may modulate the immune response by impacting a common immune regulatory pathway. To date, observed benefit across tumor types has been modest despite co-development of candidate biomarkers for identifying responders to ICI therapy, including immunohistochemistry (IHC) for PD-L1 and tumor mutation burden (TMB). These biomarkers are limited in their value as predictive markers due to inconsistent scoring methods and differing thresholds for positivity, often specifically tailored to tumor types or even clinical studies [1‐3]. Additionally, ICIs are a systemic therapy which are impacted by the TIME [4‐6]. Therefore, a biomarker that provides a phenotypic classification of the TIME may be useful to better identify tumors poised to respond to ICI therapeutics.

Atezolizumab was the first FDA approved ICI in mUC based on the initial results of IMvigor210, a non-randomized Phase II clinical trial involving two different cohorts of mUC. Cohort I (NCT02951767) was comprised of 119 patients who were ineligible to receive cisplatin [7] and Cohort II (NCT02108652) was comprised of 310 patients whose disease had progressed on or after platinum therapy [8]. Based on this study, atezolizumab was approved in May of 2016 under an accelerated approval pathway for platinum refractory bladder cancers (Cohort II) independent of PD-L1 status even though the overall response rates were more favorable in those patients who had PD-L1 staining in  ≥ 5% in immune cells (IC2/3). Unfortunately, a Phase 3 confirmatory study, IMvigor211 (NCT02302807), using the primary endpoint of OS where patients with IC2/3 status were tested first in a hierarchical sequence, failed to reach significance [9]. In 2018, Mariathasan and colleagues published the gene expression data for 348 of IMvigor210 patients, representing a mix of both cohorts, along with exploratory analyses for the association with outcome using candidate genomic biomarker signatures and clinical features in a search for biomarkers that might inform response [10].

The IO Score has been validated as a biomarker that identifies patients likely to benefit from ICIs in TNBC and NSCLC [11‐15]. These studies were conducted across tumor types using the same threshold of positivity for IO Score. The IO Score was derived from an unsupervised classification of triple negative breast cancer (TNBCtype) specimens [14, 16, 18]. Two subtypes originally thought to describe intrinsic features of the tumor (the immunomodulatory or IM and the mesenchymal stem-like or MSL, respectively) were later recognized as classifiers that distinguish tumor infiltrating lymphocytes and stromal features of a particular tumor’s TIME, respectively. A third subtype, the mesenchymal (M), likely reflects tumors that have undergone some level of epithelial-to-mesenchymal transition (EMT) [17, 19], which in turn is associated with tumor resistance to the immune system [20‐24]. Together, the unique interaction of these components defines the IO Score.

While the IO Score was developed in TNBC, its three components are not unique to TNBC but collectively measure ubiquitous features across tumors of epithelial origin. Given that urothelial carcinoma is likewise of epithelial origin, we sought to test the clinical utility of the IO Score to identify patients who benefit from the administration of ICIs in bladder cancer. We first evaluated the predefined IO Score threshold using the IM, MSL, and M subtypes to classify ICI naïve samples from TCGA-BLCA. Using the predefined threshold for IO Score we then tested its association with treatment outcome in patients from the IMvigor210 study. Finally, we compared these results to the previously published extensive exploratory analysis of candidate biomarkers, biomarker signatures, and clinical factors [10].

The TCGA-BLCA tumor samples were classified into IM, MSL, and M subtypes using the expanded TIME classifier. We then compared these classifications against the IO continuous and binary scores using the IM subtype (high inflammatory gene expression) as the surrogate classifier for IO Score positive (Fig. 2D). In total, 125 of 406 TCGA-BLCA tumor samples clustered with the high IM expressing genes. The MSL and M subtypes were surrogates for IO Score negative. In total, 151 patients clustered with the MSL genes, and 130 patients clustered with the M genes. The distribution of the IO Score positive patients within each of these clusters was 113 (90.4%), 22 (14.6%), and 25 (19.2%) for IM, MSL, and M respectively, showing a strong concordance between IO Score positive patients and the IM subtype (Fig. 3A, p < 0.0001, chi-squared test).

