No Significant Direct Causal Association Between TNF Pathway Biomarkers and Hip Fracture Risk: A Study Based on Real-World Data

Summary

This study integrates FAERS pharmacovigilance with Mendelian randomization to assess associations between TNF pathway biomarkers and hip fracture risk. Combining real-world adverse-event surveillance with genetic causal inference, the workflow offers a reproducible approach for evaluating biomarker–outcome relationships when randomized trials are unavailable, impractical, or ethically difficult to conduct in practice.

Abstract

Hip fracture is a common and serious condition in elderly individuals, often associated with altered TNF-α signaling. However, few studies have explored whether changes in TNF pathway biomarkers contribute to the risk of hip fracture. In this study, adverse events related to hip fracture reported for five TNF inhibitors were analyzed using data from the first quarter of 2014 to the fourth quarter of 2023. After data standardization and cleaning, four disproportionality methods, including the reporting odds ratio, proportional reporting ratio, multi-item gamma Poisson shrinker, and Bayesian confidence propagation neural network, were applied to assess the association between TNF inhibitor exposure and hip fracture outcomes. Complementary mendelian randomization analyses were further conducted using TNF-α, sTNFR1, and sTNFR2 as exposures and hip fracture as the outcome. Pharmacovigilance analyses revealed no significant association between TNF inhibitor exposure and hip fracture-related adverse events across the four algorithms. Mendelian randomization analyses identified no significant direct causal association between genetically predicted TNF-α and hip fracture risk, and additional analyses of sTNFR1 and sTNFR2 yielded consistent results. These findings suggest that TNF pathway biomarkers are unlikely to act as independent direct determinants of hip fracture risk and may provide a useful basis for future mechanistic studies and prevention strategies.

Introduction

Hip fracture is a common and serious clinical condition in elderly individuals, and its clinical manifestations are usually characterized by hip pain and dysfunction¹. Hip fracture not only has a serious impact on patients but also imposes a heavy burden on families and society. For patients, hip fracture is often accompanied by a long period of hospitalization and rehabilitation, which has a great impact on the physical and mental health of elderly patients^2,3,4. At the family level, patients usually require long-term care from family members, which can increase the financial and emotional burden on the family, especially for low-income families, which may have a greater impact⁵. At the societal level, the high incidence of hip fractures not only increases the consumption of healthcare resources but also poses a challenge to the health security system in an aging society^6,7. From an epidemiological point of view, the incidence of hip fracture has shown an increasing trend annually globally, which is closely related to population aging⁸. Elderly people, especially elderly women, are at high risk for hip fracture, and approximately 75% of cases of hip fracture occur in the elderly female population over 65 years of age^9,10,11. Therefore, it is necessary to carry out relevant research on hip fractures to explore the mechanism of their occurrence and development, which is conducive to better prevention and treatment of hip fractures. Previous studies have suggested that hip fracture may be accompanied by altered TNF-α expression and inflammatory activation^12,13. However, whether genetically predicted alterations in TNF pathway biomarkers are causally associated with hip fracture risk remains unclear¹⁴. Therefore, the present study investigated whether TNF pathway biomarkers are causally associated with hip fracture risk. Five TNF inhibitors, including etanercept¹⁵, adalimumab¹⁶, infliximab¹⁷, certolizumab pegol¹⁸, and golimumab¹⁹, were selected for pharmacovigilance analysis. The FAERS database was used to identify hip fracture adverse event reports associated with these agents, and complementary mendelian randomization (MR) analyses were performed to further evaluate causality.

Several methodological approaches can be used to investigate this question, but each has important limitations. Traditional observational studies, such as cohort and case-control studies, are useful for identifying associations in clinical populations, but they are often affected by residual confounding, indication bias, reverse causation, and incomplete adjustment for disease severity, comorbidities, or concomitant medications²⁰. Randomized controlled trials can provide stronger causal evidence, but they are often expensive, time-consuming, and difficult to conduct for long-term outcomes such as hip fracture, especially when the exposure of interest involves inflammatory biomarkers or post-marketing drug safety events²¹. Standalone pharmacovigilance analyses can detect potential safety signals in large real-world populations, but spontaneous reporting data cannot independently establish causality because of underreporting, duplicate reports, missing data,reporting bias, and lack of denominator information²². Conversely, standalone mendelian randomization can strengthen causal inference by using genetic variants as instrumental variables, but it does not directly capture real-world post-marketing drug safety patterns and may be limited when valid instruments are unavailable or horizontal pleiotropy is present²³.

