The Prognostic Value of Circulating Cell-Free DNA in Colorectal Cancer: A Meta-Analysis

Search Strategy and study selection

Potentially relevant studies were searched via Pubmed, Embase and Science citation prior to 3-July 2015 without applying start date limit. The term of colon cancer, rectal cancer, cell-free DNA, plasma DNA, serum DNA and prognosis were searched as topics and free words at the same time (see detailed search strategy Table S1-Table S3 in supplementary material). No language restrictions were imposed initially.

To yield relevant publications, we further evaluated the titles, abstract and author information of the collected studies. We contacted the authors for further studies if needed. For the analysis and full text review, only English language articles were considered. Non-research publications such as proceeding papers, editorials, comments, book chapter and other type uninvolved publications were excluded. However, cross references of relevant reviews were assessed for extra potential studies. Researchers would not appraised for detailed assessment unless the studies met the following inclusion criteria: (i) study should investigate the prognostic significance of cfDNA in colorectal cancer patients with at least one outcome (i.e., OS and RFS), (ii) the make-up of cfDNA was serum DNAs and plasma DNAs from peripheral blood (PB) rather than genomic and tumor DNAs, and (iii) studies from the same authors and institutions using the same cases were considered as duplicate publications, and only included the latest study to avoid re-assessment on the same cases.

For the meta-analysis, we endorsed the studies to be included in the final analysis through the following exclusion criteria: (i) duplicate publications, (ii) diagnostic studies, (iii) serum protein markers and genomic studies, (iv) studies with less than 20 patients, and (v) no survival or insufficient data to be statistically extracted.

Data extraction

Two reviewers (Basnet Shiva and Zhen-Yu Zhang) independently assessed the final set of articles and resolved disagreements by consensus. Data retrieved from the eligible study were: surname of the first author, number of patients, country of origin, year of publication, male/female ratio, mean/median age of patients, follow up durations, TNM staging (UICC), methods of detecting cfDNA, tumor locations, markers defining cfDNA, markers origin, sampling time, positive rate, endpoints and survival data including RFS, OS and other endpoints that could be estimated as RSF and OS.

Data processing and subgroup analysis

We used standard methods recommended for meta-analysis to statistically assess the prognostic effects of cfDNA on survival of CRC. HRs and accompanying 95%CIs were extracted individually if available. Apart from that, they were estimated using suggested methods by Parmar ⁴⁰ and Tierney et al ⁴¹ based on survival data or survival curves. Besides, if HRs were mentioned by both univariate and multivariate analyses, multivariate analysis was preferable because of its consideration in possible confounding effects ⁴²

Normally, HR >1 indicated a worse outcome. We pooled the extracted HRs with generic inverse variance method in Comprehensive Meta-analysis program (version2.2, Englewood, NJ, Biostat). Individual measures and overall effects were illustrated by forest plots ⁴³. We used Cochrane's Q statistics and I² Statistics to test the homogeneity of the studies ⁴⁴ and applied Random effect model only when the tests were significant (two-tailed P value≤0.1, I²>50%) ^{45, 46}.

For the studies with multiple markers, each of the markers was treated as an independent unit of analysis. If studies contained post-treatment time points, they were also meta-analyzed but also followed by sensitivity test after excluding them from the analysis. To classify the potential confounding factors related to outcomes of patients with CRC, we carried out subgroup analysis. The subgroup was stratified by tumor stages, tumor markers, methods, marker origin and size of the patients. We performed subgroup analysis only when there were two or more studies included.

Quality assessment and sensitivity analysis

Since no standardized tools could be applicable to observational studies, we used the Newcastle-Ottawa scale (NOS) ⁴⁷ to assess the quality of the studies because it was recommended by the Cochrane library. Sensitivity analysis was performed to test reliability of the results. The combined HRs after random removal of one included study were calculated and explored to observe the influences of a particular study on overall results. Cumulative meta-analysis was performed to test the impact of publication year on stability of final results.

Publication bias

We statistically tested for the potential presence of publication bias using funnel plots along with Begg's and Mazumdar rank correlation method ⁴⁸. All of the statistical techniques in the meta-analysis mentioned above had followed PRISMA checklist (Table S4).

