A comprehensive electronic search, prospectively registered in PROSPERO (CRD420251082905), was executed on 31 October 2024 and updated on 28 July 2025, from database inception, across MEDLINE (OVID-SP), Embase, Scopus, Cochrane CENTRAL, Google Scholar, and ClinicalTrials.gov. Two research registries (MedRxiv and Research Square) were also enquired. The core string combined controlled vocabulary and free-text synonyms: (“vascular endothelial growth factor” OR VEGF OR Angiogenesis) AND endometriosis AND (treatment OR management OR therap*), with field limits set to Title/Abstract. No restrictions were imposed regarding language, organism, or publication date. Reference lists of relevant systematic reviews and eligible meta-analyses were hand-searched for additional studies (Supplementary Figure S1).

Randomised and non-randomised controlled trials, comparative cohort studies, and controlled laboratory experiments (murine or primate) that evaluated a systemic or local pharmacological agent acting on VEGF signalling in surgically or histologically confirmed endometriosis were included. We defined “VEGF-directed” pharmacotherapies a priori as systemic or local agents with a primary or well-described secondary mechanism involving VEGF signalling or angiogenesis modulation (anti-VEGF antibodies, multi-kinase inhibitors targeting VEGFR signalling, mTOR-PI3K modulators, GnRH analogues/dopamine agonists with documented anti-angiogenic effects, and pleiotropic compounds with reproducible VEGF down-regulation in experimental models). We required studies to report at least one lesion-level outcome (lesion area, lesion number, or VEGF expression within ectopic tissue). Studies limited to in vitro assays without a lesion-bearing model, narrative reviews, and non-comparative series were excluded. Narrative reviews, in vitro only studies, conference abstracts without full text, and papers lacking extractable outcome data were excluded. Three reviewers (SLK, OH, and PS) independently screened titles and abstracts, followed by a full-text review. An independent fourth reviewer (OE) assessed the final set of texts included in the study (Supplementary Figure S1).

The present study aimed to quantify the effectiveness of fertility-sparing treatments, outside of the hormone suppression blanket' across clinical and histological parameters relevant to disease control. In addition, the level of evidence certainty, associated heterogeneity, and organism-level divergence was assessed.

Primary outcomes were defined as (i) lesion area, measured in mm² or as a volumetric surrogate and extracted as mean change from baseline or between-group difference at endpoint; (ii) lesion number, expressed as absolute count or percentage reduction; and (iii) VEGF expression within ectopic lesions, assessed by immunohistochemistry, ELISA, or mRNA quantification and normalised to control tissue when possible. Secondary outcomes were microvascular density (MVD), endometrial cell proliferation (Ki-67 markers), and apoptosis (TUNEL, caspase-3). Outcomes were normalised where feasible as ratios of (post-control − post-treatment)/post-control to facilitate comparison across models.

Using a piloted extraction sheet (Excel 2023), SLK and OH captured study design, population characteristics, intervention details (agent, dose, route, and duration), comparator, and outcome data [mean ± SD or median (IQR)] (Supplementary Table S1). Reasons for exclusion at the eligibility stage were documented. Data accuracy was confirmed by two reviewers (RMS and EP), who performed a 20% random audit for transcription accuracy; no discrepancies > 3% were detected.

Observational or quasi-randomised studies trials (N = 3) were assessed using the ROBINS-I tool (Supplementary Table S2) and animal studies using SYRCLE's checklist (Supplementary Table S3). Judgements were made independently by SLK and OH; consensus was sought through discussion or, if needed, through arbitration by OKE (Supplementary Figure S2). The GRADE framework was adapted for the network meta-analysis, downgrading for risk of bias, indirectness, inconsistency, imprecision, and publication bias. Summary of findings tables was generated with CINeMA.

Where ≥2 studies reported a common outcome, pairwise meta-analyses were performed in R 4.3.2 (meta and netmeta packages) using inverse-variance random-effects models (Hartung–Knapp adjustment). For network geometry with three or more competing interventions, a frequentist random-effects network meta-analysis was performed, estimating pooled relative effects as standardised mean differences and 95% confidence intervals (continuous outcomes). Surface under the cumulative ranking curve (SUCRA) values were calculated for each agent to summarise the probability of that treatment being among the best within the observed network. Notably, SUCRA values should be interpreted as relative rankings conditional on the included studies and network structure and not as absolute measures of biological efficacy or clinical usefulness. Transitivity was examined clinically and statistically; global inconsistency was probed using the design-by-treatment interaction model and local inconsistency using node splitting (Supplementary Figure S3). Heterogeneity was quantified by τ² and I², and publication bias was examined using comparison-adjusted funnel plots where ≥10 trials contributed. Pre-specified sensitivity analyses included (i) restriction to non-rodent (“human/humanoid”: human, patient-derived cells, and primates) models and (ii) exclusion of single-study nodes. These were used to assess the robustness of treatment rankings to species and network sparsity.

