Disease Monitoring Using Post-induction Circulating Tumor DNA Analysis Following First-Line Therapy in Patients with Metastatic Colorectal Cancer

Colorectal cancer represents 8.1% of all new U.S. cancer cases (1). More than 140,250 new cases and 50,630 deaths are estimated for 2018, representing the second leading cause of death from cancer in the United States (1). Mutations in a number of genes, such as N/KRAS and BRAF, have been associated with poor prognosis and resistance to therapies commonly used for colorectal cancer; in addition, many patients with tumors lacking known predictive biomarkers are also refractory to treatment (2). Although it would be beneficial to identify patients who are treatment refractory or relapse rapidly after treatment, the genomic and pathologic heterogeneity of colorectal tumors complicates the identification of additional biomarkers predictive of response (2–4). CT scans are typically performed a few months after treatment (5, 6), which may be too late to detect disease progression in patients who could have benefited from interim alternate therapies. CT scans also may not have sufficient resolution to accurately assess disease burden in patients with smaller tumors (7).

Circulating tumor DNA (ctDNA) in plasma is an alternative method to rapidly and accurately estimate disease burden, and can overcome biopsy site bias in the assessment of tumor clonality. ctDNA refers to fragmented DNA of tumor origin that is shed into the bloodstream, whereas cell-free DNA (cfDNA) encompasses all DNA freely circulating in the bloodstream and is not necessarily of tumor origin (8). Prior analyses have indicated that ctDNA levels in plasma can be used as a tool for disease monitoring, can assess treatment response, and can detect residual/recurrent disease and treatment failure more quickly and accurately than CT scans (9–13). Detection of ctDNA after colon cancer resection can identify patients at the highest risk of recurrence, and can also provide direct evidence of residual disease, as demonstrated in an analysis of 230 patients with stage II colon cancer (14). Furthermore, data from a study of 53 patients with metastatic colorectal cancer (mCRC) showed that early changes in ctDNA during first-line chemotherapy are predictive of later radiologic response (15). The prognostic value of cfDNA and ctDNA has been demonstrated in studies of other cancer types as well. For example, quantification of somatic variants in cfDNA dynamically correlated with treatment response in patients with small-cell lung cancer (SCLC; ref. 12), and ctDNA levels were an early prognostic marker for outcomes in patients with diffuse large B-cell lymphoma (DLBCL; ref. 13).

The Sequencing Triplet With Avastin and Maintenance (STEAM; NCT01765582) trial was a randomized, multicenter, open-label, U.S. based, phase II trial investigating the efficacy of bevacizumab (BEV) plus 5-fluorouracil/leucovorin/oxaliplatin (FOLFOX) and 5-fluorouracil/leucovorin/irinotecan (FOLFIRI), administered concurrently (cFOLFOXIRI-BEV) or sequentially (sFOLFOXIRI-BEV; i.e., FOLFOX-BEV alternating with FOLFIRI-BEV), versus FOLFOX-BEV in the first-line treatment of patients with mCRC (16). cFOLFOXIRI-BEV and sFOLFOXIRI-BEV were associated with numerically improved overall response rate (ORR), progression-free survival (PFS), and liver resection rates versus FOLFOX-BEV. Biosamples and data from the STEAM trial can be harnessed to identify biomarkers associated with a response, regardless of treatment arm. Here, we assessed the feasibility of using plasma ctDNA level as a prognostic marker for PFS following first-line induction therapy in patients with mCRC using data from STEAM.

STEAM was a randomized, open-label, multicenter, phase II study of patients with previously untreated mCRC (Supplementary Fig. S1; ref. 16). After a 21-day screening period, patients were randomized 1:1:1 to a 4-month treatment induction phase with cFOLFOXIRI-BEV, sFOLFOXIRI-BEV, or FOLFOX-BEV, administered in 2-week cycles. Induction could be extended up to an additional 2 months at the discretion of the treating physician. Induction therapy was followed by maintenance with 5-fluorouracil, leucovorin, and bevacizumab every 2 weeks or capecitabine plus bevacizumab every 3 weeks. Following disease progression, patients were to receive second-line bevacizumab and a fluoropyrimidine-based chemotherapy of the investigator's choice. Dosages and treatment schedules for each phase are detailed in Hurwitz and colleagues (16). Treatment continued until progression, death, withdrawal of consent, or unacceptable toxicity. Treatment efficacy was assessed by site investigators after baseline tumor measurements per RECIST version 1.1 during the pre-randomization screening period at 8-week intervals.

