Improving on-treatment risk stratification of cancer patients with refined response classification and integration of circulating tumor DNA kinetics -…

Patient cohort, clinical characteristics, and recurrence patterns

We investigated 821 patients with advanced-stage EBV-associated NPC enrolled between 2009 and 2015, who consistently received cisplatin-based NAC followed by CRT. The diagram of the study population is shown in Fig. 1. The baseline clinical characteristics are presented in Table 1. Blood samples were collected at baseline and after the completion of NAC (post-NAC) and CRT (post-CRT). On-treatment imaging evaluation was conducted post-NAC using magnetic resonance imaging (MRI). The collection schema of cfEBV DNA and MRI is presented in Additional file 1: Fig. S1. The median follow-up was 64.9 months (interquartile range [IQR]: 58.172.5 months). We recorded 109 locoregional recurrences, 143 distant metastases, and 28 synchronous locoregional and metastatic recurrences. The 5-year rates of disease-free survival (DFS), overall survival (OS), distant metastasis-free survival (DMFS), and locoregional relapse-free survival (LRFS) were 72.5%, 82.9%, 83.1%, and 87.1%, respectively.

Flowchart showing the study design and patient selection process. The medical records of 10,126 patients with non-metastatic NPC were screened, and 821 patients with LA-NPC who received NAC plus concurrent CRT and had detectable pretreatment cfEBV DNA with on-treatment circulating cfEBV DNA surveillance were selected stepwise. Abbreviations: AC, adjuvant chemotherapy; CCRT, concurrent chemotherapy; cfEBV DNA, cell-free Epstein-Barr virus DNA; IC, induction chemotherapy; LA-NPC, locally advanced nasopharyngeal carcinoma; MRI, magnetic resonance imaging

All patients (n = 821) had detectable cfEBV DNA at baseline. The distribution of pretreatment cfEBV DNA titers (median, 12.50 103 copies/mL; IQR, 2.9652.50 103 copies/mL) are shown in Additional file 1: Fig. S2A, with 661 (80.5%) patients having pretreatment cfEBV DNA higher than 2000 copies/mL. Correlation analyses revealed that pretreatment cfEBV DNA was positively associated with node (N) stage (P < 0.05, Wilcoxon test; Fig. 2A), but not tumor (T) stage, age, sex, and smoking status (P > 0.05). Additionally, in line with our previous observations [17], higher baseline cfEBV DNA load (cut-off value, 2000 copies/mL) was preferentially associated with worse survival outcomes, especially with the occurrence of distant metastasis (hazard ratio [HR] = 2.88, 95% confidence interval [CI] = 1.595.20, P < 0.01; Fig. 2B). It remained significant after correcting for clinically important covariates using the inverse probability weighting (IPW) algorithm (HRDMFS = 2.51; 95% CI = 1.464.32; P < 0.01; Additional file 2: Table S1), suggesting that in addition to their well-acknowledged reflection on tumor burden, higher pretreatment cfEBV DNA levels may also be related to tumoral biological features (i.e., sensitivity to treatment and/or tumor microenvironmental heterogeneity as were referred in previously published researches [18,19,20]). Comparisons of the baseline covariates in the unadjusted and IPW-adjusted cohorts are shown in Table 2, demonstrating that the IPW succeeded in generating balanced distributions of covariates across subgroups.

Biological responses to NAC and their correlations with radiological responses. A Comparison of pretreatment cfEBV DNA levels across N categories. B Kaplan-Meier survival plot of DMFS in patients with pretreatment cfEBV DNA 2000 copies/mL versus <2000 copies/mL. C Scatter plot showing circulating cfEBV DNA levels before treatment initiation, at NAC completion (post-NAC), and at CRT completion (post-CRT). D Changes in cfEBV DNA from baseline in patients with increased cfEBV DNA levels post-NAC (n = 33). E Kaplan-Meier survival plot of DMFS in patients with cBR post-NAC versus decreased/increased cfEBV DNA in patients with non-cBR. F RECIST groupings (columns) and cfEBV DNA biological responses (rows) of 821 patients with matched treatment-nave and post-NAC surveillance data. G Kaplan-Meier survival plot of DFS in patients with cBR versus non-cBR post-NAC. H Kaplan-Meier survival plot of DFS in patients achieving cBR at the end of CRT stratified by biological responses to NAC. Abbreviations: cBR, complete biological response; cfEBV DNA, cell-free Epstein-Barr virus DNA; CI, confidence interval; CR, complete response; CRT, chemoradiotherapy; DFS, disease-free survival; DMFS, distant metastasis-free survival; HR, hazard ratio; IC, induction chemotherapy; N, node; NAC, neoadjuvant chemotherapy; non-cBR, non-complete biological response; PD, progression disease; PR, partial responses; PreEBV, pretreatment cfEBV DNA; SD, stable disease

Upon the initiation of NAC, 586 patients (71.4%) achieved complete biological response (cBR; defined as undetectable cfEBV DNA) during the NAC phase (Fig. 2C); the distributions of post-NAC cfEBV DNA titers (median, 0 copies/mL; IQR, 00.20 103 copies/mL) are shown in Additional file 1: Fig. S2B. Among 235 patients with non-complete biological response post-NAC (non-cBR; defined as detectable cfEBV DNA; median, 1.55 103 copies/mL; IQR, 0.417.84 103 copies/mL), 33 (14.0%) had increased cfEBV DNA levels from baseline, which demonstrated worse prognosis (Fig. 2D, E).

