Study design

The study was a prospective cohort study and was approved by the Human Research Ethics Committee of Fudan University Shanghai Cancer Center. All patients provided written informed consent. From February 7, 2016 to October 31, 2017, a total of 119 LARC (cT3-4/N0-2, M0) patients were enrolled, and they received nCRT (50Gy/25 fractions; concurrent capecitabine + irinotecan chemotherapy) and 1 cycle of interval chemotherapy (CAPIRI, capecitabine + irinotecan), followed by TME and 5 cycles of adjuvant chemotherapy (CAPOX, capecitabine + oxaliplatin). Detailed information about the regimen was in S1 Text. This regimen was motivated by a Phase III randomized controlled clinical trial (CinClare, NCT02605265) conducted in our cancer center, which proved that the addition of irinotecan in nCRT could achieve a significantly higher pCR rate of 30%, compared with a pCR rate of 15% in the capecitabine arm. The toxicities were acceptable and under control [18].

Plasma samples were collected at the following time points: before nCRT (Time1), at the 15th (Time2) and the 25th (Time3) fractions of nCRT, 0 to 1 days before surgery (Time4), and 5 to 12 days after surgery (Time5). Baseline plasma samples were collected for all 119 patients, while 16 patients failed to fulfill the subsequent 4 time points of sample collection, leaving 103 patients who completed the whole course of the study. Pathological tumor regression grade (pTRG) was evaluated according to the 8th American Joint Committee on Cancer (AJCC) Staging Manual [19]. pCR was defined as pathological T0N0M0 and pTRG = 0. The pTRG and pCR statuses were evaluated by 2 independent pathologists. If their conclusions were inconsistent, it was evaluated again by a third pathologist. Pathology reviewers were blinded to MRI results. MRI assessment of TRG (mrTRG) was evaluated according to the MERCURY experience [20]. Baseline and pre-surgery MRI images were compared with the reviews of 2 independent radiologists. It was sent to a third radiologist if there were inconsistencies. MRI reviewers were blinded to pathological results. Generally, definition of cCR needs data from MRI, endoscopy, and DRE. In this study, due to lack of endoscopy and DRE information, for the convenience of comparing pCR, cCR was defined as mrTRG1 representing the complete response in MRI. The criteria for judging pTRG and mrTRG were shown in Table 1. Detailed information on study design, patient enrollment, sample collection, study flow, and patient number with various pTRG and mrTRG were shown in Fig 1 and S1 Text.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Study design, sample collection, study objectives, and work scheme of the present study.

ctDNA, circulating tumor DNA; LARC, locally advanced rectal cancer; mrTRG, magnetic resonance tumor regression grade; nCRT, neoadjuvant chemoradiotherapy; pTRG, pathological tumor regression grade.

https://doi.org/10.1371/journal.pmed.1003741.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Criteria for classifying pTRG and mrTRG.

https://doi.org/10.1371/journal.pmed.1003741.t001

ctDNA sequencing and bioinformatics analysis

A total of 531 dynamic plasma samples and 119 leukocyte germline control samples were collected and subjected to panel sequencing of 422 cancer-related genes. The 422-gene panel includes genes associated with targeted medicines approved by Food and Drug Administration (FDA) or recommended by the NCCN guideline, genes involved in the major signaling pathways regulating cancer cell survival and proliferation, and potential cancer driver genes; it covers CRC-related genes, such as those associated with CRC development or prognosis (APC, TP53, KRAS, BRAF, PTEN, PIK3CA, etc.) and hereditary CRC (MLH1, MLH3, MSH2, MSH6, etc.) (S1 Table). The panel has been used in the study of lung cancer immunotherapy [21], genomic profiling of melanoma [22], and drug resistance of pancreatic cancer [23]. The average sequencing depth was approximately 4,000X. Baseline ctDNA sequencing and analysis were applied to all 119 patients. A total of 103 patients completed the whole study (Fig 1), 89 of whom had detectable ctDNA mutations at baseline. Detailed information about sample preparation, sequencing, data processing, and bioinformatics analysis was provided in S1 Text. We tracked the dynamic change of the mutation with the highest variant allele frequency (VAF) at baseline in each patient. ctDNA clearance meant unable to detect the mutation with highest VAF at certain time points, and T234_clearance meant unable to detect the mutation in all 3 preoperative time points (Time 2, 3, and 4 persistent clearance). To reduce potential false positives, genetic alterations that were detected in at least 2 time points after baseline were used for acquired mutations analysis. Detection of colorectal driver gene mutations meant detection of mutations in any 1 of 15 potential colorectal driver genes, which were introduced by Tie and colleagues [24] (S2 Table), at certain time points (e.g., T4 and T5).

