Pulmonary arterial hypertension (PAH) exhibits phenotypic heterogeneity and variable response to therapy. The metabolome has been implicated in the pathogenesis of PAH, but previous works have lacked power to implicate specific metabolites. Mendelian randomisation (MR) is a method for causal inference between exposures and outcomes. Using GWAS summary statistics, we implemented hypothesis-free MR methodology to test for causal relationships between serum concentration of 575 metabolites and PAH. Unbiased MR causally associated five metabolites with risk of PAH after stringent multiple testing correction; of the five candidates, serine and homostachydrine were validated in a different larger PAH GWAS, and associated with clinical severity of PAH via direct measurement in an independent clinical cohort of 449 PAH patients. We used conditional and orthogonal approaches to explore the biology underlying our lead metabolites. A rare variant analysis demonstrated that loss of function (LOF) mutations within ATF4, a transcription factor responsible for upregulation of serine synthesis under conditions of serine starvation, are associated with higher risk for PAH. Homostachydrine is a xenobiotic metabolite that is structurally related to L-proline betaine, which has been previously linked to modulation of inflammation and tissue remodelling in PAH. Our MVMR analysis suggests that the effect of L-proline betaine is actually mediated indirectly via homostachydrine. Our data presents a new method for study of the metabolome in the context of PAH, and suggests several candidates for further evaluation and translational research.