In order to confirm the predefined IO Score threshold we generated a threshold specific to the TCGA-BLCA cohort by comparing the minimal distance between sensitivity and specificity for classification of IO Score positive patients (IM) versus IO Score negative patients (M or MSL) and obtained a theoretical, optimal binary classification threshold of 0.142 (Fig. 3B) [36]. The overall accuracy for this new threshold for classification of IO Score positive patients into the IM subtype was 87%. There was no statistical difference between the calculated threshold and the previously established threshold of 0.09 (p = 0.42, t-test of proportions). Based on these data, we continued to implement the pre-defined threshold of 0.09 previously validated in TNBC and NSCLC to analyze the clinical response data from the IMvigor210 clinical trial cohort.

The IMvigor210 combined platinum resistant and platinum contraindicated patients for a total of 429 subjects of which 348 patients were successfully RNA sequenced [4]. For these 348 patients, the median follow-up time was 20.6 months (95% CI 18.0–21.7) and the median survival was 8.8 months (95% CI 7.4–10.6). OS at the end of study (24.5 months) was 28.6% (95% CI 23.8–34.5%). In total, 40% of the genomic cohort had positive IO Scores. Median OS was 15.6 months (95% CI 10.0-NR) for IO Score positive patients versus 7.5 months (95% CI 6.0–9.2) for IO Score negative patients (Fig. 4A). IC2/3 positive subjects comprised 34% of the cohort. IC2/3 positive patients had a median survival of 12.8 months (95% CI 9.9-NA) versus 7.7 (95% CI 5.9–9.2) for IC2/3 negative subjects (Fig. 4B). The HRs for IO Score and IC2/3 positive subjects were similar as was the percentage of patients alive at end of study: HR for OS at the end of the study for IO Score and IC2/3 were 0.59 (95% CI 0.45–0.78, p < 0.001) and 0.62 (95% CI 0.47–0.82), respectively. At the end of the study, 41.5% (95% CI 33.5–51.4%) of IO positive subjects and 20% (95% CI 14.6–27.5%) of IO Score negative subjects were still alive. Similarly, 41.4% (33.2–51.6%) of IC2/3 positive subjects and 21% (15.4–28.7%) of IC2/3 negative subjects were alive at end of study.

Given the failure of the IMvigor211 to reach the primary endpoint of improved OS in IC2/3 patients, we performed an exploratory analysis to determine if IC2/3 positive patients who were also IO Score positive has better OS (Fig. 4C). These “double positive” patients were 24% of the total cohort, had a median OS of 17.8 months (12.8-NA) with an HR for OS of 0.48 (95% CI 0.34–0.68) and an end of study survival of 48% (95% CI 38.2–60.5%). “Double negative” patients were 50% of the population with median OS of 7.5 month (95% CI 5.7–9.6) and end of study survival of 18.7% (95% CI 12.9–27.2%).

In order to test whether IO Score was quantitatively related to objective response, the continuous IO Score was plotted against the four objective response categories (Additional file 1: Figure S1). The scores were significantly different when comparing the average IO continuous score for patients with a CR to the average of the continuous score patients with a PR, SD, and PD (p < 0.005), when response was compared to non-response (CR/PR versus SD/PD, p < 0.005), or when disease control was compared to progressive disease (CR/PR/SD to PD, p < 0.01). There was also a significant ordinal trend for IO Score across the objective response categories when ordered by decreasing response (p < 0.001).

Mariathasan et al. explored 28 different published clinical factors and biomarker signatures in the IMvigor210 cohort [10], including nineteen genomic signatures. IO Score was tested in a series of bivariate equations with each of these factors and maintained significance with every explored biomarker (Fig. 5A and B and Additional file 2: Figure S2A, Additional file 3: Figure S2B, Additional file 4: Figure S2C) and five of these biomarkers maintained significance with the IO Score. Of the standard clinical factors, ECOG status and regional versus distant metastasis maintained significance with the IO Score. Of the twenty-one genomic signatures, two maintained significance with the IO Score—a previously established, prognostic bladder cancer classifier (Lund classification [37, 38] HR = 1.08 p < 0.025 classification) and a cell cycle regulatory signature (HR = 1.34 p < 0.05). IC2/3 narrowly missed significance for OS in this cohort (HR = 0.53 p = 0.06). TMB was independent of IO Score when using either the previously FDA pan-cancer established threshold of  ≥ 10 mut/MB (TMB-high) (HR = 0.65 p < 0.01) or the study specific top quartile threshold fit to the IMvigor210 dataset (HR = 0.53 p < 0.005) (Additional file 5: Figure S3A). Of note, in the IO Score/TMB-high multivariate analysis, neither HR was significantly different from its univariate HR (univariate HRs 0.59/0.57 for IO Score/TMB-high, respectively; multivariate HRs 0.57/0.65) leading to the exploratory analysis of a decision tree model, where a patient was deemed “positive” if they were IO Score + or TMB-high and “negative” if they were neither (Additional file 5: Figure S3B). The HR for the decision tree model (n = 272) was similar, equaling 0.579 (95% CI 0.429–0.782, p < 0.001) with the key difference being the percentage of patients identified as positive: 61% by the model versus 40% and 41% for the IO Score and TMB-high respectively [Additional file 5: Figure S3C], with 22% of patients being positive for both IO Score and TMB-high.