The FDA adverse event reporting system (FAERS) is a database established by the U.S. Food and Drug Administration (FDA) to collect and monitor drug safety information²⁴. It records adverse drug event reports from multiple sources, including patients, healthcare professionals, and drug manufacturers, and is an important tool for monitoring and evaluating adverse drug reactions²⁵. The information in the FAERS database is widely used in drug safety analysis and research, and many academic studies and clinical practice guidelines base their decisions on data from this platform²⁶. An example is monitoring the risk of adverse events related to immunosuppressants and other related adverse events^27,28. Compared with conventional clinical studies, spontaneous reporting systems can capture rare or delayed adverse events in large real-world populations at relatively low cost²⁹. However, they are susceptible to underreporting, duplicate reports, missing data, and reporting bias, and therefore cannot independently establish causality^30,31.

Mendelian randomization is a statistical method used to study the causal relationship between exposure and disease by using genetic variation as an instrumental variable³². In the study of adverse drug reactions, mendelian randomization can be used to identify potential causal relationships and provide reliable evidence to support them. The validity of Mendelian randomization relies on three assumptions, as shown in Figure 1^33,34,35: the selected variants are strongly associated with the exposure; the variants influence the outcome only through the exposure pathway; and the variants are independent of major confounding factors. Compared with traditional observational studies, Mendelian randomization is less vulnerable to residual confounding and reverse causation because genetic variants are randomly allocated at conception and remain largely stable throughout life^36,37. However, this approach may be less suitable when valid instrumental variables are unavailable, when horizontal pleiotropy is substantial, or when exposure effects are strongly time-dependent³⁸.

By integrating FAERS pharmacovigilance analysis with Mendelian randomization, the present workflow combines real-world safety surveillance with genetic causal inference. This combined approach is particularly suitable when randomized controlled trials are unavailable, impractical, or unethical, and when complementary evidence is needed to evaluate biomarker-outcome relationships³⁹. In practical terms, this FAERS-Mendelian randomization workflow is most suitable when the research question meets the following conditions: first, the drug-event or biomarker-outcome relationship is clinically important but cannot be adequately tested in randomized trials⁴⁰; second, post-marketing real-world safety data are available and can be used to screen or characterize adverse event signals; third, appropriate genome-wide association study summary statistics are available for both the exposure and outcome; and fourth, the investigator aims to combine signal detection with genetic causal inference rather than relying on either method alone. Under these conditions, FAERS can provide real-world evidence on whether a potential safety signal exists, whereas Mendelian randomization can further assess whether the related biomarker or pathway is likely to have a direct causal relationship with the outcome. This combined workflow is especially useful for evaluating rare, delayed, or ethically difficult-to-study outcomes and for generating more robust evidence when conventional epidemiological designs are limited⁴¹.

Building on this rationale, this study aimed to establish and demonstrate an integrated FAERS-Mendelian randomization workflow for evaluating drug-event safety signals and biomarker-outcome causal relationships when conventional causal inference approaches are limited. FAERS pharmacovigilance analysis was first used to screen and characterize hip fracture-related adverse event reports associated with TNF inhibitors. Mendelian randomization was then applied to assess whether genetically predicted TNF pathway biomarkers, including TNF-α, sTNFR1, and sTNFR2, were causally associated with hip fracture risk. By combining post-marketing safety surveillance with genetic causal inference, this workflow provides a reproducible methodological reference for future studies investigating clinically important drug-event or biomarker-outcome relationships.

Protocol

FAERS data sources

The real-world data for this study were obtained from the FAERS database (https://fis.fda.gov/extensions/FPD-QDE-FAERS/FPD-QDE-FAERS.html). This is a publicly accessible and anonymized database, so ethical approval was not required for this study. Information about the five drugs included is shown in Table 1. The search was performed by first limiting the adverse events to hip fracture, and the search time span was from Q1 2014 to Q4 2023. To ensure reliable and stable data, the study standardized the terminology of reported adverse events via the MedDRA Dictionary version 26.1⁴². FAERS quarterly ASCII files from Q1 2014 to Q4 2023 were downloaded and imported for analysis. The extracted FAERS tables included DEMO, DRUG, REAC, THER, RPSR, and OUTC. These tables were merged across all quarters before screening. Reports were linked using CASEID and PRIMARYID to ensure consistency across demographic, drug, reaction, therapy, reporter, and outcome information. The target drugs included etanercept, adalimumab, infliximab, certolizumab pegol and golimumab. Drug names in the DRUG table were standardized by converting text to uppercase, removing extra spaces and checking spelling variants when necessary. Target drugs were identified using standardized generic names in the DRUG table, and drug-role restriction was performed using ROLE_COD = “PS”, indicating the primary suspect drug⁴³. The specific study screening technique roadmap is shown in Figure 2. After data cleaning and screening, a unique and analyzable dataset of eligible hip fracture reports was obtained for subsequent analyses.

