Paired Box 5 Methylation Detection by Droplet Digital PCR for Ultra-Sensitive Deep Surgical Margins Analysis of Head and Neck Squamous Cell Carcinoma

Oncologic evaluation of surgical margins has long depended on visible histologic diagnosis. Although intraoperative cytology or frozen section diagnosis directs surgery to be performed in a more adequate way, some histologic negative surgical margins still have genetic or epigenetic alterations associated with disease recurrence. For instance, the presence of p53 gene mutations in surgical margins was reported by our group as a predictive marker of local recurrence in head and neck squamous cell carcinoma (HNSCC; refs. 1, 2). This phenomenon may be partly explained by “field cancerization,” which is characterized as the presence of clonally related cells with malignant potential containing one or more cancer-associated genetic or epigenetic alterations in the tumor-surrounding mucosal areas. These cells may not have undergone full malignant transformation necessary for recognition as residual cancer (3). This may underlie the high propensity of treated patients to develop locoregional recurrence or second primary cancers after surgery (4). Another issue linked to recurrence of HNSCCs is undetected fully developed residual tumor cells especially at the deep surgical margins. Deep surgical margin status has been reported to be a stronger predictor of HNSCC tumor recurrence than mucosal margins (5). Histologically undetected residual cancer cells may be left behind due to the “tumor budding” phenomenon, defined as a single cancer cell or a cluster of <5 cancer cells protruding into the stroma beyond the invasive front (6–8). These are hard to detect in intraoperative frozen samples by light microscopy, which depends on the identification of visible clusters of cells having the malignant phenotype. Tumors involving deep margins are particularly problematic. Deep margins may be buried when the wound is closed or reconstructed with transferred flap tissue, allowing tumor cells to grow undetected for a time until salvage therapy is less likely to succeed. Deep margins are more difficult to orient after resection to return to the precise point of residual tumor involvement for targeted further surgery or radiation. Finally, tissues deep to epithelial cancers typically consist of muscle, nerves, and fat, which offer less barriers to tumor extension than does the surrounding mucosal epithelium with its tight junctures and cohesive structure.

To collect DNA from cells at the irregular and complex deep surgical margins, we adapted a published margin imprinting procedure (9). Briefly, we pressed nitrocellulose sheets to each of several facets of surgical tumor specimens on the back table after removal from the resection bed. After 10 seconds of contact with the margin surface, the nitrocellulose was placed into a tube with 1% SDS-PK solution for DNA extraction. Unlike standard margin tissue collection, this procedure avoids altering the surgical margin, which is thus preserved for subsequent permanent histologic evaluation. Furthermore, imprints can harvest cells from the entire cut surface, which is impractical using gross tissue harvest techniques. Using multiple pieces of nitrocellulose, this procedure can widely cover and map any uneven surface of a surgical specimen. We hypothesized that margin imprints might assess the molecular margin status more accurately than collecting tissue samples from margins in the standard manner. We confirmed this hypothesis in a recent study (10) showing that the methylation marker combination of Homeobox A9 and endothelin receptor type B was a more powerful predictor of outcome than conventional tumor factors [tumor–node–metastasis (TNM) stage, differentiation]. Locoregional recurrence-free survival (LRFS) was more highly associated with molecular margin status than was RFS or overall survival. However, the case coverage rate using this marker combination was only 56.1% (41 of 73). Additional high-quality markers are needed with higher cancer-specificity to improve case coverage rate and thus optimize a methylation marker panel of HNSCC.

