Radiation Therapy (RT) remains the single most utilized cytotoxic agent for cancer treatment. It has been estimated that approximately 50%-60% of all cancer patients receive RT. In addition, in an analysis prior to the development of immunotherapy, RT was estimated to be responsible for 40% of all cancer cures compared to 50% for surgery and 10% for chemotherapy/targeted agents.1,2 Over the last several decades, the delivery of RT has undergone a significant transformation.3 The field has transitioned from a one-size fits all RT delivery technique to an approach that is data-driven and personalized. For example, the integration of 4D patient specific anatomic information into radiation treatment planning systems and the development of highly accurate collimators have provided radiation oncologists with tools to customize RT delivery to match each individual patient's anatomy. Thus, radiation oncology has shifted from treating standard treatment fields (eg, 4-field box) to treating individualized tumor and target volumes. In addition, the added anatomical dimension has allowed the development of dose-volume estimates of normal tissue dose, which have allowed uniform tumor RT dose escalation without an increase in normal tissue toxicity.

Although from the technical aspect, the field of radiation oncology has been marching towards a personalized approach, RT dose prescription still follows a one-size fits all approach, with most RT dose based on cancer diagnosis and stage. This clinical practice where we treat most patients with a narrow range of clinically accepted doses, is in direct conflict with the realization that cancer is the most heterogenous of all diseases that affect humans.4 How can the same RT dose be optimal for every patient with a given cancer? The obvious answer is that it is not, and every practicing clinician has experienced the heterogeneity of clinical outcomes that is observed in uniformly RT-treated (eg, from the dose perspective) patients.

The central hypothesis of genomics-driven RT is that differences in tumor genomics can inform and optimize RT dose. Tumor genomic differences are now in routine clinical use in many oncologic situations including Oncotype DX, a 21-gene signature of recurrence risk in breast cancer, Decipher a prognostic signature in prostate cancer and others.5, 6, 7 In addition, tumor sequencing is commonly utilized to assess the effectiveness of targeted therapy including MEK and B-Raf inhibitors, ALK-targeted agents and immunotherapy.8, 9, 10 In contrast, RT remains uninformed about the genomic dimension of tumors in current clinical practice.

A reasonable question is: what is the optimal way to utilize omics in the field of radiation oncology? In medical oncology, the first successful clinical utilization of genomics was to exclude node-negative breast cancer patients from adjuvant chemotherapy.5,7 This strategy made clinical sense as the overall absolute clinical benefit of adjuvant chemotherapy in these patients was relatively small (eg, 5%), resulting in most patients being over-treated and exposed to the toxicity of chemotherapy with no clinical benefit. Subsequently, genomics has been utilized to identify patients that may be treated more effectively with targeted therapy or immunotherapy.8, 9, 10 This approach has led to the improvement of clinical outcomes of subsets of patients in multiple disease sites.

In radiation oncology, several efforts to utilize genomics, clinical factors and/or biomarkers to exclude patients from adjuvant radiation in breast cancer are currently underway (eg, DEBRA trial in breast cancer, NCT04852887). This approach mimics the original strategy used by medical oncologists in excluding chemotherapy. Although there is likely a group of patients that have such low risk of recurrence where adjuvant RT may be omitted, this strategy is not going to improve the overall cancer-specific outcomes of RT-treated patients. In addition, it assumes that RT dose is currently optimized. In contrast, our group has been pursuing a different hypothesis: genomics can be utilized to better understand the therapeutic benefit provided by RT and to optimize, inform and personalize RT dose.

The development of a radiosensitivity predictive assay has been a central goal of radiation oncology for many decades. Gilbert Fletcher, a pioneer in radiation oncology, declared that an accurate predictive assay was the holy grail for the field.11 However, most efforts, including ex-vivo SF2, Tpot, and hypoxic sensors, failed to reach the clinic.12, 13, 14, 15, 16, 17, 18, 19 The “omics” revolution exponentially increased the amount of biological information available for every tumor and thus provided an opportunity to assess this question from a different perspective.20 We hypothesized that differences in gene expression across tumors could predict tumor radiosensitivity. The approach was simple: identify genes that were correlated in expression with SF2 values (as determined by the clonogenic assay) across 35 cancer cell lines. Using the statistical analysis of microarrays (SAM), we identified genes that were univariately correlated to SF2 (5% false discovery rate cutoff). We utilized the identified genes to demonstrate that it was possible to develop a gene expression-based classifier to predict cellular radiosensitivity. Furthermore, as biological validation of the method, we showed that genes identified as “predictive genes” in the classifier scheme were mechanistically involved in response to radiation.21

