In this MR study, the summary-level data of TIMP-3 was obtained from a cohort study including approximately 1,000 participants available in the KORA (Cooperative Health Research in the Region of Augsburg) public database (Suhre et al., 2017). All participants of this GWAS were of European descent. The IS data were extracted from a genome-wide association meta-analysis of 29 studies, comprising approximately 440,328 participants (34,217 cases and 406,111 non-cases) of European descent (Malik et al., 2018). This meta-analysis identified 32 loci related to the risk of stroke and detected a total of 7,537,579 single nucleotide polymorphisms (SNPs). The subtypes of IS were derived from this meta-analysis. The SVS cohort had 198,048 individuals and 6,150,261 related SNPs. The CES cohort had 211,763 individuals and 7,954,834 related SNPs. The LAS cohort had 150,765 individuals and 7,992,739 related SNPs. The statistical information of ICH was acquired from a genome-wide association meta-analysis performed by the International Stroke Genetics Consortium and included 3,026 Europeans (Woo et al., 2014). The ICH cohort was further divided into LICH and NLICH cohorts.

SNPs were selected as instrumental variables from the GWAS database of TIMP-3 variants and used to investigate the causal relationship. Based on published research (Sun et al., 2020), we defined the inclusion criteria of SNPs as those with p values less than 1 × 10^–5. In addition, we clumped SNPs by setting r ² > 0.1 to exclude the influence of linkage disequilibrium on the results (Sun et al., 2020).

We conducted two-sample MR analyses to explore the potential causal role of serum TIMP-3 in stroke outcomes (IS, ICH, and their subtypes). The traditional evaluation method of inverse-variance weighting (IVW) was used as the primary method to evaluate causality in this study (Jia et al., 2021). Accurate causality can be obtained in the presence of directional pleiotropy using the IVW approach (Burgess et al., 2015). Depending on the heterogeneity, we decided whether the random-effect IVW method or fixed-effect IVW method should be used (Jia et al., 2021). The heterogeneity of instrumental variables was assessed using Cochran’s Q statistic (Greco et al., 2015). Weighted median estimation (WME) can provide a consistent assessment if more than 50% of the weights for the SNPs come from valid SNPs (Bowden et al., 2016a). The MR-Egger regression method is another method that can assist in estimating the causal relationship, although some included SNPs were atypical (Bowden et al., 2015; Bowden et al., 2016b); however, its statistical power is relatively weak (Bowden et al., 2015). In addition to assessing the causal relationship, MR-Pleiotropy RESidual Sum and Outlier (MR-PRESSO) has the advantage of detecting outliers of instrumental variables and providing evidence of directional pleiotropy (Verbanck et al., 2018). MR-Robust Adjusted Profile Score (MR. RAPS) can overcome the bias of weak instrumental variables and the effects of systematic and idiosyncratic pleiotropy to obtain a robust causal assessment (Zhao et al., 2018).

In the sensitivity analysis, the Egger regression method provided a way to detect pleiotropy, which is expressed by the value of the MR-Egger intercept. If the intercept is equal to 0 (p ≥ 0.05), it suggests that there is no pleiotropy (Hemani et al., 2018). Furthermore, the MR-PRESSO global test can also examine directional pleiotropy (Verbanck et al., 2018). If pleiotropy exists in instrumental variables, it implies that some other potential pathways may affect the pathway from instrumental variables and exposure factors to the outcome, which does not conform to the third core hypothesis of MR analysis [see Section 2.3]. The MR-PRESSO outlier test is a detection method similar to leave-one-out. It can eliminate all included SNPs one by one, correct the bias caused by outliers, and show robust causal estimation (Emdin et al., 2017; Ong and MacGregor, 2019).

To make rational use of MR analysis and improve the efficiency and accuracy of the study, all Mendelian randomisation studies must meet three core assumptions before they are conducted (Figure 1) (Emdin et al., 2017). First, instrumental variables (SNPs) should be closely related to exposure factors (TIMP-3). As shown in Supplementary Table S1, all included SNPs were statistically significantly related to TIMP-3. Second, instrumental variables (SNPs) should not be associated with any confounder which correlates with exposure (TIMP-3) or outcomes (IS, ICH). We defined confounders as carotid plaque formation, neuroinflammation, atrial fibrillation, coronary artery disease, blood pressure, type II diabetes, hypertension, systolic and diastolic blood pressure, high density lipoprotein levels, low density lipoprotein levels, and triglyceride levels (Chauhan and Debette, 2016; Cai et al., 2019). We used the PhenoScanner website (Staley et al., 2016) to query all related phenotypes of the included SNPs and found no phenotypes related to the confounding factors (p < 1 × 10^–5) (Cai et al., 2019). Finally, instrumental variables should not be other methods related to an outcome other than the pathway of exposure (Supplementary Table S2). We conducted strict tests for heterogeneity and pluripotency and found no other pathways that were statistically significant; that is, there were no other pathways from SNPs to the outcome (IS and ICH) other than the exposure (TIMP-3) (Emdin et al., 2017).