PAX5 has recently been reported by our group to be an excellent cancer-specific methylation marker with high sensitivity (80%) and specificity (94%) in HNSCCs (11). PAX5 encodes a B-cell–specific transcription factor, which is required for progression of B-cell development. It also plays a role in neural development and spermatogenesis. PAX5 gene methylation has been reported in various neoplasms, including HNSCC (11), gastric cancer (12), hepatoma (13), breast cancer, and lung cancer (14). In hepatocarcinogenesis, Liu and colleagues revealed that PAX5 directly binds to the p53 promoter and regulates p53 signaling. Alterations in p53 are nearly universal in HNSCC carcinogenesis (15). The JAK2/STAT5 pathway is also involved in PAX5 regulation (16). Activation of the pathway is reported to contribute to cellular invasion in HNSCCs (17). Palmisano and colleagues proposed that inactivation of PAX5β likely contributes to neoplastic development through loss of CD19 expression. Interestingly, they found PAX5 methylation in tumors and adjacent tissues of breast and lung cancer, but not in normal tissues and hypothesized, that this was evidence for field cancerization. These characteristics support the use of PAX5 methylation as a target for molecular HNSCC detection.

To further increase methylation detection sensitivity, we also evaluated droplet digital PCR (ddPCR) technology. In this approach, a DNA sample is fractionated into more than 10,000 droplets, and PCR amplification of the template molecules occurs in each individual droplet. Using this ultrasensitive technology, very low-frequency DNA alterations, undetected in bulk sample dilutions, can be effectively identified in droplets that contain only a few copies of the alteration (18). Histologically normal surgical margin imprint samples have low DNA quantity overall and very low levels of tumor DNA, if present at all. Although the conventional quantitative methylation-specific PCR (QMSP) assay has adequate sensitivity to detect these rare molecules, the quality of the results in later cycles is sometimes unstable and unreliable. Thus, rigorous dependence on visually distinguishable positive signals of ddPCR may reduce false positives, and analysis of more than 10,000 droplet formations per sample may reduce false negatives. We set out to test the utility of ddPCR to improve the accuracy of margin analysis.

Samples from the study cohort were prospectively collected at Johns Hopkins Hospital (Baltimore, MD) between May 2009 and December 2013. Surgical specimens were collected from 102 consecutive patients with head and neck cancer from whom written informed consent was obtained. The protocol was approved by the Johns Hopkins Hospital Institutional Review Board. As we focus on LRFS in this study, recurrent cases treated with salvage-curative intent were included. Twenty cases were excluded; 8 had histologically positive final surgical margins, 5 had unknown information about recurrence, 3 had samples with insufficient DNA concentration, 3 were found not to be squamous cell carcinoma, and 1 had only dysplasia in the tumor specimen. Thus, 82 cases were eligible for analysis (Table 1). The average time elapsed from surgery to assessment of outcome was 3.2 years. Locoregional recurrence included local recurrences and regional lymph node metastasis, excluding isolated distant metastasis. This occurred in 32 of 75 cases (43%) among non-RT imprint cases and 29 of 70 cases (41%) among non-RT tissue cases during each follow-up period.

At the time of surgical resection, frozen section samples were sent from the mucosal and deep-wound margins per routine practice. Only when these margins were read as negative was the case processed for molecular margin analysis. After removal of the tumor specimen with a rim of apparently normal margin, free liquid on the surface of the specimen was removed by blotting with cloth towels on a back table. Then, margin imprints were collected by pressing 3 × 3 cm² Hybond-C Extra nitrocellulose membranes (GE Healthcare) directly on the specimen for 10 seconds (19). The membranes were placed into coded 50-mL tubes with 5-mL 1% SDS (Sigma-Aldrich)-Proteinase K (Invitrogen) solution. Three repeat membrane samples were taken from each facet of the specimen. Thereafter, matched margin tissue samples (∼4 × 4 mm²) were sharply collected from a representative portion of the facet surface, placed in a coded tube, and stored in liquid nitrogen. To avoid possible float-on contamination of cells from the tumor, tumor imprints and tissues were collected last using the same method as for margin sampling. A fragment of normal muscle and/or normal mucosa was also collected from an area more than 5 cm distant from the tumor as a negative control. For frozen tissues, several 10-μm-thick sections were microdissected and digested with 1% SDS-Proteinase K solution. The presence of cancer in each tumor tissue sample was histologically confirmed in slides taken before and after sample harvesting.