Duplicate FAERS reports were removed before signal detection. Duplicates were identified according to CASEID and PRIMARYID. When multiple reports shared the same CASEID, the most recent report was retained according to FDA_DT. If multiple reports had the same CASEID and FDA_DT, the report with the highest PRIMARYID was retained. After deduplication, each CASEID contributed only one record to the final analytic dataset. Reports were included if they met all of the following criteria: reporting date between Q1 2014 and Q4 2023; the adverse event was coded as “Hip fracture”; at least one of the five TNF inhibitors was recorded in the DRUG table; and the drug role was coded as primary suspect. Reports were excluded if they were duplicate records, lacked valid CASEID or PRIMARYID information, had no corresponding DRUG or REAC entry, did not include the target adverse event, or listed the target TNF inhibitor only as a concomitant or secondary suspect drug.

GWAS data sources for Mendelian randomization

The exposure data for TNF-α in the Mendelian randomization for this study were obtained from the IEU OpenGWAS database (https://gwas.mrcieu.ac.uk/), with GWAS ID prot-c-3722_49_2 from the study of Suhre K et al. The study population was of European ancestry, and the number of SNPs was 501,428⁴⁴.

The exposure data for sTNFR1 in the mendelian randomization for this study were obtained from the IEU OpenGWAS database (https://gwas.mrcieu.ac.uk/), with GWAS ID prot-c-2654_19_1 from the study of Suhre K et al. The study population was of European ancestry, and the number of SNPs was 501,428⁴⁴.

The exposure data for sTNFR2 in the Mendelian randomization for this study were obtained from the IEU OpenGWAS database (https://gwas.mrcieu.ac.uk/), with GWAS ID prot-c-3152_57_1 from the study of Suhre K et al. The study population was of European ancestry, and the number of SNPs was 501,428⁴⁴.

The outcome data for hip fracture, GWAS ID GCST90161240, deposited in the GWAS Catalog (https://www.ebi.ac.uk/gwas/studies/GCST90161240), are data from a meta-analysis of a large-scale GWAS that included 11,516 hip fracture cases and 723,838 controls⁴⁵. The disease classification aligns with the International Classification of Diseases (ICD; ICD-10 codes S72.0–S72.2 and ICD-9 code 820).

Ethical approval and informed consent had been obtained in the original GWAS studies. Because the present study used publicly available, anonymized FAERS data and publicly available GWAS summary statistics, no additional ethical approval was required.

Software environment and workflow implementation

All analyses were performed using R version 4.3.2. FAERS data import, cleaning, merging, and tabulation were performed using R-based data management workflows. Data tables were imported using functions such as data.table::fread() or readr::read_delim(), merged using CASEID and PRIMARYID, and processed using dplyr functions. Descriptive statistics and 2 × 2 contingency tables were generated using custom R scripts.

Mendelian randomization analyses were conducted using TwoSampleMR version 0.5.6. Exposure instruments were extracted using a significance threshold of P < 1 × 10⁻⁵ or formatted from GWAS summary statistics using TwoSampleMR-compatible input structures. Instrument clumping was performed using clump_data() with clump_r² = 0.001 and clump_kb = 10,000. Outcome data were extracted or formatted using extract_outcome_data() or read_outcome_data(), depending on the source format. Exposure and outcome datasets were harmonized using harmonise_data(). Causal estimates were generated using mr() with the following mendelian randomization methods: MR-Egger, weighted median, inverse variance weighted, simple mode, and weighted mode. Heterogeneity was assessed using mr_heterogeneity(), and horizontal pleiotropy was assessed using mr_pleiotropy_test(). All datasets were imported, cleaned, harmonized, and analyzed within this software environment to ensure a consistent and reproducible analytical workflow.

Pharmacovigilance analysis

Descriptive analyses were used to summarize hip fracture-related adverse events associated with the five drugs. Signal detection analyses were then performed using four disproportionality algorithms, including the reporting odds ratio (ROR), proportional reporting ratio (PRR), multi-item gamma Poisson shrinker (MGPS) and Bayesian confidence propagation neural network (BCPNN). The criteria for the four major algorithms are shown in Table 2⁴⁶.

Mendelian randomization analysis

Summary statistics for TNF-α, sTNFR1, and sTNFR2 were extracted as exposure datasets, and hip fracture summary statistics were extracted as the outcome dataset. Analyses were restricted to European ancestry datasets when available to reduce population stratification bias.

To minimize bias caused by linkage disequilibrium and weak instruments, the following criteria were applied: genome-wide significance threshold P < 1 × 10⁻⁵, linkage disequilibrium threshold r² < 0.001, clumping window of 10,000 kb, and F-statistic > 20. The F-statistic was calculated for each retained instrumental variable as beta²/se² to evaluate instrument strength. SNPs with F-statistic ≤ 20 were excluded from downstream analyses.

After SNP selection, exposure and outcome datasets were harmonized to align effect alleles. During harmonization, effect alleles and other alleles were aligned between the exposure and outcome datasets. SNPs with incompatible alleles were removed, and palindromic SNPs with ambiguous allele frequencies were excluded when strand orientation could not be determined. After harmonization, the retained SNPs were checked to confirm that beta coefficients corresponded to the same effect allele in both datasets. The number of SNPs retained after clumping and harmonization was recorded for each exposure as an intermediate reproducibility checkpoint.