Statistical analysis

Analyses were performed according to a prespecified analysis plan (S1 Analysis Plan). Baseline feature analyses including the distribution of patients with baseline genomic or clinical features in pCR/non-pCR or pTRG groups and the association of baseline features with either pCR status or patients’ prognosis were performed on all 119 patients. The analyses of acquired mutation and the association between the detection of driver gene mutations and prognosis were performed on 103 patients who had serial ctDNA test data (i.e., completed the whole study). Analyses involving ctDNA clearance, such as the association of ctDNA clearance with pCR status and pCR predictive model construction, were performed on 89 patients who had both detectable baseline mutations and serial ctDNA test data.

The comparison of the proportion of patients with certain clinical or genetic features in pCR and non-pCR groups was performed by Fisher exact test (two-sided test), and their increasing or decreasing trend in different TRG groups were analyzed using Cochran–Armitage test (one-sided test). For multiple tests, the adjusted P value was calculated by Benjamini–Hochberg method.

We used univariable logistic regression to investigate the association of baseline ctDNA features such as detection of gene or Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway mutations at baseline, as well as ctDNA dynamic change (ctDNA clearance and acquired mutation) with the probability of non-pCR. The effect of baseline pathological features, such as age, sex, and disease stage, on the probability of non-pCR was also analyzed by univariable logistic regression. We selected those features that were associated with pCR/non-pCR status to construct predictive models for pCR/non-pCR prediction. Generally, we selected features with P value ≤ 0.1 in the univariable logistic regression to perform multivariable logistic regression, and only features with P value ≤ 0.25 in the multivariable regression were kept. For KEGG pathways, we used more stringent criteria. Firstly, pathways with P value ≤ 0.05 in univariable logistic regression were selected; secondly, the selected pathways were included to fit a multivariable logistic regression; thirdly, pathways maintained significance (P value ≤ 0.05) in multivariable analysis were selected for further multivariable logistic regression including other features. We constructed 3 predictive models based on multivariable logistic regression, one contained only ctDNA information, one contained only mrTRG information, and the third one contained both ctDNA and mrTRG information. For each model, multivariable logistic regression was performed to calculate the odds ratios of each feature. In logistic regression, non-pCR was defined as the positive event, and the risk score for non-pCR of each patient was calculated based on the multivariable regression equation according to the method of Pavlou and colleagues [25] (Risk score = intercept + (b_{feature 1} × feature 1 status) + (b_{feature 2} × feature 2 status) + (b_{feature 3} × feature 3 status) +…+ (b_{feature N} × feature N status). Receiver operating characteristic (ROC) curve was plotted by “pROC” package of R [26]. Confidence interval of area under the curve (AUC) was calculated by DeLong method. Comparison of AUCs of different models was performed by DeLong method. For each model, 5-fold internal cross-validation repeating 100 times was performed to evaluate the predictive performance. Different random split of the data was used in each of the 100 repeats. At the same time, stratified cross-validation was used to ensure that the proportion of pCR cases was similar in each fold (around 25%). Data split was performed by using “splitTools” R package.

Survival analyses including Kaplan–Meier method and univariable or multivariable Cox proportional hazards model were performed using the “survival” and “survminer” R packages. All statistical analyses were performed in R-3.5.1 platform (https://www.r-project.org).

This study is reported as per the REMARK guideline (S1 REMARK Checklist).

