Harnessing early postoperative MRD and VAF dynamics for precision prognosis in resected colorectal liver metastases.

Introduction
Colorectal cancer (CRC) represents a substantial global health burden, ranking among the most common cancers and leading causes of mortality (Sung et al. 2021; Xi and Xu 2021). The liver is the primary site for CRC metastasis, and the development of colorectal liver metastases (CRLM) is a major driver of patient mortalit (Adam et al. 2010). Surgical resection stands as the most effective treatment for CRLM, offering the only potential for long-term survival (Khoo et al. 2016; Engstrand et al. 2018). However, despite advancements, disease recurrence afflicts 50–70% of patients after surgery (Akgül et al. 2014; Tie et al. 2016), underscoring a critical clinical gap: the lack of precise biomarkers to stratify risk and personalize adjuvant therapies (Dasari et al. 2020; Bettegowda et al. 2014).
Circulating tumor DNA (ctDNA), acting as a non-invasive "liquid biopsy" (Zhou et al. 2025), offers a revolutionary approach by reflecting real-time tumor burden and genomics (Biller and Schrag 2021). Its potential to transform CRC management, from monitoring to prognosis, is increasingly recognize (Tie et al. 2019; Kalil et al. 2024), positioning it as a key candidate to address the biomarker deficit in CRLM.
In the perioperative CRLM setting, ctDNA analysis, particularly for minimal residual disease (MRD), has shown significant promise. There is a strong consensus that postoperative MRD detection is a powerful predictor of recurrence and survival; patients achieving ctDNA negativity post-surgery have markedly better outcomes (Kobayashi et al. 2021; Henriksen et al. 2022). Conversely, the prognostic value of preoperative ctDNA remains ambiguous, especially following neoadjuvant therapy, with studies showing inconsistent results (Parikh et al. 2021; Reinert et al. 2016). This highlights the need for careful interpretation. Furthermore, emerging data suggest that ctDNA dynamics, such as VAF changes (Parikh et al. 2021), and the timing of assessment (e.g., early 2 week testing) (Reinert et al. 2016; Malla et al. 2022), could provide crucial, time-sensitive prognostic information.
Despite this progress, how post-neoadjuvant preoperative ctDNA, early postoperative MRD, and VAF dynamics collectively inform prognosis in CRLM requires further elucidation. Crucially, it remains unknown whether VAF changes can refine risk stratification within established MRD-negative and MRD-positive groups—a vital question for optimizing patient care. This study therefore addresses this critical gap by comprehensively evaluating these key perioperative ctDNA metrics. We aim to establish their independent and combined prognostic value for DFS in resected CRLM patients, thereby providing essential tools to improve risk assessment and guide clinical decision-making towards more personalized and effective treatment strategies.

Materials and methods

Study design and patient population
This retrospective cohort study was conducted at the Sixth Affiliated Hospital, Sun Yat-sen University. The study protocol received approval from the institutional review board (IRB) of the Sixth Affiliated Hospital, Sun Yat-sen University. Written informed consent was obtained from all participating patients prior to any study-related procedures.
We retrospectively identified 111 patients with pathologically confirmed CRLM who underwent liver metastasectomy between October 2022 and February 2024. All included patients had at least one preoperative blood sample available for ctDNA analysis. Patients were excluded based on the following criteria: (1) non-R0 resection (microscopically positive margins) (n = 2); (2) postoperative mortality unrelated to cancer recurrence (n = 1); (3) incomplete follow-up data for disease-free survival (DFS) assessment (n = 11); or (4) postoperative ctDNA sample collection exceeding the 2 week time window (n = 3). Consequently, a total of 94 patients were included in the final cohort for analysis. Among these, 63 patients had evaluable ctDNA results from both preoperative and 2 week postoperative time points and were thus included in the paired perioperative analyses.

Clinical data collection and follow-up
Baseline clinical and pathological data were retrieved for all 94 patients via a review of electronic medical records. These data encompassed: age at surgery, sex, primary tumor location (colon vs. rectum), synchronicity of metastases (synchronous vs. metachronous), details of neoadjuvant chemotherapy and preoperative targeted drug use, pathological T and N stages (pT/pN), tumor differentiation grade, and microsatellite instability (MSI) status. Additionally, we recorded the number of liver metastases (solitary vs. multiple), the maximum diameter (mm) of the largest metastasis, preoperative carcinoembryonic antigen (CEA) levels (categorized as normal [≤ 5 ng/mL] or elevated [> 5 ng/mL]), and RAS/BRAF mutational status as determined from tissue samples.
Following surgery, patients followed a standardized surveillance protocol. The first postoperative imaging assessment (CT or MRI) was performed at 3 months after liver metastasectomy. Subsequent follow-up assessments, including physical examinations, serum CEA measurements, and imaging studies, were conducted at intervals of 3 to 6 months for the first 2 years, and every 6 to 12 months thereafter. DFS was defined as the time interval from the date of liver metastasectomy to the date of the first documented evidence of tumor recurrence at any site, or death from any cause, whichever occurred first.