Five Mendelian randomization methods were applied, including MR-Egger, weighted median, inverse variance weighted, simple mode, and weighted mode. Potential heterogeneity of instrumental variables was assessed using Cochran’s Q test, and P < 0.05 was considered indicative of significant heterogeneity. Potential horizontal pleiotropy was evaluated using the MR-Egger intercept, and P < 0.05 indicated pleiotropy, suggesting reduced reliability of the causal estimate⁴⁷. These analyses generated causal effect estimates together with heterogeneity and pleiotropy statistics for each exposure.

Intermediate checkpoints for reproducibility

Intermediate checkpoints were recorded after each major processing step to ensure workflow reproducibility. For the FAERS workflow, checkpoints included the number of DEMO records imported, the number of unique records after deduplication, the number of reports containing hip fracture as the target adverse event, the number of reports involving the five TNF inhibitors, and the final number of eligible reports in which TNF inhibitors were recorded as primary suspect drugs. For the Mendelian randomization workflow, checkpoints included the number of SNPs extracted for each exposure, the number of SNPs retained after linkage disequilibrium clumping, the number of SNPs available in the outcome dataset, the number of SNPs retained after harmonization, and the final number of instrumental variables used in each Mendelian randomization analysis.

Statistical reporting

Continuous results were reported with corresponding effect estimates, 95% confidence intervals (95% CI), and P values. Unless otherwise specified, statistical significance was defined as a two-sided P < 0.05. For pharmacovigilance analysis, descriptive counts and disproportionality estimates were reported for each individual TNF inhibitor and for the pooled TNF inhibitor group. For Mendelian randomization analysis, causal estimates, standard errors, 95% confidence intervals, P values, heterogeneity statistics, pleiotropy test results, and the number of retained SNPs were reported for each exposure.

Results

Results of descriptive analysis

The detailed data-cleaning and screening workflow is summarized in Figure 2. A total of 15,428,448 records were initially retrieved from the DEMO table. After duplicate removal, 13,447,292 unique DEMO records remained. After data standardization, deduplication, and cross-table linkage, the final numbers of records available in each FAERS table were 27,260,995 in DRUG, 18,851,426 in REAC, 12,573,760 in THER, 6,448,188 in RPSR, and 7,645,965 in OUTC. Among all screened records, 7,852 reports identified hip fracture as the adverse event with the drug listed as the primary suspect. After restricting the analysis to TNF inhibitors recorded as primary suspect drugs, 472 unique analyzable reports were retained for the final study population, as shown in Table 3. The number of cases of adverse events per year is shown in Figure 3. Among them, 335 patients (71.0%) were female, and 94 patients (19.9%) were male. The highest percentage of cases was 56.5% in the 65 years and above age group. The reports were mainly submitted by physicians, with 258 cases (54.7%). They were followed by health professionals (25.0%), other health professionals (16.5%), and pharmacists (3.8%). In terms of country of coverage, the United States dominated with 257 cases (54.4%), whereas other countries and unknown countries accounted for 19.6% and 20.3%, respectively.

FAERS database findings

Table 4 presents the disproportionality results for each TNF inhibitor and for the pooled analysis. Across all four signal detection algorithms, no positive safety signal for hip fracture was identified for any individual drug or for the combined TNF inhibitor group.

Validation of Mendelian randomization results

As shown in Figure 4, five Mendelian randomization methods were applied with TNF-α as the exposure and hip fracture as the outcome. All analyses yielded P values>0.05, indicating no significant association between TNF-α levels and hip fracture risk. A total of four SNPs were retained as instrumental variables. The mean F-statistic was 22.68, ranging from 21.26 to 24.93, suggesting that all selected instruments had adequate strength and were unlikely to introduce substantial weak-instrument bias. Additional Mendelian randomization analyses were performed using soluble TNF receptors as exposures. For sTNFR1, five SNPs were selected as instrumental variables, with a mean F-statistic of 22.39 (range: 20.31–24.24), as shown in Figure 5. For sTNFR2, four SNPs were retained, with a mean F-statistic of 21.39 (range: 20.30–23.07), as shown in Figure 6. Consistent with the TNF-α analysis, neither sTNFR1 nor sTNFR2 showed a significant causal association with hip fracture risk. No evidence of horizontal pleiotropy was observed in the MR-Egger intercept analyses. For TNF-α, the intercept was −0.0148 (SE = 0.0844, P = 0.877). For sTNFR1, the intercept was 0.0449 (SE = 0.0656, P = 0.543). For sTNFR2, the intercept was −0.0629 (SE = 0.0563, P = 0.380). Similarly, heterogeneity analyses did not indicate significant between-instrument heterogeneity. For TNF-α, Cochran’s Q values were 1.391 (MR-Egger, P = 0.499) and 1.422 (IVW, P = 0.700). For sTNFR1, the corresponding Q values were 1.666 (MR-Egger, P = 0.644) and 2.135 (IVW, P = 0.711). For sTNFR2, the Q values were 1.997 (MR-Egger, P = 0.368) and 3.249 (IVW, P = 0.355). These findings further support the robustness and consistency of the causal estimates.