DNA extraction from frozen tissue was accomplished using four rounds of proteinase K exposure during two overnight periods, followed by phenol/chloroform extraction and ethanol precipitation as previously described (20). DNA from margin imprints was immediately eluted, digested, and extracted by the same procedure. DNA extracted from each tissue (200 μg) or imprint sample (100 μg) was subjected to bisulfite treatment using the EpiTect Bisulfite Kit (Qiagen).

The bisulfite-modified DNA was used as a template for fluorescence-based real-time PCR as described (21). Primers and probes sequences are available on Supplementary Table S1. Real-time methylation-specific PCR reaction was performed in triplicate using the 7900HT Sequence Detector System (Applied Biosystems). Thermal cycling was initiated with a denaturation step at 95°C for 10 minutes, followed by 40 cycles of 95°C for 15 seconds and 60°C for 1 minute. Each plate included patient DNA samples, positive standards [Bisulfite Converted Universal Methylated Human DNA standard (Zymo Research)], and multiple water blanks as no-template controls. The average of triplicate samples was used for analysis. The relative level of methylated DNA (relative QMSP) for the target gene was determined as a ratio of its QMSP value to β-actin (ACTB) QMSP value × 100.

Bisulfite-treated DNA samples (2 μL) were added to ddPCR assay mixture (18 μL) including ddPCR Supermix with probes (no dUTP; Bio-Rad Laboratories). Primers and probes of target genes and reference gene (ACTB) appear in Supplementary Table S1. Eight sample mixtures (20 μL each) were loaded into each sample well of the droplet generator cartridge (Bio-Rad Laboratories). Also, 70 μL of droplet generation oil for probes (Bio-Rad Laboratories) was loaded into each of the 8 oil wells. The QX200 droplet generator (Bio-Rad Laboratories) created an average of 20,000 oil droplets per sample well and clouded droplet samples were transferred to a 96-well PCR plate (Eppendorf). The plate was heat-sealed with foil and placed in a thermal cycler under the following conditions: 95°C for 10 minutes, then 40 cycles of 94°C for 30 seconds and 60°C for 1 minute and two final steps at 98°C for 10 minutes and 4°C hold to enhance dye stabilization (22, 23). Finally, the plate was placed on the QX200 droplet reader (Bio-Rad Laboratories). The results were analyzed by QuantaSoft software (Bio-Rad Laboratories). All samples were examined in duplicate, and ACTB unexpressed samples were excluded. The average of duplicate samples was used for analysis. In effect, a sample was classified as positive if either duplicate showed a positive signal.

Surgical margin imprint and tissue data were analyzed separately. Patient clinicopathologic characteristics were compared using the Fisher exact test for categorical variables and the Student t test for continuous variables. For the target gene, methylation status was dichotomized at zero (positive vs. negative). Molecular margin analysis was considered to be positive if target gene methylation was detected in both tumor and at least one matched margin and negative if it was detected in tumor but not in any margin. If no target methylation was detected in tumor, the case was eliminated from further analysis. LRFS was defined as the time from surgery to the first documentation of locoregional disease recurrence. Distant recurrence without prior locoregional recurrence was considered as a competing-risk event and these cases were censored at the last documented date. The association of gene methylation in margins with LRFS in the imprint and tissue cohorts was evaluated using the Cox proportional hazards model with HRs and 95% confidence intervals (CI) estimated. Associations with P < 0.10 in univariate analysis were selected and further evaluated in a multivariate regression model adjusting for potential confounders (e.g., differentiation, TNM stage). All tests were two-sided and considered statistically significant and clinically promising at P < 0.05. Adjustment for multiple comparisons was not performed as the subsets for each cohort were not identical, and thus the analysis must be considered exploratory. Statistical analyses were carried out using JMP 9 software (SAS institute). The level of statistical significance was set at P < 0.05.

