Pan-cancer prediction of radiotherapy benefit using genomic-adjusted radiation dose (GARD): a cohort-based pooled analysis

Abstract

Background: Despite advances in cancer genomics, radiotherapy is still prescribed on the basis of an empirical one-size-fits-all paradigm. Previously, we proposed a novel algorithm using the genomic-adjusted radiation dose (GARD) model to personalise prescription of radiation dose on the basis of the biological effect of a given physical dose of radiation, calculated using individual tumour genomics. We hypothesise that GARD will reveal interpatient heterogeneity associated with opportunities to improve outcomes compared with physical dose of radiotherapy alone. We aimed to test this hypothesis and investigate the GARD-based radiotherapy dosing paradigm.

Methods: We did a pooled, pan-cancer analysis of 11 previously published clinical cohorts of unique patients with seven different types of cancer, which are all available cohorts with the data required to calculate GARD, together with clinical outcome. The included cancers were breast cancer, head and neck cancer, non-small-cell lung cancer, pancreatic cancer, endometrial cancer, melanoma, and glioma. Our dataset comprised 1615 unique patients, of whom 1298 (982 with radiotherapy, 316 without radiotherapy) were assessed for time to first recurrence and 677 patients (424 with radiotherapy and 253 without radiotherapy) were assessed for overall survival. We analysed two clinical outcomes of interest: time to first recurrence and overall survival. We used Cox regression, stratified by cohort, to test the association between GARD and outcome with separate models using dose of radiation and sham-GARD (ie, patients treated without radiotherapy, but modelled as having a standard-of-care dose of radiotherapy) for comparison. We did interaction tests between GARD and treatment (with or without radiotherapy) using the Wald statistic.

Findings: Pooled analysis of all available data showed that GARD as a continuous variable is associated with time to first recurrence (hazard ratio [HR] 0·98 [95% CI 0·97-0·99]; p=0·0017) and overall survival (0·97 [0·95-0·99]; p=0·0007). The interaction test showed the effect of GARD on overall survival depends on whether or not that patient received radiotherapy (Wald statistic p=0·011). The interaction test for GARD and radiotherapy was not significant for time to first recurrence (Wald statistic p=0·22). The HR for physical dose of radiation was 0·99 (95% CI 0·97-1·01; p=0·53) for time to first recurrence and 1·00 (0·96-1·04; p=0·95) for overall survival. The HR for sham-GARD was 1·00 (0·97-1·03; p=1·00) for time to first recurrence and 1·00 (0·98-1·02; p=0·87) for overall survival.

Interpretation: The biological effect of radiotherapy, as quantified by GARD, is significantly associated with time to first recurrence and overall survival for patients with cancer treated with radiation. It is predictive of radiotherapy benefit, and physical dose of radiation is not. We propose integration of genomics into radiation dosing decisions, using a GARD-based framework, as the new paradigm for personalising radiotherapy prescription dose.

Funding: None. VIDEO ABSTRACT.

Figure 1:. Patient-level association scatter plot of physical dose of radiation delivered versus the associated GARD

In this plot, all cohorts by cancer type are combined for ease of comparison. Several patients with extremely high GARD (>100) are not plotted here for ease of visualisation; they are shown in the appendix (p 3). The associated kernel density estimates (density) for each distribution are plotted on the right and top. GARD=genomic-adjusted radiation dose.

Figure 2:. Individual Cox proportional relative hazards for each cancer type and time to first recurrence

Cox proportional hazards models with GARD as a covariate for first recurrence for each site individually, and for the pooled sites (A); with sham-GARD as the covariate for the same analysis (B); and with physical dose of radiation as the covariate for the same analysis (C). Cohorts are listed by cancer site of interest and with study groups in parentheses. Numerical data for relative hazards and 95% CIs have been rounded to two decimal places for ease of presentation and plots have been drawn to a higher level of accuracy. Dashed vertical lines indicate the pooled relative hazard. EQD2=equivalent dose at 2 Gy. GARD=genomic-adjusted radiation dose. MCC=Moffitt Cancer Center. NA=not applicable. NKI=Netherlands Cancer Institute. TCC=Total Cancer Care. *Non-convergence of Cox models secondary to singular dose values.

Figure 3:. Individual Cox proportional relative hazards for each cancer type and overall survival

Cox proportional hazards models with GARD as a covariate for overall survival for each site individually, and for the pooled sites (A); with sham-GARD as the covariate for the same analysis (B); and with physical dose of radiation as the covariate for the same analysis. Numerical data for relative hazards and 95% CIs have been rounded to two decimal places for ease of presentation and plots have been drawn to a higher level of accuracy. Dashed vertical lines indicate the pooled relative hazard. EQD2=equivalent dose at 2 Gy. GARD=genomic-adjusted radiation dose. MCC=Moffitt Cancer Center. NA=not applicable. NKI=Netherlands Cancer Institute. TCC=Total Cancer Care. TCGA=The Cancer Genome Atlas. *Non-convergence of Cox models secondary to singular dose values.

Figure 4:. Relative hazard per unit GARD at 3 years after diagnosis as predicted by the stratified Cox model for both time to first recurrence (A) and overall survival (B) by cancer site

The nomogram below each waterfall plot shows that the effect of the relative hazard, as determined by GARD, can be interpreted in the context of a specific cancer site, in which predicted absolute survival probability underlies the corresponding GARD value for each cohort. The red square and vertical dashed line show the mean GARD for the cohort, with the red to blue gradient showing the gradient from highest to lowest relative hazard. Several extremely high GARD outliers have been removed from the plot for ease of visualisation, but not from the analysis, and they are included in the appendix (pp 9-10). Error bars are 95% CI. GARD=genomic-adjusted radiation dose.