ctDNA sample collection, sequencing, and analysis
Peripheral blood samples were collected at two specific time points: (1) Preoperative: within 3 days before liver metastasectomy (for patients receiving neoadjuvant chemotherapy, this sample was collected after its completion); (2) Postoperative: at 2 weeks (range 14 ± 5 days) after surgery.
Paired tumor tissue DNA (if available) and germline DNA (from white blood cells) were used alongside cfDNA for next-generation sequencing (NGS) analysis, performed by Tongshu Biotech (Shanghai, China). A targeted panel of 36 genes relevant to colorectal cancer was utilized, achieving a mean sequencing depth exceeding 7000x (Shi et al. 2021).
A patient was defined as ctDNA-positive (or MRD-positive postoperatively) if at least one confirmed somatic mutation was detected with a variant allele fraction (VAF) ≥ 0.1%. VAF was calculated as the ratio of mutant reads to total reads at a specific locus. The maximum absolute VAF Change (MaxΔVAF) was determined for each patient as the largest absolute VAF difference observed for any shared mutation between the preoperative and postoperative samples. If a patient was negative at both time points, MaxΔVAF was 0. If positive at one time point and negative at the other, MaxΔVAF was equal to the VAF at the positive time point.
For subgroup survival analysis, patients were first stratified based on their 2 week postoperative MRD status (negative vs. positive). Within each of these two subgroups (MRD-negative and MRD-positive), the respective median value of MaxΔVAF was calculated. Patients within each subgroup were then further categorized into low-MaxΔVAF (≤ median) and high-MaxΔVAF (> median) groups based on their subgroup-specific median. These MaxΔVAF-based groups were subsequently used for Kaplan–Meier survival analysis.

Statistical analysis
All statistical analyses were conducted using R software (version 4.3.0). Baseline characteristics between groups were compared using the Chi-squared test or Fisher’s exact test for categorical data, and the Student’s t-test or Wilcoxon rank sum test for continuous data. Missing values were excluded during these comparisons. DFS curves were constructed using the Kaplan–Meier method, and group differences were assessed via the log-rank test. Univariate and multivariate Cox proportional hazards models were employed to identify prognostic factors for DFS, calculating hazard ratios (HRs) and their corresponding 95% confidence intervals (CIs). Variables demonstrating a p-value < 0.2 in the univariate analysis were selected for inclusion in the multivariate Cox model. A two-sided p-value < 0.05 was considered statistically significant across all analyses.

Result

Patient characteristics
The study cohort comprised 94 patients who underwent liver metastasectomy for CRLM and had preoperative ctDNA testing. Among these, 63 patients had both preoperative and 2 week postoperative ctDNA assessments and formed the cohort for paired perioperative analysis.
We first compared the baseline characteristics of these 63 patients based on their 2 week postoperative MRD status (Table 1). The cohort consisted of 32 MRD-negative and 31 MRD-positive patients. Although most baseline characteristics, including age, sex, primary tumor location, and preoperative CEA levels, showed no significant difference between the two groups, we observed a trend towards a higher proportion of T4 primary tumors in the MRD-positive group (29%) compared to the MRD-negative group (6.9%) (p = 0.095). A non-significant trend for higher pN stage was also noted in the MRD-positive group (p = 0.14).
We also analyzed the baseline characteristics of all 94 patients based on their preoperative ctDNA status (Table S1). Patients with positive preoperative ctDNA (N = 54) were significantly more likely to have primary tumors located in the rectum (44% vs. 23%, p = 0.031) and elevated preoperative CEA levels (67% vs. 33%, p = 0.001) compared to those with negative preoperative ctDNA (N = 40). By definition, RAS/BRAF mutations in ctDNA were exclusively found in the preoperative positive group (41% vs 0%, p < 0.001). Notably, given that the preoperative sample was collected after neoadjuvant therapy, 100% of preoperative negative patients had received neoadjuvant chemotherapy, compared to 67% of preoperative positive patients (p < 0.001), suggesting a link between chemotherapy response and preoperative ctDNA negativity.