Regarding the RECIST-based radiological assessment, 56 patients (6.8%) achieved complete response (CR), 648 (78.9%) patients achieved partial response (PR) during the NAC phase, 116 patients (14.1%) had SD post-NAC, and one patient had progressive disease (PD) after receiving three cycles of docetaxel plus cisplatin (TP) NAC. Survival analysis demonstrated that patients with radiological PR had significantly worse survival compared to those with CR (HRDFS = 2.40, 95% CI = 1.135.11, P = 0.019, Additional file 1: Fig. S2C), and patients with SD/PD demonstrated the worst survival outcome. Based on this finding, we classified radiological responses into 3 subgroups: CR, PR, and SD/PD. Notably, patients with tumor stage I-II (T1-2) and tumor stage III-IV (T3-4) did not show significant differences in CR and PR rates (T1-2: 8 [7.7%] CR vs. 96 [92.3%] PR; T3-4: 48 [8.0%] CR vs. 552 [92.0%] PR; P = 0.91). The possible explanation for the comparable distribution of CR/PR in T1-2 versus T3-4 was that only locally advanced NPC (LA-NPC) (stage III-IV) patients were included in this study, thus patients with T1-2 would have more advanced N stages.

Next, we explored the relationships between biological and radiological responses and identified that they were positively correlated, with ~95% CR patients and ~75% PR patients having their cfEBV DNA dropped to zero after NAC, respectively (P < 0.01; Fig. 2F, and Additional file 1: Fig. S2D). Intriguingly, we observed an inconsistency between the biological and radiological responses in a subset of patients: of 56 and 648 patients with radiological CR and PR (radiological response) after NAC, 3 (5.3%) and 160 (24.7%) patients had detectable post-NAC cfEBV DNA (non-cBR), respectively (Fig. 2F). Moreover, across 117 patients with SD/PD (radiological responses), about 45 patients (38.5%) achieved cBR after 24 cycles of chemotherapy (Fig. 2F). These results prompted us to hypothesize that therapeutic responses evaluated by MRI and ctDNA may reflect distinct aspects of tumor biology and sensitivity to systemic treatment.

To further understand the clinical implications of cfEBV DNA-based biological responses. We first examined the correlations between post-NAC cfEBV DNA and post-CRT ctDNA. A total of 690 (84.0%) patients with matched post-NAC and post-CRT cfEBV DNA tests were included in the analysis. Among these, 51 patients had detectable post-CRT cfEBV DNA (median, 0.81 103 copies/mL; IQR, 0.334.79 103 copies/mL). The results demonstrated that detectable post-NAC DNA had 83.6% prediction sensitivity for detectable post-CRT ctDNA (95% CI = 78.088.1%). The probabilities of detectable post-CRT cfEBV DNA were 14 of 464 (3.0%) and 37 of 226 (16.4%), respectively, for patients with and without cBR after NAC (P < 0.01, 2 test; Additional file 1: Fig. S3A), suggesting that early cfEBV DNA kinetics was an informative indicator of whole-course treatment responses.

Next, we sought to determine the predictive value of post-NAC cfEBV DNA in long-term prognosis. Survival analysis revealed that cBR post-NAC was strongly predictive of long-term prognosis (HRDFS = 3.28; 95% CI = 2.554.23; P < 0.01; Fig. 2G and Additional file 1: Fig. S3B) and was independent of other clinically relevant prognostic factors in the IPW-adjusted survival analysis (Table 3). Interestingly, we identified that post-NAC cfEBV DNA was most prominently associated with distant metastasis after adjusting for clinically significant covariates (HRcfEBV DNA = 3.45 vs. HRMRI = 1.71, Pboth < 0.05; Table 3). In contrast, although post-NAC cfEBV DNA was also an independent predictor for locoregional recurrence, radiological response exhibited higher HRLRFS compared to post-NAC cfEBV DNA, suggesting that radiological response was a more preferential predictor for locoregional recurrence (HRcfEBV DNA = 1.89 vs. HRMRI(PR vs. CR) = 2.70 & HRMRI(SD/PD vs. CR) = 5.57; Table 3). This observation echoed with the above presumption that MRI and ctDNA reflected distinct aspects of tumor biology and sensitivity to systemic treatment.

