Firstly, the present network meta-analysis was registered in the International Prospective Register of Systematic Review (ID: CRD420251138090) and adhered to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA).

Studies were considered eligible for inclusion if they met the following criteria: i) Population: Adult patients aged ≥18 years with histologically determined RAS wild-type MCRC, including initially metastatic or recurrent disease; ii) intervention: Any first-line systemic anticancer treatment, including chemotherapy, targeted therapy, immunotherapy or combination regimens; iii) comparison: No restrictions on comparator arms; iv) outcomes: Reporting at least one primary outcome, including OS or PFS; and v) study design: RCTs.

Studies were excluded if they met any of the following criteria: i) Non-randomized study designs, including observational studies, case series or case reports; ii) inclusion of RAS-mutant patients without separately extractable data for RAS wild-type subgroups; iii) focus on adjuvant or neoadjuvant treatment settings; iv) inclusion of patients with resectable metastatic disease; v) duplicate publications or overlapping patient populations; vi) conference abstracts when full-text publications of the same study were available; or vii) letters, editorials, reviews or commentaries.

A comprehensive systematic literature search was conducted across MEDLINE (https://www.nlm.nih.gov/medline/medline_home.html)/PubMed (https://pubmed.ncbi.nlm.nih.gov/), Embase (https://www.embase.com/), the Cochrane Central Register of Controlled Trials through the Ovid platform (https://ovidsp.ovid.com/) and Web of Science (https://www.webofscience.com/) from inception to September 1, 2025. The search strategy combined Medical Subject Headings and free-text keywords associated with CRC, RAS wild-type status, metastatic disease and first-line treatment. No language restrictions were applied. In addition, reference lists of included studies and relevant systematic reviews were manually screened to identify additional eligible studies. The complete search strategy is provided in Table SI.

Data was extracted from eligible studies by two independent reviewers using a standardized data extraction form. Discrepancies were resolved through discussion or consultation with a third reviewer. The following information was extracted from each study: i) Trial identification, including trial ID, publication year and study region; ii) study characteristics, including trial phase and sample size; iii) patient demographics, including mean age and sex distribution; iv) tumor characteristics, including primary tumor sidedness and metastatic status; v) interventions compared; and vi) reported clinical outcomes. For efficacy outcomes, OS, PFS, ORR and duration of response were extracted, along with corresponding hazard ratios (HRs), odds ratios (ORs) and 95% CIs. Safety data, including incidence of grade ≥3 treatment-emergent adverse events (TEAEs), were also extracted when available.

Methodological quality of included studies was assessed using the Cochrane Collaboration Risk of Bias tool (22). The two independent reviewers evaluated each study across seven domains: Random sequence generation, allocation concealment, blinding of participants and personnel, blinding of outcome assessment, incomplete outcome data, selective reporting and other potential sources of bias. Each domain was classified as low risk, high risk or unclear risk of bias. Disagreements were resolved through consensus or consultation with a third reviewer. Risk of bias summary figures and graphs were generated using Review Manager (version 5.4; The Cochrane Collaboration).

A random-effects network meta-analysis was conducted within a frequentist framework using the ‘netmeta’ package in R (version 4.5.2; Posit Software, PBC) (23). This approach applied a graph-theoretical model based on electrical network theory. Network plots were generated to visualize the evidence base, where node size was proportional to total sample size for each intervention and line thickness corresponded to the number of trials for each comparison. The statistical model accounted for associated treatment effects and addressed zero-event studies by applying a continuity correction to trials with zero events in a single arm, while trials with zero events in both arms were excluded. Results were presented as ORs or HRs with corresponding 95% CIs, depending on the outcome.

Treatment rankings were determined using P-scores, representing the frequentist equivalent of the surface under the cumulative ranking curve, to estimate treatment superiority. Consistency between direct and indirect evidence within closed loops was assessed using the node-splitting method. Potential small-study effects were evaluated through visual inspection of funnel plots and formally tested using Egger's test. Heterogeneity was assessed using the Cochran Q statistic, decomposed into within-design and between-design components and quantified using the I² statistic. I² values of 25-49, 50-74 and ≥75% were interpreted as low, moderate and high heterogeneity, respectively. All statistical tests were two-sided, and P<0.05 was considered to indicate a statistically significant difference.

