PheWAS-based clustering of Mendelian Randomisation instruments reveals distinct mechanism-specific causal effects between obesity and educational attainment

Genome-wide association studies have revealed many pleiotropic variants and improved downstream analyses including causal inference via Mendelian Randomisation (MR). MR relies on three core assumptions: relevance, exchangeability (no confounder of instrument–outcome), and exclusion restriction (no pathway other than through the exposure). Horizontal pleiotropy, correlated pleiotropy via heritable confounders, dynastic effects, assortative mating, and population stratification can violate these assumptions and bias MR estimates. Family-based designs, especially within-sibling analyses, mitigate such confounding but are limited in availability. Another complication is heterogeneity of causal effects due to distinct biological mechanisms or subtypes of exposures. These issues can yield variable causal estimates depending on the instruments used. The study introduces PheWAS-driven clustering of instruments (PWC-MR) to identify instrument subgroups that may reflect distinct mechanisms or confounding pathways, and applies it to the BMI→EDU relationship, which shows a strong negative effect in population samples but is far attenuated in within-family analyses, likely due to parental socio-economic position (SEP) confounding. The goal is to dissect mechanism-specific effects and distinguish confounding pathways from true causal effects.

The paper reviews MR methodology and extensions addressing pleiotropy (e.g., MR-Egger under the InSIDE assumption) and notes the risks of correlated pleiotropy when InSIDE is violated, often due to heritable confounders. It discusses biases from dynastic effects, assortative mating, and population stratification, and the utility of family-based designs to reduce these biases. The literature also highlights heterogeneous causal effects due to different biological pathways or exposure subtypes. Prior work has questioned the large negative BMI→EDU effect seen in population-based MR, with within-family analyses showing attenuation. Existing clustering approaches such as MR-Clust (clustering variants by similar causal estimates) and NAvMix (directional clustering using selected traits) have been proposed to identify mechanisms; however, each has limitations, including sensitivity to outcome choice or computational issues with many traits. These studies motivate an approach that leverages broad PheWAS data to inform instrument clustering independent of the outcome.

Exposure and instruments: The primary exposure is adult BMI using Neale Lab UK Biobank GWAS summary statistics. Genome-wide significant SNPs (p < 5e-8) were LD-clumped (clump_kb=10000, clump_r2=0.001, EUR) to obtain independent instruments (initially 348, reduced to 324 after QC). PheWAS trait panel: 1,480 UK Biobank traits processed to 424 by removing missing, duplicates, and those with effective N < 50,000; further filtered by genetic correlation with BMI (r_g > 0.75 removed) to 407 traits. Standardized SNP–trait effect matrix constructed (SD/SD scale). QC of instruments included a Steiger-like test across all traits to exclude SNPs with significantly stronger association (explained variance) with another trait than BMI (Bonferroni p < 0.05/407), removing 24 SNPs. Instrument clustering (PWC-MR): Absolute standardized effects were used; per-SNP normalization across traits (unit variance). K-means clustering was run comparing 2–50 clusters; the Akaike Information Criterion selected six clusters. Trait enrichment: For each trait t and cluster j, compute average per-SNP squared effect σ^2_tj over SNPs in cluster j; define enrichment ratio ER R_tj = σ^2_tj / sum_k σ^2_tk across all clusters. For each cluster, top ER traits (top 10) were used to annotate clusters. Cluster-specific MR: For each cluster, perform IVW MR to estimate BMI→EDU using SNPs in that cluster; compare to the IVW estimate using all BMI instruments. Heterogeneity assessed by Cochran’s Q and average per-IV variance within clusters. Sensitivity analyses: (i) Use within-sibling (sib-regression) GWAS-based SNP effects for BMI to estimate BMI→EDU MR; (ii) Replace exposure with childhood BMI (cBMI) proxied by UKB “Comparative body size at age 10”; select genome-wide significant IVs for cBMI (171; 16 removed for stronger association with other traits; 155 retained), construct standardized SNP–trait matrix across 461 traits, and perform K-means clustering (AIC-optimal 4 clusters), compute cluster-specific MR on EDU; (iii) Replace outcome with systolic blood pressure (SBP) and compute BMI cluster-specific MR to assess homogeneity in a presumed less-confounded relationship. Systematic confounder search: From the 407 traits (excluding those highly genetically correlated with EDU), perform bidirectional MR between each trait T and BMI, and between T and EDU, using multiple MR methods and selecting the second-most significant method’s estimate to ensure robustness. Use a one-sided t-test comparing |β_A→B − β_B→A|/sqrt(SE^2) to infer dominant direction; classify traits into confounders (stronger T→BMI and T→EDU than reverse, and nominal MR p < 0.05 for both), mediators, colliders, etc. Identify candidate confounders. Multivariable MR (MVMR): Conduct stepwise MVMR across candidate confounders to select those with strong effects on EDU (Bonferroni p < 0.05/number tested) and sufficient instruments (≥3 independent genome-wide significant SNPs). Then run standard two-sample MVMR including BMI and selected confounders, compute conditional F-statistics to gauge instrument strength, and report BMI’s conditional causal effect on EDU under combinations that balance weak instrument bias and omitted confounder bias (selected combination with BMI conditional F ≈ 10.19). Tissue colocalization: For each BMI IV locus (±400 kb), perform Bayesian colocalization (PP ≥ 0.8) of BMI and eQTL signals in subcutaneous adipose and brain tissues; test overlap of cluster membership with colocalization tissue using Fisher’s exact tests. Comparison with other clustering: Apply MR-Clust to BMI→EDU, compare resulting clusters and their mean effects and trait associations with PWC-MR clusters (contingency via Fisher’s tests). Software: TwoSampleMR R package used for clumping and MR; k-means via standard algorithms; heterogeneity via Cochran’s Q.