No significant direct causal association was identified between TNF-α and hip fracture risk. Additional Mendelian randomization analyses of sTNFR1 and sTNFR2 yielded consistent results, further suggesting that TNF pathway biomarkers are unlikely to act as independent direct determinants of hip fracture. These findings provide new evidence for understanding the inflammatory mechanisms of hip fracture and may inform future mechanistic and clinical research.

Workflow verification and reproducibility materials

To enhance workflow transparency and reproducibility, we added a compact workflow-verification figure summarizing each major processing stage and the corresponding output generated at that stage for both the FAERS pharmacovigilance workflow and the Mendelian randomization pipeline, as shown in Figure 7. We also included a reproducibility table listing the exact dataset identifiers, query restrictions, filters, thresholds, and retained record or SNP counts after each major step, as shown in Table 5. In addition, we provided an analytical checklist specifying the software functions or scripts, representative code snippets, key parameters, and verification outputs for each major FAERS and Mendelian randomization step. This table also includes explicit examples of the FAERS query restriction for case identification, duplicate report removal, and SNP harmonization and filtering output within the Mendelian randomization workflow, as shown in Table 6. The retained instrumental variables for each exposure, including TNF-α, sTNFR1, and sTNFR2, together with their corresponding F-statistics, to facilitate verification of the Mendelian randomization analyses, were listed in Supplementary Table 1.

DATA AVAILABILITY:

All raw data and analysis code supporting the findings of this study have been made publicly available through Zenodo to ensure transparency and reproducibility. The deposited materials include raw and processed FAERS pharmacovigilance datasets, data-cleaning outputs, disproportionality-analysis input and output files, Mendelian randomization exposure and outcome datasets, harmonized instrumental-variable datasets, retained instrumental-variable tables, F-statistic files, sensitivity-analysis outputs, workflow-verification materials, and R scripts used for data processing and statistical analysis. The Zenodo repository is publicly available at: https://doi.org/10.5281/zenodo.20046970. This all-version DOI represents all versions of the deposited record and will always resolve to the latest version. The FAERS pharmacovigilance data were obtained from the publicly accessible U.S. Food and Drug Administration Adverse Event Reporting System database. The GWAS summary statistics for TNF-α, sTNFR1, and sTNFR2 were obtained from the IEU OpenGWAS database, with accession IDs prot-c-3722_49_2, prot-c-2654_19_1, and prot-c-3152_57_1, respectively. The hip fracture GWAS summary statistics were obtained from the GWAS Catalog, with accession ID GCST90161240. All deposited files are derived from publicly available and anonymized data sources, and no individual-level identifiable information is included.

Figure 1: Mendelian randomization analysis framework for TNF pathway biomarkers and hip fracture. Please click here to view a larger version of this figure.

Figure 2: Diagram of FAERS study screening techniques. Please click here to view a larger version of this figure.

Figure 3: Number of cases of adverse events per year. Please click here to view a larger version of this figure.

Figure 4: Mendelian randomization results for TNF-α in relation to hip fracture. Please click here to view a larger version of this figure.

Figure 5: Mendelian randomization results for sTNFR1 in relation to hip fracture. Please click here to view a larger version of this figure.

Figure 6: Mendelian randomization results for sTNFR2 in relation to hip fracture. Please click here to view a larger version of this figure.

Figure 7: Compact workflow-verification overview of the FAERS pharmacovigilance and mendelian randomization pipelines. Please click here to view a larger version of this figure.

Table 1: Summary of drugs included in this study. Please click here to download this Table.

Table 2: Four major algorithms used for signal detection. Please click here to download this Table.

Table 3: Population Characteristics of Reports of Hip Fracture Associated with TNF Inhibitors in the FAERS Database, 2014–2023. Please click here to download this Table.

Table 4: Signal intensity of 5 drugs and overall primary suspicion for hip fracture events. Please click here to download this Table.

Table 5: Reproducibility summary of datasets, filters, thresholds, and retained counts. Please click here to download this Table.

Table 6: Analytic checklist of software functions and scripts used in the FAERS pharmacovigilance and Mendelian randomization workflows. Please click here to download this Table.

Supplementary Table 1: Retained instrumental variables for each exposure in the Mendelian randomization analyses.Please click here to download this file.

