Exact mass, retention time (RT), and collision cross section (CCS) are used as identification parameters in liquid chromatography coupled to ion mobility high resolution accurate mass spectrometry (LC-IM-HRMS). Targeted screening analyses are now more flexible and can be expanded for suspect and non-targeted screening. These allow for tentative identification of new compounds, and in-silico predicted reference values are used for improving confidence and filtering false-positive identifications. In this work, predictions of both RT and CCS values are performed with machine learning using artificial neural networks (ANNs). Prediction was based on molecular descriptors, 827 RTs, and 357 CCS values from pharmaceuticals, drugs of abuse, and their metabolites. ANN models for the prediction of RT or CCS separately were examined, and the potential to predict both from a single model was investigated for the first time. The optimized combined RT-CCS model was a four-layered multi-layer perceptron ANN, and the 95th prediction error percentiles were within 2 min RT error and 5% relative CCS error for the external validation set (n = 36) and the full RT-CCS dataset (n = 357). 88.6% (n = 733) of predicted RTs were within 2 min error for the full dataset. Overall, when using 2 min RT error and 5% relative CCS error, 91.9% (n = 328) of compounds were retained, while 99.4% (n = 355) were retained when using at least one of these thresholds. This combined prediction approach can therefore be useful for rapid suspect/non-targeted screening involving HRMS, and will support current workflows.

精确的质量，保留时间（RT）和碰撞截面（CCS）在液相色谱中与离子迁移率高分辨率精确质谱（LC-IM-HRMS）耦合时用作鉴定参数。目标筛选分析现在更加灵活，可以扩展到可疑和非目标筛选。这些允许对新化合物进行初步鉴定，并在计算机上进行预测参考值用于提高置信度并过滤假阳性标识。在这项工作中，使用人工神经网络（ANN）通过机器学习对RT和CCS值进行预测。根据药物，滥用药物及其代谢物的分子描述符，827个RT和357个CCS值进行预测。分别检查了用于预测RT或CCS的ANN模型，并首次研究了从单个模型预测两者的潜力。优化的组合RT-CCS模型是四层多层感知器ANN，对于外部验证集（n = 36）和完整RT，第95个预测误差百分位数在2分钟内RT误差内，相对CCS误差在5％以内-CCS数据集（n = 357）。88。对于整个数据集，6％的预测RT（n = 733）误差在2分钟以内。总体而言，当使用2分钟的RT误差和5％的相对CCS误差时，使用至少一个这些阈值时，保留了91.9％（n = 328）的化合物，而保留了99.4％（n = 355）的化合物。因此，这种组合的预测方法对于涉及HRMS的快速可疑/非目标筛查非常有用，并且将支持当前的工作流程。