Causal Factors for Osteoarthritis: A Scoping Review of Mendelian Randomization Studies

Abstract

Background:

Mendelian randomization (MR) has increasingly been utilized as a tool for establishing causal relations between modifiable exposures and osteoarthritis (OA).

Objective:

The goal of this review was to summarize available MR studies of OA that evaluate the causal role of modifiable risk factors on OA.

Methods:

This review was performed following the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews) model. We performed a literature search for relevant studies published before December 2021 across multiple databases using the search terms “Osteoarthritis” and (“Mendelian Randomization” or Polygenic risk score”). We reported the MR estimates of causal associations between exposures and OA and then assessed methodologic quality of abstracted studies according to their efforts to validate the three key MR assumptions.

Results:

Our search identified 45 studies reporting on 141 exposure-association analyses. All studies performed a formal instrumental variable analysis to estimate the causal effect of exposure on OA. Causal associations (p < 0.05) were reported in 60 of these analyses representing 36 unique publications and MR Egger sensitivity analyses were performed in 45 of these analyses. MR studies provided support for causal associations of OA with increased levels of adiposity, coffee consumption, bone mineral density, and sleep disturbance and decreased levels of serum calcium and LDL-cholesterol.

Conclusions:

These results highlight the potential benefits of weight reduction and improvement of sleep quality to reduce risk of OA and call for a better understanding of the relation of coffee consumption of serum calcium to OA risk.

Graphical Abstract

Osteoarthritis (OA), which is characterized by progressive cartilage degradation and subchondral bone remodeling (1), is the most common musculoskeletal disease in older adults (2). It is considered one of the leading causes of pain and disability worldwide (3). Current management of OA focuses on relieving symptoms and improving physical activity and quality of life through a combination of non-pharmacologic and pharmacologic modalities (4, 5). The comprehensive management of OA, however, is shifting its focus towards preventing the disease by risk factor modification. Although epidemiologic studies have identified numerous factors associated with OA, the causal connection of many of these factors with OA is unclear (6, 7). In recent years, Mendelian randomization (MR) has been utilized as a tool for establishing causal relations between ‘exposure’ variables and disease outcomes (8, 9). MR is an analytic method that uses genetic variants influencing modifiable risk factors (i.e., ‘genetic instruments’) to test for causal associations with a disease outcome. Since genetic variants are present at birth and precede the development of the disease, the use of MR avoids biased associations that are caused by confounding or reverse causation. Evidence for an association between the genetically proxied exposure and disease outcome can thus be used for evidence of a causal relationship between exposure and outcome.

Numerous MR studies have been conducted over the past 10 years to identify causal factors associated with OA. These studies have been enabled by the rapid proliferation of large genome-wide association studies (GWAS) of osteoarthritis risk factors and the availability of summary results from these studies, allowing the construction of genetic instruments for these variables for MR studies. In this scoping review, we review all MR studies published before December 2021, organize these by type of risk factor, and provide a methodological quality assessment and brief qualitative summary of results in terms of their support for or against causal relations with osteoarthritis.

MATERIALS AND METHODS

This review was performed following the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews) model (10). Briefly, our review protocol addresses selection of articles, extraction of key study variables, summary of results, and evaluation of supporting evidence.

We selected articles through a structured review of studies that were indexed in Pubmed, Cochrane, Scopus, and Embase under the Medical Subject Heading term “Osteoarthritis” and (“Mendelian Randomization” or Polygenic risk score”). Searches were conducted to identify studies published through December 2021 and were limited to articles published in English language and human studies. Two authors (E.A. and B.D.M.) reviewed all titles for relevance to osteoarthritis and Mendelian Randomization, and a third author (M.C.H.) adjudicated any discrepancies. Articles were selected if they conducted an MR analysis with osteoarthritis as the outcome trait. Two authors (K.N. or B.D.M.) extracted the following variables from each study: base trait, genetic instrument (including # of SNPs), target dataset/study outcome, and the MR effect of exposure on outcome.