Patient characteristics

The median age of the 119 patients was 57 years old, with 71.4% of them being males. Most patients were in stages IIIB (66.4%) and IIIC (31.1%). Postoperative pathological examination showed that 41 (34.5%) patients were pCR. As for pTRG, 41 (34.5%), 12 (10.1%), 53 (44.5%), and 13 (10.9%) patients achieved grade 0, grade 1, grade 2, and grade 3, respectively. Univariable logistic regression showed that age and mrTRG were significantly associated with higher probability of non-pCR (S3A Table). For major clinical features except for age, their distribution in the pCR and non-pCR groups was not significantly different (S3B Table).

Association of baseline ctDNA features and ctDNA dynamic change with the response to nCRT

Somatic mutations were detected in 100 (84.0%) patients at baseline, and detection of baseline ctDNA was not associated with treatment response or recurrence-free survival (RFS) (S1 Fig). The most commonly detected genes were TP53, APC, and KRAS (S2 Fig). Besides the above 3 genes, genes with relatively high mutation frequency included KMT2B, NOTCH1, and POLD1. Similarly, we used univariable logistic regression to investigate the association of baseline gene mutations with the probability of non-pCR, and only genes that were detected to be mutated in at least 6 patients were included. Due to the small sample size and multiple test adjustment, the adjusted P values of 10 included genes did not achieve statistical significance, with the detection of APC, TP53, and POLD1 tending to be associated with a high non-pCR probability (P ≤ 0.1; S4A Table). We next performed univariable logistic regression analysis at the pathway level. If 1 patient had detectable mutations of any gene of a specific pathway, then the patient was considered to be mutated in that pathway. For KEGG pathways, only pathways containing at least 5 overlapping genes with the detected mutated genes in at least 8 patients of our cohort were included in the analysis. Among 125 included pathways, only 6 pathways showed significant association with pCR status (P ≤ 0.05, all adjusted P values were greater than 0.25 due to small sample size). We then included these 6 pathways into the multivariable logistic regression analysis, and only homologous recombination (HRR) and histone methyltransferase (HMT) maintained to be statistically significant (S4B Table). Subsets of HRR and HMT pathway genes that were detected in our cohort were shown in S4C Table.

We also analyzed the association between pCR status and ctDNA dynamic change, including T234_clearance and acquired mutation status. As expected, patients with T234_clearance had a lower probability of non-pCR than those with ctDNA non-clearance (odds ratio = 0.11, 95% confidence interval [95% CI] = 0.01 to 0.6, P = 0.04), and patients with acquired mutations during nCRT had a higher probability of non-pCR than those without acquired mutations (odds ratio = 5.56, 95% CI = 1.03 to 103.31, P = 0.1), although not achieving statistical significance (S4D Table).

We next studied the distribution of patients with the above key features in various pTRG groups or pCR/non-pCR groups. We found that TP53 and APC mutations were prone to be detected in patients with non-pCR and high pTRG grade (P = 0.11 of TP53 and P = 0.10 of APC for pCR, P = 0.04 of both TP53 and APC for TRG), whereas both of HRR and HMT showed an opposite trend (Fig 2A and 2B), particularly for HRR pathway (P = 0.002 for pTRG), suggesting that mutations in these 2 pathways may sensitize cancer cells to radiochemotherapy. Consistent with a recent study [27], POLD1 mutations were prone to be detected in pCR group and low pTRG grade (P = 0.05 for pCR and P = 0.03 for pTRG; Fig 2A and 2B). Detection of KRAS mutations did not show significant different distribution among various pTRG group or pCR/non-pCR group (Fig 2A and 2B).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Baseline ctDNA features, ctDNA dynamic clearance, and acquisition were associated with pTRG and pCR.

