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

Summary

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.

Introduction

Radiotherapy is the oldest and most used cytotoxic therapy in oncology and is responsible for 40% of all cancer cures.¹ Innumerable technological advances have been made since the emergence of radiotherapy, allowing for better anatomic dose targeting. However, similar advances have not been made in the physical radiotherapy dosing paradigm. A one-size-fits-all uniform prescription of doses to all patients with a given cancer diagnosis is still used, with a protracted-fractional method, based on experiments that were first done on rams and rabbits more than 80 years ago.² Physical doses of radiotherapy are prescribed on the basis of energy absorbed by tissue (measured in Gy) and historically the field has recognised that the physical interaction of radiation with tissue results in a biological effect (eg, DNA damage, clinical tumour response, and carcinogenesis).³ However, substantial interpatient heterogeneity exists in the biological effect of a given physical dose of radiotherapy; patients we treat uniformly do not have a uniform response. Previously, we have found that this difference in radiation-induced biological effect can be quantified at the patient level using tumour genomics and, subsequently, that it can be modulated by the treating radiation oncologist.^4,5

The clinical heterogeneity of radiation response, even within cancer types, is well established. This heterogeneity is at least partly driven and influenced by changes in the tumour genome, as shown by large-scale classification studies.⁶ Several research groups have tried to understand surrogate genomic or clinical measures for individual patient resistance to radiation (eg, using the 24-gene signature),^7,8 including but not limited to human papillomavirus status in head and neck cancer,⁹ and imaging-based studies in both CT-based radiomics¹⁰ and functional imaging.¹¹ Furthermore, several research groups have worked together to create genomic signatures indicative of patient groups who do or do not need radiotherapy.⁸ Additionally, efforts have been made to develop early markers of persistent microscopic disease, such as circulating tumour DNA.¹² However, none of these efforts have considered the explicit association between intrinsic tumour genomics and the physical dose of radiation, which is the standard and fundamental parameter in radiotherapy.

We previously introduced the gene expression-based radiosensitivity index (RSI), a biomarker of tumour radiosensitivity that has been validated in multiple cohorts spanning various cancer types by classifying patients as either being radiosensitive or radio-resistant.^4,5,13-23 Subsequently, we developed the genomic-adjusted radiation dose (GARD), a novel model that integrates RSI and physical dose of radiation to quantify the biological effect of a given dose in an individual patient.⁵ Use of this model allowed us the first opportunity to understand the association between intrinsic tumour radiation sensitivity and the biological effect of a given radiation dose. Furthermore, GARD allows a clinician to dose a patient’s tumour to a desired biological effect (measured in GARD units), allowing personalisation of physical radiation dose. This is a new radiotherapy prescription paradigm in which information about physical radiation dosing is enriched by a genomic dimension.

To validate the GARD-based radiotherapy dosing paradigm, we compared different outcomes and cancer types using all available clinical datasets with enough information to calculate GARD.²⁴ Here, we report our analysis of the association between biological effect, quantified by GARD, and time to first recurrence and overall survival across seven cancer types.

Methods

Study design and clinical cohorts

In this pooled pan-cancer analysis, we collected data from a series of 11 previously published datasets that had sufficient information to calculate GARD (ie, gene expression via microarray, physical radiation dose, and dose per fraction) and clinical outcome. All analysed data were collected under study specific protocols, and published previously (appendix pp 1-3). These cohorts include patients treated with radiotherapy for breast cancer (including triple-negative breast cancer),^16,17,25 glioma,¹⁴ pancreatic cancer,¹⁸ endometrial cancer,²² melanoma,¹⁹ head and neck cancer,¹³ and non-small-cell lung⁵ cancer (table). These 11 cohorts comprised 1615 patients, of whom 1218 (45·4%) were treated with radiotherapy and 397 (14·8%) were not and were included in our analyses as negative controls to determine whether GARD is specific for patients treated with radiotherapy. 1298 patients (982 with radiotherapy, 316 without radiotherapy) were assessed for time to first recurrence and 677 patients (424 with radiotherapy, 253 without radiotherapy) were assessed for overall survival. Basic descriptive outcome measures and Kaplan-Meier plots for each cohort are in the appendix (pp 4-5) and baseline demographic and clinical data for patients have been published previously.^{5,13-19,22,25} All patients in this study were treated with radiation fraction sizes of 1·8–2·0 Gy (165 received 1·8 Gy, one received 1·85 Gy and 1052 received 2·0 Gy).