Study selection processes for the present analysis are outlined in the PRISMA flow diagram (Fig. 1). The initial database search identified 3,799 records. After removal of 614 duplicates, 1,484 records were screened, resulting in retrieval of 102 full-text articles for eligibility assessment. Following detailed evaluation, 15 RCTs met the inclusion criteria and were included in the present network meta-analysis (7-10,13-16,24-30). Study characteristics are summarized in Table I. Collectively, the 15 trials included 6,298 patients. Among the included studies, 10 were phase III trials and 5 were phase II trials.

Patient characteristics were generally balanced across studies. The mean age was 55-77 years, with a predominance of male patients ranging from 61-76% across studies. Primary tumor location was predominantly left-sided, accounting for 61-100% of patients, consistent with the known favorable prognosis of left-sided tumors in RAS wild-type disease (17,18). The liver was the most common site of metastasis, reported in 12-92% of patients, followed by lung metastases in 21-33% and lymph node involvement in 20-25% of patients (Table I).

Overall methodological quality of the included studies was satisfactory (Fig. 2). All trials demonstrated low risk of bias for random sequence generation and allocation concealment, indicating appropriate randomization procedures. The majority of studies also exhibited low risk of attrition bias and selective reporting bias, indicating that outcome data were generally complete and that no major selective outcome reporting was identified according to the aforementioned prespecified risk-of-bias criteria.

The main methodological limitation was lack of blinding of participants and personnel, with 93% of studies classified as high risk for performance bias. This was largely due to open-label study designs required when comparing treatments with different administration schedules or toxicity profiles.

A notable proportion of studies (~67%) exhibited unclear or high risk of other bias, primarily associated with pharmaceutical industry funding or involvement of industry-affiliated authors. Although industry sponsorship does not necessarily compromise study validity, the potential for conflicts of interest justified classification as unclear risk in the present domain.

Evidence network for the primary analysis included 12 RCTs, seven treatment nodes and eight distinct pairwise comparisons (Fig. S1). UGT1A1-guided bevacizumab + FOLFIRI (trial ID: NCT02256800), vitamin C + bevacizumab + FOLFOX (10) and A140 + FOLFOX (trial ID: NCT04835142) were excluded because they were single-trial nodes with small sample sizes. The primary analysis comprised the following seven treatment nodes: Bevacizumab + chemotherapy (Bev + Chemo), cetuximab + chemotherapy (Cet + Chemo), panitumumab + chemotherapy (Pan + Chemo), FOLFOX, FOLFIRI, 5-fluorouracil/leucovorin (5-FU/LV) and cetuximab maintenance. The network was fully connected, with Bev + Chemo serving as the most frequently used comparator. In the network plot, node size reflects cumulative sample size and edge thickness is proportional to the number of contributing studies per comparison.

Primary network meta-analysis for OS incorporated 12 RCTs. The network demonstrated excellent statistical consistency, with no significant between-design inconsistency (Q_between=0.53; df=2; P=0.768), negligible overall heterogeneity (Q_total=2.67; df=6; P=0.849; τ²=0; I²=0%) and no statistically significant disagreement between direct and indirect evidence in node-splitting analysis for any evaluable comparison (all P>0.05; Table SII). Forest plots presenting OS HRs relative to Bev + Chemo under the random-effects model are shown in Fig. 3A. A total of two treatment regimens demonstrated statistically significant superiority compared with Bev + Chemo: Cet + Chemo (HR=0.853; 95% CI: 0.775-0.938; P=0.001) and Pan + Chemo (HR=0.855; 95% CI: 0.738-0.992; P=0.038). The remaining treatments, including cetuximab maintenance, FOLFIRI, FOLFOX and 5-FU/LV, did not differ significantly from Bev + Chemo. Complete pairwise comparisons between all treatment nodes are provided in the league table (Tables SIII and SIV). Treatment ranking based on P-scores identified Cet + Chemo as the regimen most likely to provide the highest OS (P-score=0.814), followed by Pan + Chemo (P-score=0.798) and 5-FU/LV (P-score=0.735). Bev + Chemo ranked fourth (P-score=0.420), followed by cetuximab maintenance (P-score=0.350), FOLFIRI (P-score=0.225) and FOLFOX (P-score=0.158; Fig. 4A).