Discussion

In this study, five TNF inhibitors identified from real-world pharmacovigilance data were integrated with Mendelian randomization analyses to evaluate whether TNF pathway biomarkers are causally associated with hip fracture risk. The findings consistently indicated that no significant direct causal association between TNF-α and hip fracture risk was identified, and additional analyses of sTNFR1 and sTNFR2 yielded concordant results. These results support the value of combining spontaneous-report safety data with genetic causal inference when randomized trials are unavailable or impractical. Beyond these empirical findings, this study provides a reproducible protocol framework for integrating spontaneous-report pharmacovigilance data with genetically informed causal inference. In this workflow, FAERS is used to characterize real-world adverse event reporting patterns, whereas Mendelian randomization is used to assess whether a biomarker or pathway is likely to have an independent causal effect on the clinical outcome. Therefore, this workflow is intended not only to evaluate a specific TNF-related research question but also to demonstrate a generalizable methodological strategy for safety-signal assessment and causal triangulation when randomized trials are unavailable, infeasible, or ethically difficult to conduct.

Compared with conventional cohort and case-control studies, this combined workflow is less dependent on long-term prospective follow-up and can rapidly generate real-world safety evidence from post-marketing data⁴⁸. However, observational studies may provide richer clinical information, including disease severity, medication duration, comorbidities, and confounder adjustment⁴⁹. Compared with randomized controlled trials, the FAERS-Mendelian randomization workflow is more feasible for rare or long-latency outcomes such as hip fracture, although it cannot replace trial-level evidence when randomized evidence is available⁵⁰. Compared with pharmacovigilance analysis alone, the addition of Mendelian randomization provides a genetically informed causal component that helps distinguish a reporting signal from a potentially causal biological pathway³². Compared with mendelian randomization alone, the inclusion of FAERS adds real-world post-marketing context, including drug-specific reporting patterns and safety surveillance information⁵¹. Therefore, the main value of this workflow lies in methodological triangulation rather than in replacing any single established design.

Several common failure points should be considered when applying this workflow. In the FAERS component, inconsistent raw data structures across quarterly files are a frequent source of error. Column names, file formats, and table structures may vary across reporting periods, which may lead to import errors or incorrect table merging⁵². This can be addressed by checking raw headers before merging, removing empty trailing columns, and using flexible merging functions that allow non-identical columns. Another common problem is inconsistent drug naming, because drugs may be reported using generic names, brand names, active ingredients, abbreviations, or spelling variants⁵³. Therefore, drug names and active ingredient fields should be standardized before matching, preferably using uppercase conversion, punctuation removal, and a predefined generic-name dictionary. Duplicate reporting is another critical issue. If duplicate CASEID records are not removed, disproportionality estimates may be inflated⁵⁴. A transparent deduplication rule should therefore be applied, such as retaining the report with the latest FDA receipt date and, if tied, retaining the highest PRIMARYID.

In the Mendelian randomization component, common failure points include too few available instrumental variables, weak instrument bias, linkage disequilibrium among selected SNPs, allele mismatch between exposure and outcome datasets, ambiguous palindromic SNPs, incomplete outcome SNP matching, horizontal pleiotropy, sample overlap, and ancestry mismatch^55,56. These issues may lead to unstable or biased causal estimates. To reduce these risks, investigators should prespecify the SNP selection threshold, perform LD clumping, calculate F-statistics, remove weak instruments, harmonize effect alleles carefully, exclude ambiguous palindromic SNPs when strand orientation cannot be determined, and conduct heterogeneity and pleiotropy tests. If too few SNPs remain after filtering, a relaxed threshold may be considered, but the decision should be justified and accompanied by sensitivity analyses.

This workflow can be modified according to the characteristics of the data source, exposure, outcome, and study objective. If the target adverse event is rare, investigators may broaden the adverse-event definition by including related MedDRA Preferred Terms or Standardized MedDRA Queries, while also performing sensitivity analyses using both narrow and broad definitions. If the target drug has multiple synonyms, biosimilars, or active ingredients, a curated drug dictionary should be constructed before querying FAERS. If the research question focuses on a drug class rather than a single agent, both drug-specific and pooled analyses should be reported. If reporter type, age group, sex, country, or reporting year is central to the research question, stratified disproportionality analyses may be added when the sample size is sufficient.

The Mendelian randomization component should also be adapted to the available genetic data. If GWAS data for the exact biomarker are unavailable, proxy traits, soluble receptors, cis-protein quantitative trait loci, or pathway-related biomarkers may be considered, but the biological interpretation should be limited accordingly. If exposure and outcome GWAS datasets differ in ancestry, ancestry-specific analyses should be prioritized to reduce population stratification⁵⁷. If only a small number of SNPs are available, MR-Egger may be unstable, and greater emphasis should be placed on inverse variance weighted analysis, weighted median analysis, leave-one-out analysis, and biological plausibility of individual instruments⁵⁸. If outcome data are binary, causal estimates should be interpreted on the odds-ratio scale. If sample overlap is suspected, this should be acknowledged and assessed through alternative outcome datasets or sensitivity analyses when possible⁵⁹.