(A) Distribution of mutation rates of TP53, APC, KRAS, POLD1, and 2 pathways (HRR and HMT) in pCR and non-pCR groups; Fisher exact test was used to compare the difference between pCR and non-pCR groups. (B) Distribution of mutation rates of the 4 genes and 2 pathways in different TRG groups. Y axis represents the proportion of patients carrying corresponding gene mutations accounting for total patients in the corresponding TRG group. Cochran–Armitage test was used to test the increasing trend (TP53, APC, and KRAS) and decreasing trend (POLD1, HRR, and HMT) in TRG groups. (C) Proportions of patients with T234_clearance or acquired mutations in pCR and non-pCR groups. Fisher exact test was used for intergroup comparison (two-sided). (D) Proportions of patients with T234_clearance or acquired mutations in various pTRG groups. Cochran–Armitage trend test was used to test increasing trend for acquired mutation status and decreasing trend for T234_clearance from pTRG0 to pTRG3 (one-sided). ctDNA, circulating tumor DNA; HMT, histone methyltransferase; HRR, homologous recombination; pCR, pathological complete response; pTRG, pathological tumor regression grade.

https://doi.org/10.1371/journal.pmed.1003741.g002

We then investigated T234_clearance and the acquired mutation detection rate in these LARC patients. T234_clearance rate was higher in pCR patients (P = 0.02; Fig 2C), and similar results were seen when grouping patients by their tumor regression grades, with T234_clearance rate showing a decreasing trend from complete regression (pTRG0) to poor regression (pTRG3) (95.7%, 77.8%, 71.1%, and 66.7% in pTRG 0, 1, 2, and 3 groups, respectively, P = 0.008; Figs 2D and S3A). On the other hand, the acquired mutation detection rate was higher in non-pCR patients (P = 0.11; Fig 2C) and showed an increasing trend from pTRG0 to pTRG3 (3.8%, 8.3%, 19.2%, and 23.1% in pTRG 0, 1, 2, and 3 groups, respectively, P = 0.02; Figs 2D and S3B). These results suggest that non-pCR patients not only had less clearance of baseline mutations but also acquired additional mutations during nCRT. Detailed information regarding the distribution of the patients with certain features in pTRG groups and pCR/non-pCR groups were provided in S4E Table.

ctDNA clearance supplemented MRI to identify non-pCR patients

By using the cCR (mrTRG1) standard in our study, the pCR/non-pCR prediction accuracy for all 119 patients was 70.6% (84/119), and for 89 patients who had ctDNA clearance information, the pCR/non-pCR prediction accuracy was 73.0% (65/89) (S3A Fig). Of note, 8 out of 89 patients who were determined to be cCR by MRI were confirmed to be non-pCR after surgery (indicated by 8 arrows at the bottom of S3A Fig). Among the 8 patients, 4 of whom were T234_non-clearance (indicated by 4 blue arrows in S3A Fig), including 1 patient experienced DM 138 days after surgery. This indicates that ctDNA clearance could provide additional pCR/non-pCR prediction values for 4 out of 8 cCR-misjudged patients, while the proper diagnosis of the rest 4 patients (indicated by 4 yellow arrows in S3A Fig) might be hindered by the low tumor load after nCRT. In addition, ctDNA non-clearance seemed to be enriched in non-pCR group. Only 1 of 23 patients with pCR was ctDNA non-clearance (S3A Fig). Patients with detectable acquired mutations were also enriched in non-pCR patients (S3B Fig).

Clearance of HRR and HMT mutations during nCRT

Among 89 patients who had detectable mutations at baseline and completed the whole sample collection and sequencing procedures, a total of 19 HRR mutations and 16 HMT mutations were detected at baseline. If clearance was defined as being cleared at all of the 3 time points before surgery (Time2, Time3, and Time4), only 1 HRR mutation [BRCA2:c.2525T>C(p.V842A)] and 1 HMT mutation [NSD1:c.2903A>G(p.K968R)] were not cleared during nCRT, representing a 5.3% non-clearance rate of HRR mutations and 6.2% non-clearance rate of HMT mutations (S4A and S4B Fig). The overall non-clearance rate was 11.9% (51 out of 429 mutations were not cleared), while HRR and HMT were among the top 5 KEGG pathways with the lowest non-clearance rate (S4C Fig), suggesting that tumor cells carrying HRR or HMT mutations may be more sensitive to nCRT, thereby, were largely cleared by nCRT. This result was in agreement with the previous finding that patients with HRR or HMT mutations were enriched in the pCR group and low pTRG group.

