Mendelian randomization in ovarian research: applications in PCOS, ovarian aging, and assisted reproductive technology.

To address this challenge, we outline a framework centered on two complementary analytical strategies, both underpinned by methodological rigor. Figure  2 summarizes the overarching logic. Fig. 2 A strategic framework for mendelian randomization in ART. This schematic presents a strategic framework synthesized to address the ART phenotype gap. The long-term solution (top pathway) involves establishing dedicated ART genomic consortia with standardized deep phenotyping and ethical governance. Interim analytical strategies (bottom pathways) include upstream exposure analysis (green) and proxy outcome analysis (orange). Both approaches necessitate rigorous MR methodologies (gray box), including multivariable MR, colocalization, CAUSE, and comprehensive sensitivity testing. Arrows illustrate how these complementary strategies collectively target the phenotype gap from distinct analytical directions A strategic framework for mendelian randomization in ART. This schematic presents a strategic framework synthesized to address the ART phenotype gap. The long-term solution (top pathway) involves establishing dedicated ART genomic consortia with standardized deep phenotyping and ethical governance. Interim analytical strategies (bottom pathways) include upstream exposure analysis (green) and proxy outcome analysis (orange). Both approaches necessitate rigorous MR methodologies (gray box), including multivariable MR, colocalization, CAUSE, and comprehensive sensitivity testing. Arrows illustrate how these complementary strategies collectively target the phenotype gap from distinct analytical directions The solution to the phenotype gap lies in the establishment of large-scale, deeply phenotyped ART genomic consortia. This goal will require international collaboration and funding. However, such infrastructure remains a long-term goal rather than an immediate solution. Common data elements must be adopted for embryology, clinical outcomes, and biospecimen collection. These should build on frameworks such as the Vienna Consensus and models like the Endometriosis Phenome Harmonization Project [ 43 ]. This effort must be accompanied by governance structures that protect participant privacy. Ethical and data governance infrastructure must be established as a priority, adapted from practices in sensitive data contexts [ 44 – 47 ]. Reproductive data carry sensitivities. They encompass embryological assessments, pregnancy outcomes including miscarriage and termination, and child health information. Such information may reveal paternity, genetic relatedness, or carrier status for heritable conditions. International consortia must develop governance frameworks that reconcile two competing demands. One is the open science principle that data should be findable, accessible, interoperable, and reusable (FAIR). The other is the privacy requirements of reproductive health data, particularly when linking parental and offspring genotypes. Cross-border data transfers must comply with regulatory regimes. For instance, the European Union’s General Data Protection Regulation (GDPR) imposes requirements that may conflict with data access norms in other jurisdictions. Consortia should accordingly adopt federated analysis models. In these models, sensitive individual-level data remain within national boundaries and only summary-level results are shared [ 44 ]. Dynamic consent models are suited to ART research. In this context, a couple’s reproductive journey evolves and their willingness to share embryo or child health data may shift [ 45 ]. Protocols for managing incidental findings should be developed in consultation with clinical geneticists, reproductive endocrinologists, and patient advocacy groups. These findings include genetic variants with reproductive implications, such as Y-chromosome microdeletions, FMR1 premutations, and karyotype abnormalities. Gamete donor data governance presents challenges. The increasing use of expanded carrier screening in gamete donation [ 46 , 47 ] generates genetic information that straddles the boundary between donor anonymity and recipient child health interests. This requires consent and disclosure policies. Early models such as the Endometriosis Phenome Harmonization Project [ 43 ] and initiatives like GEMINI and IVF-GWAS provide templates that can be adapted. However, the ART field must invest in governance structures that anticipate the ethical complexities inherent in reproductive genomic research. A commitment to FAIR principles will ensure the resulting resource benefits the global scientific community. While these efforts are underway, causal inference in ART must proceed using existing data. The strategies outlined in sect.  " Interim strategy I: focusing on upstream exposures "–" Critical considerations: assumptions, limitations, and a validation framework " therefore represent the primary mode of causal inquiry for the foreseeable future. While consortia are being developed, causal research must proceed in parallel. A productive approach involves shifting MR analyses toward upstream exposures that are clinical risk factors. Studies can interrogate whether genetically predicted higher serum testosterone exerts a causal influence on PCOS risk or intermediate phenotypes such as ovarian AFC [ 7 , 24 ]. A large-scale MR study exemplifies this strategy. It demonstrated that genetically predicted overall and central adiposity causally elevates the risk of PCOS, uterine fibroids, and pre-eclampsia [ 41 ]. The study discussed visceral fat and insulin resistance as plausible mediators, highlighting the need for mediation MR in future studies. Such findings bolster the biological rationale for targeting these exposures in ART populations. For