生物活性天然产物的发现速度缓慢可归因于在复杂混合物如植物提取物中快速鉴定它们的困难。为了克服这些障碍，我们探索了两种机器学习技术的实用性，即，弹性网络和随机森林，用于鉴定含有数百种天然产物的啤酒花（Humuluslupulus）花序提取物的个体抗炎原理。我们通过柱层析分离了啤酒花提取物，获得了40个不纯的馏分，使用基于巨噬细胞的生物测定法确定其抗炎活性，该生物测定法测量iNOS介导的一氧化氮形成的抑制作用，并通过流动注射HRAM质谱和LC-MS/MS表征了馏分的化学成分。生物活性的前10个预测因子是异戊烯化类黄酮和腐殖质。生物活性的最高随机森林预测因子，黄腐酚,在相同的生物测定中以纯形式进行测试，以验证预测结果（IC50为7μM）。使用全球天然产品社交网络（GNPS）算法，通过与已知啤酒花天然产品的光谱相似性来鉴定其他生物活性预测因子。我们的机器学习方法表明，可以识别单个生物活性天然产物，而无需对植物提取物进行广泛和重复的生物测定指导分馏。
The slow pace of discovery of bioactive natural products can be attributed to the difficulty in rapidly identifying them in complex mixtures such as plant extracts. To overcome these hurdles, we explored the utility of two machine learning techniques, i.e., Elastic Net and Random Forests, for identifying the individual anti-inflammatory principle(s) of an extract of the inflorescences of the hops (Humulus lupulus) containing hundreds of natural products. We fractionated a hop extract by column chromatography to obtain 40 impure fractions, determined their anti-inflammatory activity using a macrophage-based bioassay that measures inhibition of iNOS-mediated formation of nitric oxide, and characterized the chemical composition of the fractions by flow-injection HRAM mass spectrometry and LC-MS/MS. Among the top 10 predictors of bioactivity were prenylated flavonoids and humulones. The top Random Forests predictor of bioactivity, xanthohumol, was tested in pure form in the same bioassay to validate the predicted result (IC50 7 µM). Other predictors of bioactivity were identified by spectral similarity with known hop natural products using the Global Natural Products Social Networking (GNPS) algorithm. Our machine learning approach demonstrated that individual bioactive natural products can be identified without the need for extensive and repetitive bioassay-guided fractionation of a plant extract.