Descriptive analysis of the data revealed that a higher proportion of hip fracture reports was observed in women than in men. This pattern may be related to sex differences in bone mineral density, hormonal changes after menopause, and age-related skeletal fragility^60,61,62. In particular, the present study revealed that hip fracture accounted for the largest proportion of people over 65 years of age when the vast majority of women stopped menstruating, with a significant decline in estrogen levels, abnormal bone metabolism, etc., leading to a decrease in bone density, which further validates that the risk of hip fracture may be greater in women than in men^63,64,65. The reporting population is mostly medicine-related, and although the participation of professionals can improve the accuracy and standardization of reporting, the lack of public participation may lead to some unreported side effects and potential drug safety hazards due to low public awareness and sensitivity to adverse drug reactions. In the future, more attention needs to be given to improving ADR monitoring, especially how to increase public participation, which can be accomplished by increasing public education, improving communication channels, and setting up a convenient reporting mechanism^66,67. The analysis of the reporting countries reveals that most of the reporting countries are in the United States, which suggests that international cooperation and information sharing may need to be strengthened in the future. The number of ADR reports in the United States may reflect higher reporting activity, broader utilization of the FAERS system, or stronger pharmacovigilance infrastructure, but globally, ensuring drug safety still requires close collaboration among countries^68,69. By strengthening the integration and sharing of global adverse drug reaction data and promoting the standardization of international drug safety assessments, drug safety risks can be better prevented and controlled and public health can be improved⁷⁰.

According to existing studies, TNF-α is an important mediator of inflammatory bone remodeling and may directly promote osteoclast activity, bone resorption and skeletal fragility. However, the presence of these biological effects does not necessarily indicate that circulating TNF-α levels alone constitute an independent direct causal determinant of hip fracture risk in the general population^71,72,73. For example, studies have shown that the expression level of IL-1β, a more common inflammatory factor, may change during fracture^74,75. The upregulation of IL-1β further activates osteoclasts and increases bone resorption, which has the potential to lead to an increased risk of hip fracture^76,77,78. Moreover, TNF-α can enhance the inflammatory response by inducing the release of IL-1β from immune cells such as macrophages and monocytes^79,80,81. In addition, TNF-α can also increase IL-1β expression by activating the NF-κB pathway, which in turn promotes IL-1β transcription^82,83. These findings suggest that changes in circulating TNF-α levels alone may not constitute an independent direct determinant of hip fracture risk. However, TNF-α may still contribute to skeletal fragility indirectly through regulation of IL-1β expression and broader inflammatory pathways. Moreover, the increased risk of fracture is more likely related to changes in hormone levels, age, and long-term use of medications such as steroids rather than to a single factor such as TNF-α alone^84,85,86. This framework may also help explain why TNF inhibitors have been associated with reduced fracture risk in inflammatory diseases such as rheumatoid arthritis and ankylosing spondylitis⁸⁷. The protective effects of TNF inhibitors may reflect multiple mechanisms beyond direct blockade of TNF-α alone, including improved disease control, reduced chronic systemic inflammation, lower glucocorticoid exposure, improved physical function, and attenuation of inflammation-related bone loss^88,89. Therefore, treatment benefits observed in specific inflammatory populations should not be directly interpreted as evidence that TNF-α alone is a universal direct causal determinant of hip fracture risk.

The present findings have important practical implications. No significant direct causal association was identified between TNF pathway biomarkers and hip fracture risk, suggesting that circulating TNF-α, sTNFR1, and sTNFR2 may have limited value as independent predictors or direct intervention targets for hip fracture in clinical practice. This result indicates that fracture risk assessment should not rely excessively on a single inflammatory biomarker, but should instead continue to focus on established determinants such as age, sex, bone mineral density, hormonal status, falls, comorbidities, and medication exposure. In addition, these findings suggest that the clinical benefits of TNF inhibitors observed in certain inflammatory diseases may reflect broader anti-inflammatory and disease-control effects rather than a direct one-to-one causal relationship between TNF-α and hip fracture. Therefore, the present study provides a more cautious basis for interpreting TNF-related biomarkers in fracture research and may help guide future studies toward broader inflammatory networks and multifactorial risk mechanisms. From a workflow perspective, this protocol also provides a structured approach for evaluating whether post-marketing safety concerns are supported by genetically informed causal evidence. It may be particularly useful for hypothesis prioritization, safety signal interpretation, and mechanistic triangulation.

