Two decades of Mendelian randomization mapped: pleiotropy tops the field’s methodological worries

Mendelian randomization has become one of the most widely used tools in genetic epidemiology, and a new combined bibliometric and content analysis now offers the most structured picture yet of what the field actually worries about, how it responds to those worries, and where it says it is heading. The study, published in Frontiers in Genetics, screened more than 21,000 publications on the method and then coded a focused subset of 307 methods-oriented papers in detail. Its headline finding is that horizontal pleiotropy, the phenomenon in which a genetic variant influences an outcome through pathways other than the exposure of interest, was the most frequently represented methodological challenge, appearing as the primary concern in 90 of the 307 coded publications, or 29.3 percent.

The research team, led by Jiapeng Zhang and Ying Xiao of Macau University of Science and Technology together with colleagues from several Chinese institutions, built two linked corpora with distinct purposes. The first, called the A-corpus, preserved breadth: the researchers searched the Web of Science Core Collection and Scopus for English-language articles and reviews containing the exact terms “Mendelian randomization” or “Mendelian randomisation” published between 1 January 2003 and 31 December 2025, with the databases searched on 15 February 2026. After field harmonization and hierarchical deduplication, the A-corpus comprised 21,346 publications spanning 2,407 journals, 53,126 authors, 109 countries and regions, and 3,828 institutions.

The second corpus, the B-corpus, was designed for interpretive depth. The team developed a prespecified terminology dictionary covering methodological challenges, solution pathways, and future priorities, drawing on canonical sources such as the MR-Egger estimator, weighted median and mode-based estimators, MR-PRESSO, weak-instrument bias, sample overlap, winner’s curse, Steiger directionality, CAUSE, colocalization, within-family designs, and reporting standards. Applying this dictionary to titles, abstracts, and keywords flagged 10,598 candidate records from the A-corpus. Manual eligibility review then excluded 10,291 of them, leaving 307 publications that explicitly addressed methodological problems, evaluated or proposed solutions, discussed reporting or reproducibility, or articulated future directions. Routine exposure-outcome applications that merely reported standard sensitivity analyses were deliberately excluded.

Each of the 307 publications was independently coded by two reviewers, with the second blinded to the first’s decisions, across four dimensions: methodological challenge, solution pathway, future priority, and solution orientation. Inter-coder reliability, measured with Cohen’s kappa before consensus adjudication, was moderate for methodological challenges at 0.434 and solution pathways at 0.531, substantial for future priorities at 0.703, and almost perfect for solution orientations at 0.897. All disagreements were resolved through consensus discussion against the source text and operational definitions. The authors are candid that the moderate kappa values for the first two dimensions indicate residual subjectivity in the taxonomy, and they caution readers to weigh this when interpreting category-level frequencies.

Beyond pleiotropy, the coded challenges formed a clear hierarchy. A general methodological framework category accounted for 48 publications (15.6 percent), molecular attribution and colocalization for 41 (13.4 percent), reverse causation and directionality for 32 (10.4 percent), and reporting, reproducibility, and transparency for 31 (10.1 percent). Weak instruments, a long-standing concern in instrumental-variable analysis, remained highly visible at 26 publications (8.5 percent). Less frequent but still represented were family-structure-related bias, multivariable confounding decomposition, population stratification and ancestry, nonlinear and lifecourse effects, and sample overlap with winner’s curse.

On the solution side, the dominant pathway was what the coding manual calls the mixed sensitivity framework, a panel of complementary estimators, diagnostics, and robustness checks rather than reliance on any single method. This category covered 107 publications, or 34.9 percent of the B-corpus. Reporting guideline standards followed at 45 publications (14.7 percent), then omics integration frameworks at 33 (10.7 percent) and bidirectional or directionality frameworks at 28 (9.1 percent). Within-family designs, triangulation with external validation, multivariable MR, mediation MR, nonlinear MR, cis-MR and drug-target MR, colocalization, and MR-Egger each occupied smaller shares. The challenge-solution cross-tabulations revealed telling pairings: horizontal pleiotropy was most often paired with mixed sensitivity frameworks, reporting standards, and MR-Egger, while molecular attribution questions clustered with omics integration and colocalization approaches, and family-structure bias was paired almost exclusively with within-family designs.