instance, pre-treatment anti-androgen therapy could improve ART outcomes, even when the final endpoints cannot yet be modeled genetically. Recent MR evidence illustrates the validity of this approach. The evidence confirms that genetically predicted elevated fasting insulin, adiposity, and dyslipidemia are causal drivers of PCOS risk [ 23 ]. These findings were synthesized in a recent perspective in Cell Metabolism. They demonstrate that causal inference on upstream metabolic components can yield biological insights and prioritize therapeutic targets, such as insulin sensitizers and weight management [ 23 ]. This serves as a proof-of-concept. Upstream exposure analysis represents a promising pathway to derive causal insights for complex reproductive conditions. Another interim strategy uses genetically correlated proxy outcomes. Researchers can use well-powered GWAS for heritable traits. These traits are genetically correlated with ART success. Reproductive lifespan and age at first birth are two examples. Reproductive lifespan is defined as the time from menarche to menopause. Both are shaped by ovarian reserve and fecundity [ 27 ]. Recent MR evidence demonstrates that genetically predicted older age at first birth and higher sex hormone-binding globulin levels are associated with increased longevity, while pre-eclampsia liability decreases lifespan [ 28 ]. These findings strengthen the biological plausibility of using reproductive timing traits as proxies for health outcomes relevant to ART populations. Population-based live birth data deserve attention among the available proxies. Large-scale biobanks such as the UK Biobank and FinnGen provide GWAS summary statistics for traits like number of live births or probability of ever having children. These phenotypes serve as integrated measures of natural fecundity. They reflect the function of the reproductive system, spanning gamete quality, hormonal environment, embryo viability, and uterine receptivity. An MR analysis may show that a genetically predicted exposure, such as higher BMI, causally reduces live birth likelihood in the general population. This provides indirect evidence that the same exposure may also hinder ART success. The rationale is that if a factor disrupts the biological processes necessary for natural conception, it will likely also impair those same processes in the IVF setting. For example, a recent study demonstrated that genetically predicted fasting insulin levels were causally associated with risk of female infertility [ 26 ]. This result reinforces the hypothesis that insulin resistance is a causal contributor to subfertility. It highlights interventions such as metformin or lifestyle changes as priorities for study in ART populations, even without a direct GWAS on IVF live birth. This strategy cannot replace the need for ART-specific phenotypic data. However, it provides an immediate and biologically relevant means of screening for causal relationships. Researchers can use existing genetic datasets to identify which modifiable risk factors are most promising for clinical intervention. Researchers seeking to implement proxy outcome MR can access publicly available GWAS summary data through the OpenGWAS platform. From an ART perspective, relevant proxies include age at menopause (a marker of ovarian aging; GWAS ID ebi-a-GCST90029037, N  = 156,364) [ 27 ], age at first birth (a marker of reproductive timing; GWAS ID ebi-a-GCST90000050, N  = 542,901) [ 48 ], and number of children ever born (a population-level measure of fecundity; GWAS ID ieu-b-5122, N  = 785,604) [available from OpenGWAS]. All exposures and outcomes should be selected based on demonstrable genetic correlation (sect. " Critical considerations: assumptions, limitations, and a validation framework " ) and biological plausibility. These GWAS identifiers are provided to lower the entry barrier for reproductive researchers new to MR analysis. Supplementary Table S2 provides a reference table of these and additional GWAS datasets. The table includes sample sizes, ancestry backgrounds, genetic correlation evidence where available, and guidance for their use as proxies. Researchers should match the ancestry background of exposure and outcome GWAS wherever possible. They should verify instrument strength (F-statistic > 10) for each exposure-outcome pair before proceeding to analysis. They should also consult established power calculation tools, such as the mRnd web application, to confirm adequate statistical power for the planned MR design [ 49 ]. While the proxy outcomes discussed above are accessible through public GWAS, their relevance to ovarian research requires explicit mapping. Below we map available proxies to specific ovarian phenotypes of interest. While the GWAS data presented in sect. " Publicly available GWAS data for proxy outcome MR " are introduced from an ART perspective, they are valuable for ovarian research. Key ovarian phenotypes that can be proxied by these data include ovarian reserve, ovarian response, and ovarian aging. Age at menopause (GWAS ID ebi-a-GCST90029037) serves as a proxy for ovarian reserve. Its genetic correlations with AMH levels and AFC have been confirmed [ 27 , 50 ]. Researchers investigating determinants of ovarian reserve can use this proxy to generate indirect evidence when GWAS on AMH or AFC are unavailable. Number of children ever born (GWAS ID ieu-b-5122) can proxy for oocyte yield in studies of ovarian response. This measure reflects cumulative fertility potential. However, socio-behavioral confounding such as educational attainment must be considered. Where possible, it should be adjusted for using multivariable MR [ 51 ]. Age at first birth (GWAS ID ebi-a-GCST90000050) is genetically correlated with reproductive lifespan. It can inform studies of premature ovarian insufficiency [ 28 ]. Its genetic