- Overall population-based MR for BMI→EDU: IVW α = −0.19 (95% CI: −0.22, −0.16). - PWC-MR identified six BMI instrument clusters (sizes: 32, 98, 35, 41, 69, 49). Trait enrichment: cluster #2 enriched for lean/lean-mass traits (e.g., trunk/whole-body fat-free mass); cluster #3 for blood/body stature; cluster #4 strongly enriched for socio-economic position (SEP)-related traits (e.g., job physical work, time outdoors, fluid intelligence); cluster #6 for nutrient/supplement traits (e.g., folate, potassium). - Cluster-specific BMI→EDU MR estimates were highly heterogeneous (Q = 130.61, p < 1e−300). Notable estimates: • Cluster #2 (lean-mass enriched): −0.09 (p = 1.23×10−5) • Cluster #5: −0.12 (p = 5.22×10−5) • Cluster #1: −0.44 (p = 7.78×10−20) • Cluster #4 (SEP-enriched): −0.49 (p = 1.63×10−11) All clusters were less heterogeneous internally than using all IVs combined. - Within-sibling GWAS MR for BMI→EDU: −0.05 (95% CI: −0.09, −0.01), significantly different from population-based MR (difference p < 0.001). - Childhood BMI (cBMI) PWC-MR: 155 IVs, four clusters (sizes 37, 42, 32, 44). Cluster enrichment limited; cluster #4 enriched for body-measurement/fat-mass traits; cluster #2 had mild SEP-related enrichment. Cluster-specific effects were homogeneous (Q = 3.84, p = 0.43). Notable cluster estimates: cluster #2 −0.09 (95% CI: −0.1638, −0.0148); cluster #4 −0.04 (95% CI: −0.0823, −0.0024). Overall cBMI→EDU effect attenuated (e.g., −0.03 [−0.06, −0.002]). - BMI→SBP as a positive-control outcome: cluster-specific estimates were homogeneous (Q = 4.49, p = 0.61); IVW using all IVs ≈ 0.15 (p = 1.09×10−28). - Systematic confounder search identified 19 candidate confounders (many environmental/behavioral proxies linked to SEP). After requiring adequate instruments, 12 traits remained; stepwise MVMR retained four with Bonferroni-significant effects on EDU (including time spent watching TV, usual walking pace, past tobacco smoking, cereal type: muesli). The final MVMR specification used three confounders (TV watching, past tobacco smoking, muesli consumption) achieving BMI conditional F ≈ 10.19; conditional MVMR estimates (Table 1): • Time spent watching TV: −0.2771 (SE 0.0256; p = 4.63×10−25) • Past tobacco smoking: 0.1592 (SE 0.0218; p = 7.85×10−13) • Cereal type: muesli: 0.2930 (SE 0.0383; p = 7.96×10−14) • BMI conditional effect on EDU: −0.0455 (SE 0.0106; p = 2.07×10−5) This represents strong attenuation relative to univariable MR (−0.19; p = 7.11×10−41). - MR-Clust comparison (BMI→EDU): two main clusters (means −0.496 and −0.246) plus a null cluster; trait associations suggest stronger SEP enrichment in MR-Clust cluster #1. Cross-tab with PWC-MR showed partial concordance, especially between MR-Clust cluster #1 and PWC-MR clusters #1 and #4 (SEP-enriched). - Tissue colocalization: No significant association between PWC-MR cluster membership and colocalization in adipose or brain tissues (Table 3).