The literature review identified 162 titles; 82 duplicate titles were removed. Eighty abstracts were screened; 15 irrelevant abstracts were excluded (articles not related to OA, not using MR or polygenic risk scores, not using primary data). We also excluded review articles, correspondence and replies, and abstracts (19 articles) and one article that considered OA as a exposure variable. Forty-five full-text articles were included in the final review (see PRISMA 2020 flow diagram, Figure 1). For each included article, we extracted the exposure(s) or risk factor(s) tested, information about the genetic instrument, the target dataset (i.e., population to which the genetic instrument was applied), and the main study result. Information about the genetic instrument included the source of the GWAS upon which the genetic instrument was based, the ancestry of this population, and the number of single nucleotide polymorphisms (SNPs) included in the instrument. We also provide a brief evaluation of the evidence supporting associations reported between each risk factor and outcome.

Figure 1.

Identification of studies via databases and registers

The three assumptions underlying MR studies are that: 1) the genetic variants are associated with the risk factor of interest (the relevance assumption); 2) there are no unmeasured confounders that are associated with both the genetic variants and the outcome (the independence assumption); and 3) the genetic variants affect the outcome only through their effect on the risk factor of interest (no horizontal pleiotropy assumption) (11). The relevance assumption relates to the strength of the instrument, which can be estimated by its correlation with the risk factor or by an F statistic, which is a function of the proportion of the variance explained by the genetic instrument, the number of instruments (variants), and the sample size. The presence of unmeasured confounders (assumption 2) can be addressed by evaluating associations of the instrument with study covariates, assessing for heterogeneity of causal effects, and consistency of associations across subgroups. The independence assumption refers to whether the genetic variants affect the risk of OA through pathways independent of the considered exposure. The presence of (horizontal) pleiotropy is typically evaluated using sensitivity analyses such as weighted median MR and MR Egger regression that evaluate whether genetic variants have pleiotropic effects on the outcome that differ on average from zero (directional pleiotropy) and that provide consistent estimates of the causal effect. In the description of study findings, we comment on the degree to which these assumptions were addressed.

Because there are no standardized tools to assess quality of MR studies, we followed the approach of several prior reviews that assessed quality of MR studies based on validation of the Mendelian randomization assumptions through sensitivity analyses (12, 13). For assumption 1 (the relevance assumption evaluating the strength of the genetic instrument), we rated validation to be “good” if the instrument had been associated with the exposure through a prior GWAS and met genome-wide thresholds for statistical significance (i.e., 5 × 10 ⁻⁸) or if each variant included in the instrument was assessed to be strong by an F-statistic > 10. We rated validation of the relevance assumption to be “moderate” if strength of the instrument was validated in some other way, and “poor” if the instrument was not validated or strength was not reported.

Violations of assumptions 2 (no confounding) and 3 (no horizontal pleiotropy) are difficult to rule out but sensitivity analyses can be performed that can detect overt violations of these assumptions. Because the same sensitivity analyses can detect violation of either assumption, we lumped evaluation of these 2 assumptions together. We considered four types of sensitivity analyses in our evaluation: (1) formal MR-Egger/PRESSO analysis to detect potential pleiotropy; (2) evaluation of heterogeneity of variant effects; (3) leave-one-out analysis to detect outlier variants; and (4) other methods (e.g,, whether the study used positive/negative controls or employed subgroup analyses). For each analysis in our review, we recorded whether a formal MR-Egger/PRESSO analysis was performed. We rated validation of assumptions 2 and 3 as “good” if formal MR-Egger analysis was performed and the totality of results (directionality of effect sizes, strengths of effects, and Egger intercept) were consistent with those obtained from the primary IV analysis. We also rated the validation of assumptions 2 and 3 as “good” if formal MR-Egger was not performed but the variants were biologically selected and at least one other sensitivity analysis was performed and provided no evidence for violation of these assumptions. We rated validation of the assumptions 2 and 3 as “possible” if consistency of the MR Egger analyses was equivocal or if MR Egger analysis was not carried out but at least one other type of sensitivity analysis was performed (any of criteria 2–4 above) and provided no evidence for violation of these assumptions. We rated validation of the assumptions 2 and 3 to be “poor” if there was insufficient assessment of MR assumptions or if any of the sensitivity analyses provided evidence that the association could be explained by violation of MR assumptions 2 and 3 (eg, heterogeneity among instrumental variable effects from MR Egger or leave-one-out analyses, inconsistent or unexpected results from analyses of subgroups or positive/negative controls, etc.).