Predicting pCR/non-pCR by combining ctDNA and mrTRG information

Next, we investigated whether baseline genetic features and ctDNA clearance can be used to improve the prediction of patient’s pCR status. We firstly selected features with P value ≤ 0.1 in previous univariable logistic regression to perform multivariable logistic regression analysis. Because POLD1 is a member of HRR pathway, only HRR mutation was selected. APC mutation and TP53 mutation were largely overlapped, and the effect of TP53 mutation on chemoradiotherapy has been confirmed by many studies; therefore, TP53 mutation instead of APC mutation was selected. Features with P value ≤ 0.25 in multivariable regression were then selected to construct predictive models. Of note, age was not included in the models because its P value was greater than 0.25 in multivariable regression, although it was significant in univariable regression. Besides statistical significance, we also took account of the biological significance of these features. For example, HRR and HMT pathways have been extensively reported to be associated with radio- and chemo-sensitivity [28–30]. Lastly, in addition to the mrTRG information, 5 ctDNA features were selected, including TP53 mutation, HRR mutation, HMT mutation, T234_clearance, and the acquired mutation status. Based on multivariable logistic regression, we constructed 3 scoring pCR/non-pCR prediction models (Statistical analysis of the Materials and methods section), which were derived from ctDNA information only, mrTRG information only, or the combination of both aspects. A total of 89 patients, consisting of 23 (25.8%) pCR patients and 66 (74.2%) non-pCR patients, who had ctDNA clearance data, were included in the model construction and evaluation. Detailed information regarding model construction and calculation of the risk score were provided in S5 Table. As shown in Fig 3A, by measuring the risk score that quantifies the chance of a patient to be non-pCR, the model incorporating both ctDNA and mrTRG (i.e., combining model) had the most significant differences between pCR and non-pCR patients (P = 3.40e-08, Wilcoxon test), implying its superiority in pCR/non-pCR discrimination. Consistent with this observation, the combining model had higher pCR/non-pCR prediction performance (AUC: 0.886, 95% CI: 0.810 to 0.962) than the mrTRG only model (AUC: 0.729, 95% CI: 0.641 to 0.816) (P < 0.001) or the ctDNA only model (AUC: 0.818, 95% CI: 0.725 to 0.912) (P = 0.008) (Fig 3B). Five-fold cross-validation (repeating 100 times) showed that training AUC of the combining model was 0.89 ± 0.02 and testing AUC was 0.81 ± 0.12 (Fig 3C).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Predicting pCR/non-pCR by combining ctDNA and mrTRG information.

(A) Distribution of risk scores obtained from the 3 predictive models in pCR and non-pCR groups. Wilcoxon rank sum test was used for intergroup comparison (two-sided). (B) AUC analysis of the 3 models. 95% CI of AUC was calculated by “DeLong” method. (C) Predictive performance of the 3 models was evaluated by internal 5-fold cross-validation and 100 times repeats. The numbers on the top of the bars were average AUC ± standard deviation. Construction of the 3 models refers to Materials and methods section. AUC, area under the curve; ctDNA, circulating tumor DNA; mrTRG, magnetic resonance imaging tumor regression grade; pCR, pathological complete response; 95% CI, 95% confidence interval.

