A reference map of potential determinants for the human serum metabolome

The collection of metabolites circulating in the human blood, termed the serum metabolome, contains a plethora of biomarkers and causative agents. Although the origin of specific compounds is known, we have a poor understanding of the key determinants of most metabolites. Here, we measured the levels of 1251 circulating metabolites in serum samples from a healthy human cohort of 491 individuals, and devised machine learning algorithms to predict their levels in held-out subjects based on a comprehensive profile consisting of host genetics, gut microbiome, clinical parameters, diet, lifestyle, anthropometric measurements and medication data. Notably, we obtained statistically significant predictions for over 76% of the profiled metabolites. Despite using the strict out-of-sample prediction metric, which is a lower bound for the explained variance, diet and microbiome each explained hundreds of metabolites, with over 50% of the variance explained in some metabolites. We further validated the robustness of the microbiome related associations by showing a high replication rate in two geographically independent cohorts that were not available to us when developing the algorithms. We also demonstrate that some of these interactions are causal, as some metabolites we predicted to be positively associated with bread increased in level following a randomized clinical trial of bread intervention. Microbiome-explained metabolites were enriched with unnamed metabolites, and we devised an algorithm that accurately predicts their biological pathway, finding that they mainly include food components, aromatic amino acids and secondary bile acid derivatives. Overall, our results unravel potential determinants of over 800 metabolites, paving the way towards mechanistic understanding of alterations in metabolites under different conditions and to designing interventions for manipulating circulating metabolite levels.EGA study EGAS00001004512循环于人体血液中的代谢物集合，即血清代谢组（serum metabolome），蕴藏海量生物标志物与致病相关因子。尽管部分特定化合物的来源已被阐明，但我们对绝大多数代谢物的关键调控决定因素仍缺乏深入认知。本研究对491名健康受试者的血清样本中的1251种循环代谢物完成定量检测，并依托宿主遗传学、肠道微生物组（gut microbiome）、临床指标、饮食模式、生活习惯、人体测量数据与用药记录构成的多维度特征集合，开发机器学习算法以预测预留测试受试者的代谢物水平。值得注意的是，本研究对超过76%的检测覆盖代谢物实现了具有统计学显著性的预测效果。尽管采用了严格的样本外预测指标（该指标为解释方差的下界），饮食模式与肠道微生物组仍各自对数百种代谢物具有显著解释力，部分代谢物的方差解释度甚至超过50%。我们进一步验证了微生物组相关关联的稳健性：在开发算法阶段未获取数据的两个地理独立队列中，该关联展现出极高的复现率。此外，我们证实部分关联具有因果性：在一项面包干预的随机临床试验后，我们预测与面包摄入呈正相关的部分代谢物水平确实出现了显著升高。受微生物组解释的代谢物中未命名代谢物占比偏高，我们开发了可精准预测其生物学通路的算法，发现这类代谢物主要涵盖食物成分、芳香族氨基酸与次级胆汁酸衍生物。综上，本研究揭示了800余种代谢物的潜在调控决定因素，为解析不同生理病理状态下代谢物变化的机制，以及设计可调控循环代谢物水平的干预手段奠定了坚实基础。EGA研究EGAS00001004512