The search strategy is provided in Supplementary File S2. In 31 July 2023, we conducted a comprehensive search on PubMed, EMBASE, Cochrane Library, and ClinicalTrials.gov to find relative RCTs and published studies. There were no restrictions on the publication date, and language was limited to English. Moreover, abstracts from the European Society for Medical Oncology, American Society of Clinical Oncology since 2021 were also included in the search.

Two researchers (YJ and TS) independently screened all articles identified through the database search by title and abstract. Articles that met the inclusion criteria were then subjected to full-text screening. Discrepancies were sorted out through discussions involving other researchers (YJ, TS, MZ, YX, and XZ). The eligibility criteria based on the PICOS framework were as follows:

(1) Population: Adult patients with confirmed CRLM, diagnosed either histologically or cytologically. Patients also need to meet the requirement of receiving first-line treatment or maintenance treatment after prior system treatment stabilization. No limitations were imposed regarding individual-level characteristics. Due to the fact that some RCTs only reported results for mCRC patients with 2 or more organ metastases, we assumed that such patients had liver metastasis, considering liver metastasis is present in more than 90% of these patients (Riihimäki et al., 2016; Reboux et al., 2022).

To avoid redundancy, we focused on trials that provided the most recent and significant insights. Moreover, we dismissed trials that explored treatments unrelated to any comparisons. Additionally, trials that specifically investigated varying dosages but implemented the same administrations were also eliminated.

Two independent researchers (YJ and TS) were responsible for extracting the required data. The extracted information encompassed the characteristics of eligible trials (publication year, registration information, etc.), characteristics of populations (age, sample size, countries, etc.), and characteristics of the program (interventions, outcomes of endpoints, etc.). The clinical outcomes extracted included OS, PFS, ORR, SAEs, and R0 liver resection. For studies that only published Kaplan-Meier curves without providing hazard ratios (HRs) or a 95% confidence interval (CI), Liu et al.'s tool was used to extract OS or PFS rates, and number-at-risk from Kaplan-Meier curves. Individual patient data (IPD) were then reconstructed, and HR and their 95% CI were calculated based on reconstructed IPD (Liu et al., 2021).

Cochrane Collaboration’s risk of bias (ROB) tool was used to evaluate the quality of the studies included (Higgins et al., 2011). The eligible studies were categorized into three groups: high, low, or unclear risk (Lin and Chu, 2018). To evaluate the publication bias, the Egger regression test was utilized, with p-values <0.05 being interpreted as evidence of bias.

The primary outcomes analysed are OS and PFS, while the secondary outcomes included ORR and SAE, as well as R0 liver resection. Network plots were created to compare and visually represent the different treatment options. Pooled hazard ratios (HRs) with 95% CI were computed for OS and PFS. Pooled odds ratios (ORs) with 95% CI were calculated for ORR, SAE, and R0 resection rate. The analysis of synthesized HRs or ORs utilized the Bayesian approach, taking into account that the majority of direct evidence stemmed from a single trial. Therefore, the fixed effects consistency model was chosen (Zhao et al., 2019). The Bayesian network meta-analysis (NMA) using the R statistical packages Gemtc was carried out by employing four sets of Markov chains. Each set consisted of 50,000 samples with 10,000 burn-in samples. Non-informative prior distributions were utilized: specifically, a uniform prior distribution (Uniform (0, 1)) was used for parameter theta, and a normal prior distribution (Normal (0, 10^6)) was used for parameter mu (Sutton et al., 2008). In addition, we calculated the probability ranking for each available treatment and represented it using the surface area under the cumulative ranking (SUCRA). A higher SUCRA value indicated a greater rank.

The I² statistic was used to evaluate heterogeneity among studies, with a value greater than 50% indicating a moderate level of heterogeneity (Zhao et al., 2019). The edge-splitting method was used to assess the inconsistency of models, taking into account direct and indirect evidence (Zhao et al., 2019). To ensure the robustness of this study, several comparisons were conducted using pairwise meta-analysis. To confirm the convergence of Markov chains, trace plots and Gelman-Rubin diagnostic statistics were utilized (Brooks and Gelman, 1998).

To assess the influence of the number of metastatic organs, subgroup and sensitivity analyses were conducted to evaluate the dependability of the findings. We classified the population receiving first-line treatment into two categories: those with liver-limited metastasis and those with multiple metastasis sites, and then conducted subgroup analysis separately. Due to limited evidence for patients receiving maintenance treatment, we chose not to classify this group. In the sensitivity analyses, we studied the potential impact of mutation target levels on the efficacy of intervention strategies.

A total of 4979 records were identified from database searches. 3614 records were excluded during the title and abstract screening, and 1365 records were screened in full text. Sixty-two articles, comprising 56 RCTs, were included in the review. The flow chart can be seen in Figure 1. Details about the included studies are presented in Supplementary File S3.