RESULTS

Description of included studies

A total of 45 MR studies met our inclusion criteria and are included in this review. Summary characteristics of these studies are provided in Table S1, which indicates for each study the exposure (risk factor) whose causal effect on OA is being tested, the number of variants used in the genetic instrument to characterize that exposure and the source of the SNPs included in the instrument, size of the base genome-wide association analysis from which the instruments were extracted, and the target population to which the instrument has been applied (e.g., OA cases and controls). Thirty-three of the 45 studies utilized OA cases and controls from the UK Biobank, with cases defined as self-reported or hospital-diagnosed OA, or history of joint replacement.

We reviewed a total of 141 exposure-association analyses reported across the 45 MR studies. All studies performed a formal instrumental variable analysis to estimate the causal effect of exposure on OA. Of the 141 analyses reviewed, 105 employed a 2-sample inverse- variance-weighted (IVW) MR approach, in which the genetic instrument was derived from one sample and then applied to subjects from another sample, typically the UK Biobank. Two studies employed a 1-sample MR design, and the remaining 34 studies employed a 2 sample generalized summary data-based Mendelian randomization (GSMR) design in which the analyses require summary-level data only and leverage linkage disequilibrium between genetic variants from a reference sample (14). Two-sample MR designs are attractive when very large GWAS of the exposure have previously been carried out so that the effect sizes of variants included in the instrument developed from this resource are estimated in an independent sample, typically with more precision, than the target population to which the instrument is to be applied. With only a few exceptions, studies used as genetic instruments only variants that were reported as being of genome-wide significance from a prior genome-wide association study (GWAS). For nearly all cases, the genetic instruments were derived from populations that were predominantly of European ancestry and the target populations were also of European ancestry.

Assessment of MR assumptions

We reviewed a total of 141 exposure-association analyses reported across the 45 MR studies. Association results for these analyses are summarized in Table S2. Causal associations (p < 0.05) were reported in 60 of these analyses representing 36 unique publications. Results for adiposity-related exposure associations (n = 19) are summarized in Table 1, results for cardiovascular, inflammation, hormone, and nutrient-related exposure associations (n = 20) in Table 2, and results for lifestyle-related and other exposure associations (n = 21) in Table 3. Assessment of the methodologic quality of these studies is also provided in these tables. We rated the evidence that MR assumption 1 held as “good” for 51 of these analyses and “poor” for only 1 of the analyses. MR Egger sensitivity analyses were performed in 45 of these analyses. We found no evidence no evidence for violation of MR assumptions 2 and 3 for 31 of these associations, “some” evidence for assumption violations for 21 associations, and strong evidence for assumption violations for 8 associations.

Table 1.

Quality assessment of the three Mendelian randomization assumptions for studies reporting associations for adiposity exposures at p > 0.05

					MR Assumption
Exposure	Study (Ref)	Effect of exposure on outcome (per SD increase)	p-value		1	2/3
						MR Egger performed	Support
Childhood obesity	6 (19)	1.07 (1.05 – 1.10)	7.20E-08		++	++	0
Body mass index	9 (20)	1.57 (1.44 – 1.71)	2.99E-24		++	++	++
Body mass index	10 (21)	1.82 (1.73 – 1.92)	5.3E-118		++	++	++
Body mass index	13 (22)	1.03 (1.02 – 1.04) *	1.25E-13		++	++	++
Body mass index	14 (23)	1.39 (1.05–1.86)	0.025		++	++	++
Body mass index	17 (24)	1.52 (1.37 – 1.68)	8.10E-16		++	++	++
Body mass index	22 (25)	1.72 (1.37 – 2.16)	2.00E-06		++	++	++
Body mass index	37 (26)	1.61 (1.35 – 1.92)	8.34E-07		++	++	++
Body mass index	45 (14)	1.50 (1.34 – 1.68)	4.08E-12		++	0	+
Body Fat Percentage	22 (25)	1.82 (1.73 – 1.91)	1.0E-124		++	++	++
“favourable” adiposity	22 (25)	1.45 (1.19 – 1.76)	2.00E-04		++	++	+
“unfavourable” adiposity	22 (25)	2.20 (1.64 – 2.95)	1.00E-07		++	++	++
Metabolically ‘unfavourable” profile (high BMI and high WHR)	23 (27)	1.56 (1.31 – 1.85)	NP		+	++	+
Metabolically “neutral” (high BMI, but no assoc with WHR)	23 (27)	1.60 (1.15 – 2.23)	NP		+	++	+
Extreme body mass index	37 (26)	1.04 (1.02 – 1.07)	3.04E-03		++	0	0
Overweight	37 (26)	1.10 (1.06 – 1.15)	2.41E-04		++	++	++
Height	45 (14)	1.09 (1.05 – 1.12)	1.83E-06		++	0	+
Waist circumference	37 (26)	1.51 (1.21 – 1.88)	4.74E-04		++	++	++
Hip circumference	37 (26)	1.33 (1.04 – 1.68)	0.021		++	++	++