Table:

Cancer cohorts included in the recurrence and survival pooled analysis by treatment, event, and evidence category

	Evidence category	Participants	Time to first recurrence		Overall survival
			Events	Length of follow-up (years)	Events	Length of follow-up (years)
Breast (Erasmus)16	C	··	··	··	··	··
With radiotherapy	··	282/344	91	7·19 (3·09–9·04)	··	··
Without radiotherapy	··	62/344	12	7·30 (5·50–8·82)	··	··
Breast (Karolinska)16	C	··	··	··	··	··
With radiotherapy	··	77/159	19	7·28 (5·86–7·78)	··	··
Without radiotherapy	··	82/159	21	6·29 (4·28–7·80)	··	··
Breast (NKI)17	D	··	··	··	··	··
With radiotherapy	··	285/285	99	10·00 (4·90–10·00)	··	··
Triple-negative breast cancer (MCC)25	C	··	··	··	··	··
With radiotherapy	··	55/55	9	8·36 (3·30–9·62)	9	8·36 (5·17–9·62)
Triple-negative breast cancer (NKI)17	D	··	··	··	··	··
With radiotherapy	··	58/58	20	10·00 (2·12–10·00)	··	··
Endometrium (TCC)22	C	··	··	··	··	··
With radiotherapy	··	63/204	11	3·14 (1·60–6·35)	33	4·20 (2·31–6·62)
Without radiotherapy	··	141/204	7	2·84 (0·72–6·14)	29	3·58 (0·95–6·47)
Glioma (TCGA)14*	D	··	··	··	··	··
With radiotherapy	··	188/244	··	··	134	1·07 (0·68–1·74)
Without radiotherapy	··	56/244	··	··	56	0·17 (0·08–0·32)
Head and neck (NKI)4	B	··	··	··	··	··
With radiotherapy	··	92/92	28	1·40 (0·57–2·99)	··	··
Melanoma (TCC)19†	C	··	··	··	··	··
With radiotherapy	··	10/41	5	1·40 (0·89–3·08)	7	2·17 (1·24–4·52)
Without radiotherapy	··	31/41	23	1·83 (0·96–2·96)	22	3·44 (2·00–5·43)
Non-small-cell lung (MCC)5	C	··	··	··	··	··
With radiotherapy	··	60	23	1·44 (0·72–4·28)	38	3·11 (1·95–5·27)
Pancreas (TCC)18	C	··	··	··	··	··
With radiotherapy	··	48/73	··	··	33	1·97 (1·04–4·92)
Without radiotherapy	··	25/73	··	··	17	2·29 (1·14–2·80)

Data are n, n/N, or median (IQR). Cohorts are listed by cancer site of interest and with study groups in parentheses. Evidence categories were defined as follows: category A was defined as a prospective trial that directly assesses the biomarker of interest; category B as a prospective study designed to assess a different clinical question and the biomarker is assessed prospectively or retrospectively; category C as a study that is based on a prospective observational registry, but the treatment and follow-up of patients are usually per standard of care; and category D as a retrospective observational study in which all tissue and clinical features are collected retrospectively. GARD=genomic-adjusted radiation dose. MCC=Moffitt Cancer Center. NKI=Netherlands Cancer Institute. TCC=Total Cancer Care. TCGA=The Cancer Genome Atlas. *20 patients were censored from the TCGA glioma cohort because they were given short-course palliative hypofractionated regimens. ^†One patient from the melanoma cohort was censored because they were prescribed short-course palliative or hypofractionated radiation doses, because GARD was not designed for hypofractionated regimens.

Simon and colleagues²⁶ proposed categories to define levels of evidence for biomarker studies. Category A was defined as a prospective trial that directly assesses the biomarker of interest; category B as a prospective study designed to assess a different clinical question and the biomarker is assessed prospectively or retrospectively (very often in this type of study the biomarker question is statistically underpowered); category C as a study that is based on a prospective observational registry but the treatment and follow-up of patients are usually per standard of care; and category D as a retrospective observational study in which all tissue and clinical features are collected retrospectively. On the basis of these criteria, we assigned a category to each of the cohorts in this study. These data represent all available cohorts with the information available to calculate the variable of interest, and taken together represent a high level of evidence for biomarker-based studies.²⁶

Outcomes

We had two clinical outcomes of interest, time to first recurrence and overall survival, both of which were calculated from the end of treatment for all cohorts except for the glioma and endometrial cancer cohorts, which were both calculated from date of pathological diagnosis, and the head and neck cancer cohort, which was calculated from date of randomisation. Time to first recurrence included time to local, regional, and distant metastasis. We used time to first recurrence because of the heterogeneity of the available data, and this outcome represented what we thought was the most conservative choice.