Furthermore, we found that patients with non-cBR post-NAC that finally achieved cBR at the end of the CRT still sustained worse prognoses compared to those with cBR post-NAC (HRDFS = 2.70; 95% CI = 2.003.64; P < 0.01; Fig. 2H), suggesting that early biological responses were informative and that delayed ctDNA response conferred unfavorable outcomes. Moreover, among 242 patients with disease progression events, detectable cfEBV DNA post-NAC encompassed over half (122/242) of all failures, while detectable post-CRT ctDNA encompassed only 18% (39/211) of all failures (P < 0.05). Together, these data indicated that unfavorable biological cfEBV DNA responses at early treatment course identified an at-risk subgroup that encompassed large proportions of long-term failures.

Given the above observations, we asked whether early ctDNA kinetics provided additional clinical utility beyond imaging response assessments. To answer this question, we first stratified patients according to their radiological response and investigated whether patients with RECIST CR or PR had an unfavorable prognosis when they had detectable post-NAC ctDNA. Interestingly, we identified that post-NAC cfEBV DNA further stratified PR subgroup, with non-cBR patients having significantly worse DFS (HRDFS = 3.17, 95% CI = 2.364.25, P < 0.01; Fig. 3A). Unfortunately, the survival outcomes for CR subgroup (non-cBR vs. cBR) were not depicted due to the limited sample size in CR+non-cBR subgroup (n = 3). In addition, across patients with RECIST SD/PD, patients who achieved cBR post-NAC had more favorable DFS compared with those who did not (HRDFS = 2.32; 95% CI = 1.284.20; P < 0.01; Fig. 3A).

Biological responses provide additional prognostic information to RECIST. A Top panel: Kaplan-Meier survival plot of DFS in patients achieving RECIST PR stratified by biological responses to NAC. Bottom panel: Kaplan-Meier survival plot of DFS in patients with RECIST PD/SD stratified by biological responses to NAC. B Top panel: Kaplan-Meier survival plot of DFS in patients achieving cBR stratified by RECIST (CR vs. PR vs. SD/PD). Bottom panel: Kaplan-Meier survival plot of DFS in patients who did not achieve cBR stratified by RECIST (PR vs. SD/PD). C Kaplan-Meier survival plot of DFS, OS, DMFS, and LRFS across response phenotypes based on biological plus radiological responses to NAC. G1: cBR+CR, G2: non-cBR+CR, G3: cBR+PR, G4: non-cBR+PR; G5: cBR+SD/PD, and G6: non-cBR+SD/PD. Abbreviations: cBR, complete biological response; cfEBV DNA, cell-free Epstein-Barr virus DNA; CR, complete response; DFS, disease-free survival; DMFS, distant metastasis-free survival; HR, hazard ratio; LRFS, locoregional relapse-free survival; non-cBR, non-complete biological response; OS, overall survival; PD, progression disease; PR, partial responses; SD, stable disease

Next, we determined whether radiological responses can further stratify patients with or without cBR. We found that radiological response further stratified patients with cBR and that patients with SD/PD had significantly worse DFS compared to those with CR (HRDFS = 4.93; 95% CI = 2.2510.82; P = 0.02; Fig. 3B) and PR (HRDFS = 2.06; 95% CI = 1.522.78; P = 0.04), whereas the difference was not significant between CR versus PR (P > 0.05), possibly attributed to the limited events, given that cBR patients had superior prognosis compared to the overall cohort (Additional file 1: Fig. S2C). In addition, across the non-cBR subgroups, although patients with PR demonstrated better prognosis compared to those with SD/PD, the differences did not reach statistical significance for DFS (P > 0.05; Fig. 3B), suggesting that patients who did not successfully achieve biological response (non-cBR) would have equally inferior long-term tumor control regardless of radiological PR or SD/PD.