Building on this success, we expanded our database to include 48 cancer cells lines and integrated systems biology principles to our strategy in identifying reliable “predictive genes.”22,23 The systems approach facilitates the integration of scales; multiple biological levels of information (eg, molecular, cellular, tissue) can be integrated into algorithms.24, 25, 26 Based on this, we integrated multiple terms (gene expression, mutational information on RAS and TP53, tissue of origin) into an algorithm designed to fit individual gene-based models and SF2. After all genes in our database were assessed (7168 probesets), we developed a biological network of the top 500 genes based on known mechanistic relationships. The structure of the resulting radiosensitivity network was consistent with a scale-free network, where there are “hub” nodes that have a larger influence over the network. The ten hub genes identified by our approach were: (1) STAT1, (2) C-JUN, (3) NFkB, (4) IRF1, (5) AR, (6) SUMO1, (7) PKC, (8) HDAC1, (9) C-ABL, (10) CDK1. We used these 10 genes to train a linear regression gene expression rank model to predict SF2 which we called the radiosensitivity index or RSI. The algorithm to generate RSI has been unaltered since its original publication in 2009 (Fig. 1).22

The RSI network captures a diverse and complex biology that includes critical pathways involved in cellular proliferation, apoptosis, DNA damage repair (DDR), metabolism, stress adaptation, immune response, as well as others.22 Since the RSI network discovery was integrated among various cell line origins, this measured biology may exist within a transcriptional network supercluster, which is agnostic to cancer type. Importantly, the ten RSI network hub genes consist of 5 transcription factors, 3 kinases and 2 regulators of post-translational modification, which enables additional pathway crosstalk and/or amplification of downstream signaling diversity. Though the expression level and rank of the 10 hub genes describe the readout of the RSI algorithm, it should not be neglected that these were identified in a context of hundreds of other network genes with both stimulatory and inhibitory impacts, and in some instances having their own ‘hub’ domains. Also, the initial clustering analysis identified RAS mutation status as a major contributor to radiosensitivity, more so than tissue of origin and TP53, suggesting the Ras oncoprotein family and modulators may be integral to RT response dynamics.27 Thus, on a cellular level, RSI takes the beginning molecular composite of a cell line and describes stress-response features initiated within diverse subcellular compartments after exposure to 2 Gy of radiation.

As RSI-related biology was derived from cell lines, it was also important to describe the biology inferred within cellularly diverse patient tumors. Within an individual cell line, the RSI is a readout of cell autonomous biology related to radiosensitivity, but when measured in patient tumors, RSI is now the composite of gene expression patterns from heterogeneous cell types. To this end, single cell sequencing in breast cancer demonstrated individual cell variability in RSI, which was also associated with diversity in mutation profiles.28 Although, cancer cell autonomous RSI signal may be dampened by other cellular inputs when performing bulk RNA expression analyses, tumors with predominant biology reflective of increased or decreased radiosensitivity as estimated by RSI are still measurable; we and other investigators have demonstrated RSI is predictive of RT response across various tumor types using this approach (see below for further discussion).

A unique component of the RSI is the connection with immunobiology. Though our in vitro experiments modeling cellular responses to radiotherapy were not expected to identify immunomodulatory mechanisms, characterization of RSI-related biology in patient tumors suggests measurement of an immune element. In support of this, Strom et al. found that tumors estimated to be radiosensitive by RSI were simultaneously characterized as being replete with a panel of 12 chemokines.29 Also, we previously analyzed 10,469 primary tumors (representing 31 tumor types) and found that RSI identified distinct tumor immune microenvironments, which were enriched with interferon pathway activity, specific immune cell infiltrates (eg, CD8+ T cells, activated natural killer cells), and coordinated STAT1-CCL4/MIP-1β-IRF1 signaling networks.30 Notably, the strength of an RSI-immunobiology relationship varied by tumor type with only 20%-30% of RSI variability being explained by immune infiltration in the correlated tumor types; this implies that RSI measures more than immune activity in patient tumors. This work was also complemented by Dai et al.31 which evaluated the Cancer Genome Atlas and found that RSI was strongly associated with various immune-related features. In total, RSI may represent a biological readout that integrates multiple factors involved in mediating radiation response including intrinsic radiosensitivity and host immune response.

