Background
Cancer patients demonstrate significant intertumoral heterogeneity [
1,
2]. Such biological variations contribute to dramatically different outcomes. Over the past decades, increasing efforts have been made in unraveling this heterogeneity through histological, radiographic, and molecular dissection [
3‐
5]. However, published evidence has largely focused on pretreatment parameters and generally does not consider the dynamic response information, partly due to the difficulty in obtaining serial tumor samples. Nevertheless, there is substantial evidence that tumor dynamic response to systemic therapy harbors critical clinical implications, which demonstrates significant heterogeneity even across patients sharing identical pretreatment characteristics [
6‐
8]. For example, a subset of high-risk patients can achieve exceptional responses and be cured with frontline therapy, while other patients with identical pretreatment risk factors do not respond well to frontline therapy and develop disease recurrence soon after treatment completion.
Clinically, tumor responses are routinely evaluated by conventional imaging examinations based on the Response Evaluation Criteria in Solid Tumors (RECIST) or pathological/biopsy materials. However, these conventional approaches do not fully capture the tumor biological features or the dynamics of clinical benefits over time. For example, over 30% of patients with lung cancer who had stable disease (SD) at the first imaging scanning can ultimately achieve durable clinical benefits [
9,
10]. On the other hand, repeated tumor biopsies are not clinically feasible and may not capture timely information during the treatment course. On this note, there is a considerable unmet need for early response assessment methods that can reflect tumor biology in a more accurate way; meanwhile, identify patients with long-term tumor control in a timely manner.
Liquid biopsy of circulating tumor DNA (ctDNA) in the peripheral blood provides noninvasive access to cancer-specific genomes and biology [
11]. Accumulating evidence suggests that ctDNA quantification and on-treatment changes provide informative information on therapeutic responses, tumor biology, and risk stratification [
12,
13]. Despite the promising data, their clinical utility to date has remained limited in various cancers, given the small study cohorts and heterogeneous treatment modalities, especially from the aspects of response assessment and risk-adapted treatment guidance. Moreover, the associations between ctDNA and RECIST-based response assessment remain poorly understood.
Head and neck cancer is a heterogeneous epithelial tumor with subsets of tumors demonstrating strong associations with virus infection [
14]. Nasopharyngeal carcinoma (NPC) is one of the most aggressive head and neck cancer originated from nasopharynx and is typically associated with Epstein-Barr virus (EBV) infection in patients from endemic areas [
15]. Plasma circulating cell-free EBV DNA (cfEBV DNA) is a sensitive and specific biomarker for EBV-associated NPC, which consists of short DNA fragments released by NPC cells and can be detected using ultrasensitive polymerase chain reaction (PCR)-based assays [
16]. In this vein, NPC represents a suitable model here for addressing the value of on-treatment ctDNA kinetics.
Here, we hypothesized that biological response evaluated by ctDNA kinetics added critical information to RECIST, and integrating on-treatment biological and radiological response information refined patient risk stratification for personalized clinical decision-making. In this investigation, we tested this hypothesis in a large-scale patient cohort with nasopharynx of head and neck cancer consistently treated with sequential neoadjuvant chemotherapy (NAC) and chemoradiotherapy (CRT), who had pretreatment and on-treatment cfEBV DNA and imaging surveillance. We demonstrated that biological response to NAC assessed by on-treatment cfEBV DNA demonstrated 25.3% discordance with RECIST-based response evaluation; moreover, they harbored important prognostic information not only complementary to but also beyond the conventional radiological responses. In addition, on-treatment cfEBV DNA kinetics conferred early determination of treatment benefits, and delayed ctDNA responses indicated unfavorable outcomes. Finally, we established risk prediction models and demonstrated that introducing on-treatment ctDNA significantly refined risk stratification of cancer patients. Our findings help advance the implementation of ctDNA-based testing in therapeutic response evaluation for a refined risk stratification. Consequently, the dynamic and refined on-treatment risk profiling would inform future risk-based therapeutic adaptation for personalized medicine in cancer patients.