This research study included a total of 21,323 patients diagnosed with metastasis CRC (mCRC). Of these, 50 RCTs are applicable for the comparative analysis of first-line treatment, and six RCTs were specifically designed for maintenance treatment. To form a complete indirect comparison, we classified chemotherapy into single-drug chemotherapy (SingleCT), fluorouracil-based combination chemotherapy (CTFU), capecitabine-based combination (CTCA), and intensified CTFU (ICTFU) which contained four drugs. We uniformly referred to the best supportive treatment, observation, or placebo as BSC without distinction. Briefly, 12 mechanisms have been included, including: BSC, immune checkpoint inhibitor (ICI)+multi-drug chemotherapy (MultiCT), ICI + targeted therapy (TAR)+MultiCT, local treatment (defined as hepatic artery infusion [HAI], Selective internal radiation therapy [SIRT], transcatheter arterial chemoembolization [TACE], resection or ablation [R/A]), local treatment + MultiCT, local treatment + SingleCT, MultiCT, multi-targeted therapy (MultiTAR), RNA therapy + TAR + MultiCT, SingleCT, TAR + MultiCT, TAR + SingleCT. Besides, 29 treatments were involved, comprising of Adavosertib, Atezolizumab + Bevacizumab + ICTFU, Bevacizumab, Bevacizumab + CTCA, Bevacizumab + CTFU, Bevacizumab + Erlotinib, Bevacizumab + ICTFU, Bevacizumab + SingleCT, Cediranib + CTCA, Cetuximab + CTFU, Cetuximab + ICTFU, CTCA, CTFU, CTFU + SIRT, GOLFIG (defined as combination of gemcitabine, oxaliplatin, levofolinate, 5-fluorouracil, granulocyte-macrophage colony-stimulating factor, and interleukin-2), ICTFU, R/A + CTFU, R/A + SingleCT, panitumumab + CTFU, panitumumab + ICTFU, pelareorep + bevacizumab + CTFU, S1 (combined chemotherapy of tegafur, gimeracil, and oteracil), S1+bevacizumab + singleCT, S1+singleCT, SingleCT, sorafenib + CTFU, TAS-102 (trifluridine/tipiracil) + bevacizumab, tivozanib + CTFU, TSU-68 (Orantinib) + S1 + singleCT. We did not classify capecitabine and fluorouracil as the same drug category. Despite similar mechanisms, their differences in clinical application and side effect management warranted separate analyses (Twelves et al., 2005; Cassidy et al., 2008). This yielded more accurate results and better supported clinical decisions.

The assessment of ROB is presented in Supplementary File S4. Overall, ROB in all RCTs was generally low. However, multiple RCTs were open-label in our study (Cassidy et al., 2008; Ducreux et al., 2011; Hoff et al., 2012; Hong et al., 2012; Ruers et al., 2012; Yamada et al., 2013; Correale et al., 2014; Heinemann et al., 2014; Lee et al., 2014; Loupakis et al., 2014; Gruenberger et al., 2015; Simkens et al., 2015; Tournigand et al., 2015; Benson et al., 2016; Luo et al., 2016; Stintzing et al., 2016; Aparicio et al., 2018; Jonker et al., 2018; Qin et al., 2018; Yamada et al., 2018; Aranda et al., 2020; Cremolini et al., 2020; Van Cutsem et al., 2020; Avallone et al., 2021; Kanemitsu et al., 2021; Antoniotti et al., 2022; Modest et al., 2022; Rossini et al., 2022; Ychou et al., 2022; Bond et al., 2023; Watanabe et al., 2023), this raised concerns about the blinding of participants and personnel, assessment of outcomes, and concealment of allocation. Furthermore, several RCTs were found to have potential bias because of insufficient availability of outcome data. The network’s results from the Egger test showed no publication bias, and the funnel plots can be found in Supplementary File S5.

This study is the first to systematically evaluate the efficacy and safety of different treatment options for patients with CRLM. The key findings are summarized as follows:

1. For CRLM patients, the optimal treatment options were local treatment + chemotherapy and TAR + chemotherapy. In terms of overall survival (OS), the best choices were R/A+ SingleCT or CTFU, Cetuximab + ICTFU or CTFU, Bevacizumab + ICTFU or CTFU, and Panitumumab + CTFU. For progression-free survival (PFS), the top options were Immune + TAR + chemotherapy, MultiTAR, local treatment + chemotherapy, and TAR + chemotherapy. For patients with liver-limited metastasis, Cetuximab, Bevacizumab, and Panitumumab + chemotherapy were the best choices for both OS and PFS. For those with multiple metastatic sites, Bevacizumab + ICTFU, Cetuximab or Panitumumab + CTFU were the best for OS, while TAS-102 + Bevacizumab, Bevacizumab + SingleCT, and GOLFIG were optimal for PFS.