The co-primary endpoints were investigator-assessed ORR for first-line treatment with cFOLFOXIRI-BEV versus FOLFOX-BEV, and PFS for first-line treatment with pooled FOLFOXIRI-BEV (i.e., concurrent and sequential) versus FOLFOX-BEV. The evaluation of biomarkers associated with treatment outcomes was an exploratory endpoint. Key inclusion criteria included age 18 to 75 years, histologically confirmed mCRC and ≥1 measurable lesion considered unresectable at baseline, Eastern Cooperative Oncology Group (ECOG) performance status of 0 or 1 for patients <71 years old or ECOG performance status of 0 for patients 71 to 75 years old, and a willingness to comply with the study protocol, including tissue and blood sampling. Key exclusion criteria included prior systemic treatment for mCRC, except for use of palliative radiosensitizers and adjuvant chemotherapy completed <12 months prior to study enrollment. Additional inclusion and exclusion criteria are described in the published article (16).

STEAM conformed to the principles outlined in the Declaration of Helsinki, the International Conference on Harmonisation E6 guideline for Good Clinical Practice, and applicable local laws. The study protocols were approved by the Institutional Review Boards at each participating study site, and patients provided written informed consent.

Biomarker-evaluable patients were defined as those who provided optional consent to participate in the biomarker program and who had 1 tissue sample, 1 pre-induction plasma sample, and 1 post-induction plasma sample. Tissue and pre-induction plasma samples had to be analyzable and collected within 90 days of each other. The 90-day cutoff was chosen based on examination of the distribution of relative days of collection between tissue and pre-induction plasma (Supplementary Fig. S2). In addition, post-induction plasma samples had to be collected within 60 days of last drug induction date. Archival or fresh baseline (cycle 1, day 1, pretreatment) tumor tissue for each patient was collected. Blood plasma for biomarker analysis was collected at pre-induction and at the end of the induction phase (end of cycle 8, with possible extension to cycle 12, per the discretion of the investigator).

cfDNA was extracted from 4 mL of plasma using the cobas cfDNA Sample Preparation Kit (Roche), and tumor DNA was isolated from formalin-fixed paraffin-embedded tumor tissue sections using the cobas DNA Sample Preparation Kit (Roche). DNA yields were quantified by a Qubit HS assay (ThermoFisher). cfDNA samples were prepared for sequencing using the AVENIO ctDNA Expanded Kit (For Research Use only; Not for use in diagnostic procedures; Roche) using 1 to 208 ng (median 50 ng) of cfDNA. Tumor DNA samples were prepared for sequencing using a prototype of the AVENIO Tumor Tissue Expanded Kit (For Research Use only; Not for use in diagnostic procedures; Roche); 12 to 350 ng (median 110 ng) of tumor DNA was used as input. Sequencing was performed on an Illumina HiSeq 4000.

The AVENIO Expanded Kits (For Research Use only; Not for use in diagnostic procedures) are based on the previously described CAPP-Seq technology (17, 18), which utilizes molecular barcoding and a hybrid-capture methodology to interrogate selected regions from 77 genes with known prognostic or predictive value as therapeutic biomarkers, or as emerging markers in clinical trials. ctDNA variants were identified using the AVENIO ctDNA Analysis Software (For Research Use only; Not for use in diagnostic procedures; Roche) which incorporates integrated digital error suppression to remove stereotypical sequencing errors (18). Tumor tissue variants were identified using a precursor of the AVENIO Tumor Tissue Analysis Software (Roche). Somatic variants were called using a machine learning model (19) that leverages variation in allele fraction (AF) from the same variant found in different samples of the same patient and information on common germline variants in public databases. Additional filters were applied to further refine the list of tissue somatic variants, as follows: (1) tissue variants were restricted to nonrepetitive regions of targeted exons, (2) tissue variants in public germline databases (present in >0.001 population frequency in subpopulations in ExAC release 0.3.1 or 1000 Genomes phase3v5b, or annotated as a common SNP in dbSNP build 144) were removed, (3) tissue variants in the AVENIO Expanded Kit whitelist below 3% AF were removed, and (4) tissue variants not in the AVENIO Expanded Kit whitelist below 5% AF were removed.