PFS analysis included the same 12 RCTs and 7 treatment nodes, with moderate between-study heterogeneity (I²=44.8%; Q_total=10.87; df=6; P=0.092). Between-design inconsistency was not statistically significant (Q_between=1.308; P=0.520) and node-splitting determined no statistically significant direct-indirect disagreement for any evaluable comparison (all P>0.05; Tables SII and SV). Forest plots for PFS are presented in Fig. 3B. Under the random-effects model, Cet + Chemo showed a non-significant trend toward PFS benefit over Bev + Chemo (HR=0.880; 95% CI: 0.766-1.011; P=0.072). Notably, FOLFIRI alone was significantly inferior to Bev + Chemo for PFS (HR=1.631; 95% CI: 1.294-2.056; P<0.001), reflecting the marked PFS benefit conferred by bevacizumab addition to chemotherapy. No other treatment differed significantly from Bev + Chemo in PFS. P-score ranking placed Cet + Chemo as the top-ranked treatment for PFS (P-score=0.914), followed by Bev + Chemo (P-score=0.654), 5-FU/LV (P-score=0.598) and cetuximab maintenance (P-score=0.561; Fig. 4B).

ORR analysis incorporated 8 RCTs and 7 treatment nodes, with moderate heterogeneity (I²=40.9%) and borderline between-design inconsistency (Q_between=2.784; P=0.095). Forest plots for ORR are presented in Fig. 3C, with estimates expressed as ORRs relative to Bev + Chemo (OR >1 indicates higher ORR compared with Bev + Chemo). A total of three regimens demonstrated significantly higher ORR than Bev + Chemo under the random-effects model: Cet + Chemo (OR=1.612; 95% CI: 1.087-2.390; P=0.017), cetuximab maintenance (OR=2.321; 95% CI: 1.011-5.327; P=0.047) and 5-FU/LV (OR=2.989; 95% CI: 1.012-8.828; P=0.047). Pan + Chemo, FOLFIRI and FOLFOX did not differ significantly from Bev + Chemo. P-score ranking identified 5-FU/LV (P-score=0.915) and cetuximab maintenance (P-score=0.830) as top-ranked, followed by Cet + Chemo (P-score=0.650) and Pan + Chemo (P-score=0.562; Fig. 4C).

A pre-specified subgroup analysis was conducted stratifying treatment effects by primary tumor sidedness (left-sided vs. right-sided). Subgroup-specific HRs were extractable from 5 trials for OS and 6 trials for PFS. Forest plots are provided in Fig. 5. For OS, anti-EGFR-containing regimens demonstrated consistent benefit in left-sided tumors across all 5 contributing studies. The pooled HR for anti-EGFR vs. control in left-sided tumors was 0.750 (95% CI: 0.672-0.836; I²=0%), representing a 25.0% mortality risk reduction. In right-sided tumors, the corresponding pooled HR was 1.097 (95% CI: 0.909-1.325; I²=0%), indicating no OS benefit from anti-EGFR therapy and a non-significant trend favoring bevacizumab-based or chemotherapy-only approaches. For PFS, a broadly consistent directional pattern was observed. In left-sided tumors, the pooled PFS HR for anti-EGFR vs. control was 0.840 (95% CI: 0.687-1.027; I²=68.6%), reflecting a non-significant trend toward benefit. In right-sided tumors, the pooled PFS HR was 1.143 (95% CI: 0.851-1.536; I²=54.9%), suggesting no PFS advantage and a non-significant numerical disadvantage, for anti-EGFR agents. Notably, the PARADIGM trial reported a statistically significant PFS detriment for panitumumab plus mFOLFOX6 vs. bevacizumab + mFOLFOX6 in right-sided tumors (HR=1.43; 95% CI: 1.03-1.97) (13).