https://doi.org/10.1371/journal.pmed.1003741.g003

Recurrence risk assessment of LARC patients undergoing nCRT by ctDNA monitoring

We then studied the prognosis of these LARC patients. The median follow-up after surgery was 644 days (35 to 925 days), and 21 out of 119 (17.6%) patients progressed. As expected, pCR status was associated with RFS (P = 0.0025; Fig 4A), and only 1 out of 41 pCR patients relapsed during follow-up, compared with 20 relapses in 78 non-pCR patients. Of note, the detection of TP53 or KRAS mutations at baseline indicated a high recurrence risk (P < 0.001 and P = 0.02, respectively; S5A and S5B Fig). Besides, several pathological features, including perineural invasion (PNI), tumor deposits, vascular invasion, and lymph node metastasis, were also associated with a high risk of recurrence (P = 0.014, 0.0016, 0.003 and P < 0.001, respectively; S5C–S5F Fig). We thereby defined a “high-risk feature (HR_feature) positive” status to indicate at least 1 of the above 6 features being positive. As a result, the majority (18/20) of non-pCR patients with recurrence were HR_feature positive (Fig 4B), suggesting that HR_feature was a key factor to determine non-pCR patient recurrence.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Recurrence risk assessment of LARC patients undergoing nCRT by ctDNA monitoring.

(A) Kaplan–Meier analysis of RFS stratified by pCR status. All 119 patients with pCR/non-pCR information were included in the analysis. (B) Kaplan–Meier analysis of RFS stratified by pCR status plus HR_feature. HR_ features included baseline-detectable TP53 or KRAS mutation, tumor deposits, PNI, vascular invasion, and lymph node metastasis. HR_feature (+) was defined as positive in at least 1 feature, otherwise, HR_feature (−). All 119 patients with pCR/non-pCR information were included in the analysis. (C) Kaplan–Meier analysis of RFS stratified by T4_driver_gene mutation detection status (Time4, after nCRT). T4_detectable was defined as at least 1 mutation of the 15 CRC driver genes could be detected at Time4 point, otherwise, T4_undetectable. A total of 103 patients who completed the whole study were included in the analysis. (D) Kaplan–Meier analysis of RFS stratified by T5_driver_gene mutation detection status (Time5, after surgery) plus HR_feature. A total of 103 patients who completed the whole study were included in the analysis. CRC, colorectal cancer; ctDNA, circulating tumor DNA; HR, hazard ratio; HR_feature, high-risk feature; HR_feature (+), HR_feature positive; LARC, locally advanced rectal cancer; nCRT, neoadjuvant chemoradiotherapy; pCR, pathological complete response; PNI, perineural invasion; RFS, recurrence-free survival; 95% CI, 95% confidence interval.

https://doi.org/10.1371/journal.pmed.1003741.g004

Lastly, we investigated the value of ctDNA monitoring for risk assessment and patient stratification. We focused on 2 time points: after nCRT (Time4) and after surgery (Time5). For ctDNA clearance, the Time4_clearance did not show significant impacts on patients’ recurrence (P = 0.65; S6A Fig). However, as more pathological information was available after surgery, the concurrence of ctDNA non-clearance (Time5) and HR_feature positive was found to be associated with the worst RFS (S6B Fig). Detection of potential CRC driver gene mutations after nCRT was associated with a much worse RFS (hazard ratio [HR]: 9.29; 95% CI: 3.74 to 23.10, P < 0.001; Fig 4C). Similarly, in combination with HR_feature, driver ctDNA detection could make a more precise risk evaluation (Fig 4D), and all 6 patients with the concurrence of detectable driver mutations and positive HR_feature relapsed. Thus, patients could be stratified into 3 risk groups. The low-risk group included double-negative patients with a 2-year RFS of 95.5%. The intermediate-risk group included patients with positive HR_feature but undetectable driver mutations; their 2-year RFS was 76.4% and HR was 5.71 (95% CI: 1.28 to 25.54; P = 0.02, relative to the low-risk group; Fig 4D). The high-risk group included double-positive patients; their 2-year RFS was 0 and HR was 90.29 (95% CI: 17.01 to 479.26, P < 0.001, relative to the low-risk group; Fig 4D). There was only 1 patient who was driver mutation detectable but HR_feature negative; therefore, the risk of this group could not be evaluated. Multivariable Cox analysis showed that HR_feature and Time5 driver mutation detection were independent risk factors for postoperative recurrence (S6 Table). Overall, these data support the potential prognostic value of ctDNA when combining with other risk factors/features.