Global characteristics

From the extensive search made on 3^rd July 2015, a total no. of 1282 articles was retrieved. Among the outcomes, 1023 studies were recognized as non-English publication, duplicates and studies out of scope and another 223 articles were reported as non-research articles, all of which were excluded from detailed assessment. Notably, three out of four studies from Spindler et al ^49-51 and one out of two studies from Philipp et al ⁵² were excluded because of significant overlapping (the same author, same institutions, time period and part data were repeated), and only most recent studies (latest submission) were given priority. After the independent full review of the rest 36 studies, nine publications ^53-61 dealing with quantitative and qualitative analysis of cfDNA with the prognostic purpose of CRC were considered as eligible studies (Figure 1). The 9 studies included a total of 1022 colorectal patients published between the year 2003 and 2014, eight from Europe ^{53-59, 61} and one from China ⁶⁰. The mean patient no. per study was 114. The samples were collected before treatment (n=8) ^{53, 55-61} and after treatment (n=1) ⁵⁴ One study reported cancer-related survival ⁵⁷ which could be regarded as a subset of OS. Molecular techniques like PCR followed by sequencing ^{53, 54}, Spectrophotometry ⁴¹, quantitative PCR (q-PCR) ^{42, 46} mutant-enriched PCR (ME-PCR) ⁵⁷ and real-time PCR (rt-PCR) ^58,59,61 were used to detect cfDNA from PB samples. Quality assessment was performed by NOS scale (Table S5) and all of the studies scored as high quality with 5 points or more. In total, all main characteristics of the eligible studies (19 units of analysis) were summarized (Table 1 and Table S6).

Heterogeneity and overall effects

There was no significant heterogeneity observed in the statistical tests among study group of RFS (I²=32.23%, P-value=0.1820) and OS (I²=29.94%, P-value=0.1530). Thus we used fixed effects model. The pooled results demonstrated that the detection of cfDNA had significant value in predicting RFS (HR [95%CI]=2.78[2.08-3.72], n=7) and OS (HR [95%CI]=3.03[2.51-3.66], n=12) in patients with CRC (Figure 2).

Sensitivity analysis

To test whether the results were stable, we therefore performed sensitivity analysis. First of all, when the fixed effects model was switched to random effects model to re-compute the pooled HRs, the results showed no significant changes in overall effects for RFS (HR [95%CI]=2.99 [2.00-4.47]) and OS (HR [95%CI]=3.02[2.38-3.83]). Secondly, when the single study by Ryan (2003)⁵⁴ with post-treatment data from RFS group was removed from the analysis, the result remained stable (HR [95%CI]=2.58, [1.90-3.50], n=6, I²=17.86, P-value=0.298). The results showed similar effect after the removal of the study by Schwarzenbach (2008)⁵⁵ with the method of spectrophotometry from the RFS group because the result was only slightly changed (HR[95%CI]=2.97[2.18-4.04], n=6, I²=29.89%, P-value=0.211). Similarly, sensitivity analysis showed no significant changes when removing total cfDNA from the analysis (HR [95%CI]=3.34, [2.22-5.02], n=5, I²=37.26, P-value=0.173). As per the study in OS group, we excluded the study by Trevisiol (2012) ⁵⁷ which reported cancer-related survival. The analysis did not lead any obvious changes in the total effects (HR[95%CI]=3.00[2.49-3.62], n=11, I²=27.14%, P-value=0.186). Likewise when total cfDNA was removed from the analysis, no obvious changes was seen in the combined results (HR [95%CI]=3.18, [2.57-3.94], n=10, I²=39.01, P-value=0.098). Finally, the combined effects of the analysis were stable even randomly removing one particular study from RFS and OS groups (Figure S1 and Figure S2). The situations were similar with cumulative meta-analysis of both groups cumulated by the year of publication (Figure S3 and Figure S4).

Subgroup Analysis

Although no significant heterogeneity was observed in our including studies either on RFS or OS. It was reported that, tumor markers, tumor stage, marker origin, methods and population size for the presence on cfDNA acted as confounding factors which might exert considerable influences on outcomes of cancer patients⁶². Therefore, we performed subgroup analysis stratified according to the above mentioned factors. And the findings had confirmed the sensibility. In subgroup analysis (Table 2) followed by patient size, the results remained stable. Particularly, the combined HRs and associated CIs of both group was >1 with narrower CIs in large patient group. The results implied studies containing larger patient number précised a better prediction in survival effects. Similarly, when stratified by stage, we found a similar direction of pooled HRs in predicting survival of patients in both localized (UICC stage I-III) and metastatic (UICC stage IV) cancer patients. However heterogeneity became significant when pooling them together (M₀+M₁).The subgroup analysis stratified by methods demonstrated that the combined HRs and associated CIs of both RFS and OS seemed stable, nevertheless, qPCR technique showed reliable results with narrower CIs. Finally subgroup analysis based on type of tumor markers showed both quantitative and qualitative detection of cfDNA were predictive in RFS and OS, while contributing to considerable inter-study heterogeneity, the situation was similar with marker origin. In depth analysis showed that methylated, mutant markers and total cfDNA determination were of prognostic significance.

Publication bias

The funnel plots were symmetric. Meanwhile, results of Begg's and Mazumdar rank correlation were not significant for both group (RFS: two-tailed P-value=0.5480, OS: two-tailed P-value=0.4507) (Figure 3). Hence, there were no obvious publication biases in our included studies.