correlation with ovarian aging pathways makes it a useful proxy. However, attention to socio-behavioral pleiotropy is again required. Researchers investigating ovarian biology should prioritize these proxies based on demonstrable genetic correlation (sect. " Critical considerations: assumptions, limitations, and a validation framework ") and biological plausibility. For example, an MR analysis testing whether fasting insulin affects ovarian reserve would use age at menopause as the outcome, as illustrated conceptually in sect. " Publicly available GWAS data for proxy outcome MR ". This approach generates hypotheses about ovarian function that can later be validated in ART-specific cohorts. As with all proxy-based analyses, findings should be framed as supportive evidence requiring future confirmation using direct ovarian phenotypes. The proxy approach depends on two criteria. We refer to these as the shared pathway conditions. One condition is a demonstrably significant genetic correlation between proxy and target ART biology, typically established via LD score regression and yielding a statistically significant rg [ 52 ]. Another condition is a plausible shared biological pathway linking the proxy to the ART outcome. These two conditions, while conceptually straightforward, can fail in ways that are not immediately apparent. One failure mode involves distinct tissue-specific regulatory mechanisms. The proxy and target outcomes may be influenced by different pathways even when both are linked to ovarian function. Age at menopause is determined by ovarian follicular depletion governed by DNA damage response pathways in oocytes [ 53 ]. In contrast, blastocyst formation rate, an ART endpoint, is influenced by oocyte-intrinsic factors, granulosa cell function, cumulus cell metabolism, and the intra-follicular microenvironment. A genetic instrument for fasting insulin may affect age at menopause through hypothalamic-pituitary-ovarian axis modulation, yet affect blastocyst formation through granulosa cell insulin signaling. In this case, the shared pathway assumption is violated and the proxy-based estimate cannot be transported to the ART endpoint. Another failure mode is socio-behavioral pleiotropy, which can create spurious proxy–target correlations. Age at first birth is genetically correlated with educational attainment, risk-taking behavior, and socioeconomic position [ 48 , 51 ]. An MR analysis using age at first birth as a proxy for natural fecundity may therefore capture genetic effects operating through educational and career pathways rather than through reproductive biological mechanisms. Interpreting such effects as evidence of impaired ART success would conflate socially mediated behaviors with biological fertility. A further failure mode is temporal discordance between proxy and target outcomes. Age at menopause is typically measured decades after the ART window. Genetic variants that accelerate ovarian aging may show different effect magnitudes on these temporally distinct phenotypes if compensatory mechanisms or age-dependent modifiers are operative. Given these failure modes, researchers should validate proxy outcomes before interpreting their findings. A significant genetic correlation via LD score regression [ 52 ] should serve as an initial screening criterion. However, this is necessary but not sufficient for validity. A weak or non-significant genetic correlation should disqualify the proxy unless an explicit biological rationale is provided. Colocalization analysis using the coloc R package [ 54 ] can test whether the genetic signals for exposure and proxy outcome share a common causal variant at lead loci. A posterior probability of shared causal variant (PP.H4) above 0.8 provides supportive evidence of shared genetic architecture. Mediation MR can also be employed where GWAS data exist for intermediate phenotypes on the hypothesized causal pathway. For example, AMH may serve as a potential mediator of the insulin–ovarian aging relationship. Mediation MR can test whether the intermediate explains a significant proportion of the exposure–proxy effect, thereby strengthening the plausibility of pathway overlap. Finally, all proxy-based MR findings should be triangulated with independent lines of evidence. These include observational data, experimental findings from model systems, and clinical trial results where available. Convergence across methodological approaches increases confidence that the estimate reflects biology rather than an artifact of proxy selection. Without such convergence, causal claims should be suspended until direct phenotypic data become available. To guide researchers through these requirements, we present a practical decision framework (Fig.  3 ). This framework outlines a pathway that begins with defining a target ART outcome. It then proceeds through proxy screening and validation of genetic correlation and biological plausibility. Finally, it reaches MR analysis and interpretation, with explicit consideration of limitations. Fig. 3 A decision framework for validating and interpreting proxy outcomes in ART MR studies. This stepwise flowchart guides researchers through proxy screening, validation (genetic correlation, colocalization, and biological pathway assessment), MR analysis with sensitivity testing, and interpretation with explicit limitation statements. The final step requires stating the counterfactual estimand and acknowledging that inference applies to shared biological pathways rather than directly to specific ART endpoints A decision framework for validating and interpreting proxy outcomes in ART MR studies. This stepwise flowchart guides researchers through proxy screening, validation (genetic correlation, colocalization, and biological pathway assessment), MR analysis with sensitivity testing, and interpretation with