Discussion
Tumor responses following systemic treatment demonstrate significant intertumoral heterogeneity over time, which informs the dynamic risk probabilities for individual patients. The conventional approach for capturing phenotypic responses has been RECIST-based imaging examinations, which are inherently limited by the difficulty of on-schedule collection, expensive cost, and stratification inaccuracy. Our analyses of the performance of on-treatment cfEBV DNA to systemic treatment in a large cohort of patients with NPC sheds light on this issue by demonstrating the feasibility and enhanced prognostication of ctDNA-based biological response evaluation and risk stratification, which provides critical information not only complementary to but also beyond RECIST. Our findings add to the body of prior knowledge that patient risk probabilities are dynamic and can be refined with the introduction of biological response information, which will facilitate future risk-adapted clinical trial designs for personalized therapeutic approaches.
Previous risk stratification of patients with cancer has primarily focused on pretreatment factors, and therapeutic decisions are made before treatment initiation and remain unchanged throughout the whole treatment course [
21,
22]. Nevertheless, in the present research, we demonstrated and proposed that early on-treatment tumor responses, not only radiological response but also ctDNA-based biological responses, harbored critical prognostic information and that incorporating these factors greatly enhanced patient risk stratification. The refined and dynamic on-treatment risk stratification strategy would consequently open the door to risk-adapted treatment intensification/de-intensification for personalized medicine. Taking the NPC model as an example, stage III–IV
A NPC are classified as locally advanced NPC with high risks, and standard treatment for LA-NPC entails NAC plus CRT [
23,
24]. Although a large body of phase III clinical trials supports its clinical efficacy [
25‐
27], in clinical settings, it cannot be neglected that a subset of patients who responded well to NAC had decreased recurrence risk and would be overtreated during the subsequent intensive CRT. In contrast, another proportion of patients who harbored chemotherapy-resistant tumor clones did not respond well to NAC, and experienced treatment failure soon after the completion of CRT. These suggest that the current pretreatment-determined, one-size-fits-all treatment strategy is suboptimal. Based on our on-treatment cfEBV DNA-based risk model, we would carefully propose a risk-adapted clinical approach. Specifically speaking, when patients with predicted 5-year overall survival rates over 90% based on the on-treatment risk model, it would be intuitive to de-intensify treatment to avoid unnecessary adverse events and improve quality of life, given that the accumulating evidences support the premise that LA-NPC with relatively low risk may be less beneficial to intensive concurrent chemotherapy [
28], and two cycles of concurrent cisplatin share equal survival outcomes but significantly reduced adverse events with three cycles [
29]. Moreover, for at-risk patients with predicted 5-year overall survival rates less than 90% after NAC, it is promising to intensify treatment in appropriate ways (i.e., additional adjuvant chemotherapy). Especially, for those with predicted 5-year overall survival rates lower than that of pretreatment, which may suggest insensitivity to prior chemotherapy, it is of urgent need to alter chemotherapeutic regimens during the following therapeutic phases and/or integrate novel treatment modalities (i.e., immune checkpoint inhibitors and/or target therapy) in order to circumvent chemoresistant clones and improve survival [
30,
31]. Notably, the clinical feasibility of the above interventions ought to be carefully tested in prospective clinical trials.
