Deep Learning for the Precise Peak Detection in High-Resolution LC–MS Data

Abstract

This letter is devoted to the application of machine learning, namely, convolutional neural networks to solve problems in the initial steps of the common pipeline for data analysis in metabolomics. These steps are the peak detection and the peak integration in raw liquid chromatography–mass spectrometry (LC–MS) data. Widely used algorithms suffer from rather poor precision for these tasks, yielding many false positive signals. In the present work, we developed an algorithm named peakonly, which has high flexibility for the detection or exclusion of low-intensity noisy peaks, and shows excellent quality in the detection of true positive peaks, approaching the highest possible precision. The current approach was developed for the analysis of high-resolution LC–MS data for the purposes of metabolomics, but potentially it can be applied with several adaptations in other fields, which utilize high-resolution GC− or LC–MS techniques. Peakonly is freely available on GitHub () under an MIT license.

Deep metabolome: Applications of deep learning in metabolomics

“…Numerous free and commercial software are available for MS-based data processing such as MZmine [47] , XCMS [45] , metaMS [48] , and metAlign [49] , to name a few. However, key challenges, such as false positive signals, co-eluting compounds and non-linear retention time shift, still need to be addressed [50] , [51] , [52] . With the complexity of MS data, DL approaches have been proposed to solve this key data pre-processing step and major bottleneck of MS-based metabolomics pipelines.…”

“…These profiles were initially modelled by PARAllel FACtor analysis2 (PARAFAC2) [50] , [54] and subsequently delineated into chemical peaks (metabolite), baselines and other non-related peak areas by the CNN model, which resolved which peak component were most suitable for selection or integration. Similarly, Melnikov et al proposed ‘peakonly’ algorithm [51] for both peak detection and integration that used a CNN model to classify raw LC-MS data into regions of noise, chemical peaks, and uncertain peaks, which was then used to determine peak boundaries for integration. Automated and high accuracy peak classifiers would greatly improve efficiency in these critical steps, which often heavily rely on domain experts.…”

Food Phenotyping: Recording and Processing of Non-Targeted Liquid Chromatography Mass Spectrometry Data for Verifying Food Authenticity

“…Numerous algorithms exist for the peak detection of LC-MS data [ 85 , 130 , 131 , 132 , 133 , 134 ]. Most of them first convert profile spectra to line spectra, if necessary, and then perform spectra smoothing.…”

MS-CleanR: A feature-filtering approach to improve annotation rate in untargeted LC-MS based metabolomics

“…A few packages have been developed to deal with feature degeneracy: CAMERA 16 is based on adduct relationships; RAMClust 17 correlates features in multiple samples; MS-FLO 18 uses Pearson’s correlation and peak height similarity to identify adducts, duplicate peaks and isotope features of the main monoisotopic ion, and MZunity 19 which confronts adducts or neutral loss index to decipher relationship among the acquired high resolution pseudo-molecular ions list. Deep-learning approaches have also been developed based on LC-MS spectral peak shape filtering 20,21 . All these packages focus on a single type of degeneracy, however, and they are difficult to implement in a unified workflow.…”