A key challenge in human genetics is the discovery of modifiable causal risk factors for complex traits and diseases. Mendelian randomization (MR) using molecular traits as exposures is a particularly promising approach for identifying such risk factors. Despite early successes with the application of MR to biomarkers such as low-density lipoprotein cholesterol and C-reactive protein, recent studies have revealed a more nuanced picture, with widespread horizontal pleiotropy. Using data from the UK Biobank, we illustrate the issue of horizontal pleiotropy with 2 case studies, one involving glycolysis and the other involving vitamin D synthesis. We demonstrate that, although the measured metabolites (pyruvate or histidine, respectively) do not have a direct causal effect on the outcomes of interest (red blood cell count or vitamin D level), we can still use variant effects on these downstream metabolites to infer how they perturb protein function in different gene regions. This allows us to use variant effects on metabolite levels as proxy exposures in a cis-MR framework, thus rediscovering the causal roles of histidine ammonia lyase (HAL) in vitamin D synthesis and glycolysis pathway in red blood cell survival. We also highlight the assumptions that need to be satisfied for cis-MR with proxy exposures to yield valid inferences and discuss the practical challenges of meeting these assumptions.