Many of the commercially successful signatures have been developed as prognostic signatures, by training their algorithms to predict the clinical outcome of the population of interest.5, 6, 7 In contrast, RSI was never trained or optimized to predict clinical outcome; it was only trained to predict cellular SF2. Since the clinical interest in the development of RSI was in the capturing of the therapeutic benefit provided by RT, we hypothesized that RSI would predict clinical outcome in RT-treated populations and not in patients treated without RT. In addition, as a genomic surrogate for SF2, we hypothesized that the distribution of RSI would be similar to the distribution of clinical tumor radiosensitivity.

To date, RSI has been clinically validated across 12 disease sites (21 independent cohorts) in over 4000 RT-treated patients as a predictor of outcome in RT-treated patients. Table presents a summary of all validation cohorts to date including those performed by our group and others.32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49 The validation studies include a Phase 3 clinical trial in bladder cancer, data from Phase 2 prospective trials in prostate cancer and head and neck cancer, a multi-institutional cohort in oropharynx cancer as well as data from large public and private repositories. Although most of the studies have been performed by our group, there are now 7 independent studies performed by groups from England (University of Manchester),46,47 Italy (Instituto Tumori),44 South Korea,48 a collaborative European consortium45 and Sweden (University of Lund).49 Clinical endpoints assessed include overall survival, recurrence-free survival, metastasis-free survival, local control and tumor response. Consistent with its nature as a predictor of cellular radiosensitivity, RSI is not prognostic in patients treated without RT (> 1700 patients, 10 independent cohorts, 6 different disease sites). We are only aware of 1 reported non-RT cohort where RSI was prognostic (Bladder).48 So, at a minimum the association of RSI with clinical outcome is stronger in RT-treated patients when compared to non-RT treated patients.

One interesting observation is that the ability of RSI to predict outcome is not uniform across disease sites. This could be reflective of the differences in RT therapeutic impact across disease sites or that RSI captures biology that is more important in specific biological subtypes. For example, in breast cancer, RSI is highly predictive of clinical outcome in ER negative breast cancer patients. Multiple studies have shown that RSI-sensitive ER negative patients have improved local control, recurrence-free survival and overall survival when compared with RSI-resistant patients.43,48,49 In contrast, most data in ER-positive tumors by us and others have been negative for prediction of clinical outcome for RSI.43,48,50 This may reflect that RSI does not capture the therapeutic benefit in these patients or that patients in this group are receiving more RT dose than they need, thus abrogating the impact of differences in radiosensitivity on clinical outcome. It is also possible that ER+ patients that are RSI-sensitive have a poorer prognosis at baseline and that this is improved by the therapeutic impact of RT as suggested by data from the University of Lund.49

In Figure 2, we show the distribution of RSI in 8,271 patients across 18 disease sites. Interestingly, RSI ranks known radiocurable diseases as HPV-related oropharynx and cervical cancer as 2 of the most sensitive solid tumors. On the radioresistant spectrum, gliomas, sarcomas and melanomas ranked 1, 3, and 4 respectively in the order of resistance. Another key observation is that there is RSI heterogeneity across and within disease sites, with sometimes 3 and 4-fold differences between the most sensitive and most resistant sample in our analyses.

Although our data is consistent with RSI being a predictive biomarker, this can only be demonstrated in a clinical trial where patients are randomized between treatment/no treatment with RT or between 2 different RT doses (eg, RTOG 0617, RTOG 0126). However, in several retrospective analyses that included patients treated with and without RT, we and others have shown an interaction between RSI x RT.38,49 Furthermore, in another independent retrospective analysis of breast cancer patients, we demonstrated that RSI-resistant patients benefitted from higher RT doses.43 All these data put together make a compelling case that RSI is a predictive biomarker of RT benefit.

Over the last several decades, the field of radiation oncology has been transitioning from an empiric discipline to one that is data driven. Field design has evolved from a 1-size fits all approach defined by anatomical landmarks (eg, 4 field box) to volumetric-based treatments designed on patient-specific 4D anatomical information. In contrast, RT dose is still prescribed to maximally tolerated dose (MTD) based on empiric standards defined in the 1960s-1970s.