^*effect size reported for a one unit increase in BMI.

MR assumption 1: genetic variant is associated with exposure. ‘++’ = good evidence that this assumption holds; ‘+’ = moderate evidence; ‘0’ = assumption not validated or evidence not reported.

MR assumptions 2 and 3: MR Egger analysis performed: ‘++’ = yes; ‘0’ = no.

Support: ‘++’ = no evidence from sensitivity analyses for violation of assumptions; ‘+’ = some evidence from sensitivity analyses for violation of assumptions; ‘0’ = strong evidence from sensitivity analyses for violation of assumptions (see text for details).

Table 2.

Quality assessment of the three Mendelian randomization assumptions for studies reporting associations for cardiovascular, inflammation, hormone, and nutrient-related exposures at p > 0.05

					MR Assumption
Exposure	Study	Effect of exposure on outcome (per SD increase)	p-value		1	2/3
						MR Egger performed	Support
Cardiovascular
Systolic blood pressure	9 (20)	0.76 (0.60 – 0.96)	0.019		++	++	0
LDL-cholesterol	10 (21)	0.94 (0.91 – 0.98)	0.003		++	++	++
LDL-cholesterol	14 (23)	0.86 (0.75 – 0.98)	0.029		++	++	++
LDL-cholesterol	14 (23)	0.86 (0.79 – 0.95)	0.002		++	0	+
LDL-cholesterol	45 (14)	0.96 (0.92 – 1.00)	0.04		++	0	+
Hormone
Leptin	8 (28)	2.40 (1.13 – 5.09)*	0.023		++	++	0
Insulin-like growth factor	11 (29)	1.49 (1.21 – 1.83)	0.0001		++	++	++
Parathyroid hormone	16 (30)	0.67 (0.50 – 0.90)	0.008		++	++	+
Parathyroid hormone	30 (31)	0.71 (0.61 – 0.82)	6.23E-06		++	++	++
SHBG	28 (32)	1.09 (1.01 – 1.17)	0.027		++	++	+
Inflammation
CCL23	7 (33)	0.98	p ≤ 0.05		+	0	0
MMP−1	7 (33)	1.02	p ≤ 0.05		+	0	0
LAP−TGF− B−1	7 (33)	0.94 (0.91 – 0.98)	0.0009		+	++	+
C-reactive protein	27 (34)	1.17 (1.01 – 1.36)	0.04		++	0	0
Nutrients
Calcium	29 (35)	0.71 (0.60 – 0.85)	1.84E-04		++	++	++
Calcium	36 (36)	0.77 (0.61 – 0.98)	0.032		++	++	++
Calcium	42 (37)	0.67 (0.51 – 0.88)	0.004		++	++	++
Copper	43 (38)	1.07 (1.01 – 1.13)	0.01		++	0	+
Copper	44 (39)	1.07 (1.02 – 1.13)	0.01		++	0	+
Zinc	44 (39)	1.07 (1.01 – 1.13)	0.02		++	0	+

^*effect size reported for a one unit increase in natural log-transformed leptin

MR assumption 1: genetic variant is associated with exposure. ‘++’ = good evidence that this assumption holds; ‘+’ = moderate evidence; ‘0’ = assumption not validated or evidence not reported.

MR assumptions 2 and 3: MR Egger analysis performed: ‘++’ = yes; ‘0’ = no.