Incidence of grade ≥3 TEAEs across included treatments is summarized in Table I. Rates ranged from 23.4-83.3% across treatment arms. Anti-EGFR-based combinations were associated with the highest toxicity burden: Cetuximab biosimilar plus FOLFIRI reported 83.3% grade ≥3 TEAEs (trial no. NCT03206151) and panitumumab plus FOLFOX 81.3% (trial no. NCT01328171). Bevacizumab-containing regimens demonstrated comparatively more favorable toxicity profiles (grade ≥3 TEAE rate: 23.4-68.3% across trials). FOLFIRI alone was associated with grade ≥3 TEAEs in 66.9% of patients. Due to heterogeneous adverse event reporting across trials, a formal network meta-analysis of safety outcomes was not performed. Descriptively, anti-EGFR agents were consistently associated with skin toxicity, hypomagnesemia and infusion reactions, while bevacizumab-based regimens were more commonly associated with hypertension and thromboembolic events. Detailed adverse event data are provided in Table SVI.

A sensitivity analysis incorporating all 15 RCTs and 9 treatment nodes, including UGT1A1-guided bevacizumab + FOLFIRI, vitamin C plus bevacizumab + FOLFOX and A140 + FOLFOX, was conducted. The expanded evidence network is presented in Fig. S2, with full numerical results in Tables SI and II.

The primary analysis conclusions were demonstrated to be robust: Cet + Chemo and Pan + Chemo maintained statistically significant OS benefits in comparison with Bev + Chemo. In the nine-node sensitivity network, the UGT1A1-guided bevacizumab + FOLFIRI node ranked highest for both OS (P-score=0.932) and PFS (P-score=0.992), consistent with the favorable efficacy estimates reported in the single PURE FIST trial (OS HR=0.693; 95% CI: 0.503-0.955; PFS HR=0.539; 95% CI: 0.398-0.730) (9). However, as these estimates derive from a single study employing a non-genotyped control arm, their comparability with other nodes is limited and they are reported as supplementary findings. A140 + FOLFOX and vitamin C + bevacizumab + FOLFOX showed no statistically significant benefit compared with any comparator.

For PFS (I²=44.8%; τ²=0.010) and ORR (I²=40.9%; τ²=0.041), where moderate between-study heterogeneity was observed, meta-regression analyses were conducted incorporating publication year, geographic region (Eastern vs. Western), chemotherapy backbone (FOLFOX vs. FOLFIRI) and trial phase (II vs. III) as study-level covariates.

For PFS, none of the examined moderators reached statistical significance: Publication year (P=0.420), region (P=0.523), trial phase (P=0.686) and chemotherapy backbone (P=0.151). Chemotherapy backbone explained the largest, though non-significant, proportion of PFS heterogeneity (R2 analog=9.9%), with FOLFOX-based regimens showing a non-significant trend toward higher HR in comparison with FOLFIRI-based regimens (coefficient=0.128; P=0.151).

For ORR, geographic region was the strongest, though non-significant, moderator (R2 analog=24.5%; P=0.150), with Eastern-region trials showing higher pooled ORR for anti-EGFR regimens compared with Western-region trials (coefficient=0.600). Publication year (P=0.347), trial phase (P=0.359) and chemotherapy backbone (P=0.610) did not explain ORR heterogeneity. Bubble and dot plots illustrating the association between study-level covariates and treatment effects for PFS and ORR are provided in Fig. S3.

Comparison-adjusted funnel plots for OS, PFS and ORR are presented in Fig. S4. The funnel plots for all three outcomes demonstrated broadly symmetric distributions around the null, with no consistent pattern suggestive of small-study effects or publication bias. Formal statistical asymmetry tests were not performed owing to the limited number of studies per individual comparison (k<10 for the majority of direct comparisons).