Advanced HNSCCs were dominant in this cohort. The 82 eligible HNSCC cases are characterized in Table 1. In total, 529 samples were collected from 82 cases. In detail, 179 margin imprints and 75 tumor imprints were collected comprising the imprint sample set (n = 75), whereas 170 margin tissues, 70 tumor tissues, 25 normal muscle, and 10 normal mucosa samples were collected as the tissue sample set (n = 70). There were 63 cases with both imprints and tissues available, 12 cases with only imprints and 7 cases with only tissues. There was no significant difference in clinicopathologic features between the imprint samples (n = 75) and tissue samples (n = 70). Postoperative radiation therapy was given to patients according to standard clinical practice patterns without regard to study participation. Among the 75 imprint margin cases, 28 patients received postoperative radiotherapy [radiotherapy (Rx), 15; chemoradiotherapy (CRx), 13; pTNM stage I, 2; II, 1; III, 4; IV, 20; unknown, 1] and 43 patients did not receive any additional therapy (pTNM stage I, 9; II, 7; III, 7; IV, 16; unknown, 4). Among the 70 tissue margin cases, 26 patients received postoperative radiotherapy [radiotherapy (Rx), 13; chemoradiotherapy (CRx), 13; pTNM stage I, 1; II, 1; III, 4; IV, 20] and 41 patients did not receive any additional therapy (pTNM stage I, 9; II, 7; III, 7; IV, 14; unknown, 4). Late-stage cases did not receive radiation for a variety of reasons, including distant past history of head and neck radiation, complete surgical removal with clear margins and no nodal metastases, patient refusal, and early postoperative recognition of distant disease.

Although the PAX5 gene has already been established as one of the tumor-specific methylation markers in HNSCC (11), we confirmed the marker sensitivity and specificity in our samples. Relative QMSP values of all samples are plotted in Supplementary Fig. S1. For tumor sensitivity and case coverage, PAX5 methylation was found in 68.0% (51 of 75) of tumors in the imprint samples and 71.4% (50 of 70) of tumors in the tissue samples. For specificity, we examined normal muscle in addition to normal mucosa as a negative control because deep surgical HNSCC margins typically consists of muscle tissue. The tumor specificity compared with normal mucosa was 100.0% (10 of 10) and normal muscle was 95.0% (24 of 25). We also checked 11 serum samples from the samples, and no PAX5 methylation of serum DNA was detected.

As an initial step, univariate and multivariate analyses were performed exploring the association of clinicopathological factors with LRFS. For PAX5 methylation in the tumor, QMSP values were dichotomized as positive or negative. Results for the imprint samples (n = 75) are shown in Supplementary Table S2 and tissue samples (n = 70) in Supplementary Table S3. In each subset of cases, smoking history, tumor differentiation, TNM stage, close margin (surgical margin of the main tumor specimen <1 mm at the final histopathologic diagnosis), and postoperative radiation therapy (RTx) were significant or marginal predictive factors in univariate analysis. The presence of PAX5 methylation in tumor was not a significant predictor of LRFS. By multivariate analysis, poor differentiation, advanced TNM stage (III or IV), close margin, and lack of postoperative RTx were significant poor prognostic factors for LRFS. In particular, lack of postoperative RTx had an extremely strong association with LRFS in both imprint samples (HR, 6.13; 95% CI, 2.53–17.19; P < 0.001) and tissue samples (HR, 4.44; 95% CI, 1.87–11.76; P < 0.001), which is consistent with previously reported analyses (24, 25). Therefore, we analyzed molecular margin data for the samples after dividing them into RTx and non-RTx groups.

Before proceeding with molecular margin analysis, the methylation detection technique by QMSP was piloted using 6 HNSCC surgical margin tissues, which had been previously collected and found to be cancer-positive by light microscopy (positive controls; Supplementary Table S4). PAX5 gene showed methylation in 5 of 6 samples. As histologically positive margins display microscopically visible tumor cells, the result of molecular analysis should confirm the presence of methylated DNA if the marker is positive in the main tumor. This requires highly robust (sensitive) methylation methods to demonstrate molecular margin involvement.