Support: ‘++’ = no evidence from sensitivity analyses for violation of assumptions; ‘+’ = some evidence from sensitivity analyses for violation of assumptions; ‘0’ = strong evidence from sensitivity analyses for violation of assumptions (see text for details).

Table 3.

Quality assessment of the three Mendelian randomization assumptions for studies reporting associations for lifestyle-related and other exposures at p > 0.05

					MR Assumption
Exposure	Study	Effect of exposure on outcome (per SD increase)	p-value		1	2/3
						MR Egger performed	Support
Sleep
Insomnia	24 (40)	1.22 (1.15 – 1.30)	8.05E-10		++	++	++
Sleep duration	24 (40)	0.99 (0.99 – 1.00)	2.75E-02		++	++	++
Short sleep duration	24 (40)	1.04 (1.02 – 1.07)	2.20E-03		++	++	++
Smoking
Lifetime smoking	10 (21)	2.23 (1.85 – 2.68)	1.24E-17		++	++	++
Smoking quantity (# cigs)	18 (41)	0.84 (0.74 – 0.96)	0.008		++	++	++
Smoking behavior	20 (42)	0.94 (0.90 – 0.99)	0.035		+	++	+
Education
Education (yrs schooling)	15 (43)	0.95 (0.91 – 0.97)	NP		++	++	++
Education (yrs schooling)	37 (26)	0.78 (0.60 – 1.00)	0.048		++	++	+
Education (yrs schooling)	45 (14)	0.84 (0.73 – 0.95)	0.007		++	0	+
Educational attainment	10 (21)	0.59 (0.54 – 0.64)	3.08E-37		++	++	++
Coffee
Coffee consumption	19 (44)	1.46 (1.04 – 2.04)	0.025		+	++	0
Coffee consumption	39 (45)	1.01 (1.00 – 1.02)*	0.006		++	++	++
Coffee consumption	25 (46)	1.23 (1.11 – 1.35)	4.46E-05		++	++	++
Other
Bone mineral density	9 (20)	1.18 (1.04 – 1.35)	2.92E-08		++	++	++
Bone mineral density	12 (18)	1.27 (1.03 – 1.55)	0.024		++	++	+
Ischemic stroke	3 (47)	1.03 (1.00 – 1.07)	0.036		++	++	+
Major depression	1 (48)	1.21 (1.15 – 1.28)	2.67E-13		++	0	+
Major depressive disorder	38 (49)	1.27 (1.18 – 1.37)	2.36E-08		++	0	+
Gut microbiota	35 (50)	NP	< 0.05		+	++	+
Plasma ADAMTS5 levels	40 (16)	0.98 (0.97 – 0.99)	0.005		++	0	++
Social isolation	41 (51)	1.20 (1.10 – 1.31)	NP		++	++	++

^*effect size reported for a 1% increase in coffee consumption.

See footnote in Table 1 and 2 for assessment of MR assumptions.

Summary of association results

MR has been used to test causal associations of OA with a wide range of exposures (Table S2). Summarized below are key results from these analyses. Study numbers are referenced in Table S1.

Adiposity-related exposures:

There is consistent evidence that genetically predicted higher body mass index (BMI) is associated with higher OA risk. Odds ratios among seven studies evaluating the casual association of BMI on OA ranged from 1.39 to 1.82 (Figure 2 and Table 1). One study (#22) created a PRS for subcategories of adiposity and concluded that adiposity is causally associated with OA regardless of whether the adiposity is considered metabolically favorable, metabolically neutral, or metabolically unfavorable, with these categories based on the association of BMI with a panel of other metabolic health indicators, including waist-hip ratio, triglycerides, high density lipoprotein (HDL) and low-density lipoprotein (LDL) cholesterol, glucose and type 2 diabetes mellitus, and coronary artery disease. A second study (#23) also reported associations of ‘metabolically unfavorable’ and ‘metabolically neutral’ adiposity with OA, although the genetic instruments used for these exposures were considered to be of moderate strength only. There is also causal evidence for an effect of genetically predisposed childhood obesity on OA (OR = 1.07; 1.05–1.10), although this association was attenuated and no longer statistically significant after accounting for adult BMI (study #6). Genetically predisposed body fat percentage has also been associated with OA risk (OR = 1.82; 1.73–1.91; study #22), a result perhaps expected given the high correlation of this trait with BMI. Genetically predicted height was associated with increased OA risk in one study (OR = 1.09; 1.05–1.12; study #45), although sensitivity analyses could not rule out potential violation of MR assumptions. Height was but not significantly associated with OA in a second (OR = 0.99; 0.93–1.06; study #37). Genetically predicted waist circumference (OR =1.51; 1.21–1.88; study #37) and hip circumference (OR =1.33; 1.04–1.68; study #37) have each been associated with OA, although not waist-to-hip ratio or birthweight (Figure 2 and Table 1). Taken together, there is strong support for a causal association of adiposity with OA, although it cannot be determined whether the metabolic sequelae accompanying increased adiposity contribute to this association.

