MR has experienced explosive growth in research output over the past decade. As shown in Fig. 1, the number of MR-related publications from 2015 to 2024 has risen exponentially, reflecting the increasing adoption of this method in epidemiological and genetic studies [13]. This surge underscores the growing recognition of MR as a powerful tool for causal inference.

Fig. 1

Annual number of Mendelian Randomization (MR) and tumor-related MR publications from 2015 to 2024

The plot illustrates the yearly publication counts of all MR studies (in red) and those specifically related to cancer (in blue). Both categories exhibit a steady upward trend, with total MR publications increasing sharply after 2018 and peaking in 2024. Tumor-related MR studies have also grown substantially, reflecting the expanding adoption of MR methodology in oncology research.

In recent years, a substantial proportion of Mendelian Randomization (MR) publications have originated from China. This surge may be partly attributed to the Chinese translation and dissemination of the influential book Mendelian Randomization: Methods for Using Genetic Variants in Causal Estimation, which has significantly raised awareness and interest in MR methodologies among Chinese researchers. The growing accessibility of MR concepts through translated materials has likely contributed to its widespread adoption across various biomedical fields in China [4].

Several factors contribute to this rapid expansion. One primary driver is the growing availability of large-scale genetic and health datasets, which provide fertile ground for MR analyses [7]. Additionally, advancements in analytical tools and computational methods, particularly the development of two-sample MR techniques, have significantly lowered the technical barrier for conducting MR studies, making the method more accessible to researchers [1]. Artificial intelligence (AI) and machine learning are beginning to influence MR research in both positive and problematic ways. On one hand, AI-based tools can automate and enhance MR analyses. For example, new algorithms like MRAID (MR with Automated Instrument Determination) use machine learning to automatically select and fine-map instrumental SNPs, improving power and accounting for pleiotropy​[5]. Such tools streamline data processing and might increase analysis accuracy by handling complexities (correlated instruments, non-linear effects) that traditional methods struggle with. AI can also help sift through huge biobank datasets to identify potential causal relationships more efficiently than classicalapproaches.On the other hand, there are concerns that AI-driven ease of analysis may exacerbate the flood of MR studies, including many of questionable quality. Commentators have noted an “overflow” of simplistic two-sample MR papers in recent years[28]. With user-friendly software (often powered by built-in algorithms and large public GWAS databases), researchers can now run MR with minimal effort—potentially leading to MR study proliferation without sufficient critical evaluation. Indeed, a decade ago only ~ 100 MR papers per year were published, whereas in 2023 there were over 3000 [28]. This surge, partly enabled by automated analysis pipelines, has raised concerns about many results being generated (or even mass-produced) without proper scrutiny of assumptions [12]. In summary, AI can boost MR research quality by handling big data and complex analyses, but it can also lower the barrier to performing MR, contributing to an abundance of studies (some low-quality). The net impact of AI thus depends on its responsible use: ideally as a means to augment rigor (through better instrument selection, sensitivity analyses, etc.) rather than just to amplify quantity.

Despite the widespread adoption of two-sample MR enabled by summary statistics, concerns have arisen regarding methodological rigor. In particular, weak instrument bias—when genetic variants are only weakly associated with the exposure—can lead to inflated standard errors and biased causal estimates. Horizontal pleiotropy, where genetic variants affect the outcome through pathways unrelated to the exposure, also violates key assumptions and undermines validity. These issues highlight the need for critical evaluation and appropriate analytical strategies in MR studies. [24].

To address these challenges and improve the robustness of findings, researchers have developed several methodological refinements in recent years:

i.

Development of robust analytical techniques

To address key methodological challenges such as horizontal pleiotropy and weak instrument bias, a suite of robust analytical techniques has emerged and matured in recent years. These innovations have significantly enhanced the reliability of MR results and broadened its applicability(See Tables 1A, B and for summary of platforms and tools).

Table 1 A genetic association and MR platforms, B. Functional annotation and systems tools

Sensitivity-Robust Estimators

Several estimators now address violations of instrumental variable assumptions. MR-Egger regression accounts for directional pleiotropy, while weighted median and mode-based estimators remain valid even when up to 50% of instruments are invalid [22]. MR-PRESSO detects and corrects outliers, and the RAPS (Robust Adjusted Profile Score) method improves estimates in the presence of weak instruments.

Multivariable and Mediation MR

Beyond univariable MR, multivariable MR (MVMR) allows researchers to adjust for pleiotropic pathways and estimate direct effects of exposures[10]. For instance, MVMR is used to disentangle the effects of childhood versus adult BMI[24]. Mediation MR is increasingly applied to explore indirect effects through intermediate phenotypes, which is especially useful in cancer pathway research[12] [19].

AI-Driven Tools and Omics Integration

Recent developments have integrated MR with artificial intelligence and single-cell omics. Algorithms like MRAID automate instrument selection using machine learning[32], and deep learning frameworks assist in fine-mapping causal SNPs. Other studies combine MR with single-cell transcriptomics to identify cell-type-specific effects[9].

These advances reflect a shift in MR—from a niche causal inference tool to a versatile, high-dimensional framework in systems epidemiology.

ii.