Prognostic value of preoperative and postoperative ctDNA status
We assessed the association between ctDNA status at different time points and DFS. Kaplan–Meier analysis revealed no significant difference in DFS between patients with negative (N = 40) and positive (N = 54) preoperative ctDNA status (log-rank p = 0.77) (Fig. 1A). This lack of association was further confirmed by the univariate Cox analysis for all 94 patients (HR = 1.08, 95% CI: 0.65–1.79, p = 0.8) (Table S2).
In stark contrast, the 2 week postoperative MRD status demonstrated a strong and highly significant association with DFS in the 63 patients with paired samples (Fig. 1B). Patients who were MRD-positive (N = 31) experienced significantly shorter DFS compared to those who were MRD-negative (N = 32) (log-rank p < 0.001). Notably, the prognostic robustness of 2 week postoperative MRD status was confirmed through subgroup analyses stratified by treatment history. We observed that MRD positivity remained a significant predictor of inferior DFS regardless of whether patients had received neoadjuvant chemotherapy or targeted therapy (all p < 0.001) (Fig. S1).

Independent predictors of disease-free survival
To identify independent prognostic factors for DFS, we performed univariate and multivariate Cox proportional hazards regression analysis on the 63 patients with paired ctDNA data (Table 2). Univariate analysis identified only multiple metastases (HR = 2.56, 95% CI: 1.11–5.90, p = 0.028) and postoperative MRD positivity (HR = 6.17, 95% CI: 2.96–12.86, p < 0.001) as significantly associated with worse DFS.
These significant factors, along with others meeting the p < 0.2 inclusion criterion, were entered into a multivariate Cox model. Both postoperative MRD positivity (HR = 5.52, 95% CI: 2.46–12.39, p < 0.001) and multiple metastases (HR = 3.42, 95% CI: 1.42–8.25, p = 0.006) remained significant independent predictors of DFS. Postoperative MRD status thus emerged as the most potent prognostic factor within our cohort.

Perioperative dynamics of ctDNA mutations and VAF
We explored the changes in the ctDNA mutational landscape and VAF between preoperative and postoperative samples in the 63 paired cases (Fig. 2). Preoperatively, TP53 (57%), APC (54%), and KRAS (49%) were the most frequently mutated genes. Following surgery, the prevalence of most mutations decreased substantially; TP53 (32%), APC (41%), and KRAS (28%) remained the most common, albeit at lower frequencies (Fig. 2A, B, C). Concordantly, VAF levels for nearly all detected genes showed a marked decrease postoperatively (Fig. 2D).

Prognostic stratification using VAF dynamics (MaxΔVAF)
Finally, we investigated whether the magnitude of VAF change (MaxΔVAF) could further stratify patients within the postoperative MRD subgroups (Fig. 3). In the MRD-negative subgroup (N = 18 evaluable), patients with a high MaxΔVAF exhibited a trend towards worse DFS compared to those with a low MaxΔVAF (log-rank p = 0.053) (Fig. 3A). However, within the MRD-positive subgroup (N = 31), MaxΔVAF did not show any association with DFS (log-rank p = 0.99) (Fig. 3B).