Figure 2.

Summary of Mendelian randomization studies assessing associations of adiposity with OA

odds ratios correspond to a 1-SD unit change in the exposure

Cardiovascular, inflammation, hormonal, and nutrient-related exposures:

Lipids:

Four (of five) studies have reported OA risk to be statistically associated with lower genetically predicted levels of LDL cholesterol, with odds ratios of all five studies ranging from 0.85 to 0.96 (Figure 3 and Table 2). Analyses from two of these studies were robust with respect to the MR assumptions, while the other two studies did not include extensive sensitivity analyses that would have addressed more thoroughly the potential for confounding and pleiotropy. OA risk was not associated with genetically predicted levels of HDL cholesterol, total cholesterol, or triglycerides.

Figure 3.

Summary of Mendelian randomization studies assessing associations of lipids with OA

odds ratios correspond to a 1-SD unit change in the exposure

Blood pressure:

MR studies provide little support for a causal association of blood pressure on OA risk (Figure S1). Two studies (#10 and 45) reported virtually no association between genetically determined systolic blood pressure (SBP) and OA risk (Table S2), and a third (#14) provided only modest support (odds ratio (OR) for a 1 standard deviation increase in SBP = 0.89; 0.65–1.23) that was greatly diminished after accounting for potential pleiotropy. A fourth study (#9) observed an initial association of lower SBP with knee OA (OR = 0.76; 0.60–0.96), although the magnitude of effect was reduced and no longer significant after removing SNPs also associated with BMI.

Diabetes:

Neither the genetic risk of diabetes nor genetically predicted glucose levels have been associated with OA. The ORs for diabetes ranged from 0.99 to 1.01 in three studies, and the ORs for glucose levels ranged from 0.96 to 1.20 for a 1-SD change in glucose levels (Figure S1).

Hormones and adipokines:

Higher genetically predicted levels of leptin have been associated with OA risk, although leptin is highly correlated with body mass index and sensitivity analyses could not rule out the possibility of confounding.

In contrast, genetically predicted levels of IGF-1 were robustly associated with higher OA risk, with sensitivity analyses (OR = 1.49; 1.21–1.83), including adjustment for a BMI PRS, suggesting that the IGF-1 PRS association with OA was unlikely to be due to pleiotropy with BMI (study #11). Higher genetically predicted levels of parathyroid hormone have been associated in two studies with lower OA risk (OR = 0.67; 0.50–0.90 and OR = 0.71; 0.61–0.82) with sensitivity analyses in one of these studies also indicating the association unlikely to be due to pleiotropy (studies #16, #30). There is modest support for a causal association of sex hormone-binding globulin with OA risk at the hip, but not the knee (OR = 1.09; 1.01–1.17), although again, further sensitivity analyses were unable to rule out the possibility of the association being secondary to horizontal pleiotropy (study #28).

There was little support for a causal association of OA risk with genetically predicted levels of testosterone, dehydroepiandrosterone sulfate (DEAS), or dihydrotestosterone (DHT) (study #32).