AF for a given somatic mutation was calculated as the number of deduplicated reads with that mutation divided by the total number of deduplicated reads covering that genomic position. Mutant molecules per milliliter (MMPM) for a given somatic mutation was calculated from the AF of that variant by multiplying by the extracted mass (ng), dividing by the plasma volume (mL), and adjusting by a factor of 330 haploid human genome equivalents per 1 ng. The mean allele fraction (mAF) or mean MMPM (mMMPM) for each plasma sample was calculated across somatic variants (single-nucleotide variants) predefined by the matched tissue sample. Multiple cutoffs for classifiers to identify good versus poor molecular responders were determined and tested based on distribution of post-induction mAF and mMMPM values. ctDNA levels were evaluated via mAF and mMMPM after ≥8 cycles of chemotherapy (a median of 15 days after the last dose in the induction phase).

Distribution of baseline characteristics were compared between the biomarker-evaluable population (BEP; n = 54) and the remaining STEAM-enrolled population (n = 226) by Wilcoxon rank-sum test, Pearson's χ² test, and Fisher's exact test as appropriate. Continuous characteristics were summarized by median, minimum, and maximum values, and categorical characteristics were summarized by counts and percentages. A Spearman's correlation test was used to determine the correlation between mAF and mMMPM. PFS was defined as time from start of treatment to the date of disease progression or death. Subjects were censored based on the last evaluable date of progression-free assessment during first-line treatment. Tumor assessments that were non-evaluable were considered progression-free assessments only if the subject had curative surgical resection prior to second-line induction. Kaplan–Meier method and log-rank tests were used to estimate the median PFS and log-rank P-value. Cox proportional hazards models were used to evaluate the association between PFS and mAF and mMMPM classifiers. Unadjusted Cox proportional hazards models were used to assess association of PFS and potential adjustment factors (age, gender, cancer type at diagnosis, ECOG at baseline, extent of metastatic disease, location of primary tumor, prior cancer surgery, and surgical resection during study). Significant variables with P-values < 0.05 were included as multivariable adjustments.

Among the 280 patients enrolled in STEAM between January 23, 2013 and December 26, 2014, 183 had tumor tissue that was sequenced and evaluable, 118 had matched pre-induction plasma, and 54 patients (i.e., the BEP) had ctDNA-evaluable sequencing data for both pre- and post-induction plasma (Fig. 1). Median follow-up duration was 19.3 (range, 6.7–28.0) months in the 54 patients included in the BEP. Median PFS after first-line induction therapy (calculated from date of randomization) in the BEP (n = 54) was similar to median PFS after first-line induction therapy observed in the remaining STEAM population [n = 226; 11.2 vs. 10.9 months, respectively; HR, 0.92; 95% confidence interval (CI), 0.65–1.31; log-rank P = 0.6361].

No patient characteristics were significantly different when compared between BEP (n = 54) and the remaining population (n = 226), suggesting that the BEP is representative of the intent-to-treat population (N = 280; Supplementary Fig. S3, Table 1). In addition, no potential adjustment factors were significantly associated with PFS (Table 2). A summary of pre-induction mAF, post-induction mAF, and post-induction mMMPM values for patients in the BEP is provided in Supplementary Table S1, and individual patient data are shown in Supplementary Table S2. In the BEP (n = 54), there was a decrease in mean mAF and median mAF values from pre-induction to post-induction (Supplementary Table S1). Two patients had pre-induction and post-induction mAF values of 0, and 25 patients (including these 2 patients) in the BEP had post-induction mAF values of 0 (Supplementary Table S2). ctDNA variations between pre- and post-induction for individual patients in the BEP are shown in Supplementary Fig. S4.