To investigate the relative efficacy of molecular analysis between the margin imprinting procedure and conventional margin tissue collection, QMSP-positive rates of tumor collected in the two sampling methods were compared. The PAX5 QMSP-positive rate of tumors in the imprint samples (68.0%, 51 of 75) was almost identical to that in the tissue samples (71.4%, 50 of 70; P = 0.719, Fisher exact test). Margin analysis among cases with PAX5 QMSP-positive tumors was similar for the imprint samples (25.5%, 13 of 51) and tissue samples (20.0%, 10 of 50; P = 0.636, Fisher exact test). Although the margin imprinting procedure only collects cells from the cut surface and produces very low gDNA concentrations, it did not appear to have a disadvantage compared with traditional sharp resective tissue sampling.

As postoperative RTx strongly affected the LRFS in both imprint and tumor sets (Supplementary Tables S2 and S3), we divided each into an RTx group and non-RTx group and examined the effect of molecular margin positivity on LRFS. The imprint samples (n = 75) had 47 non-RTx cases and 28 RTx cases (Table 2). Among these cases, 31 in the non-RTx group and 20 in the RTx group had PAX5 methylation in tumor and were thus eligible for comparative assessment. In the tissue samples (n = 70), 44 non-RTx cases and 26 RTx cases were available and 29 of the non-RTx and 21 RTx cases harbored PAX5 methylation in the tumor (Table 2). In the non-RTx group alone, the presence of PAX5 methylation in the surgical margins was significantly associated with poor LRFS in both 31 imprint samples (HR, 3.05; 95%, CI, 1.11–7.71; P = 0.031) and 29 tissue samples (HR, 3.27; 95% CI, 1.08–9.27; P = 0.037) by univariate analysis. Twenty-two cases were overlapping between these two groups. The number of locoregional recurrences during the follow-up period was 20 of 31 cases (65%) in non-RTx imprint cases and 15 of 29 cases (52%) in non-RTx tissue cases. One of the reasons for these high relapse rates after surgery is that the majority of cases were locally advanced cases. In the 31 imprint cases, stage I/II/III/IV was 6/4/4/13 (4 unknown), whereas in the 29 tissue cases, stage I/II/III/IV was 5/3/5/11 (5 unknown). Around 60% cases were stage III/IV in both groups. Another reason is that these subjects had no adjuvant radiotherapy after the surgery either because of patient choice or because of a history of radiation to the head and neck in the past. No association of PAX5 methylation with clinical outcome was found in RTx group (Table 2).

In an effort to increase the methylation detection sensitivity, we used ddPCR on the basis of the QMSP assay (ddQMSP). First, we optimized primers and probes for PAX5 and ACTB. The same sequencing of QMSP primers and probes were used with the dark quencher (BHQ; Supplementary Table S1). Results were evaluated using 100%, 50%, 25%, and 0% methylated bisulfite standards, which were made of the mixture of EpiTect PCR Control DNA Set (Qiagen; Fig. 1). The PAX5 methylation–positive cluster (blue dots) was clearly divided from the negative cluster (black dots), and concentrations of amplified copies were linearly associated with methylation status. For ACTB, the positive cluster (green dots) of the 100% unmethylated sample was relatively small. However, it still worked as a proof of the existence of bisulfate-converted DNA copies in the sample. As the target regions of PAX5 (chromosome 9p13) and ACTB (chromosome 7) are located on different chromosomes, a tiny cluster that indicates the existence of both positive droplets (orange dots) was observed in some samples.