An ideal biomarker for predicting ICI benefit would perform well across many tumor types and treatment indications, would be independent of clinicopathologic features, and would improve on currently established biomarkers. The IO Score has been previously shown to be associated with improved response to ICI therapy in NSCLC and TNBC [11‐15]. Consistent with the hypothesis that the IO Score classifies three components of the TIME that are common to tumors of epithelial origin, IM, MSL (linked with the expression of cancer associated fibroblasts) and M, we tested whether the IO Score would be applicable to classification of bladder cancer.

We first confirmed the TIME classification function, using our previously established threshold for binary classification, to distinguish those strongly expressing the IM signature from those which were more enriched for MSL or M using TCGA-BLCA gene expression data in ICI-naïve patients. We then tested this validated algorithm and threshold for an association with clinical response to ICI therapy in the IMvigor210 bladder cancer clinical trial cohort. The strong association of the IO Score with OS was independent of other explored biomarkers, including both TMB and PD-L1.

The IMvigor210 study resulted in an accelerated, non-biomarker contingent approval for atezolizumab in platinum-refractory patients but was later withdrawn based on a negative confirmatory trial in the IC2/3 biomarker selected patient population (IMvigor211) [9]. This withdrawal highlights the need for careful selection of a biomarker that demonstrates consistent biology when being used for patient selection in a trial. We hypothesize that using a biomarker such as IO Score to enrich for likely responders would minimize the risk of clinical trial failures. Interestingly, one potential use of the IO Score is to combine it with other biomarkers. In the case of IC2/3, the combination of “double positives” increased median survival by 4.9 months, decreased the HR by 0.12, and increased the percentage of patients alive at end of study by 7% over IC2/3 alone. In the case of TMB and IO Score, where the bivariate analysis showed both markers were independent with little change in HR, the analysis was the inverse; a patient was deemed “positive” if he or she were positive for at least one of the biomarkers. This decision tree model led to an increase in the number of patients deemed positive overall (61% versus the 40% and 41% for IO Score and TMB-high, respectively) with little loss in clinical performance. Mariathasan et al. [10] attempted to identify biomarkers from the IMvigor210 trial to enrich for likely responders and yielded several potential candidates. While several of these signatures were significant in univariate models, one striking observation from our study is that of the 21 signatures tested, only two signatures, the Lund signature, and a cell-cycle regulator signature, maintained independence with IO Score (Fig. 5 and Additional file 2: Figure S2).

The demonstrated clinical utility of the IO Score may be due to the unsupervised clustering performed to first identify robust classifiers before proceeding to building the algorithm. This approach to biomarker development is similar to the approach taken by Perou and colleagues who demonstrated the use of hierarchical clustering of gene expression data to identify likely responders to endocrine therapy within hormone positive breast cancer. Their work led to the development of a targeted biomarker panel, PAM50 [39], now an FDA 510 (k) cleared assay to help inform use of adjuvant chemotherapy in addition to endocrine therapy. Partially because PAM50 was built on a robust classifier based on shared components between cancer types, it was confirmed and further defined as an informative biomarker in hormone-sensitive prostate cancer [40, 41].

Given our prior approach and published data in TNBC and NSCLC, we believe that by assessing the IO Score as a classifier of the TIME in ICI naïve patients, we can better inform clinical decision making for mUC patients due to the IO Score’s strong association with response to ICI therapy in multiple tumor types. Future studies in both bladder cancers and additional tumor types using randomized clinical trial samples are warranted.