The frequencies of somatic variants in tumor tissue and pre-induction plasma in patients with at least 1 tissue sample sequenced and analyzable (n = 183) and in patients in the BEP are shown in Supplementary Fig. S5. The top 3 most common somatic variants for patients with ≥1 tissue sample sequenced and analyzable and for those patients in the BEP, respectively, were TP53 (tissue, 60.7% and 66.7%; plasma, 39.3% and 66.7%), APC (tissue, 59.0% and 53.7%; plasma, 37.7% and 55.6%), and KRAS (tissue, 47.0% and 50.0%; plasma, 24.0% and 42.6%). Genes with variant frequency ≥10% generally aligned between tumor tissues and pre-induction plasma. The most common somatic variants observed in patients in sequencing-evaluable (matched tissue and pre-induction plasma, n = 118), non–mAF-evaluable (n = 226), and mAF-evaluable (>median mAF, n = 27; ≤ median mAF, n = 27) patients are shown in Supplementary Fig. S6.

Patients with less than or equal to median versus greater than median pre-induction mAF levels had similar median PFS (10.7 vs. 11.2 months, HR, 0.78; 95% CI, 0.41–1.47; log-rank P = 0.4453), suggesting that pre-induction levels of ctDNA did not appear to be associated with PFS following induction therapy (Fig. 2).

Lower-than-median post-induction mAF levels were associated with longer median PFS relative to higher-than-median mAF (17.7 vs. 7.5 months, HR, 0.33; 95% CI, 0.17–0.63; log-rank P = 0.0005; Fig. 3A). To evaluate the robustness of post-induction mAF as a prognostic biomarker, we assessed PFS using different cutoffs for mAF in plasma (0%, 0.1%, 0.2%, 0.3%, 0.4%, 0.5%, and 1.0%; Supplementary Fig. S7). Post-induction mAF levels showed a statistically significant inverse correlation with median PFS, regardless of the cutoff used. The most significant difference was observed in patients with <0.3% versus ≥0.3% post-induction mAF (16.9 vs. 7.4 months, HR, 0.25; 95% CI, 0.12–0.53; log-rank P = 0.0001; Fig. 3B). There were no statistically significant differences in PFS between treatments groups in patients with post-induction mAF <0.3% or ≥0.3% (Supplementary Fig. S8).

As with differing cutoffs for mAF percentage, ctDNA levels stratified by a 10-fold or 100-fold reduction in mAF between pre- and post-induction plasma were inversely correlated with median PFS (>10-fold vs. ≤10-fold reduction: 13.4 vs. 6.9 months, HR = 0.24; 95% CI, 0.10–0.60; log-rank P = 0.0008; >100-fold vs. ≤100-fold reduction: 16.9 vs. 7.4 months, HR = 0.24; 95% CI, 0.11–0.51; log-rank P = 0.0001; Supplementary Fig. S9).

As an alternative method of estimating ctDNA levels in plasma, we evaluated the association of absolute variant molecules per unit plasma, calculated as MMPM, with PFS. The distribution of post-induction mMMPM is shown in Supplementary Fig. S10. Similar to post-induction mAF, post-induction mMMPM levels were inversely associated with PFS at multiple cutoffs, suggesting that mMMPM levels are also prognostic for clinical outcomes. In patients with ≤ median versus > median mMMPM, median PFS was 17.7 versus 8.1 months, respectively, and HR was 0.39 (95% CI, 0.20–0.75; log-rank P = 0.0034; Supplementary Fig. S11). Following testing of multiple cutoff values, we found that an mMMPM value of 8 MM/mL showed clinically significant separation of good versus poor molecular responders. In patients with <8 MM/mL versus ≥8 MM/mL, median PFS was 16.9 versus 7.4 months, respectively, and HR was 0.29 (95% CI, 0.15–0.57; log-rank P = 0.0001; Fig. 4). There were no statistically significant differences in PFS between treatment groups in patients with post-induction mMMPM <8 MM/mL or ≥8 MM/mL (Supplementary Fig. S12).