Statistical analysis

GARD was calculated as previously reported,⁵ scaling RSI within each cohort (details on this calculation and the genes involved in RSI are in the appendix p 1).⁵ RSI was derived as a surrogate for the surviving fraction at 2 Gy (SF2), and therefore our results are true only for doses near to this. We report all total doses in the form of biologically equivalent doses (EQD2) for more appropriate comparison, assuming an α/β=10, although we will continue to refer to this dose as the physical or total radiation dose for clarity of exposition. Spatial intratumoural heterogeneity, now known to be a possible confounder, was not considered here.

We plotted the joint distributions of radiotherapy dose and GARD to show the association between the two for each patient, by cancer type, with the associated distributions of each association plotted as histograms. In this plot, we combined each cohort by cancer type for ease of clinical comparison. For each cohort, we plotted the distributions of individual dose and GARD in stacked histograms.

We analysed individual cohorts using Cox proportional hazards models. To compare cohorts that are heterogeneous in outcome and disease type, we used a Cox proportional hazards model, stratified by cohort, with GARD, physical radiation dose, or sham-GARD (ie, patients who did not receive radiotherapy) as the only covariate. This approach allowed each cohort to have a different baseline hazard function, yet still enabled us to determine a common GARD effect. Patients who did not receive radiotherapy were pooled together to form a cohort, sham-GARD, for a similar analysis. To calculate GARD for patients who did not receive radiotherapy (ie, sham GARD), we assumed that they all received the standard-of-care radiation dose (50 Gy in 25 fractions for breast; 50 Gy in 25 fractions for pancreas; 60 Gy in 30 fractions for glioma, melanoma and lung; and 54 Gy in 27 fractions for endometrium; there was no sham GARD group for head and neck cancer). We acknowledge that this is a strong assumption, however, this dose seemed reasonable to determine the association of GARD with recurrence and survival in non-radiotherapy treated patients. Notably, this sham-GARD analysis is functionally equivalent to testing for the RSI as a prognostic biomarker in patients not treated with radiotherapy, something we have previously studied in-depth,¹⁶ because GARD and RSI are rank preserved when all patients are given equivalent physical doses of radiotherapy.

We did the pooled analysis using a Cox proportional hazards model stratified by cohort, and a χ² statistic calculated using the Wald test. To determine whether GARD is a pan-cancer predictor of radiotherapy benefit, we calculated the interaction between GARD and whether patients received radiotherapy or not for both patient pools (ie, time to first recurrence and overall survival) using the Wald statistic. To assess the association between radiotherapy and GARD, we created Cox models using all patients (treated with and without radiotherapy) with GARD and radiotherapy as predictors, along with their interaction term. We tested the resulting models, for both time to first recurrence and overall survival, for significance using one-way ANOVA and the Wald statistic. We fit an additional model to explore potential non-linear effects using a restricted cubic spline with three knots (at 13·1, 23·0, and 46·9) to model GARD. Finally, we did an exploratory analysis for recurrence by excluding patients with low GARD on the basis of a previously published GARD cutpoint (GARD <21).

To test how GARD compares directly to the physical dose of radiation (EQD2), we did a second Cox regression analysis for each cohort using physical dose of radiation as the covariate. To understand the effect of increasing GARD (increased biological effect delivered, measured in units of GARD) for any patient, we fit a relative hazard function stratified by cohort (n=11) to assess whether GARD has an association with the outcomes of interest.

To investigate the clinical significance of our results, cancer type and baseline hazard functions were reintroduced to predict absolute survival probabilities at 3 years after starting treatment using the fitted Cox models, plotted as a nomogram beneath each outcome model.

A result was considered to be significant if the p value was less than 0·05. All statistical analyses were done with R (version 3.6.2) and rms package (version 6.2-0).

Role of the funding source

There was no funding source for this study.

Results

We calculated GARD for each patient who received radiotherapy in each cohort. Although the range of physical radiation doses in these modern cohorts was limited to values near those of standard of care and were delivered in standard fraction sizes, GARD reveals a wide range of predicted biological effects, enabling understanding of outcomes at a higher resolution (figure 1; appendix p 3).

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.

The pooled analysis of all available data showed that GARD is associated with time to first recurrence (HR 0·98 [95% CI 0·97–0·99]; p=0·0017; figure 2A; appendix p 5) and overall survival (0·97 [0·95–0·99]; p=0·0007; figure 3A; appendix p 5) in patients who received radiotherapy. As a control, the same analysis in patients not treated with radiotherapy (sham-GARD) showed no association between GARD and either time to first recurrence or overall survival (HR of 1·00 [95% CI 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; figure 2B, 3B).

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.