Inflammatory mediators:

To date, no MR studies have reported robust associations with OA risk. In one study, OA was marginally associated with higher genetically predicted levels of C-reactive protein (CRP) based on an 18-SNP genetic instrument (OR = 1.17; 1.01–1.36), but this study did not include a rigorous test of pleiotropy (study #27). However, as a sensitivity analysis that used a more specific instrument containing only 4 SNPs within the CRP gene, thus minimizing the potential for pleiotropy, CRP was not associated with OA (OR = 0.94; 0.78–1.13; #27), nor was CRP associated with OA risk in a third study that excluded SNPs associated with BMI as instrumental variables (study #9). OA risk was also associated with lower genetically predicted levels of transforming growth factor, beta 1 proprotein (LAP-TG-β−1), encoded by TGFB1, although this instrument was based on a GWAS of a small number of individuals (n=872) and no opportunity to replicate associated variants (study #7). In this same study, no associations were reported with genetically predicted levels of a large panel of other inflammatory mediators.

Nutritional factors:

Lower genetically predicted levels of calcium have been associated with OA risk in four studies (ORs ranging from 0.64 to 0.87), with the associations achieving statistical significance in three of these studies (Supplemental Table 2). At least two other MR studies detected significant associations of with higher genetically predicted levels of serum copper (OR = 1.07 in both; studies #43,44), although neither study performed extensive sensitivity analyses. One of these studies (#44) also observed an association between genetically determined zinc levels and higher OA risk (OR = 1.07; 1.01–1.13), although again extensive sensitivity analyses were not performed. Two studies, both utilizing a 6-SNP instrument, reported no association between genetically determined 25-hydroxyvitamin D levels and OA (ORs = 1.02 and 1.03, respectively). Very modest evidence for an association with genetically predicted iron levels with OA were reported (ORs = 1.11; 0.96–1.28 and OR = 1.05; 0.97–1.13 in study #s 44 and 29, respectively).

Lifestyle-related and other exposures:

Sleep:

OA risk has been associated with both genetic risk of insomnia (OR = 1.22; 1.15–1.30) and shorter genetically determined sleep duration (OR = 1.04; 1.02–1.07) (study #24). These analyses were obtained using robust instruments and the associations persisted after adjusting for potential covariates, including BMI, depression, and diabetes. Further sensitivity analyses provided no evidence for pleiotropy. In contrast, no association was observed with genetically determined long sleep duration or in another study, genetically determined sleep duration and OA (study #37).

Smoking:

The causal relation of smoking to OA risk is unclear. The most comprehensive MR study to date (#10) used a 117-SNP genetic instrument and found genetically predicted lifetime smoking to be associated with higher OA risk (OR = 2.23;1.85–2.68). In contrast, two other studies reported protective effects of smoking on OA risk, one using a 4-SNP instrument to proxy smoking quantity (cigarettes per day) (OR = 0.94; 0.90–0.99; #20), and the second finding that among smokers, genetically predicted smoking quantity was associated with lower risk of total joint replacement (OR = 0.84; 0.74–0.96; study #18). To further support this result, this latter study, which used only a single SNP to proxy smoking quantity, found no association of predicted smoking quantity with total joint replacement in non-smokers.

Coffee consumption:

Three studies reported that higher levels of genetically predicted coffee consumption is associated with increased risk of OA, although in one of these studies (#19) the potential for MR assumption violations could not be ruled out. The two remaining studies reported odds ratios of 1.23 (1.11–1.35) for habitual coffee consumption (study #25) and 1.01 (1.00–1.02) for a 1% increase in coffee consumption.

Education and social factors:

Four studies, which utilized instruments including from 66 to 527 variants, have reported genetically predicted levels of education to be associated with lower OA risk (ORs ranging from 0.59 to 0.95; studies #15, 37, 45, and 10). Further analyses revealed these associations to be partially, although not completely, mediated by BMI and smoking.

Genetically predicted levels of social isolation were associated with higher OA risk in one study (OR = 1.20; 1.10–1.31; #41). This study utilized a 5-SNP instrument to proxy social isolation, and sensitivity analyses revealed no evidence for pleiotropy. Mediation analyses were not performed to evaluate mechanism through which social isolation might increase OA risk.

Bone mineral density:

There is strong evidence that genetically predicted higher bone mineral density (BMD) is associated with higher OA risk. These findings are consistent with the clinical observations of a high bone mass OA phenotype and the inverse correlation between osteoporosis and OA in older adults. Two of these studies used 2-sample MR to enable generation of PRS scores from the large GEFOS consortium meta-analysis of BMD, for which ORs of 1.18 (1.04–1.35) and 1.27 (1.03–1.55) were estimated for knee and hip OA, respectively (study #s 9 and 12; Table S2). One of these studies utilized multivariate MR analysis to show that the association of genetically predicted BMD with OA was at least partially independent of BMI (study #12). A third study reported modest evidence between lumbar spine BMD and OA (OR = 1.06; 0.95–1.17).

