CCS Predictor 2.0: An Open-Source Jupyter Notebook Tool for Filtering Out False Positives in Metabolomics

Abstract

Metabolite annotation continues to be the widely accepted bottleneck in nontargeted metabolomics workflows. Annotation of metabolites typically relies on a combination of high-resolution mass spectrometry (MS) with parent and tandem measurements, isotope cluster evaluations, and Kendrick mass defect (KMD) analysis. Chromatographic retention time matching with standards is often used at the later stages of the process, which can also be followed by metabolite isolation and structure confirmation utilizing nuclear magnetic resonance (NMR) spectroscopy. The measurement of gas-phase collision cross-section (CCS) values by ion mobility (IM) spectrometry also adds an important dimension to this workflow by generating an additional molecular parameter that can be used for filtering unlikely structures. The millisecond timescale of IM spectrometry allows the rapid measurement of CCS values and allows easy pairing with existing MS workflows. Here, we report on a highly accurate machine learning algorithm (CCSP 2.0) in an open-source Jupyter Notebook format to predict CCS values based on linear support vector regression models. This tool allows customization of the training set to the needs of the user, enabling the production of models for new adducts or previously unexplored molecular classes. CCSP produces predictions with accuracy equal to or greater than existing machine learning approaches such as CCSbase, DeepCCS, and AllCCS, while being better aligned with FAIR (Findable, Accessible, Interoperable, and Reusable) data principles. Another unique aspect of CCSP 2.0 is its inclusion of a large library of 1613 molecular descriptors via the Mordred Python package, further encoding the fine aspects of isomeric molecular structures. CCS prediction accuracy was tested using CCS values in the McLean CCS Compendium with median relative errors of 1.25, 1.73, and 1.87% for the 170 [M − H] − , 155 [M + H] + , and 138 [M + Na] + adducts tested. For superclass-matched data sets, CCS predictions via CCSP allowed filtering of 36.1% of incorrect structures while retaining a total of 100% of the correct annotations using a Δ CCS threshold of 2.8% and a mass error of 10 ppm.

Unknown Metabolite Identification Using Machine Learning Collision Cross-Section Prediction and Tandem Mass Spectrometry

“…To predict CCS values, two data sets were created using the Unified CCS Compendium, one containing all entries for [M + H] + adduct species and another set for all [M – H] − entries ( n = 644 and 582, respectively). These data sets were then randomly split 75% into training sets and 25% into test sets, which were used to construct a support vector regression (SVR)-based ML model using CCSP 2.0, as illustrated in Figure . Using the optimized SVR models, CCS values for any given candidate structure could be predicted for [M + H] + and [M – H] − species from their neutral InChI code.…”

Prediction of a Large-Scale Database of Collision Cross-Section and Retention Time Using Machine Learning to Reduce False Positive Annotations in Untargeted Metabolomics

“…A CCS index filter defined as mean error +/− 2 SD, i.e., maximum 16.16 Å 2 , could be used as the threshold for excluding false positives. This match tolerance, reflecting the deviation of analytes or family of analytes or type of adducts, is relatively large compared to other work demonstrating that median relative errors as low as 3 to 5% are reachable using other models [ 22 , 24 , 25 , 26 , 27 ]. However, excluding false positive identifications with a CCS match higher than the defined threshold remains of great importance when considering the number of possible matches when using m/z match, isotope similarity, and fragmentation score only.…”

“…The resulting performances could have been further validated by performing a side-by-side comparison with other existing machine-learning tools. Such a comparison has already been described elsewhere [ 22 , 24 , 25 , 26 , 27 ]. Instead of that, the chosen strategy consisted of emphasizing the usefulness of our workflow with concrete application on biological data.…”

“…Several studies have developed, or applied machine learning-based prediction approaches [ 23 ]. Softwares like AllCCS [ 22 ], CCS Predictor [ 24 ], DeepCCS [ 25 ], MetCCS Predictor [ 26 ], or LipidCCS predictor [ 27 ] can efficiently generate a model when molecular descriptors are provided [ 15 ]. Molecular descriptors are numeric information generated by mathematical treatment of compound structures that characterize the physico–chemical properties of metabolites (ex: polarity, LogP…) [ 28 ].…”

Ion Mobility Mass Spectrometry (IM-MS) for Structural Biology: Insights Gained by Measuring Mass, Charge, and Collision Cross Section

“…Such comparisons allow the researcher to gain better insights into the conformational landscape adopted. − These methods have been well-reviewed ,,− and broadly fall into three categories. In the first category are those that fully evaluate the trajectory of the ion as it interacts with the buffer gas so call trajectories methods (TM) including (MOBCAL-TM and IMoS). ,,,,− The second category includes those that consider the projected area of the candidate structure and use empirical data to determine a CCS (PA, PSA, IMPACT). ,,,, The last category considers the recently emergent machine learning approaches. ,− The first two approaches rely on a reasonable starting structure, and commonly with proteins, molecular dynamics methods, both atomistic and coarse-grained that can be used to provide candidate gas-phase geometries. Such molecular dynamics (MD) evaluation can be computationally very expensive, although refinements to this have been made that integrate CCS values into the conformational searching for suitable candidate geometries.…”