We found a significant interaction between GARD and radiotherapy treatment status for overall survival (Wald statistic p=0·011; appendix p 7), but not for time to first recurrence (Wald statistic p=0·22; appendix p 7). Using a previously defined cutpoint (GARD <21),²⁵ the interaction for recurrence on the linear model was also not significant (Wald statistic p=0·060). In an exploratory analysis using non-linear restricted cubic splines, we found varying associations for different ranges of GARD and time to first recurrence (appendix p 8).

We found no association between physical dose of radiotherapy (EQD2) and time to first recurrence (HR 0·99 [95% CI 0·97–1·01]; p=0·53; figure 2C; appendix p 6) or overall survival (1·00 [0·96–1·04]; p=0·95; figure 3C; appendix p 6).

We found GARD to be a significant linear variable for both time to recurrence and overall survival, with an increase in GARD corresponding with a decrease in the GARD-specific relative hazard (time to first recurrence: p=0·0017; overall survival: p=0·0007; appendix p 9). In an analysis of 3-year probability of recurrence and 3-year overall survival we found GARD to be a significant, continuous, predictive biomarker of survival and recurrence in patients treated with radiotherapy (figure 4).

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.

Discussion

Radiotherapy is the most commonly used anticancer treatment and is prescribed on the basis of an empiric, one-size-fits-all dosing paradigm. In this pan-cancer pooled analysis, we found that the biological effect of radiotherapy, as quantified by GARD, is associated with improvements in time to first recurrence and overall survival in patients with cancer treated with radiotherapy, but physical dose is not. We found that the interaction of GARD and radiotherapy is significant for overall survival, establishing GARD to be predictive of benefit from treatment with radiotherapy.

The European Organisation for Research and Treatment of Cancer has summarised the clinical data supporting RSI and GARD as representing “near level one evidence” for its use in clinical practice.²⁷ Furthermore, Khan and colleagues have published a validation study of RSI in bladder cancer samples from a phase 3 prospective clinical trial (ie, a category B study).²⁸ Although Simon’s classification of biomarker studies does not specify a role for meta-analyses, we propose that the combined evidence developed for RSI and GARD represents level one evidence (ie, consistent results in multiple category B, C, and D studies including the current pooled analysis). However, we acknowledge that definitive evidence can only come from a randomised clinical trial.

Our findings make sense in the setting of clinical trials of dose escalation of radiotherapy in unselected patients,²⁹ which has not been found to improve overall survival, but does not mean that physical dose of radiation is not an important parameter. How do you reconcile the results of these trials with the proposal that each unit increase in GARD positively affects clinical outcome? Although this finding stands in contrast with these negative trials of dose escalation, it is not dissonant. First, given our new understanding of heterogeneous response to homogeneous physical dose of radiation, we realise that uniform dose escalation without understanding intrinsic radiosensitivity is actually not uniform. Because the biological effect of a given dose varies widely between individual patients, the effects can be obfuscated in unselected patients. A GARD-based approach allows for the understanding of this non-uniform response to dose escalation: some patients benefit while others are potentially harmed when they receive unnecessary extra doses and associated toxicity, such as is the leading hypothesis for the results of the RTOG 0617 trial.²⁹ Instead of rejecting the hypothesis that dose escalation is beneficial, GARD allows us see which patients would benefit most from dose optimisation, permitting next-generation trials of personalised radiotherapy dose. To make progress, much like in medical oncology, incorporation of the extra dimension of genomics might provide a higher resolution of understanding the biological effect of the radiation dose being delivered to patients and how this can affect their outcomes. Thus, shifting the radiotherapy dosing paradigm from one that is based on physical dose of radiation alone to one that combines physical dose and biological effect, as quantified by GARD, is crucial.

GARD is commonly mistaken to be a model of clinical outcome; however, GARD was never developed, trained, or optimised to predict clinical outcome. GARD is a model of the biological effect of radiotherapy and is capable of predicting outcome because it captures the differential clinical benefit of radiotherapy across patients. But an important caveat is that the effect size of its impact on outcome is limited to the therapeutic effect of radiotherapy. Although this fact might result in a less than impressive model of clinical outcome, the clinical use of such a model is paramount because it provides clinicians with information on the relative benefit of the treatment they are providing.