The target-mapping workflow was initiated by harmonising all chemical–gene associations of Homo sapiens for each analysed index compound, drawing from the Comparative Toxicogenomics Database (9), STITCH v5.0 (10), the DrugBank open subset (11), ChEMBL v33 (12), and the complementary patent- and GWAS-driven resources DGIdb v5 (13) and Open Targets 23.12 (14). The most potent measurement per drug–gene dyad was retained and duplicate HGNC symbols were collapsed into a single non-redundant list. The resultant sets were imported into a Python 3.13/Miniforge environment, intersected by straightforward set algebra and rendered through the UpSet algorithm (15), with a putative human background of 20,000 protein-coding genes. Genes present in at least two drugs (N = 37, Supplementary Table S2) were interrogated for functional over-representation against KEGG (16), Reactome (17), and WikiPathways (18) via clusterProfiler 4.10 (19) and cross-validated in g:Profiler2 (20), adopting a Benjamini–Hochberg FDR <0.05 (Supplementary Table S4). The analysis converged on a 16-gene kernel encompassing caspase-mediated apoptosis, MAPK/VEGF-downstream signalling, p53 control, and a sirtuin node (Figure 3, Supplementary Table S5).

To prioritise fertility-sparing candidates, we mapped every drug-specific target list onto the KEGG and Reactome pathways that emerged as significantly enriched in the common-gene analysis (Supplementary Table S2). Gene–drug matrices were constructed in Python 3.13 (pandas 2.2) and interrogated using the GSEApy wrapper for clusterProfiler 4.10 and ReactomePA. In enrichment statistics, we used a one-tailed hyper-geometric test against the 20,000-gene human universe with Benjamini–Hochberg correction (α = 0.05). For each compound, we calculated a “pathway-coverage score”—the proportion of hall-mark genes (N = 16) captured within the apoptosis/caspase, MAPK, PI3K-Akt/mTOR, and VEGF modules (STRINGApp, edge-confidence ≥ 0.7). The drugs were then ranked by coverage score, after which safety descriptors (pregnancy registry data and EMA/FDA labels) were overlaid in R 4.3 using dplyr and ggplot2 to flag regimens compatible with conception (Supplementary Table S6).

A total of 31 studies, six conducted in humans (21–26), six in vitro studies with patient isolated cells (27–32), two conducted in baboons (33, 34), and 17 performed in rodents (4, 35–47, 63, 64) met the eligibility criteria and contributed to the network meta-analysis (Supplementary Figure S1, Supplementary Table S1). Across all models, the primary outcomes were expressed as mean difference (MD) in absolute lesion area (mm²), lesion number reduction (%), VEGF expression reduction (%), vs. placebo. The secondary outcomes were expressed as microvessel density, apoptosis, and proliferation reduction (−%) or increase (−%). The results were reported first for the complete network (“all organisms,” Figures 1A,C,E, 2A,B,D) and then for the restricted network that excluded rodent data to probe inter-species robustness (Figures 1B,D,F, 2C). Risk of bias was assessed in accordance with human-based studies through the Robins-I framework (Supplementary Table S2), with an overall moderate risk of bias, and in accordance with an animal-based study using the SYRCLE checklist, with an overall fair quality of included studies (Supplementary Table S4). Pooled SUCRA values placed bentamapimod, ganirelix, cetrorelix, curcumin, and sorafenib in the top quintile of efficacy, whereas antioxidant vitamins and non-selective NSAIDs clustered near placebo (Supplementary Figure S4); individual study results were analysed to correlate study-level outcomes (direct) with NMA-level cumulative (mixed) outcomes (Supplementary Figure S5).

In an exploratory in silico analysis based on curated drug–gene interaction databases, the Venn–STRING composite suggested a star-shaped overlap centred on CASP3, with 22–36 peripheral genes per drug (Figures 3A,B, Supplementary Tables S4, S5). Bentamapimod, curcumin, and sorafenib shared targets in MAPK1/3, RAF1, and AKT1, which is consistent with their observed anti-proliferative signals in vivo. Ganirelix and cetrorelix contributed a GNRHR–CASP3–MAPK1 sub-module, while anti-VEGF antibodies selectively docked VEGFA. Pathway enrichment mapped these intersections to MAPK signalling, PI3K-Akt/mTOR, apoptotic caspases, and VEGF-driven angiogenesis. These findings should be interpreted as hypothesis-generating rather than as mechanistic proof.

