LipoCLEAN:\nA Machine Learning Filter to Improve Untargeted\nLipid Identification Confidence

In untargeted lipidomics experiments, putative lipid\nidentifications\ngenerated by automated analysis software require substantial manual\nfiltering to arrive at usable high-confidence data. However, identification\nsoftware tools do not make full use of the available data to assess\nthe quality of lipid identifications. Here, we present a machine-learning-based\nmodel to provide coherent, holistic quality scores based on multiple\nlines of evidence. Underutilized metrics such as isotope ratios and\nchromatographic behavior allow for much higher accuracy of identification\nconfidence. We find that approximately 50% of tandem mass spectrometry-based\nautomated lipid identifications are incorrect but that multidimensional\nrescoring reduces false discoveries to only 7% while retaining 80%\nof true positives. Our method works with most chromatography methods\nand generalizes across a family of MS instruments. LipoCLEAN is available\nat https://github.com/stavis1/LipoCLEAN.