Although mAF and mMMPM generally correlated with each other (ρ = 0.987, P < 0.0001), 3 patients in this analysis with very high (>100 ng/mL) plasma cfDNA levels had low post-induction mAF levels (0.21%, 0.36%, and 0.03%) but high mMMPM (211.62, 793.43, and 10.93). In these outlier cases, mMMPM was a better prognostic biomarker than mAF, as mAF may have underestimated disease burden (Supplementary Fig. S13).

In this extended biomarker analysis of the STEAM trial, higher-than-median post-induction mAF correlated with poor PFS, and post-induction mAF levels stratified by different cutoff points consistently showed inverse correlation with PFS. Hence, post-induction ctDNA level appears to be quantitative for disease burden and may be a robust prognostic biomarker in patients with mCRC following induction therapy. Pre-induction mAF levels did not correlate with post-induction PFS. Decreases of >10× or >100× in ctDNA levels between pre- and post-induction plasma were associated with better clinical outcomes, supporting the notion that mAF levels reflect prognostic status.

Levels of MMPM, which can be used as an alternative method of quantifying ctDNA in plasma, generally correlated with mAF levels in this study. Similar to mAF, mMMPM levels consistently showed an inverse correlation with PFS despite the use of different cutoff points, suggesting that it can also act as a robust prognostic biomarker. In patients with extremely high cfDNA levels per unit of plasma, mAF may underestimate disease burden, as observed in a small sample size (n = 3) of these outlier cases in our study. In these cases, mMMPM, as an absolute (rather than relative) indicator of variants per unit plasma, appears to be a better indicator of disease burden than mAF, but this needs further validation.

There is a continued need for additional measures of somatic variant profile that can distinguish patients and reveal underlying biology. In this analysis, the most common somatic variants observed in tumor tissue and pre-induction plasma aligned with previously reported variants for colorectal cancer, including a high prevalence of mutations in TP53, APC, and KRAS from the Cancer Genome Atlas Network project (TCGA; ref. 20). In our study, TP53 was the most commonly mutated gene, followed by APC and KRAS, whereas APC was the most commonly mutated gene in the TCGA analysis (20). Inactivating TP53 mutations have been associated with worse prognosis in advanced colorectal cancer and are more frequent in advanced stage tumors (21). The higher prevalence of TP53 mutations in our analysis might be due to a greater number of patients with late-stage disease and/or the relatively small patient number in our study.

Previous studies in patients with mCRC and other cancers have demonstrated that detectable ctDNA levels in plasma correlate with worse clinical outcomes compared with patients without detectable ctDNA (12, 13, 22, 23). In a systematic review evaluating the prognostic value of ctDNA in patients with colorectal cancer, 8 of 13 studies showed that detectable ctDNA level in the plasma was an independent predictor of worse overall survival (24). Similar to our study, a prospective study of patients with mCRC by Garlan and colleagues found that higher ctDNA concentration [assessed in plasma collected before the first chemotherapy cycle (C0)] was associated with shorter PFS (25). Although STEAM included patients who received only first-line chemotherapy, the Garlan and colleagues study included patients who received both first-line (83%) and second-line (17%) chemotherapy. Our study assessed mAF and mMMPM of ctDNA with tumor burden, whereas Garlan and colleagues measured ctDNA levels directly from plasma, expressed as ng/mL. In addition, Garlan and colleagues investigated baseline ctDNA concentration and change in ctDNA concentration from C0 to cycle 1 or 2, whereas our study evaluated ctDNA levels after ≥8 cycles of chemotherapy (a median of 15 days after the last dose in the induction phase). Furthermore, Garlan and colleagues stratified patients according to slope change in ctDNA (≥80% vs. <80%) and ctDNA concentration (<0.1 ng/mL, 0.1 ng/mL to 10 ng/mL, or >10 ng/mL) while we used cutoffs of 8 mMMPM/LMPM and 10- or 100-fold reduction. Our study also investigated post-induction plasma only, whereas the Garlan and colleagues' study did not. Both our study and the study by Garlan and colleagues present informative methods to study correlation of ctDNA level and outcomes in different treatment settings.