The R studio software with two packages, “Mendelian Randomisation” and “TwoSampleMR”, was used to conduct the MR analysis. The principal analyses for TIMP-3, IS, and ICH were performed using the random-effects model IVW approach. The statistical tests to estimate causal relationships were regarded as statistically significant at a Bonferroni corrected p < 0.007 (0.05/7; p = 0.05 adjusted for seven tests). p-value was set at 0.05 for statistical significance for both pleiotropy and heterogeneity tests.

As shown in Supplementary Table S1, from the GWAS data of TIMP-3, we identified 10 statistically significant instrumental variables, namely rs135150, rs137487, rs17303757, rs2097326, rs241890, rs34843069, rs3788495, rs5754249, rs5754266, and rs9422881. However, we could not find rs17303757 in the IS GWAS data. rs17303757 or rs34843069 in SVS data or rs17303757 or rs5754266 in ICH subtypes data. Therefore, IS had only nine valid instrumental variables while SVS, ICH, and the ICH subtypes had only eight instrumental variables. Although a more relaxed p-value threshold (p < 1 × 10^–5) for the inclusion of instrumental variables was adapted, as the F statistic was >10 (Supplementary Table S1), the potential of weak instrumental variable bias was ruled out.

Genetically proxied serum TIMP-3 level was positively associated with IS, especially SVS. The odds ratio (OR) of the genetically evaluated causal role of TIMP-3 in the risk of IS was 1.0262 (95% CI: 1.0072–1.0455, p = 0.0067) in the random-effect IVW model (Figure 2; Figure 3). There was also robust evidence of a positive correlation in MR. RAPS (OR = 1.0240, 95% CI: 1.0050–1.0430, p = 0.011) and MR-PRESSO (OR = 1.0260, 95% CI: 1.0070–1.0450, p = 0.0266). However, lower precision was found in WME (OR = 1.0164, 95% CI: 0.9980–1.0351, p = 0.0801) and MR-Egger regression (OR = 1.0117, 95% CI: 0.9718–1.0531, p = 0.589).

Likewise, in the SVS subtype, the random-effect IVW model suggested that increased serum TIMP-3 increases the risk of SVS (OR = 1.0454, 95% CI: 1.0157–1.0758, p = 0.0025). The MR. RAPS (OR = 1.0398, 95% CI: 1.0101–1.0703, p = 0.0082) and MR-PRESSO (OR = 1.0450, 95% CI: 1.0150–1.0703, p = 0.0192) approaches also supported the causal relationship between serum TIMP-3 and SVS. We found weak evidence for the relationship between TIMP-3 and the risk of CES (OR = 1.0495, 95% CI: 1.0063–1.0945, p = 0.024; Figure 2 and Supplementary Figure S1). However, after Bonferroni correction for multiple comparisons, the relationship was not significant. The causal role of serum TIMP-3 in LAS (OR = 1.0027, 95% CI: 0.9755–1.0306, p = 0.849) could not be confirmed by MR analysis.

In addition to finding a causal relationship between TIMP-3 and IS and its subtypes, we also identified genetic loci closely associated with TIMP-3, which may help with the prevention of IS in the future. Among them, two genetic loci, rs135150 and rs3788495 were positively related to IS. Moreover, rs135150 was also associated with SVS. Four genetic loci, rs135150, rs34843069, rs3788495, and rs241890 were positively associated with CES (Supplementary Figure S2).

In contrast to IS, the causal correlation between serum TIMP-3 and ICH was not found by IVW (OR = 0.9900, 95% CI: 0.9403–1.0423, p = 0.701), MR. RAPS (OR = 0.9925, 95% CI: 0.9051–1.0881, p = 0.8729), MR-PRESSO (OR = 0.9890, 95% CI: 0.9400–1.0420, p = 0.7131), WME (OR = 1.0054, 95% CI: 0.9063–1.1154, p = 0.918), or MR-Egger regression (OR = 0.9855, 95% CI: 0.8224–1.1809, p = 0.8791). The genetically predicted causal role of TIMP-3 in ICH subtypes was similar to that in ICH, and no causal relationship was observed in any one of the five models (Figure 3).

As shown in Table 1, heterogeneity was not found in ICH (p = 0.936) or in either of its subtypes, LICH (p = 0.954) and NLICH (p = 0.263). In ischaemic stroke and its subtypes, we could not identify heterogeneity in IS (p = 0.069), LAS (p = 0.709), or SVS (p = 0.385), but heterogeneity was revealed in the CES subtype (p = 0.022). Therefore, the random effects model, IVW, was implemented in this study to analyse causality.

The pleiotropy of causality in TIMP-3 with IS (intercept = 0.0077, p = 0.454) and ICH (intercept = 0.0026, p = 0.956) were not detected in the MR-Egger regression (Table 2). Moreover, in the MR-PRESSO global test, the pleiotropy of the causal relationship between TIMP-3 and IS (p = 0.134) and ICH (p = 0.945) was also not evidenced.