Next, we applied the ddQMSP assay to our surgical margin samples. Results of representative normal mucosa and muscle samples, margin, and tumor samples are displayed in Fig. 2. Most normal samples showed no PAX5 methylation signal (blue dots) in channel 1 with dense ACTB signals (green dots) in channel 2 (Fig. 2A), whereas all of tumor samples in which PAX5 methylation was detected by conventional QMSP showed both PAX5 and dense ACTB signals (Fig. 2C). As for surgical margin tissue samples, some of them demonstrated positive blue dots with relatively weak ACTB signals due to the small amount of DNA (Fig. 2B). When blue dots were clearly apart from negative clusters as in the positive control, we considered samples (D07 and F07) as PAX5 methylation positive and samples in G07 as negative. We then checked the association of PAX5 methylation and LRFS shown in Fig. 3. To focus on the difference of methylation detection sensitivity, we compared conventional QMSP with ddQMSP in the same samples. In the non-RTx cases from the imprint samples (Fig. 3A and B, n = 31), 8 of 9 QMSP-positive margins were also positive in ddQMSP assay. An additional 14 cases were revealed to be positive by ddQMSP that were scored as negative by traditional QMSP. Thus, a total of 22 margin samples were positive, and they showed significantly poorer LRFS (P = 0.026, log-rank test; Fig. 3B). In the non-RTx cases from the tissue samples (Fig. 3C and D, n = 29), all 7 QMSP-positive margins were also positive in ddQMSP and an additional 5 cases were revealed to be positive. Again, a total of 12 margin samples were positive, and they showed significantly poorer LRFS (P = 0.044, log-rank test; Fig. 3D). Thus, the ddQMSP assay not only confirmed the QMSP results but also reduced the number of false-negative cases especially in the samples obtained through the imprinting technique.

QMSP and ddQMSP results of molecular margin imprint analysis were compared with conventional clinical factors (TNM stage and tumor differentiation) as LRFS predictors (Table 3). Among non-RTx cases with PAX5-positive tumors, conventional QMSP result failed to demonstrate PAX5 methylated imprint margin as an independent prognostic factor in multivariable analysis (Table 3, Conventional QMSP). However, ddQMSP dramatically improved the results by increasing the identification of margin-positive cases, and PAX5 methylated margin became one of the significant independent prognostic factors of LRFS (Table 3, ddQMSP; HR, 3.89; 95% CI, 1.19–17.52; P = 0.023). Also in margin tissue analysis, ddQMSP improved the HR of PAX5 methylation factor to (Table 3; HR, 5.52; 95% CI, 1.56–25.75; P = 0.008).

Surgical margin status is a major criterion for selection of postoperative chemotherapy and radiation therapy. This fact has increased the intraoperative scrutiny of margins as surgeons attempt to render the patient free of disease and avoid the need for more intensive regimens. Frozen section analysis largely depends on the pathologist's visual impression of histopathologic slides taken from selected sampling sites. Although diagnosis of frozen samples gives the most reliable information available, tumor-specific gene alterations may indicate the presence of tumor without visible patterns of morphologic change (as with field cancerization or isolated invasive tumor cells). These rogue cancer or precancer cells may be the reason that histologically negative surgical margins do not guarantee local RFS (26). Intraoperative, real-time molecular evaluation of surgical margins can be considered a “next-generation” strategy. QMSP is a well-established technique for detecting low levels of tumor DNA in various body fluids such as saliva (27, 28) and urine (29). Practically, the adaptation of this approach to HNSCC surgical margin analysis using imprint methods also has been shown to have acceptable quality in our recent study (10). To realize general clinical adoption of imprint molecular assessment, several refinements are needed. We have already optimized the time needed for processing from DNA extraction to QMSP assay analysis so that it may be accomplished within 3 hours by one person (30). While this assay still exceeds the time needed for frozen section analysis, it may be completed during subsequent neck dissection(s) and/or the reconstructive phase of the operation, allowing surgeons to resect additional margins if the QMSP margin is positive. With the change to the ddQMSP assay, the total time required was again 3 hours or less. While ddQMSP needs additional droplet-generating time, the plate preparation time and sample consumption can be reduced because of multiplex procedure using two different probes in the same sample well. This time frame can surely be shortened in fully automated laboratories.

The result of PAX5 methylation assessment of surgical margin clearly identified high-risk patients for locoregional recurrence. However, there are still false-negative cases even with ddQMSP (Fig. 3). These missed cases might be identified by other molecular markers. Hence, our ongoing strategy begins with preoperative examination of the tumor biopsy sample. Once the most influential methylation marker(s) (simply detected by the largest relative QMSP value) are selected from a HNSCC-specific panel of genes before the surgery, we can focus on those genes for intraoperative surgical margin analysis. Future personalized oncology workflows may use methylation arrays or methylation sequencing approaches to identify, preoperatively, the complete methylome of each patient and design a custom panel of methylated primers and probes that can be used for intraoperative surgical margin analysis.