Integration of high-throughput data sources

The emergence of large-scale biobanks and computational platforms has transformed MR research. Resources such as MR-Base and EpiGraphDB facilitate high-throughput MR analysis across thousands of traits, enabling researchers to rapidly test multiple exposure-outcome relationships [7].

Phenome-wide MR (PheWAS-MR): This approach systematically evaluates the impact of a genetic instrument across numerous health outcomes, helping to identify potential pleiotropic effects and unintended associations [26].

iii.

Improving Study Design and Instrument Selection

Biological validation of instruments

In recent MR research, there is a growing emphasis on selecting functionally validated genetic variants rather than relying solely on genome-wide significance thresholds. This shift reflects a concern that statistically strong instruments may not always reflect true biological mechanisms, thereby violating core instrumental variable assumptions.

A notable example is the PNPLA3 I148M variant, commonly used in MR studies investigating nonalcoholic fatty liver disease (NAFLD). Functional experiments have confirmed its causal role: knockout mouse models demonstrated that deletion of Pnpla3 alters hepatic lipid metabolism and fibrosis [17], thus validating its biological relevance as an instrument.

In oncology, similar functional confirmation is emerging. For instance, variants in the TERT gene, frequently used as instruments in cancer MR studies, have been shown to directly influence telomere maintenance and tumorigenesis. TERT promoter mutations are functionally active in cancer cell lines and enhance transcriptional activity, linking them causally to various malignancies, including glioma and bladder cancer [30, 30].

Another example involves rs6983267 at 8q24, a well-known cancer susceptibility locus. Functional assays revealed that this SNP lies in a MYC enhancer region, altering Wnt signaling and promoting tumorigenesis in colorectal cancer models[20].

These cases underscore the importance of biological annotation and experimental validation in MR. Integrating tools like GTEx for eQTL analysis, CRISPR screening data, and functional genomics repositories helps ensure that instruments are not only statistically robust but mechanistically credible—an essential step to enhance the interpretability and trustworthiness of MR findings, particularly in cancer research.

Sibling comparisons and within-family MR

Within-family Mendelian randomization (WF-MR), including sibling comparison designs, has emerged as a powerful strategy to address key biases that traditional MR cannot fully eliminate—particularly population stratification, dynastic effects, and assortative mating. Unlike classic MR that analyzes unrelated individuals, WF-MR restricts comparisons to genetically related individuals (typically siblings), who share similar family environments and genetic backgrounds.

This design helps eliminate confounding arising from differences across families. For example, dynastic effects—where parental genotypes influence offspring outcomes through environmental pathways (e.g., educational investment or health-related behavior)—can bias classic MR estimates. WF-MR minimizes this by holding family-level environments constant, isolating the effect of within-family genotype variation on outcomes.

A key illustration comes from the landmark study by Brumpton et al. (2020), which systematically compared classic MR and within-family MR across several traits in the HUNT and UK Biobank cohorts [18]. They showed that traditional MR often overestimated effects (e.g., the association between BMI and education) due to residual confounding, while within-family estimates were more conservative but arguably more accurate.

However, WF-MR comes with trade-offs. It suffers from reduced statistical power, since within-family genotype differences are smaller than population-wide differences. Moreover, it requires family-structured data, which is less readily available and may limit generalizability.

Despite these limitations, WF-MR represents a critical advancement in causal inference, particularly in socially sensitive exposures (e.g., education, income, lifestyle behaviors) or outcomes influenced by familial environments. In cancer epidemiology, this approach remains underutilized but holds promise, especially when combined with rich biobank resources such as the Norwegian HUNT Study, FinnTwin, or family-structured data within UK Biobank.

iv.

Standardizing Reporting and Quality Control

STROBE-MR guidelines

To improve transparency, reproducibility, and methodological rigor in Mendelian Randomization (MR) studies, the STROBE-MR guidelines were introduced in 2021 by Skrivankova et al. as an extension of the widely used STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) framework[2].

The STROBE-MR checklist comprises 20 items tailored specifically to MR studies, covering key aspects of study design, data sources, assumptions, statistical methods, and interpretation of results. It emphasizes the importance of:

Clearly articulating the three instrumental variable assumptions (relevance, independence, and exclusion restriction)

Describing the selection and validation of genetic instruments

Reporting sensitivity analyses, including heterogeneity tests, pleiotropy-robust methods, and leave-one-out analyses

Justifying causal interpretations and acknowledging limitations such as weak instruments or horizontal pleiotropy

Importantly, the guideline encourages pre-registration and the use of open data and code, aligning with best practices in open science. Several high-impact journalsnow require STROBE-MR compliance for MR submissions.

The adoption of STROBE-MR has significantly contributed to improving the quality and interpretability of MR literature, offering both authors and reviewers a standardized framework to assess methodological soundness. As MR research continues to expand—particularly in complex fields like cancer epidemiology and multi-omics—STROBE-MR provides an essential tool for maintaining scientific rigor and credibility.

Bias assessment tools

Frameworks such as MR-GRADE are being explored to systematically evaluate the reliability of MR studies, enhancing their credibility in systematic reviews and meta-analyses[31].

These methodological advancements have significantly strengthened the validity of MR research, addressing concerns about pleiotropy, weak instruments, and misinterpretation of results. As the field continues to evolve, integrating new statistical techniques, data resources, and rigorous reporting standards will be crucial to ensuring the continued reliability and impact of MR studies.