Abstract
Mendelian Randomization (MR) enables estimation of causal effects while controlling for unmeasured confounding factors. However, traditional MR’s reliance on strong parametric assumptions can introduce bias if these are violated. We introduce a new machine learning MR estimator named Quantile Instrumental Variable (IV) that achieves low estimation error in a wide range of plausible MR scenarios. Quantile IV is distinctive in its ability to estimate nonlinear and heterogeneous causal effects and offers a flexible approach for subgroup analysis. Applying Quantile IV, we investigate the impact of circulating sclerostin levels on heel bone mineral density, osteoporosis, and cardiovascular outcomes in the UK Biobank. Employing various MR estimators and colocalization techniques that allow multiple causal variants, our analysis reveals that a genetically predicted reduction in sclerostin levels significantly increases heel bone mineral density and reduces the risk of osteoporosis, while showing no discernible effect on ischemic cardiovascular diseases. Quantile IV contributes to the advancement of MR methodology, and the case study on the impact of circulating sclerostin modulation contributes to our understanding of the on-target effects of sclerostin inhibition.
Introduction
Mendelian Randomization (MR) serves as a critical tool in epidemiology, utilizing genetic variants as instrumental variables to infer causal relationships between exposures and outcomes. However, traditional MR methods often assume linear relationships, which may fail to capture the biological complexity of thresholds, diminishing returns, or heterogeneous effects across different populations. This paper introduces Quantile IV, a machine learning-based estimator designed to overcome these limitations by modeling non-linear and heterogeneous causal effects without the rigid parametric assumptions of classical methods.
Methodology: The Quantile IV Estimator
Quantile IV is a non-parametric estimator built on neural network architectures. It represents a streamlined evolution of previous deep-learning approaches like DeepIV. While traditional non-parametric methods attempt to model the entire conditional distribution of an exposure—a process often plagued by computational noise—Quantile IV focuses on predicting specific quantiles (e.g., the 10th, 50th, and 90th percentiles) of the exposure.
The estimator operates in two stages:
- First Stage: A neural network learns the relationship between genetic instruments and the exposure using quantile regression.
- Second Stage: These predicted values are utilized to estimate the causal effect on the outcome. This framework allows for the estimation of Conditional Average Treatment Effects (CATE), enabling researchers to identify how causal impacts might vary across subgroups, such as by sex or age.
Simulation Benchmarking
To validate the performance of Quantile IV, the researchers conducted extensive simulations across 15 distinct scenarios involving varying sample sizes, instrument strengths, and causal relationship shapes (linear, quadratic, and threshold).
As shown in the simulation results, Quantile IV significantly outperformed the state-of-the-art DeLIVR estimator in 12 out of 15 scenarios, particularly in cases involving non-linear causal shapes. The method demonstrated lower Root Mean Square Error (RMSE) and greater stability, suggesting it is a more robust tool for complex biological modeling.
Case Study: Sclerostin Inhibition and Cardiovascular Safety
The researchers applied Quantile IV to a pressing clinical question: the safety of sclerostin inhibition. Sclerostin inhibitors are effective treatments for osteoporosis, but clinical trials (such as ARCH and BRIDGE) raised concerns regarding a potential increase in cardiovascular events.
Using data from the UK Biobank and large-scale consortia like CARDIoGRAMplusC4D, the study evaluated the causal link between sclerostin levels and health outcomes.
Key Findings:
- Efficacy Confirmed: All estimators, including Quantile IV, confirmed that a reduction in sclerostin significantly increases heel bone mineral density and reduces the risk of osteoporosis.
- Cardiovascular Safety: After adjusting for genetic pleiotropy and the "Linkage Disequilibrium (LD) trap"—where a variant linked to heart risk was found to affect a neighboring gene (MPP3) rather than sclerostin itself—the analysis found no evidence of a causal link between sclerostin reduction and heart attack or stroke risk. This suggests that the cardiovascular events observed in trials may be "off-target" effects rather than a direct consequence of the drug's primary mechanism.
- Biological Heterogeneity: The model suggested that sclerostin inhibition might be more effective in women than in men, mirroring observations from clinical trial data.
Discussion and Limitations
While Quantile IV offers superior flexibility and accuracy, it requires individual-level data, which is more restrictive than methods utilizing summary statistics. Additionally, the computational cost of bootstrapping for confidence intervals is significant. Despite these challenges, Quantile IV represents a major step forward in integrating machine learning with causal inference, providing a more nuanced lens through which to evaluate drug safety and efficacy.
Summary updated