explicit limitation statements. The final step requires stating the counterfactual estimand and acknowledging that inference applies to shared biological pathways rather than directly to specific ART endpoints Ultimately, MR analyses relying on proxy outcomes yield indirect evidence. Findings should be framed as supportive causal evidence for a hypothesis. They should not be treated as definitive proof. All causal inferences from proxy-based MR remain provisional. They require future validation through large-scale GWAS on direct ART phenotypes. Mendelian randomization is a counterfactual framework. It aims to estimate what would happen to an outcome if an exposure were intervened upon [ 12 ]. When a proxy outcome is used, the counterfactual being estimated becomes ambiguous. Consider a two-sample MR where the exposure (X) is fasting insulin and the outcome (Y) is age at menopause (a proxy for ovarian reserve). The MR estimate answers one counterfactual question: what would be the effect on age at menopause if we intervened on fasting insulin? This differs from the clinically relevant question: what would be the effect on blastocyst formation rate if we intervened on fasting insulin? The gap between these two counterfactuals is bridged by an untested assumption. This assumption holds that the biological pathways from X to Y (the proxy) are identical to those from X to Z (the target ART outcome). We refer to this as the shared pathway assumption. The assumption can fail in practice. For example, the exposure may affect the proxy and target outcomes through distinct mechanisms. When this occurs, the proxy MR estimate cannot be transported to the target outcome. Researchers using proxy outcomes should therefore state two things explicitly. They should state which counterfactual their analysis estimates. They should also state what additional assumptions are required to transport that estimate to the clinically relevant endpoint. Any MR analysis must follow best practices to avoid spurious conclusions. This applies whether the analysis uses upstream exposures or proxy outcomes. Multivariable MR should be used to estimate the independent effect of an exposure while conditioning on correlated traits. For example, it can isolate the effect of androgens from that of adiposity [ 29 – 32 ]. Sensitivity analyses should test how robust the findings are under different assumptions about pleiotropy. These analyses include MR-Egger, weighted median, and MR-PRESSO [ 33 – 36 ]. As discussed in sect. " Critical considerations: assumptions, limitations, and a validation framework ", colocalization should be employed to assess shared genetic architecture between exposure and outcome [ 33 ]. Recent reviews provide detailed methodological guidance [ 33 – 37 ]. Consistency of estimates across multiple methods strengthens confidence in causal findings. A distinction should be drawn among sensitivity methods. Some methods address uncorrelated horizontal pleiotropy, where individual genetic variants influence the outcome through pathways independent of the exposure. Others also account for correlated horizontal pleiotropy, which arises when variants affect both exposure and outcome through a shared heritable factor. Standard methods such as MR-Egger regression, the weighted median estimator, and MR-PRESSO address uncorrelated pleiotropy. They assume that pleiotropic effects are not systematically shared across variants. However, correlated pleiotropy is a concern in reproductive epidemiology. Traits such as obesity, insulin resistance, and hyperandrogenemia are genetically correlated and tightly clustered in conditions like PCOS. A shared genetic architecture between exposure and outcome may produce a causal signal that reflects confounding by a heritable common factor, rather than a direct causal pathway. The CAUSE method (Causal Analysis Using Summary Effect Estimates) models both correlated and uncorrelated pleiotropic effects within a Bayesian framework. It compares a causal model against a model in which all observed association is attributable to pleiotropy [ 55 ]. Simulation studies have shown that CAUSE generates fewer false positives than alternative methods when correlated pleiotropy is present. This makes it a useful complement to the sensitivity toolbox for MR studies in ovarian and ART research. We recommend that researchers include CAUSE alongside traditional sensitivity analyses. This is particularly important when the exposure and outcome share genetic correlations and when conventional methods such as MR-Egger and the weighted median estimator yield discordant results that may signal underlying correlated pleiotropy. Researchers should attend to several patterns when interpreting sensitivity analyses. These patterns may indicate violations of MR assumptions. A significant MR-Egger intercept ( P  < 0.05) suggests directional pleiotropy. In this case, the primary inverse-variance weighted (IVW) estimate should be interpreted cautiously. Attenuation of the weighted median estimate relative to the IVW estimate may indicate that a subset of invalid instruments is driving the association. The weighted median estimate is more robust to violations by a minority of variants and may therefore be preferred. MR-PRESSO can identify outlier variants. These variants should be examined for pleiotropic associations using resources such as PhenoScanner. Results should be reported both with and without outlier removal. Multivariable MR should be used when the exposure of interest is strongly correlated with a second trait that has a well-powered GWAS. For example, estimating the effect of fasting insulin independent of BMI can disentangle the independent contributions of tightly clustered metabolic and endocrine traits. The same applies to estimating the effect of testosterone independent of adiposity [ 31 ]. CAUSE and conventional sensitivity methods may produce divergent conclusions. For instance, IVW and weighted median estimates may be statistically significant while CAUSE favors the pleiotropic model. In such cases, the possibility of correlated pleiotropy should be considered and causal claims should be tempered. MR is more robust to population stratification than conventional observational epidemiology. Genetic variants are assigned at conception and are generally not confounded by postnatal environmental factors. However, this protection is not absolute. Residual stratification can bias MR estimates when exposure and outcome GWAS are drawn from populations with different ancestry compositions. The same applies when summary statistics include admixed populations. We recommend that researchers prioritize exposure and outcome GWAS from the same or related ancestry groups. Analyses should be restricted to homogeneous ancestry strata when individual-level data are available. Caution should be exercised when combining trans-ancestry GWAS for phenotypes with documented population differences in prevalence or genetic architecture, such as age at menopause and PCOS. Most currently available reproductive GWAS are derived from European-ancestry populations. This limits the generalizability of MR findings to other ancestry groups and underscores the need for diversity in future ART genomic consortia. Researchers working with non-European ancestry populations face smaller GWAS sample sizes for reproductive traits and limited statistical power. We recommend that they prioritize well-powered European-ancestry instruments for exposures with known cross-ancestry genetic similarity, such as BMI and fasting insulin. They should acknowledge the uncertainty introduced by ancestry mismatch. Sensitivity analyses should be conducted on ancestry-homogeneous subsets where individual-level data permit. We recommend the following practices for instrument selection in two-sample MR studies on ovarian and ART outcomes. Instrumental variants should be selected at genome-wide significance ( P  < 5 × 10⁻⁸). They should be subjected to LD clumping ( r ² < 0.001, window 10,000 kb) using a reference panel matched to the GWAS population. The F-statistic should be calculated for each variant and for the combined instrument set to assess instrument strength. The threshold of F > 10 provides a screening benchmark, but it does not guarantee the absence of weak instrument bias. Researchers should be aware that bias increases as the expected F-statistic decreases. This is particularly true when the observed F-statistic in a given dataset exceeds expectation [ 56 ]. Parsimonious genetic models and the avoidance of over-parameterized allele scores can further reduce bias. For binary exposures or outcomes, researchers should verify that the chosen MR method accommodates binary variables appropriately. The causal estimand in such settings may differ from the conditional odds ratio estimated by conventional multivariable regression [ 57 ]. Power calculations should be performed using established tools such as the mRnd web application or the analytical formulae [ 49 ]. Proxy-based MR studies with modest genetic correlation or outcomes with low heritability may require larger sample sizes to achieve adequate power. Detailed step-by-step guidance is available [ 16 ] and the STROBE-MR reporting guidelines [ 58 ]. These resources should be consulted during study planning to ensure methodological transparency and reproducibility.

Despite its rationale, widespread application of MR in ART faces a barrier: the disconnect between large-scale genomic resources and the detailed, cycle-specific phenotyping essential to modern reproductive medicine (Fig.  1 ). Public biobanks such as the UK Biobank and FinnGen have advanced research in medical areas, but from an ART perspective, their phenotypic data remain limited [ 20 , 21 ]. They may include indicators like infertility status or self-reported family size, but they lack the cycle-level data for reproductive endocrinology and embryology. Fig. 1 The ART phenotype gap: a barrier to causal inference. This figure illustrates the disconnect between standard biobank data and deep ART phenotypes that currently prevents direct MR analysis. Left panel: phenotypes widely available in public biobanks. Right panel: detailed ART phenotypes that are absent from public repositories, including those lacking standardized definitions (e.g., AMH). The central ART Phenotype Gap represents the missing link between genetic instruments and clinically relevant ART outcomes, necessitating alternative analytical strategies The ART phenotype gap: a barrier to causal inference. This figure illustrates the disconnect between standard biobank data and deep ART phenotypes that currently prevents direct MR analysis. Left panel: phenotypes widely available in public biobanks. Right panel: detailed ART phenotypes that are absent from public repositories, including those lacking standardized definitions (e.g., AMH). The central ART Phenotype Gap represents the missing link between genetic instruments and clinically relevant ART outcomes, necessitating alternative analytical strategies To quantify this gap, we interrogated the OpenGWAS database ( https://opengwas.io ), a primary resource for applied Mendelian randomization research, as of early 2026. The search revealed a disparity in the availability of genome-wide association study (GWAS) summary data for reproductive traits. For conditions with genetic infrastructure, studies were available: age at menopause (7 studies, maximum N  = 156,364), polycystic ovary syndrome ( N  = 141,355), and endometriosis (16 studies). In contrast, for outcomes central to ART clinical practice and research, the database returned zero results for blastocyst formation rate, cumulative