Findings indicate that the large negative BMI→EDU effect observed in population-based MR is driven by subsets of instruments associated with SEP-related traits, likely reflecting dynastic or environmental confounding (correlated pleiotropy), violating MR assumptions. PWC-MR separated instruments into biologically interpretable clusters: a lean-mass/body-measurement cluster yielded a modest, near-zero effect consistent with within-sibling MR and with MVMR conditioning on SEP-proxy traits, whereas SEP-enriched clusters generated large negative effects. Sensitivity analyses support specificity of the approach: for BMI→SBP, where confounding is less problematic, cluster-specific estimates were homogeneous and matched the overall effect, indicating PWC-MR does not artifactually induce heterogeneity. Childhood BMI analyses further suggest that childhood adiposity genetics are less entangled with SEP, yielding homogeneous, attenuated estimates. Comparative analyses with MR-Clust underscore that PWC-MR’s use of outcome-agnostic PheWAS information can produce clusters that map onto plausible mechanisms or confounding pathways, offering interpretability and guidance for follow-up causal modeling. Overall, results address the research question by revealing mechanism-specific and confounder-linked heterogeneity in MR instruments, refining causal inference for BMI→EDU.

The study introduces PWC-MR, a PheWAS-driven instrument clustering framework that reveals distinct, mechanism-specific causal effects and highlights confounding pathways. Applied to BMI→EDU, PWC-MR discovered six instrument clusters with heterogeneous effects; clusters enriched for SEP traits accounted for the unrealistically large negative effects seen in standard MR, while a lean-mass/body-measurement cluster yielded a more plausible, attenuated effect close to within-sibling and MVMR estimates. Childhood BMI analyses reinforced the near-null conclusion and displayed homogeneous cluster effects, and BMI→SBP showed homogeneity consistent with minimal confounding. PWC-MR complements within-family and multivariable MR by not requiring family data or chronologically ordered exposures and by offering multiple, interpretable cluster-specific estimates. Future work could expand trait panels for PheWAS to capture additional biological mechanisms, integrate richer tissue- or cell-type-specific functional genomics to aid cluster interpretation, and apply PWC-MR to other exposure–outcome pairs to systematically identify mechanism-specific effects and confounding pathways.

- Dependence on availability and choice of PheWAS traits: limited trait coverage may miss key pathways, influencing detected clusters. - Treatment of binary traits as continuous; although effects correlate in large samples, residual differences could affect clustering. - Arbitrary thresholds (e.g., genetic correlation cutoff r_g ≤ 0.75; p-value thresholds at 5%) may affect trait inclusion and cluster structure. - Candidate confounders used in MVMR likely act as proxies for true confounders (e.g., proxies of parental SEP), not the causal factors themselves. - Ambiguity in identifying which clusters capture true causal mechanisms versus confounding in general applications without external validation. - Alternative exclusion tests or MR methods could be used; choices here were pragmatic and may influence pre-selection but are unlikely to change main conclusions. - Colocalization analyses may be underpowered due to limited eQTL sample sizes and high false-negative rates, limiting tissue-level validation. - Computational constraints prevented application of some alternative clustering methods (e.g., NAvMix) at scale with many traits.