In addition to these colorectal cancer data, previous studies in patients with SCLC and DLBCL suggest that ctDNA levels, estimated by quantifying somatic variants in cfDNA, dynamically correlate with disease burden and can therefore be an early indicator of response to treatment (12, 13). In particular, the analysis of patients with DLBCL relied on targeted sequencing using CAPP-Seq, the same readily available technology as used in the current analysis, indicating the accuracy and versatility of quantifying ctDNA in plasma using this digital sequencing technology (13). Interestingly, the Kurtz and colleagues' analysis of patients with DLBCL indicated that pre-treatment ctDNA levels were independently prognostic of outcomes (13), whereas our analysis in a different indication (colorectal cancer) showed no correlation between pre-induction ctDNA and post-induction PFS. The cause of the difference between the prognostic values of pre-induction ctDNA may be a reflection of the etiology of the cancer types, the size of the cohorts studied, or other factors.

This study had several limitations. The sample number for this study was relatively small, as only 54 (19%) of the 280 patients enrolled in STEAM were included in this biomarker analysis. This high attrition rate indicates the importance of mandating sample collection for biomarker analysis, and to take into account the association between ctDNA analysis and patient engagement throughout the treatment process. As this study did not include training and testing cohorts, validation of these findings in an independent cohort would be required.

In conclusion, our data suggest that the quantification of ctDNA in post-induction plasma can potentially serve as a prognostic biomarker for outcomes following treatment in patients with mCRC. This study describes a clinical research tool that can help evaluate the extent of induction therapy response via ctDNA levels, and the correlation between ctDNA levels and survival. This and other studies help demonstrate the use of ctDNA for disease monitoring in the assessment of patients failing treatment across multiple cancer types and may prompt closer management of patients with appropriate clinical validation.

H.I. Hurwitz is an employee/paid consultant for, reports receiving commercial research grants from, and is an advisory board member/unpaid consultant for Genentech/Roche. C. Ju is an employee/paid consultant for Roche. A. Lovejoy is an employee/paid consultant for Roche Sequencing Solutions. C. Mancao is an employee/paid consultant for, holds ownership interest (including patents) in, and reports receiving other remuneration from Hoffmann–La Roche. A. Nicholas is an employee/paid consultant for Genentech. R. Price is an employee/paid consultant for Roche. N. Sommer is an employee/paid consultant for Genentech and holds ownership interest (including patents) in Roche. N. Tikoo is an employee/paid consultant for Roche Molecular Systems, Inc. S.J. Yaung is an employee/paid consultant for Roche Sequencing Solutions, Inc. J.F. Palma is an employee/paid consultant for and reports receiving other remuneration from Roche. No potential conflicts of interest were disclosed by the other authors.

Conception and design: J.J. Lee, C. Mancao, R. Price, J.F. Palma

Development of methodology: A. Lovejoy, J.F. Palma

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): J.C. Bendell, H.I. Hurwitz, N. Sommer

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): X. Max Ma, H.I. Hurwitz, C. Ju, J.J. Lee, A. Lovejoy, C. Mancao, A. Nicholas, R. Price, N. Tikoo, L. Yao, S.J. Yaung, J.F. Palma

Writing, review, and/or revision of the manuscript: X. Max Ma, J.C. Bendell, H.I. Hurwitz, C. Ju, J.J. Lee, A. Lovejoy, C. Mancao, A. Nicholas, R. Price, N. Sommer, L. Yao, S.J. Yaung, J.F. Palma

Study supervision: J.J. Lee, J.F. Palma

The authors would like to acknowledge Aarthi Balasubramanyam and Sandeep Gattam for assistance with statistical analysis.

This study was funded by F. Hoffmann–La Roche, Ltd./Genentech, Inc. Support for third-party writing assistance was provided by CodonMedical, an Ashfield Company, part of UDG Healthcare plc.

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.