live birth rate per cycle, and ovarian response to gonadotropin. These three outcomes, each relevant to embryology, clinical success, and cycle management, remain uncharacterized at the genome-wide level. This absence reflects a structural deficit in phenotype collection and data sharing specific to the ART field, rather than a lack of sample size or research interest. By contrast, even a conservative search for coronary artery disease returns six GWAS studies in the same database. Coronary artery disease is a phenotype more complex than an IVF cycle outcome, yet it has the genetic infrastructure that has enabled decades of causal inference in cardiology. The ART field remains at a pre-GWAS stage for its clinically relevant outcomes. A deeper dimension of the ART phenotype gap concerns the standardization of phenotype definitions themselves. This challenge extends beyond data availability to the consistency of the measures that would populate future GWAS. This is illustrated by the controversy surrounding AMH as a potential diagnostic biomarker for PCOS. A comprehensive review [ 42 ] demonstrated that AMH correlates strongly with antral follicle count and is consistently elevated in women with PCOS. However, its clinical adoption faces barriers, including significant inter-assay variability, the absence of standardized cut-off values across populations and age groups, and confounding by common comorbidities such as obesity. These unresolved issues have resulted in proposed AMH diagnostic thresholds ranging from 3.5 to 5 ng/mL, with uncertain biological significance and clinical utility. This controversy carries an implication for genomic resource development. If a candidate phenotypic marker such as AMH lacks a consensus definition in clinical practice, GWAS based on inconsistently defined AMH measurements are unlikely to yield reproducible findings when employed as MR outcomes or exposures. This argument aligns with the systematic review [ 22 ] that documented substantial inconsistency in the definition and reporting of pregnancy outcomes across infertility trials. Thus, bridging the ART phenotype gap requires larger sample sizes, richer data dimensions, and a consensus on phenotype standardization. This prerequisite has yet to be addressed by the field. Supplementary Table S1 provides a systematic comparison of available versus absent reproductive GWAS phenotypes. This comparison documents the disparity in genomic data infrastructure between well-characterized reproductive traits and the cycle-specific outcomes central to ART clinical practice. The absent phenotypes are those of clinical and mechanistic relevance. They include embryological parameters such as morphokinetic timings, blastocyst formation rates, and quality grades. They also include cycle efficiency metrics such as gonadotropin dosing and ovarian response patterns, as well as key clinical endpoints like cumulative live birth rate per initiated oocyte retrieval. Cumulative live birth rate is the gold standard measure of ART success. This absence constitutes an obstacle in the causal inference pipeline. Researchers can access genetic instruments for traits like serum testosterone from large-scale GWAS involving hundreds of thousands of individuals [ 24 ]. However, they cannot test the causal effect of these instruments on outcomes such as cumulative live birth failure or low blastocyst conversion rate. No GWAS currently exists for these endpoints. This phenotype gap necessitates a reliance on indirect evidence and demands strategic methodological adaptations. This phenotype gap is consequential for ovarian research. Key ovarian phenotypes, including AMH levels, antral follicle count, and histologically defined ovarian aging, remain uncharacterized at the genome-wide level. Without GWAS for these direct measures of ovarian function, researchers must rely on indirect strategies. The framework proposed here is designed to support causal inference in ovarian biology under these constraints.

This study is a methodological review and conceptual framework synthesis. It does not report primary data or novel causal findings. The objective is to provide an approach for applying Mendelian randomization in assisted reproductive technology research under current data constraints. The strategic framework was synthesized through an iterative process involving three components. We examined recent MR studies in reproductive medicine published between 2018 and 2025, with attention to studies that navigated phenotypic data limitations. Exemplars included studies on PCOS [ 23 , 24 ], female infertility [ 25 , 26 ], and reproductive aging [ 27 , 28 ]. This analysis identified common strategic approaches, including the use of upstream exposures and proxy outcomes, which formed the foundation of our framework. We synthesized contemporary methodological developments in MR, including instrumental variable selection [ 16 ], multivariable MR [ 29 – 32 ] sensitivity analysis [ 33 – 37 ] and colocalization [ 33 ], ensuring that the proposed strategies align with current practices. We also examined cross-disciplinary case studies from cardiology [ 38 , 39 ] and psychiatry [ 40 ] where MR was successfully implemented despite complex phenotype landscapes. These parallels validated that the proposed approach is feasible and offered lessons for avoiding pitfalls. This review does not provide empirical validation of the proposed framework, and the strategies described should be regarded as conceptual proposals rather than established protocols. All strategies require case-by-case assessment in future studies, and their practical utility remains to be demonstrated. The framework is focused on female ART outcomes; its extension to male infertility, discussed in sect.  " Learning from parallel frontiers ", follows the same conceptual logic.