Other exposures:

MR analyses of OA with a host of other exposures have been carried out, including with gout and urate levels, periodontitis, telomere length, the microbiome, cardiovascular disease and cancer, and major depression. Results from these studies have revealed little support for a causal association of these risk factors with OA. In contrast, missense variants in the gene, A disintegrin and metalloproteinase with thrombospondin motifs 5 (ADAMTS5), have been associated with higher ADAMTS35 levels and lower OA risk (OR = 0.98; 0.97–0.99), a finding that is consistent with animal models showing that ADAM-TS5 is necessary for the development of OA (15) under the assumption that higher ADAMTS35 levels reflect reduced function (16).

DISCUSSION

To our knowledge, this report provides the first scoping review and systematic compilation of Mendelian randomization studies that have evaluated associations between OA risk factors and OA. Nearly all studies used 2-sample Mendelian randomization in which the genetic instruments were constructed from one population and then tested against OA cases and controls from another – generally, the UK Biobank, arcOGEN, or the Genetics of Osteoarthritis Consortium. In general, MR studies provided support for causal associations of OA with increased levels of adiposity, coffee consumption, bone mineral density, sleep disturbance and decreased levels of serum calcium and LDL-cholesterol. These results highlight the potential benefits of weight reduction and improvement of sleep quality to reduce risk of OA and call for a better understanding of the relation of coffee consumption of serum calcium to OA risk.

There are some caveats to interpreting these findings. First, although some risk factors were evaluated in multiple studies, the data sets utilized and types of analyses were not always completely independent. For example, some studies used the same instruments but applied them to different target populations or to different subsets of the same target population. Thus, multiple studies of the same risk factor cannot necessarily be regarded as replication analyses. Second, there was a wide variation in the number of variants included in the different genetic instruments, so that some instruments were stronger than others, thus providing greater power to detect associations. Moreover, available GWAS results for some risk factors (eg, nutritional factors, inflammatory mediators, hormones, microbiome, and sleep) were based on relatively small sample sizes, again limiting the power of the genetic instruments. Future studies based on more powerful instruments will provide more informative tests of the risk factor – OA associations.

There was also some variability in how (and whether) studies addressed the assumptions underlying MR. For example, not all studies provided metrics about instrument strength, and different metrics and analytic approaches were used to evaluate the potential for horizontal pleiotropy. Importantly, this review highlights the need for future studies to provide explicit evaluations of MR study assumptions.

The application of Mendelian randomization analysis to OA (as well as to other disorders) is in the early stages, and there are numerous opportunities for progress in this area. First, as noted above, the growing availability of more powerful genetic instruments will increase power to detect risk factor – OA associations. Second, the growing availability of GWAS of new risk factors will provide opportunities to test new hypotheses regarding the genetic determinants of OA. Third, there remains much methodologic work to be done in sorting out the relationships of OA risk factors with each other and how they may influence OA risk in combination. There are pleiotropic relationships among many of these risk factors, particularly those related to sex hormones, body weight and its distribution and bone mass, and detailed mediating analyses are required to sort out which factors drive OA risk. Examples of the issues that have been recently addressed in the literature include sorting out the independent contributions of smoking and obesity on OA risk and bone mineral density and obesity on OA risk (17, 18). Such determinations are important in the choice of factors that will be the focus of ongoing and future primary prevention public health initiatives. In conclusion, Mendelian randomization studies provide useful insights into the causes of OA, with the hope that these insights may point to new therapeutic targets for OA treatment or prevention.

Supplementary Material

Fig S1

Figure S1. Summary of Mendelian randomization studies assessing associations of blood pressure and diabetes/glucose with OA

odds ratios correspond to a 1-SD unit change in the exposure (or odds of T2D)

Significance and Innovations.

• Mendelian randomization is increasingly being used to assess causal associations of risk factors with osteoarthritis

• This review fills a gap by summarizing results of Mendelian randomization studies carried out to date