Personalization of Radiotherapy Dose in HPV Positive Oropharynx Cancer Using GARD

Abstract

A central clinical goal for patients with HPV-positive oropharynx cancer has been to reduce radiation doses while maintaining cure rates. Recent results of Phase 3 prospective trial HN005 demonstrated that RT dose de-escalation can not be safely done based on clinical factors alone. We have previously shown that the genomic adjusted radiation dose (GARD) is predictive of radiation treatment benefit and can be used to guide RT dose selection. We hypothesize that GARD can be used to guide RT dose de-escalation in HPV-positive OPSCC patients. Gene expression was analyzed for 191 formalin-fixed paraffin-embedded samples from HPV-positive OPSCC patients within an international, multi-institutional, prospective/retrospective observational study including patients with AJCC 8th edition stage I-III. Two RT dose fractionations were utilized for the majority of primary RT cases (70 Gy in 35 fractions or 69.96 Gy in 33 fractions). Median RT dose was 70 Gy (range 51.0-74.0), survival at 36 months and 60 months was 94.1% and 87.3%, respectively. Cox proportional hazards analyses were performed with GARD as a continuous variable and time-dependent ROC analyses compared the performance of GARD to clinical variables alone. Despite near-uniform RT dosing, GARD reveals significant heterogeneity (range 15.4 - 71.7) in predicted effect. In univariate analysis, GARD was associated with an improvement in OS (HR = 0.941 (0.888, 0.998), p = 0.041). In multivariable analysis, each unit increase in GARD was associated with an improvement in OS (HR = 0.943 (0.891, 0.999), p = 0.046) where stage was not (T stage HR = 1.992 (0.711-5.576), p=0.190, N stage HR = 2.367 (0.867-6.460), p=0.093). ROC analysis for GARD at 36 months yielded an AUC of 78.26 (65.14, 91.38) compared with 71.20 (54.47, 87.93) for standard clinical variables. We identify two GARD-based strategies to RT dose personalization which are predicted to yield improved clinical outcomes, while delivering an average lower RT dose. In this multi-institutional cohort of patients with HPV-positive OPSCC, GARD associates with OS, outperforms standard clinical variables and provides a novel genomic strategy to RT dose personalization. We propose that GARD should be incorporated in the diagnostic workup of HPV-positive OPSCC patients.

Keywords: GARD; genomics; mathematical model; personalized medicine; radiation therapy.

Figure 1.. GARD exhibits large underlying genomic heterogeneity in radiation effect compared to radiation dose alone.

Left: EQD2 (median 70.0, IQR 0.7) is plotted against associated GARD (median 39.1, IQR 12.6) for each patient in the cohort. Kernel density estimates are plotted on each edge to show the distributions of the individual variables. Patients that received an EQD2 of 69–71 Gy (standard dosing) are indicated in yellow color. Right: The GARD value of all patients receiving an EQD2 of 69–71 Gy (standard of care) highlight GARD’s ability to stratify patients by their genomic heterogeneity. Data points are overlaid with a box-whisker plot with box representing quartiles and whiskers extending to 1.5 times IQR.

Figure 2.. GARD is a continuous predictor of OS in radiation treated patients with HPV-positive OPSCC.

(a) Cox proportional hazards analysis demonstrates significant continuous association between GARD and OS for patients treated with primary definitive RT and censored at 60 months (p = 0.041, HR = 0.941 (0.888, 0.998) per unit GARD). (b) Cox proportional hazards analysis demonstrates significant continuous association between GARD and OS for the subset of patients treated with standard of care primary definitive RT (EQD2 69–71Gy) (p = 0.019, HR = 0.920 (0.857, 0.986) per unit GARD).

Figure 3.. AUC analysis shows GARD outperforms standard clinical variables.

In this AUC analysis we compare Cox regression of clinical variables (red) to GARD alone (blue) and a combined model using all factors (black). The combined model shows dramatic improvement compared to clinical variables alone. This model included all 191 definitive primary RT patients, and analysed outcome at 3 years.

Figure 4.. GARD predicts that uniform RT dose de-escalation in HPV-positive patients would result in an inferior clinical outcome compared to standard of care.

(A) GARD identifies HPV-positive patient subsets with differential risk of failure. An exploratory analysis identified one cut-point which group patients in two risk levels. Patients that achieve the lowest GARD (< 42) have a higher risk of failure (3 year OS = 90.5%). (B) In silico clinical trial designs. Left: We simulate unselected RT dose de-escalation (cisplatin + 60 Gy vs cisplatin + 70 Gy). We utilize the RSI distribution of the BD2Decide cohort to generate GARD for 400 virtual patients randomized to either 60 or 70 Gy, and repeat this 100 times. Right: GARD-based De-escalation. In one example of a potential trial, we simulate a GARD-selected trial where only patients with GARD ≥ 42 are eligible for randomization to de-intensification. (C) Simulation of unselected RT dose de-escalation (cisplatin + 60 Gy), results in inferior OS compared to standard of care (cisplatin + 70 Gy). The unselected in silico clinical trial predicts that patients treated with RT dose de-escalation experience a statistically significantly worse overall survival when compared to standard of care (3-yr OS 92.7% vs 94.6%, non-overlapping confidence intervals). (D) Selective de-intensification produces similar OS.

Figure 5.. GARD-based opportunities for equipoise targeted RT dose reduction: maintaining equivalent outcomes to standard of care.

(A) GARD targeted equipoise RT dose reduction. KM curves show that patients in the BD2Decide cohort that achieve a GARD of at least 32 achieve isocurative outcomes compared with the unselected cohort. Each patient can then have a prescription RT dose to match the target GARD (at least 32). Comparing this to SOC provides equivalent outcomes. (B) A histogram depicting the difference between the dose predicted to be required for each patient compared to the dose delivered. Red suggests patients were underdosed (39/175), offering opportunities to increase oncologic outcomes, and blue overdosed, indicating opportunities to decrease toxicity. (C) Calculating the difference between fractions delivered in SOC compared to the number predicted on a per patient basis reveals a large potential for toxicity reduction at the population level. This averages approximately 5 fractions (one week of radiotherapy) per patient, but with large heterogeneity.