There have been increasing efforts to combine on-treatment response-based parameters with pretreatment factors for refined risk stratification. In routine clinical practice, physicians often utilize serial anatomical imaging or biopsy information for response assessments [
32,
33]. Unfortunately, these parameters do not easily lend themselves to longitudinal examinations due to the difficulty of on-schedule collection, expensive cost, and invasiveness. Moreover, we determined that RECIST-based imaging evaluation did not fully capture patients with favorable prognosis and clinical benefit, which is in line with recently published research [
9,
10]. To overcome this barrier, we demonstrated the feasibility and enhanced prognostication of minimally invasive ctDNA-based biological response evaluation in a large cohort of patients with cancer being treated consistently. Furthermore, we found 25.3% discordance between the ctDNA and imaging-based response evaluations and that early ctDNA kinetics provided critical information not only complementary to but also beyond RECIST for identifying patients with long-term tumor control, especially those with SD/PD but cBR, who would still derive meaningful clinical benefit from treatment. This finding is clinically informative, as current strategies for identifying a responsive population with clinical and radiological methods at early timepoints are suboptimal. Notably, in the context of immunotherapy, patients who met the criteria for PD based on RECIST were noted to have late but deep and durable responses [
34]. Further studies are warranted to examine whether serial ctDNA surveillance may facilitate the timely recognition of patients who would benefit from ICIs or other treatment modalities.
In recent years, ctDNA has revolutionized the management of patients with cancer due to its minimally invasive nature and access to cancer-specific information. Nevertheless, previous studies have largely focused on the implications of pretreatment ctDNA. Although low baseline ctDNA levels are widely reported to be associated with favorable clinical outcomes, we and others have found that the effect size is generally modest compared to the on-treatment counterpart [
12,
35]. Moreover, we identified that post-NAC cfEBV DNA levels, but not baseline cfEBV DNA levels strongly correlated with RECIST, suggesting that other mechanisms underpinning clonal sensitivity to treatment are likely implicated in the on-treatment, but not baseline, ctDNA. Therefore, it is not surprising that the model incorporating on-treatment cfEBV DNA changes and pretreatment cfEBV DNA factors demonstrated the highest prognostication effect compared to the model incorporating pretreatment cfEBV DNA alone. On this note, we propose that on-treatment cfEBV DNA kinetics provide critical information not only complementary to pretreatment cfEBV DNA, but also beyond it, which ought to be included in future prognostic/predictive models. Here, it is also worth noting that although we adopted cfEBV DNA in NPC as a model in this research to unveil the application of on-treatment ctDNA, we cannot exclude the possibility that there are nasopharyngeal carcinoma cells that do not have EBV integrations and that except for EBV DNA, other types of cancer-associated DNA (i.e., point mutations, copy number aberrations, alterations in DNA methylation) can also be released into the plasma of NPC patients. Especially, for NPC patients that are not associated with EBV infections, detections of genetic and epigenetic markers other than cfEBV DNA for risk monitoring are challenging and merit future investigations.
Limitations of this study included lack of sequencing data in the cfEBV DNA assay design. However, this is beyond the scope of our study. According to previously published research, such information is important when distinguishing healthy controls from patients with NPC, given that the EBV DNA fragments of patients have longer fragment lengths and higher methylation levels compared to that of people without NPC [
36,
37]. Second, the radiological response evaluation in NPC using RECIST criteria is rather tricky compared to other solid tumors. The skull base was not included as target lesions here, given their inferiority in reproducible repeated measurements, and the well-documented challenges of MRI in evaluating the shrinkage of skull base lesions [
38]. Third, the number of patients with non-cBR+CR was limited, which made it difficult to perform survival analysis. Hence, further investigation on this subgroup is warranted to determine whether it is attributed to the false-positive cfEBV DNA test or is truly clinically informative in this subgroup. Fourth, the prognostic model in the present study was not externally validated. Therefore, the applicability of this model should be interpreted with caution; external validation cohorts and prospective studies are warranted to further confirm the clinical applicability of our model. Except for the above limitations, we revealed the values of on-treatment cfEBV DNA kinetics and their clinical utility for response evaluation and dynamic risk stratification in a cohort with homogeneous treatment modality and mature follow-up data. An additional strength is that we adopted the IPW algorithm to examine and compare the prognostic values of cfEBV DNA and other variables, which takes the advantage of efficiently simulating a randomized trial in time-to-event analyses without sacrificing sample size and statistical power, which enhanced the validity of our findings.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.