Discussion
This study investigated the prognostic significance of perioperative ctDNA dynamics in patients undergoing curative-intent resection for CRLM. Our findings underscore the paramount importance of early postoperative ctDNA assessment for risk stratification. We demonstrated that 2 week postoperative MRD status is a potent and independent predictor of DFS, significantly outperforming preoperative ctDNA status. Furthermore, we provided preliminary evidence suggesting that VAF dynamics might offer additional prognostic refinement, particularly within the MRD-negative subgroup.
The central finding of our study is the strong association between 2 week postoperative MRD positivity and a significantly higher risk of disease recurrence (Fig. 1B, Table 2). Patients with detectable ctDNA 2 weeks after surgery had more than a five-fold increased risk of recurrence compared to those who were MRD-negative (HR = 5.52, p < 0.001). This aligns robustly with a growing body of literature highlighting postoperative MRD as a cornerstone in predicting outcomes across various stages of CRC, including the metastatic setting (Tie et al. 2019; Kalil et al. 2024). Our study contributes by confirming this prognostic power at a very early time point (2 weeks), suggesting that crucial risk information can be obtained promptly after surgery. This early insight is clinically valuable, as it can potentially inform decisions regarding the necessity and intensity of adjuvant chemotherapy earlier than conventional or later assessments (Henriksen et al. 2022; Martínez-Castedo et al. 2025). The independence of postoperative MRD from other factors, including the number of metastases (which was also an independent predictor in our cohort), emphasizes its unique biological and clinical significance (Bolhuis, et al. 2021; Nakamura et al. 2024).
In contrast, we found that preoperative ctDNA status, assessed after neoadjuvant chemotherapy, did not correlate with DFS (Fig. 1A, Table S2). This finding contributes to the ongoing debate regarding the utility of preoperative ctDNA in CRLM. While some studies have shown preoperative ctDNA to be prognostic (Kobayashi et al. 2021; Hoang et al. 2025), others, particularly those including patients treated with neoadjuvant therapy, report conflicting results (Tie et al. 2021; Liu et al. 2024). Our results suggest that in a post-neoadjuvant setting, the preoperative ctDNA status might primarily reflect the immediate response to chemotherapy rather than the underlying residual disease burden that ultimately drives recurrence after surgery. It reinforces the notion that the clearance of ctDNA following definitive surgery is a more critical determinant of long-term outcomes than the pre-surgical (post-chemo) snapshot (Malla et al. 2022; Tie et al. 2015).
We also observed a trend linking higher primary tumor T stage (pT4) with postoperative MRD positivity (Table 1). Although this did not reach statistical significance in our cohort (p = 0.095), it is biologically plausible. Higher T stage often signifies more aggressive tumor biology and a greater propensity for micrometastatic spread (Abidoye et al. 2025), which could contribute to the persistence of ctDNA even after resection of macroscopic liver lesions (Ganesh et al. 2019). This potential link warrants further investigation in larger cohorts.
A novel aspect of our study was the exploration of VAF dynamics (MaxΔVAF) for further risk stratification. Interestingly, we found a borderline significant trend (p = 0.053) in the MRD-negative group, where patients with a higher MaxΔVAF (suggesting a larger drop from a potentially higher preoperative VAF) paradoxically showed a trend towards worse DFS (Fig. 3A). This counter-intuitive finding might imply that even among patients achieving MRD negativity, a higher initial tumor burden (reflected by a high preoperative VAF leading to a large ΔVAF) could harbor a residual risk not captured by our 0.1% VAF detection limit (Parikh et al. 2021; Ignatiadis et al. 2021). Conversely, MaxΔVAF held no prognostic value in the MRD-positive group (Fig. 3B), likely because the mere presence of detectable MRD post-surgery is such a strong adverse factor that it overshadows any subtler prognostic information from the magnitude of VAF change (Pereira et al. 2017). These findings, particularly the MRD-negative subgroup (N = 18) analysis, are exploratory and require validation, but they open avenues for more sophisticated risk modeling. Our OncoPrint analysis (Fig. 2) further detailed the molecular landscape changes, confirming the overall reduction in tumor burden while highlighting the persistence of key driver mutations like TP53, APC, and KRAS in MRD-positive patients (Cancer Genome Atlas Network 2012).
This study possesses several strengths, including the focus on a well-defined CRLM cohort undergoing resection, the use of paired perioperative samples, and the specific assessment at an early 2 week postoperative time point. However, we acknowledge several limitations. Firstly, the retrospective design carries inherent biases. Secondly, the sample size, especially for the subgroup analyses (N = 63 for paired, N = 18 for MRD-negative MaxΔVAF), is relatively small, which may limit statistical power and warrants caution in interpreting borderline findings (like p = 0.053). Thirdly, it is a single-center study, potentially limiting the generalizability of our findings. Lastly, while our 36-gene panel covers key CRC drivers, it might not capture all relevant mutations, and its VAF detection limit could misclassify some low-level MRD-positive cases as negative.
Despite these limitations, our findings have significant clinical implications. The robust predictive power of 2 week postoperative MRD strongly supports its integration into routine clinical practice for CRLM patients. It can identify high-risk (MRD-positive) patients who may benefit most from intensified adjuvant chemotherapy or novel therapeutic strategies (Taniguchi et al. 2021; Nakamura et al. 2021), and low-risk (MRD-negative) patients for whom treatment de-escalation or even active surveillance could be considered, thus minimizing treatment-related toxicity (Dasari et al. 2020; Osumi et al. 2019). The potential, albeit preliminary, role of MaxΔVAF in MRD-negative patients suggests future research could develop combined clinical-molecular scores for even more granular risk assessment.

Conclusions
In conclusion, our study confirms that early postoperative MRD, detected 2 weeks after surgery, is a powerful independent predictor of DFS in patients undergoing resection for CRLM. It provides a more accurate risk assessment than preoperative ctDNA status in the post-neoadjuvant setting. While VAF dynamics show potential for further refining prognosis, especially in MRD-negative patients, this requires validation in larger, prospective studies. Integrating early postoperative MRD assessment into clinical workflows holds significant promise for optimizing personalized treatment strategies and improving outcomes for CRLM patients.

Supplementary Information
Below is the link to the electronic supplementary material.