However, like all pooled data analyses and meta-analyses, our study has some limitations, most of which are secondary to the heterogeneous nature of the cohorts and interventions. In particular, comparing patients across cancer types makes more standard clinicopathological comparisons impossible because of the differences in these variables between disease sites, and there is a subsequent paucity of overlap in collected data between cohorts. Full analyses were completed in each individual cohort in previous publications, which showed the independent value of RSI and GARD controlling for other (disease-specific) variables, including chemotherapy. Additionally, because GARD is currently limited to understanding dosing near standard fractionation, we are unable to confidently analyse or model some of the newer hypofractionated or hyperfractionated regimens. Furthermore, because of the heterogeneity of our cohorts, we are unable to address chemoradiation-specific questions here. Finally, our survival analysis does not account for differences in salvage therapy. These limitations aside, a paradigm driven by GARD defines, for the first time, an actionable measure of the biological effect of radiotherapy in a given patient. This knowledge allows a reframing of our thinking, enabling us to quantify, modulate, and personalise the biological effect of the radiotherapy given to patients, not just the physical dose, and establishes the first clinically validated approach to genomic-based radiation oncology.

The fundamental goal of radiotherapy planning is to deliver the prescribed physical dose of radiotherapy to the target volume while minimising the dose to normal tissue. The integration of three-dimensional anatomy into radiation treatment planning systems, brought about by the invention of the CT scanner, enabled the geometric optimisation of radiation fields. This new method of planning led to a shift from one-size-fits-all radiotherapy field shapes to ones that are anatomically personalised, improving the ability of radiation oncologists to further spare normal tissue for each individual patient. The resultant techniques (eg, intensity modulated radiation therapy and stereotactic body radiation therapy) have been shown to decrease toxicity in clinical trials and changed the standard of care in several cancers, including lung and prostate. In other words, the development of optimisation techniques in radiation planning has led to substantial clinical gains for patients treated with radiotherapy. Similarly, we propose that GARD can be used to maximise the biological effect of physical dose to the tumour, while respecting standard-of-care guidelines for normal tissue. Previously, we found that GARD-based optimisation of the physical dose of radiation to a tumour can be achieved within standard-of-care guidelines for up to 75% of patients with non-small-cell lung cancer who receive post-operative radiotherapy.³⁰ In this previous study, we found that GARD can be used to rationally optimise (through escalation or de-escalation) radiation benefit and toxicity, something that has not been possible with physical dose of radiotherapy alone.

As the first genomic framework to predict radiotherapy benefit, we believe the clinical use of GARD will be paramount. To our knowledge, GARD is the first model to provide clinical radiation oncologists with decision support information to modulate the potential benefit of radiotherapy for each individual tumour. Additionally, it allows quantification of the effect size for each individual patient. This information will provide radiation oncologists with a completely new set of tools to optimise radiation dose using genomics. Importantly, GARD has been established in the Clinial Laboratory Improvement Ammendments laboratory at Moffitt Cancer Center (Tampa, FL, USA), and clinical trials using GARD are expected to start later this year. Finally, we are not suggesting abandoning physical dose of radiation, but instead, like the CT scanner did for x-ray, we suggest enhancing dose with another dimension—genomic data—allowing us to see each individual patient’s potential for radiotherapy benefit at a higher resolution.

Research in context.

Evidence before this study

Although the dose of radiation can be directly measured, the biological effect can only be quantified using the linear quadratic model. This effect varies across tissue models and patients and we do not have the methods to measure it directly; thus clinicians have had to assume that a given dose of radiation results in the same clinical effect. Because the null hypothesis is that all patients have the same likelihood of benefit from radiotherapy, the radiation dose prescription paradigm is to prescribe a uniform dose of radiation on the basis of clinical factors including disease site and stage. In a previous study, we developed the genomic-adjusted radiation dose (GARD), a model to predict individual patient radiotherapy effect on the basis of the linear quadratic model and a previously developed and clinically validated gene expression-based biomarker of tumour cell radiosensitivity, the radiosensitivity index (RSI). GARD provides the first model to quantify and optimise the effect of radiation for each individual patient.

Added value of this study

Using a pooled analysis of 11 cancer cohorts, we found that GARD predicts the therapeutic benefit of radiotherapy, quantifies the relative benefit of radiotherapy for each individual patient, and outperforms radiation dose, the current standard of care, in terms of recurrence and survival. GARD provides a view of the effect of radiotherapy for each individual patient, providing radiation oncologists with crucial information to maximise the potential benefit of radiotherapy for each individual patient. A GARD-based prescription framework rejects the null hypothesis that radiotherapy benefit is homogenous and highlights that there is a clinically significant opportunity for the optimisation of radiation dose and effect for each individual patient.

Implications of all the available evidence

A paradigm based on radiation dose and GARD provides radiation oncologists with a full view of both the physical dose and the biological effect for each individual patient. We propose that the incorporation of genomics into radiotherapy planning becomes the new paradigm for radiation dose prescription and clinical trial design.

Data sharing

A statement regarding availability of code and data from this study is in the appendix (p 1).