Unlike the TNM staging or cancer cell differentiation, only intraoperative margin testing can direct effective intraoperative treatment, including additional resection, intraoperative radiotherapy (IORT; ref. 31), or other treatment (32). Even with IORT, Scala and colleagues reported that patients with histologically negative surgical margins had superior disease-free survival compared with those with positive margins (33). Effective intraoperative resection seems to have the best possibility of achieving local control, as even rare residual cancerous cells would be removed. Intraoperative margin testing with ddQMSP may play an important role in improving effective intraoperative treatment.

Bernier and colleagues compared postoperative chemoradiotherapy with radiotherapy in an EORTC and RTOG samples (34). The result showed that chemoradiotherapy was superior to radiotherapy only in cases that had histologically involved margins or nodal extracapsular spread. The results suggest that advanced treatment provides incrementally superior outcomes only in cases with residual cancer cells. We compared outcomes in RTx cases with non-RTx cases among the PAX5 imprint–positive margin group (Fig. 4A and C) and -negative margin group (Fig. 4B and D). ddQMSP results had 21 more molecular margin-positive cases than conventional QMSP results due to its high sensitivity and demonstrated a more distinctive statistical difference between RTx cases and non-RTx cases (Fig. 4C). Actually, 16 of 22 non-RTx cases recurred within 1 year, whereas only 3 radiation cases recurred in the first 2 years. On the other hand, there was almost no statistical difference between RTx cases and non-RTx cases in the molecular margin-negative group (Fig. 4D). These findings indirectly substantiate the presence of cancerous cells in the molecular margin-positive samples.

One of the difficulties of molecular margin analysis is the lack of established cutoff points for margin QMSP. QMSP cutoff points are usually established by analyzing a number of tumor samples after microdissection of more than 70% to 80% cancer cells and paired normal tissues. However, margin samples are expected to contain only very rare cancer cells. For this reason, we used the dichotomized paradigm for scoring QMSP and ddQMSP results. As a result, PAX5 methylation had outstanding tumor specificity (>95%) and favorable tumor sensitivity (∼70%) in our primary QMSP data (Supplementary Fig. S1). This result depends on the lack of false-positive signal in normal samples. Another problem is potential contamination at the sample collection step. If detached cancer cells from tumor were deposited on the surrounding surface of the specimen and collected by margin tissue or imprint procedure, a false-positive result would be scored. However, in fact, at least some molecular margin samples were free of tumor signal, indicating that there was not systematic or widespread contamination. Also, when methylation signal was present, it was limited to one or two sampling regions. The observed association with clinical outcome further substantiates the biologic legitimacy of cancer-related methylation signal in these margin samples.

In conclusion, the detection of PAX5 methylation is a powerful tool for identifying patients at high risk for HNSCC locoregional recurrence, after surgical treatment to clear histologic margins with no postoperative radiation therapy. Moreover, ddQMSP analysis resulted in increased sensitivity to permit detection of tumor that would recur even after postoperative radiation, an effect that remained a significant prognostic factor in multivariable analysis. Although further confirmation using larger sample sets is necessary, molecular surgical margin analysis with ddQMSP appears to be a powerful and useful intraoperative diagnostic tool.

No potential conflicts of interest were disclosed.

Conception and design: M. Hayashi, R. Guerrero-Preston, D. Sidransky, W.M. Koch

Development of methodology: M. Hayashi, R. Guerrero-Preston, W.M. Koch

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): M. Hayashi, W.M. Koch

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): M. Hayashi, D. Sidransky, W.M. Koch

Writing, review, and/or revision of the manuscript: M. Hayashi, R. Guerrero-Preston, D. Sidransky, W.M. Koch

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): M. Hayashi

Study supervision: W.M. Koch

W.M. Koch received grants from the NIH and the National Institute of Dental and Craniofacial Research (R01 DE013152-11).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.