The therapeutic objective of any RT dose protocol is to maximize the individual patient's clinical outcome with the least possible risk of normal tissue toxicity.51 Thus the “optimal dose” of RT can be defined as the dose that achieves that goal. However, the field has always viewed dose as a monolith, where all patients have the same opportunity to benefit from a given RT dose. Based on this assumption, the field has pursued a technological solution to the achievement of an optimal dose by developing more precise physical delivery of RT. While this resulted in an improved technical ability to safely deliver higher total doses to tumors, most prospective Phase 3 studies testing standard to escalated RT dose failed to show an improvement in clinical outcome and in some cases even detriment with the higher dose.52,53

One approach that is showing promise is stereotactic body radiotherapy (SBRT), whereby reducing the volume of treatment, the delivery of extremely high tumoricidal doses of RT is possible (eg, 50 Gy in 5 fractions).54 The reduction in treatment volume minimizes the normal tissue toxicity while the high RT doses delivered can abrogate any individual biological differences in radiosensitivity between tumors. Although still empiric from the dose perspective (all patients still get similar doses of RT), this is an approach that has been utilized to approach a “uniform optimal” RT dose. However, for standardly fractionated RT, if differences in tumor radiosensitivity do exist, then the idea of a single RT dose being optimal for all patients, in our view, defies logic.

To develop an approach to use genomic-data to personalize RT dose, we invented the genomic-adjusted radiation dose (GARD).55 GARD provides the first approach to quantify the effect of a given dose of RT for an individual patient. The mathematical derivation of GARD is shown in Figure 3A. In simple terms, GARD is calculated by integrating RSI into the LQ model. Since RSI is a genomic surrogate for SF2 it can be utilized to generate a tumor-specific alpha parameter, where n = 1 and d = 2 in the classic LQ equation. GARD is then generated using the classic equation for effect (E), where alpha is calculated using the RSI for each patient tumor, a constant beta and the dose and fractionation received by each patient. Thus, in the GARD mathematical construct, giving the same RT dose results in a higher GARD value for a radiosensitive tumor than a radioresistant tumor.

A classic principle of the linear quadratic model is that a higher dose always results in a higher effect.51 However, in the GARD construct, a higher dose not always results in a higher GARD. The distribution of calculated GARD for 8271 patients modelled for 45, 60, and 70 Gy is shown in Figure 3B. While on average, patients modelled at 70 Gy have a higher average GARD value, at an individual level, there are patients modelled at 45 Gy that achieve a higher GARD than other patients modelled at 70 Gy.

To validate GARD, we performed a meta-analysis where we compiled all publicly available data of RT-treated patients with genomic information and clinical outcome.56 In addition, we included all patients treated without RT that had been reported in the same cohort with the RT patients. The total cohort included 1615 patients (1218 RT-treated and 397 non-RT treated) in a total of 7 disease sites. We then tested GARD as a continuous variable as a predictor of clinical outcome including overall survival and recurrence-free survival. As shown in Figure 4, each unit increment in GARD is associated with an improved overall survival (HR = 0.97 (0.95, 0.99) P = 0.0007) and recurrence-free survival (HR = 0.98 (0.97, 0.99) P = 0.0017) in RT-treated patients but not in patients treated without RT (Sham-GARD, P = 0.87, 1). In contrast, RT dose as measured by equivalent dose at 2 Gy (EQD2) did not predict outcome for either endpoint measured (P = 0.53, P = 0.95). Therefore, the effect of RT as quantified by GARD units provides more information about the clinical outcome of RT-treated patients than the RT dose (Gy) delivered.

To determine whether GARD was capturing the therapeutic benefit of RT, we performed an interaction analysis (GARD x RT). As expected, GARD's ability to predict clinical outcome was observed only in RT-treated patients and the interaction analysis (GARD x RT) for overall survival was significant (P = 0.017), consistent with GARD being predictive of RT therapeutic benefit. While the pan-cancer analysis includes GBM patients where RT has been shown to improve OS, it also includes other disease sites where RT has no impact on OS. How can GARD predict the therapeutic benefit of RT for OS in cohorts where RT has not been shown to improve OS? One possibility is that unselected clinical trials may mask the therapeutic impact of RT in genomically-identifiable subpopulations. Similar to targeted drugs, the benefit of RT may be highest in a specific sub-population of patients. For example, approximately 4% of NSCLC patients are positive for the ALK fusion gene and these patients benefit from treatment with crizotinib. However, if crizotinib had been tested in unselected clinical trials, its therapeutic benefit in the small sub-population of ALK-fusion positive patients would have likely gone undetected. GARD is proposing a similar concept for RT; its therapeutic benefit is heterogeneous across an unselected population. For example, in a recent study (further described below) a GARD-based model predicts that unselected dose escalation from 60 Gy to 74 Gy in NSCLC patients, as performed in RTOG 0617, would result in no overall clinical gains for RT dose escalation. However, it also proposes that a small sub-population of about 17% of patients would derive clinical benefit, while the rest would be harmed by the additional dose, suggesting that accounting for genomic subpopulations may impact our overall understanding of the therapeutic impact of RT.57

In summary, GARD outperforms RT dose (Gy) because it predicts clinical outcome and RT treatment benefit. Thus, we think that as a prescription parameter, it is an improvement over RT dose and provides an important step towards a data-driven approach to RT prescription.