Based on the above observation, we further combined the radiological and biological response subgroups and yielded 6 response phenotypes: G1 (cBR+CR, n = 53, 6.5%), G2 (non-cBR+CR, n = 3, 0.4%), G3 (cBR+PR, n = 488, 59.4%), G4 (non-cBR+PR, n = 160, 19.5%); G5 (cBR+SD/PD, n = 45, 5.5%), and G6 (non-cBR+SD/PD, n = 72, 8.8%). Across diverse phenotypes, we next mainly focused our following analysis on phenotypes with contradictory biological and radiological response evaluations (G4 [non-cBR+PR] and G5 [cBR+SD/PD]). For G4, one important issue here was whether non-cBR was potentially confounded by false-positive cfEBV DNA tests. To address this point, we further compared their baseline characteristics with G3 (cBR+PR) and identified that patients with non-cBR+PR response phenotype tended to have higher clinical stages and baseline cfEBV DNA load (Additional file 2: Table S2). Interestingly, even adjusting for clinical covariates in multivariate analysis, patients with non-cBR+PR still had significantly worse prognosis in all endpoints compared to cBR+PR (Table 4), suggesting that detectable cfEBV DNA for patients with PR was clinically informative, rather than just confounded by false-positive tests. Analogously, to further address whether cBR was potentially confounded by false-negative cfEBV DNA tests for patients with SD/PD in G5, we further compared their baseline characteristics with G6 (non-cBR+SD/PD) and observed that they had lower pretreatment cfEBV DNA compared to G6 (Additional file 2: Table S3). Interestingly, after adjusting for clinically relevant covariates, patients with cBR+SD/PD (G5) still harbored significantly better prognosis in OS, DFS, and DMFS compared to non-cBR+SD/PD (G6) (Table 5). Notably, the differences in DMFS were most prominent (HRDMFS = 5.81, 95% CI = 2.0916.18, P < 0.01), whereas the difference in LRFS did not reach statistical significance (P > 0.05). These data indicated that undetectable cfEBV DNA for patients with SD/PD was clinically informative rather than just confounded by false-negative tests, especially in forecasting better distant control across patients with SD/PD, but not for local control. Collectively, we revealed that the contradictory biological and radiological responses bred additional valuable prognostic information.

Finally, we asked whether patients with biological cBR plus radiological SD/PD (G5) would have comparable survival with patients who achieved radiological PR plus biological non-cBR (G4). To our surprise, G5 had significantly more favorable long-term prognosis compared to G4, especially in the control of distant metastasis (Pall < 0.05; Fig. 3C).

Given the above findings that cfEBV DNA harbored critical biological information and that its on-treatment clearance kinetics identified preferentially at-risk populations beyond the traditional imaging evaluations, we presumed that inclusion of ctDNA testing would refine the risk estimates across patients with similar initial risks based on clinically relevant factors; moreover, as therapy is introduced, further risk stratification considering the on-treatment ctDNA measurement, radiological response, and therapeutic information would refine personalized dynamic risk estimates. To test this hypothesis, we established five risk prediction models incorporating clinically important factors with/without ctDNA and on-treatment parameters (Fig. 4A). The models were constructed based on Cox proportional hazard regression (CpH) model.

The combinations of biological and radiological responses refine risk groupings. A Bar plot showing the C-index and 95% CI for predicting the 5-year DFS by five models incorporating pretreatment risk factors with/without ctDNA and on-treatment parameters using the CpH method. B Nomogram for predicting the 3- and 5-year DFS, which integrated conventional pretreatment risk factors with pretreatment ctDNA, radiological and ctDNA-based response phenotypes, and therapeutic information. The total point values were independently calculated and then applied to the corresponding probability scale. C Calibration plots showing the actual risk probability by decile (y-axis) over the nomogram-predicted risk probability (x-axis). Abbreviations: cBR, complete biological response; cfEBV DNA, cell-free Epstein-Barr virus DNA; CR, complete response; DFS, disease-free survival; non-cBR, non-complete biological response; PD, progression disease; PR, partial responses; SD, stable disease

In the first model, three parameters (sex, age, clinical stage) established from prior literature or datasets were initially incorporated (Model-I: pretreatment clinical [Model-I_preCLI]). We determined the performance of the model for predicting 5-year DFS, a clinically relevant milestone and standard endpoint in cancer, and identified a bias-corrected Harrells concordance index (C-index) of 0.57. Importantly, the predictive accuracy of 5-year DFS significantly improved when pretreatment cfEBV DNA was incorporated (Model-II: pretreatment clinic-biological [Model-II_preCLIBIO]), with the C-index reaching 0.60. Next, we introduced treatment information and radiological/biological response parameters into the model (Model III-V). Model-III_postMRI, incorporating treatment information and radiological responses, had a significantly improved C-index of 0.65, and the C-index of Model IV (Model-IV_postctDNA), which incorporated treatment information and biological responses, was 0.68. Finally, given the above observation that on-treatment MRI and ctDNA reflected distinct aspects of tumor biology and sensitivity to systemic treatment, we established Model-V (Model-V_INTEGR), which integrated pretreatment factors with radiological and ctDNA-based response phenotypes, and therapeutic information. The C-index of Model-V reached 0.69.

Given that Model-V_INTEGR outperformed the models using pretreatment risk factors or on-treatment radiological assessments, we further developed a nomogram for quantifying the 3- and 5-year risks of disease progression in patients with diverse pretreatment and on-treatment features (Fig. 4B). The calibration plots indicated good agreement between the models predicted and observed survival estimates (Fig. 4C).

More here:
Improving on-treatment risk stratification of cancer patients with refined response classification and integration of circulating tumor DNA kinetics -...