Among agents with lesion regression signals, curcumin and quinagolide had the broadest coverage of the 16-gene hub without direct mTOR inhibition or systemic endocrine suppression (Figure 3A, Supplementary Table S5). Tacrolimus, when modelled alongside curcumin in this framework, theoretically expanded pathway coverage, while avoiding some of the gonadotoxicity associated with sirolimus. We further compared fertility-sparing vs. non-sparing agents against the drug–gene overlapping pathway to detect putatively clinically meaningful, pharmaceutical targets (Supplementary Table S6). With regard to fertility-sparing agents, curcumin ranked first (pathway coverage 87%), while sorafenib ranked first from the fertility non-sparing agents (pathway coverage 38%). These patterns are purely in silico and were presented only to prioritise combinations for further pre-clinical evaluation. Therefore, they do not constitute evidence that such regimens would be effective or safe in humans.

Combinatorial therapy appears a promising option, although no current studies have explored this therapeutic avenue. Add-on low-exposure tacrolimus would supply NFAT-cytokine blockade that curcumin lacks, covering the remaining two untargeted genes in the hub without imposing teratogenic potential (53, 54). Alternatively, quinagolide may present a second, endocrine-neutral partner that down-regulates VEGF transcription via dopaminergic signalling. Importantly, future studies should build on the observations of the present analysis in a staged manner. Initial work will be required to optimise dosing, formulation, and delivery of candidate agents such as curcumin and also in relevant animal models, including a rigorous assessment of reproductive and embryonic safety. Only after reproducible lesion-level and mechanistic effects are demonstrated in multiple species should early-phase human trials be contemplated. If such trials are pursued, they should prioritise patient-centred outcomes such as pain, quality of life, and fertility as well as long-term safety, rather than relying solely on anatomical endpoints.

This synthesis is constrained by the biological and methodological limitations of the underlying evidence. Fifty-eight per cent of the pooled data were derived from rodent models, which do not recapitulate human stromal–immune crosstalk and typically rely on ectopic transplantation under high-oestrogen conditions. Resultant lesions evolve over weeks rather than years and may not capture the chronic inflammatory, fibrotic, and neuropathic features of human disease. Even among non-rodent and human-derived models, there was substantial heterogeneity in lesion induction, dosing schedules, and outcome definitions. Most experiments were short term (<12 weeks) and focused on surrogate anatomical endpoints, with almost no data on nociceptive behaviour, quality of life, fertility, or long-term safety. These species- and experiment-level constraints mean that our effect estimates should be viewed as preliminary indicators of biological activity rather than as predictors of clinical benefit. Risk-of-bias appraisal was frequently non-trivial, with only about one-third of animal studies clearly reporting randomisation, allocation concealment, or blinded outcome assessment, and non-randomised human cohorts were prone to residual confounding, including by disease stage. Several network nodes were informed by few and often small studies, yielding wide and overlapping confidence intervals. In this context, the SUCRA-based hierarchy reflected relative rankings within a sparse and heterogeneous dataset and should be interpreted as a heuristic to prioritise candidates for further investigation and not as evidence of inherent biological superiority or clinical precision. Comparison-adjusted funnel plots were compatible with possible small-study and publication bias, particularly for lesser-studied agents, although the limited number of trials per comparison undermined formal detection, and dose–response data were too sparse to support modelling of optimal human dosing. Finally, our drug–gene intersection and pathway-enrichment analyses were strictly in silico overlays that depended on incomplete, curated interaction databases. Therefore, they could not be used to establish causality nor to confirm target engagement at clinically achievable exposures, and their role is limited to prioritising mechanistic hypotheses for future pre-clinical and translational work.

Despite these limitations, our meta-network integrated animal, organoid, and first-in-human data into a hypothesis-driven, efficacy analysis yet for anti-angiogenic drugs in endometriosis treatment. Rigorous duplicate screening, full bias scoring, and design-by-treatment tests supported the identification of putative therapeutic targets. Furthermore, the present NMA classified higher-ranking compounds into drug–gene maps to explore biologically relevant and clinically plausible recommendations for future research.