The historical basis of the empiric RT prescription paradigm relies on the original experiments by Regaud over 100 years ago where he observed that delivering small daily doses (fractions) of RT could achieve a desired biological effect with lower normal tissue toxicity.58 These observations were incorporated in the clinic by Henri Coutard in the 1920s-1930s, who demonstrated that the “protracted fractionated method” of RT prescription resulted in long-term cure of solid tumors.59 Subsequently in the 1960s the standard doses to control subclinical (45 Gy), microscopic (60 Gy) and macroscopic (70 Gy) disease were established.60

In parallel, there were efforts to try to understand the relationship between dose and radiation-induced toxicity. The linear-quadratic model was the result of these efforts, a mathematical formalism developed by Lea.61 In the LQ model, radiosensitivity is represented by the alpha/beta parameter where alpha is proportional to the dose (linear component) and beta to the quadratic of the dose.62 Although the LQ model has been criticized because of its lack of biological mechanistic insight into radiation response, to be fair the LQ model was developed before there was any understanding of the molecular biology underlying radiosensitivity. Importantly, the LQ model has been widely used to generate equivalent fractionation schedules that have been demonstrated to be safe and effective in clinical trials (eg, hypofractionation in prostate cancer).63

In summary, the empiric basis of RT prescription today remains based on experiments and clinical experiences first reported almost 100 years ago. It assumes that patient tumors are homogenous and that every single tumor has the same opportunity to benefit from RT. The advance introduced by GARD into the LQ model is the ability to account for an individual genomic surrogate for cellular SF2 (RSI) which allows the calculation of a tumor-specific alpha. This updates the LQ model to account for the clinical reality of tumor genomic heterogeneity. GARD no longer assumes a homogenous tumor and no longer assumes that all patients have the same clinical benefit opportunity from RT.

In Figure 5, we demonstrate how GARD can help optimize RT dose for an individual patient. In this particular example we are using data from a cohort of NSCLC patients treated with postoperative RT. We defined the RxRSI as the “optimal RT dose” in Gy, predicted by GARD to maximize the clinical outcome for a patient.57,64 Patients that achieved a GARD value of 33 have a superior local control compared to those patients with GARD below 33. In the GARD RT prescription parameter, the goal is no longer to deliver a uniform RT dose (eg, 60 Gy) but rather to achieve a uniform GARD (eg, 33). The distribution of RxRSI for lung cancer is shown in Figure 5D.

We determined the clinical potential for dose optimization using GARD-based RT prescription. As shown in Figure 5D, the current approach of uniform RT dose within a range (50-70 for postoperative RT in NSCLC) fails to achieve RxRSI in more than half of the clinical situations. However, the GARD-based RT prescription can achieve RxRSI in up to 75% of patients without prescribing RT doses outside the standard of care. Thus, there is a large opportunity to utilize GARD to optimize RT dose for patients.

One key observation is that in order to achieve GARD = 33, the RxRSI for each patient is different. Therefore, the toxicity cost to achieve an “optimal RT dose” is different for each patient. To further understand this, we developed a penalized model that incorporates both GARD and a toxicity estimate to quantify the “excess toxicity” delivered to patients when they receive higher doses than their RxRSI.57 It also helps estimate the toxicity cost of achieving RxRSI when a higher RT dose than standard is required. We tested this GARD-based model with a toxicity parameter by determining whether it would predict the results of RTOG 0617; that uniform unselected RT dose escalation to 74 Gy would result in worse local recurrence than 60 Gy. We developed an in-silico clinical trial where we virtually randomized 400 patients to either 60 Gy or 74 Gy. RSI was randomly assigned to each patient from the measured distribution of RSI for NSCLC. GARD-based predictions of clinical outcome for each patient were then calculated and penalized for any potential “excess toxicity.” The outcome predictions by the GARD model were within the confidence interval of the observed outcome in the clinical trial.57