The phenotype gap is not unique to ART. Several other fields have confronted similar disconnects between genetic instruments and clinical endpoints and have navigated them successfully. Early MR studies established the causal role of LDL-cholesterol in coronary artery disease using widely accessible serum measurements [ 38 ]. This evidence supported targeted interventions, such as statins, before genetic instruments for endpoints like plaque vulnerability were developed. Researchers then adopted proxy outcomes. For example, atrial fibrillation, a defined electrical endpoint, served as a proxy for stroke risk in MR studies investigating causal factors such as blood pressure and obesity [ 39 ]. The lesson for ART is to construct causal chains incrementally. This process should begin with upstream exposures and proxies, and build a rationale for deep phenotyping of endpoints. Psychiatric diagnoses such as depression are heterogeneous syndromes rather than distinct biological entities. MR has played a key role in unpacking these complex categories. It does so by turning to endophenotypes, which are measurable and heritable traits. These traits are closer to underlying biological pathways than clinical diagnoses. For example, a recent MR study tested the causal effects of elevated blood pressure on 3,935 brain imaging-derived phenotypes. The study linked these phenotypes to cognitive function and pinpointed white matter structures that mediate vascular-related cognitive impairment [ 40 ]. A similar approach could be applied in ART. Parameters such as embryological morphokinetics or follicular fluid proteomic profiles could serve as endophenotypes. They would represent the outcome of live birth while being practical to assess at scale. These parallels outline a pathway. Causal discovery in ART should be seen as an iterative process. It begins with indirect evidence from upstream exposures and proxy outcomes. It then refines toward direct causal estimates as deeply phenotyped data become available. The framework presented here supports this iteration. Each MR study on an upstream factor or proxy outcome builds a foundation for future investigations. Collaboration is essential. Landmark GWAS in cardiology and psychiatry succeeded because international consortia established shared phenotyping standards. Initiatives like GEMINI and other ART projects should follow this model [ 59 , 60 ]. This review focuses on female reproductive health and ART outcomes. However, the core methodological challenges and strategic solutions described here apply with equal force to male infertility. Male infertility contributes genetic influences that extend beyond fertilization success. These influences affect embryo development, implantation potential, and offspring health. Male reproductive genetics confronts a phenotype gap that mirrors the one described for female ART. Public biobanks lack GWAS for detailed semen parameters, sperm DNA fragmentation indices, testicular histopathology phenotypes, and ART-specific male outcomes. Examples of such outcomes include fertilization rate per oocyte and blastocyst formation rate stratified by paternal genotype [ 61 ]. The two interim strategies proposed in this framework are transferable. Upstream exposure analysis can interrogate whether genetically predicted metabolic traits causally affect male fertility endpoints for which GWAS exist. These endpoints include diagnosed male infertility and reproductive hormone levels [ 62 ]. Proxy outcome analysis can leverage genetically correlated traits, such as number of children fathered in population-based biobanks, as integrated measures of male fecundity. Adjustment for socio-behavioral pleiotropy is needed. The validation framework described above includes genetic correlation assessment, colocalization, and uncertainty acknowledgment. It applies without modification. An understanding of ART success at the couple level will require phenotyped genomic resources for both partners. We encourage parallel methodological development in male reproductive genetics alongside the female-focused infrastructure advocated here. Several limitations must be acknowledged when interpreting this framework. The interim strategies are indirect. They rely on assumptions of genetic correlation and shared biological pathways. Findings should be considered supportive evidence rather than proof. In addition, this framework has not been empirically validated and remains a conceptual proposal. The proxy outcome strategy has not been systematically benchmarked against direct ART phenotypes, and its sensitivity and specificity for detecting true causal relationships in reproductive settings are unknown. The long-term solution of building deeply phenotyped ART genomic consortia faces practical obstacles, including the cost and logistics of standardizing phenotype collection across international sites, the challenge of sustaining long-term funding, and unresolved questions regarding data governance across jurisdictions with differing regulatory frameworks. Finally, as noted in sect. " Methodological rigor: a prerequisite for credible inference ", the generalizability of the framework beyond European-ancestry populations is limited by the current lack of diverse reproductive GWAS data. This review focuses on female reproductive health and ART outcomes. However, the core methodological challenge, the phenotype gap, and the strategic analytical logic apply equally to male infertility. Male infertility accounts for a substantial proportion of couple subfertility. It also confronts an analogous scarcity of integrated genomic and phenotypic data [ 61 , 62 ]. As advocated in the preceding discussion, we encourage parallel methodological efforts in male reproductive genetics. Ultimately, achieving couple-level causal inference will depend on the parallel development of deeply phenotyped genomic resources for both sexes. These parallel trajectories hold direct lessons for ovarian research. Cardiology moved from serum lipids to plaque phenotypes. Psychiatry moved from diagnostic categories to neuroimaging endophenotypes. Ovarian research must now move from crude proxies toward direct biological measures. The framework presented here is designed to support this progression.