The present study also has the following limitations. First, potential confounding introduced by underlying autoimmune diseases and concomitant glucocorticoid use among patients treated with TNF-α inhibitors could not be fully assessed. Because FAERS does not comprehensively capture indications, disease severity, cumulative glucocorticoid exposure, or co-medication information, subgroup and sensitivity analyses were not feasible⁹⁰. However, the complementary Mendelian randomization analyses were conducted using genetic variants as instrumental variables for TNF-α, sTNFR1, and sTNFR2 levels. As genetic variants are randomly allocated at conception and remain largely stable throughout life, this design is less susceptible to conventional confounding factors and reverse causation⁹¹. Second, FAERS reports data from a variety of sources, and the quality of the data may vary depending on the professional competence of the reporter, the attention given to the report, and the level of detail in the report. Therefore, this study standardized the data via the MedDRA lexicon version 26.1 and validated Mendelian randomization. Finally, FAERS data rely mainly on the reporting of adverse drug reactions, which often do not fully establish the causal relationship between drugs and adverse reactions; therefore, this paper used four algorithms for the correction and testing of FAERS data, and Mendelian randomization was used for further validation with the aim of strengthening the reliability and stability of the results. Nevertheless, the FAERS and Mendelian randomization components should be interpreted as complementary rather than interchangeable. FAERS provides real-world safety-surveillance evidence, whereas Mendelian randomization provides genetically informed causal inference under specific instrumental-variable assumptions.

This workflow can be adapted to other drug classes, adverse events, biomarkers, and disease outcomes. In pharmacovigilance research, it can be applied to other spontaneous reporting systems or extended by integrating electronic health records, claims databases, disease registries, or hospital-based adverse event monitoring systems. In causal inference research, the Mendelian randomization component can be extended to multivariable Mendelian randomization, two-step Mendelian randomization, drug-target Mendelian randomization, cis-Mendelian randomization, colocalization analysis, or phenome-wide Mendelian randomization to investigate pathway-specific mechanisms. The workflow may also serve as a screening framework for prioritizing safety signals that require prospective validation, mechanistic experiments, or enhanced clinical surveillance. Future studies could further improve this protocol by automating FAERS data cleaning, standardizing drug dictionaries, embedding reproducibility checkpoints, and developing open-source pipelines that automatically generate workflow-verification outputs.

In summary, this study did not identify a significant direct causal association between TNF pathway biomarkers and hip fracture risk. More importantly, it provides a reproducible FAERS Mendelian randomization workflow for integrating post-marketing safety surveillance with genetic causal inference. The value of this protocol lies in its ability to evaluate safety signals, test biological plausibility, identify common failure points, and guide methodological adaptation across different data environments and research questions.

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
Generic laboratory computer workstation	Equipment	Generic laboratory computer	N/A
FDA Adverse Event Reporting System database	Public pharmacovigilance database	U.S. Food and Drug Administration	https://fis.fda.gov/extensions/FPD-QDE-FAERS/FPD-QDE-FAERS.html
FAERS quarterly ASCII files	Public pharmacovigilance data files	U.S. Food and Drug Administration	Q1 2014 to Q4 2023; DEMO, DRUG, REAC, THER, RPSR, and OUTC tables
MedDRA Dictionary version 26.1	Medical terminology resource	MedDRA Maintenance and Support Services Organization	https://www.meddra.org/
IEU OpenGWAS database	Public GWAS database	MRC Integrative Epidemiology Unit, University of Bristol	https://gwas.mrcieu.ac.uk/
TNF-α GWAS summary statistics	Exposure GWAS dataset	IEU OpenGWAS database	GWAS ID: prot-c-3722_49_2
sTNFR1 GWAS summary statistics	Exposure GWAS dataset	IEU OpenGWAS database	GWAS ID: prot-c-2654_19_1
sTNFR2 GWAS summary statistics	Exposure GWAS dataset	IEU OpenGWAS database	GWAS ID: prot-c-3152_57_1
GWAS Catalog	Public GWAS database	European Bioinformatics Institute	https://www.ebi.ac.uk/gwas/
Hip fracture GWAS summary statistics	Outcome GWAS dataset	GWAS Catalog	Accession ID: GCST90161240; https://www.ebi.ac.uk/gwas/studies/GCST90161240
R version 4.3.2	Statistical computing environment	R Foundation for Statistical Computing	https://www.r-project.org/
data.table package	R package	R package repository / R community	https://cran.r-project.org/package=data.table
dplyr package	R package	R package repository / R community	https://cran.r-project.org/package=dplyr
readr package	R package	R package repository / R community	https://cran.r-project.org/package=readr
stringr package	R package	R package repository / R community	https://cran.r-project.org/package=stringr
vroom package	R package	R package repository / R community	https://cran.r-project.org/package=vroom
TwoSampleMR version 0.5.6	R package for Mendelian randomization	MRC Integrative Epidemiology Unit	https://mrcieu.github.io/TwoSampleMR/
Custom R scripts for FAERS processing	Analysis script	Prepared by the study authors	Provided as supplementary files
Custom R scripts for Mendelian randomization	Analysis script	Prepared by the study authors	Provided as supplementary files