Semi-Quantitative LC/ESI/MS Analysis Using Predictive Models of ESI Ionization Efficiencies Jaanus Liigand (Presenter) University of Tartu Presenter Bio: I am third-year analytical chemistry PhD student from University of Tartu, studying standard substance free quantification in LC/ESI/MS analysis

Authors: Jaanus Liigand (1), Piia Liigand (1), Mari Ojakivi (1), Karl Kaupmees (1), Anneli Kruve (1)(2)
(1) University of Tartu, Estonia; (2) Freie Universität Berlin, Germany

Until now, in the discovery of metabolites and in the absence of standard substances in LC/MS analyses equal ionization efficiencies are assumed. This may lead to misunderstandings of the processes occurring in organisms as concentrations of some metabolites can be up to 5 orders of magnitude over- or underestimated. By prediction of ionization efficiencies in both positive and negative electrospray ionization and in biological matrices the accuracy of such predictions can be improved, the best prediction being a 4-times mismatch with reality allowing for more accurate semi-quantitative analysis. This prediction method is user-friendly, as it uses 2D structures of the analytes and a small set of calibration compounds incorporated in the analytical run; thus, enabling quicker and more accurate estimation of the abundance of compounds of interest.

Introduction

Increased sensitivity and resolution of liquid chromatography-mass spectrometers (LC/ESI) reveal more information about the transformations occurring in our bodies. However, different analytes ionize with different efficiency in electrospray ionization source, which means that authentic standards are needed to obtain quantitative information about these transformations. Ionization efficiency (IE) depends on the structure of the compound, on the solvent and on the used setup. Up to now, in the absence of standard substances the relative abundances has mostly been evaluated based on the peak areas. We propose an approach where we combine ionization efficiency measurements, computational modelling and machine learning to overcome the need for authentic standard substances via predicting the ionization efficiencies in both ESI positive and negative mode on various mass spectrometers and solvents and in biological matrices (blood, plasma, cerebrospinal fluid, urine, brain and liver tissue). This approach improves the prediction accuracy by two orders of magnitude (from up to 10 000-fold concentration mispredictions based on peak areas to 4-fold misprediction with ionization efficiency prediction) and could, therefore, lead to the possibility of semi-quantitative analyses.

Methods

Altogether, ionization efficiencies (presented as logIE values) of 400 compounds covering a wide range of chemical space (logP =-4.1 … 7.7; pKa = -4.5 … 12) were measured in electrospray ionization positive and negative mode in 29 mobile phases. In studied mobile phases, the percentage of acetonitrile or methanol was changed from 0-100% and water phase additives covered a variety of the most used additives. The measurements were carried out in flow injection mode and MS1 signals were recorded. The ionization efficiencies of a subset of 20 compounds were measured on 11 different instrumental setups covering different ESI source designs with different types of mass analyzers from all major vendors to demonstrate the transferability of IE values between different instrumental setups. A subset of 10, predominantly pharmaceutical, compounds were measured in the biological matrices (blood, plasma, cerebrospinal fluid, urine, brain and liver tissue) as well as in acetonitrile/0.1% ammonia aqueous solution 80/20 in negative ESI/MS. Altogether 2000 ionization efficiencies were measured. All logIE values measured on different setups were transformed to universal and quantitatively comparable logIE values. For model development, descriptors from open source software PaDEL were calculated based on 2D structures. Different machine learning approaches e.g. multilinear regression, random forest regression and artificial neural networks were applied to obtain a model to predict ionization efficiencies which can be used to estimate concentrations. The obtained models for both modes were validated with results obtained on a different instrument and for compounds and conditions not included in model development. Moreover, it was demonstrated that the approach is also applicable in LC/MS analysis with gradient separation.

Results

In positive ESI mode, the span of measured ionization efficiencies (expressed as logIE values) cover 10 orders of magnitude.

The universality of the ionization efficiency scales on different instruments was verified. We incorporated different ESI sources with different mass analyzers from all the main mass spectrometer vendors and saw that in general picture the ionization efficiency scales are universal between different instruments. The order of compounds’ ionization efficiencies does not change significantly between instruments and correlation is good (R2 0.6-0.98, average 0.8). However, the spans of logIE values differ remarkably. This means that ionization efficiencies can be transferred from one instrument to another via a small set of calibration compounds (e.g. 6 compounds). Using this strategy, we can easily transform the results into universal quantitatively comparable logIE values

The best model for ESI positive mode, accounting also the various mobile phase compositions was obtained with regularized random forest regression with RMSE for training set 0.62 for test set 0.54 and for external validation set 0.75 logIE units. This means 4.2-, 3.5- and 5.6-fold mismatch in concentration prediction respectively.

In negative mode, the span of measured logIE values covered 7 orders of magnitude. Again, the best model was obtained with regularized random forest regression with RMSE for training set 0.38 and for test set 0.42 logIE units. This means 2.4- and 2.6-fold mismatch in concentration prediction respectively. Similar trends in the transferability of logIE values between instruments in positive mode were also confirmed in ESI negative mode. The model was validated on chromatographic analysis of 16 acids in gradient elution. The concentration prediction resulted in an average 3.4-fold miss.

As both modes gave promising results in different eluent compositions and on different instrumental setups we brought this approach one step further. Namely, we tested the approach for the analysis of complicated matrices. For that, a set of 10, predominantly pharmaceutical, compounds were measured in the biological matrices in ESI negative mode in flow injection mode. The logIE values in different biological matrices were in good correlation with the logIE values in a solvent. This good correlation suggests that ionization efficiencies can be predicted in the matrices. Different physicochemical parameters were used to fit the models for predicting logIE values in biological matrices. The best predictive power of the model was observed in liver tissue homogenate (RMSE = 0.49) and the lowest predictive power in urine (RMSE= 0.80). The correlation between measured and predicted logIE values over all matrices is high, R2 = 0.83 and RMSE = 0.67 logIE units meaning an average mismatch of 8-fold. Estimating the concentrations based on peak areas would lead to more than a 660-fold misprediction. This shows that the accuracy of standard substance free quantitation in biological matrices was improved by almost two orders of magnitude with the use of predicting ionization efficiencies.

Conclusions & Discussion

Until now quantitation in LC/ESI/MS in lack of authentic standards may lead to a mismatch up to 4 orders of magnitude with reality. We demonstrate that ionisation efficiencies can be predicted in both electrospray modes in different eluent compositions and moreover in biological matrices. Using predicted ionization efficiencies, the accuracy improves significantly and results in lower than 4-fold mismatch for different mobile phases, and an 8-fold mismatch for biological matrices This approach requires only 2D structures of the compounds as input and a small set of compounds to transform results into universal logIE values.

Kruve and Kaupmees Anal Chem 89(9), 2017

Liigand et al Anal Chem 89(11), 2017

Liigand et al J Am Soc Mass Spectrom 28(3), 2017

Liigand et al J Am Soc Mass Spectrom 26(11), 2015

Liigand et al J Am Soc Mass Spectrom 25(11), 2014