The future-priority dimension produced a striking result: the single most frequent code was “none explicitly stated,” covering 94 publications, or 30.6 percent. The authors emphasize that this label means only that no future-priority statement meeting the coding criteria appeared in the analyzed text, not that the publications had no unresolved issues. Among papers with an explicitly codable future direction, a reproducible reporting pipeline led with 45 publications (14.7 percent), followed by mechanism-aware inference (36, 11.7 percent), omics integration (34, 11.1 percent), clinical translation and public health application (21, 6.8 percent), and family-based designs (19, 6.2 percent). Drug-target translation, triangulation with external evidence, multi-ancestry validation, and nonlinear lifecourse MR were also named.

The analysis also introduced a descriptive framework of solution orientations. Validity-first work, focused on bias and robustness, dominated with 138 publications (45.0 percent), followed by mechanism-aware work on biological interpretation and molecular attribution (78, 25.4 percent), a hybrid code for publications emphasizing multiple orientations comparably (42, 13.7 percent), generalizability-enhanced work on ancestry, family structure, and external validity (32, 10.4 percent), and translation-ready work on drug-target, clinical, or public-health applications (17, 5.5 percent). The authors stress that this framework summarizes coded emphases and should not be read as a ranking of methods or as evidence that any orientation is scientifically superior.

The macro-level bibliometric results add context to the coded findings. Annual publication output was relatively low between 2003 and 2011, increased after 2012, accelerated after 2018, and rose further after 2020, a trajectory the authors link to the accumulation of genome-wide association study summary statistics, open two-sample MR resources, and expanding applications in drug-target and multi-omic contexts. Using whole counting, in which each multi-country publication contributes one count to every participating country, China recorded the largest number of country-level participation records at 14,695, followed by the United States at 3,364 and the United Kingdom at 3,043. Harvard University (1,087 whole-count records), the University of Bristol (1,083), and the University of London (962) led at the institutional level, while Smith GD was the most frequently represented author at 472 records. The most common journals were Medicine, Scientific Reports, and Frontiers in Endocrinology. The authors repeatedly caution that these whole-count values and network centrality measures describe participation and connectedness, not methodological leadership or scientific superiority. Keyword burst analysis traced the field’s shifting attention, with early bursts around coronary heart disease, myocardial infarction, and instrumental variables giving way to more recent bursts including the UK Biobank, pleiotropy, causal effect, and anxiety.

The study’s limitations are spelled out in detail. The analysis was restricted to two databases and to English-language articles and reviews, so publications indexed elsewhere or in other languages may have been missed. The exact-phrase search prioritized specificity over comprehensiveness and may have excluded early or related instrumental-variable genetic studies that did not use the standard terminology. Dictionary-based screening may likewise have missed publications expressing methodological issues only implicitly or with uncommon terms. The one-primary-code-per-dimension rule simplified multidimensional papers and may have underrepresented secondary concerns. The authors also note that the 2025 publication count, which was lower than the 2024 count, may reflect incomplete database indexing at the February 2026 retrieval date and should be interpreted cautiously. Most fundamentally, they write, bibliometric indicators and content coding characterize publication patterns and reported methodological emphases; they cannot determine the comparative validity of specific MR methods or the causal or clinical credibility of individual findings. What the analysis delivers is a structured, auditable map of where a rapidly growing field concentrates its attention: on the exclusion-restriction problem that pleiotropy poses, on layered sensitivity analyses as the default response, and, increasingly, on reproducible pipelines, mechanistic interpretation, and broader population coverage as the priorities it names for the years ahead.

Subject of Research: Bibliometric and content analysis of methodological challenges, solutions, and future priorities in Mendelian randomization research

Article Title: Methodological challenges, solution pathways, and future priorities in Mendelian randomization research: a bibliometric and content analysis

Article References: Zhang, J., Xiao, Y., Feng, F., Feng, M., Jiang, D., & Wang, L. (2026). Methodological challenges, solution pathways, and future priorities in Mendelian randomization research: a bibliometric and content analysis. Frontiers in Genetics, 17, Article 1853737. https://doi.org/10.3389/fgene.2026.1853737

Image Credits: AI Generated

DOI: 10.3389/fgene.2026.1853737

Keywords: Mendelian randomization, bibliometrics, horizontal pleiotropy, causal inference, sensitivity analysis, genome-wide association studies, reproducibility, colocalization, instrumental variables, research trends, Methodological, challenges