The ART phenotype gap is not an insurmountable barrier for ovarian research. Progress toward causal understanding may be achievable today. Researchers must adhere to proxy validation, sensitivity analysis, and acknowledgment of remaining uncertainties. This review outlines a framework for generating provisional causal evidence within existing constraints. In the near term, researchers can implement indirect approaches. They can investigate upstream exposures and employ validated proxy outcomes. These approaches should be supported by multivariable MR, colocalization, and sensitivity testing. In parallel, the field must build toward the long-term solution. This solution involves large-scale, deeply phenotyped ART genomic consortia with standardized data elements and ethical governance. The translational power of this framework lies in its ability to prioritize and design clinical studies. Emerging MR evidence is generating testable hypotheses. For instance, genetically predicted higher testosterone may causally increase PCOS risk. This implicates androgens in its pathophysiology and supports investigation of anti-androgen therapies in ART contexts [ 24 ]. Multivariable MR has disentangled the effects of correlated risk factors. It has distinguished the impact of obesity from that of insulin resistance on reproductive outcomes [ 41 ]. This causal understanding offers insights for designing stratified or combination intervention studies. Achieving this vision requires coordinated efforts across multiple levels. Researchers should implement and refine the MR strategies outlined here. Professional societies, funding agencies, and research networks must prioritize the development of phenotyped ART genomic resources. They should champion data standardization and international collaboration. Journals and peer reviewers can encourage this paradigm shift. They can promote methodologically sound causal inference studies. Clinicians and patient advocacy groups must be engaged. Their involvement will ensure that research priorities align with clinical needs. Progress must be pursued on both strategic fronts. Researchers should leverage methodological innovation to use existing data. At the same time, infrastructure must be built to support long-term research. This dual-track strategy has the potential to generate causal evidence that may inform clinical practice and trial design. It establishes a foundation for the advancement of mechanism-driven, personalized reproductive medicine [ 59 – 65 ]. The ovarian research community should prioritize the collection of deeply phenotyped ovarian data in large-scale cohorts and consortia. This will enable the transition from indirect to direct causal inference.

The clinical trajectory of assisted reproductive technology (ART) is influenced by a complex interplay of patient physiology, pharmacological stimulation, laboratory precision, and embryonic potential [ 1 , 2 ]. Decades of observational research have identified factors associated with success or failure, including endocrine profiles, metabolic health, and follicular fluid biomarkers [ 3 – 5 ]. However, a question remains: which of these associations represent causal relationships? Distinguishing causality from correlation is essential for identifying therapeutic targets, optimizing research resources, and advancing toward mechanism-driven personalized care [ 6 ]. This challenge is acute in polycystic ovary syndrome (PCOS) [ 7 , 8 ]. Patients with PCOS present with a cluster of tightly linked traits: hyperandrogenemia, insulin resistance, and obesity [ 9 ]. When such a patient experiences embryo development or implantation failure, identifying the pathogenic driver, whether high testosterone, metabolic dysfunction, excess weight, or a combination, is difficult. Standard observational studies cannot provide answers due to residual confounding and reverse causation. Mendelian randomization (MR) offers a way forward [ 10 , 11 ]. MR uses genetic variants as instrumental variables for modifiable exposures [ 12 ]. Because these variants are assigned at conception and fixed throughout life, they are not affected by disease processes or environmental confounders, mimicking the design of a randomized controlled trial [ 13 – 16 ]. This framework has advanced causal inference in fields from cardiology to psychiatry [ 17 , 18 ]. Yet its application in ART remains limited [ 19 ]. The lag arises from a structural bottleneck in data availability. Public biobanks such as the UK Biobank and FinnGen lack the cycle-level phenotypes for reproductive medicine: embryological parameters (morphokinetic timings, blastocyst formation rates), cycle efficiency metrics (gonadotropin dosing, ovarian response), and critical clinical endpoints (cumulative live birth rate) [ 20 , 21 ]. We term this the ART phenotype gap: the shortage of large-scale GWAS for cycle-specific, embryological, and cumulative reproductive outcomes. This gap extends beyond a logistical shortfall; it is rooted in the inconsistent definition and reporting of outcomes in infertility research [ 22 ]. This review confronts this gap directly. We argue that it should be viewed as a call for strategic research, rather than an insurmountable barrier. We present a strategic framework that aims to support causal inference now while planning the infrastructure needed for the future. We begin by providing a guide for conducting MR studies using existing genetic data. We then outline the architecture of future deeply phenotyped resources necessary to achieve causal understanding.

Supplementary Material 1. Supplementary Material 1.