In the base-case analysis, heterogeneity was observed in some networks, such as the PFS of the overall population, likely due to lack of limitation on target mutation levels. Sensitivity analysis confirmed that controlling for target expression levels reduced heterogeneity across all networks.

TAR + chemotherapy can produce a higher remission rate and improve resectability. EGFR or VEGF inhibitors combined with chemotherapy are the best choices for patients with unresectable CRLM. EGFR inhibitors, such as Cetuximab and Panitumumab, are associated with higher tumor response and expedited symptom relief. Anti-EGFR therapy may induce tumor-specific adaptive immune responses and immunogenic cell apoptosis. VEGF inhibitors, such as Bevacizumab and Aflibercept, normalize tumor vasculature, increasing tumor blood supply (Xie et al., 2020). Combining chemotherapy with targeted therapy can enhance patient survival, but it also increases SAEs, making it challenging to find a solution that is both effective and safe. Surprisingly, adding Cetuximab to ICTFU did not worsen safety compared to CTFU, and Bevacizumab + ICTFU is also acceptable in terms of safety. Thus, Bevacizumab + ICTFU and Cetuximab + ICTFU are optimal for multiple metastasis sites and liver-limited metastasis, respectively. For the WT population, Cetuximab + CTFU is the best choice. New therapies like ICI + TAR and MultiTAR showed great PFS performance but had poor safety profiles, and more OS data are needed to confirm their efficacy. More RCTs are needed to identify the applicable population for these new mechanisms in precision treatment for CRC.

Despite using fewer drugs, monotherapy has been relatively effective in maintaining treatment efficacy for both PFS and OS. Bevacizumab combined with SingleCT is the most effective for OS, but its safety has been poor (Simkens et al., 2015). On the other hand, Adavosertib has demonstrated the best PFS, but more data is needed to establish its safety profile. Overall, monotherapy may still be the ideal choice at present.

Right- and left-sided colorectal tumors exhibit distinct epidemiological, clinicopathological characteristics, gene expression profiles, genetic alterations, and prognoses. Thus, efficacy of EGFR inhibitors is significantly influenced by the primary tumor site. Moretto et al. (2016) demonstrated that anti-EGFR therapies are less effective in right-sided tumors compared to left-sided ones. Takayuki’s meta-analysis (Yoshino et al., 2024) further confirmed the superior efficacy of EGFR inhibitors in patients with left-sided primary tumors. Due to data limitations, our study could not differentiate the primary tumor site in patients with liver metastases. Future research should address this aspect more thoroughly.

Recent trials have explored the use of anti-EGFR monoclonal antibodies in maintenance therapy. Filippo et al.'s phase II trial indicated that using panitumumab alone was less effective for PFS compared to a combination of panitumumab with fluorouracil-leucovorin (Pietrantonio et al., 2019). Similarly, the ERMES study concluded that cetuximab alone was not as effective for maintenance following FOLFIRI/cetuximab induction and thus is not recommended (Pinto et al., 2024). Due to the focus on liver metastasis in our study and limited available data, we could not evaluate additional maintenance therapies, including anti-EGFR monoclonal antibodies.

The advantage of this research is clear. It is the first study to systematically compare the efficacy and safety of all first-line and maintenance treatment regimens for CRLM patients, providing significant reference value for clinical practice and guidelines. Unlike previous studies, this study meticulously analyzed the type and mechanism of chemotherapy to minimize heterogeneity. The low heterogeneity of the study population enhances the reliability of the conclusions. Sensitivity analysis and in-depth examination of heterogeneity sources confirmed the robustness of the basic analysis. Multiple subgroup analyses, including distinctions between liver-limited and multiple-site metastases and analyses of wild-type patients, support precision treatment for CRC. Our innovative conclusions offer a new direction for future clinical research and provide substantial evidence for clinical decision-making.

Due to the availability of data, this study has some limitations. First, we cannot analyze the RAS/KRAS or RAF mutation population. Second, limited by a lack of individual data, the majority of studies only reported HR. Therefore, we utilized the time-invariant HR methods for indirect comparison, as opposed to using other risk variable models. Third, in order to form more comparisons, we consider patients with multiple-site metastases to have liver metastasis, even though this proportion exceeds 90%, it also introduces some uncertainty. Fourth, the relative efficacy between a certain number of schemes is obtained through indirect comparison, and more direct evidence from RCTs is needed to validate our findings. Fifth, for the results related to immunotherapy, a more cautious interpretation is required, as the evidence is based on a small subset of patients. Additionally, more clinical evidence is needed to validate the conclusions of this study.