Toward Machine Learning Electrospray Ionization Sensitivity Prediction for Semiquantitative Lipidomics in Stem Cells

Abstract

Specificity, sensitivity, and high metabolite coverage make mass spectrometry (MS) one of the most valuable tools in metabolomics and lipidomics. However, translation of metabolomics MS methods to multiyear studies conducted across multiple batches is limited by variability in electrospray ionization response, making batch-to-batch comparisons challenging. This limitation creates an artificial divide between nontargeted discovery work that is broad in scope but limited in terms of absolute quantitation ability and targeted work that is highly accurate but limited in scope due to the need for matched isotopically labeled standards. These issues are often observed in stem cell studies using metabolomic and lipidomic MS approaches, where patient recruitment can be a years-long process and samples become available in discrete batches every few months. To bridge this gap, we developed a machine learning model that predicts electrospray ionization sensitivity for lipid classes that have shown correlation with stem cell potency. Molecular descriptors derived from these lipids' chemical structures are used as model input to predict electrospray response, enabling quantitation by MS with moderate accuracy (semiquantitation). Model performance was evaluated via internal and external validation using cultured cells from various stem cell donors, achieving global percent errors of 40% and 20% for positive and negative electrospray ion modes, respectively. Although this accuracy is typically insufficient for traditional targeted lipidomics experiments, it is sufficient for semiquantitative estimation of lipid marker concentrations across batches without the need for specific chemical standards that many times are unavailable. Furthermore, the precision for model-predicted concentrations was 16.9% for the positive mode and 7.5% for the negative mode, indicating promise for data harmonization across batches. The set of molecular descriptors used by the models described here was able to yield higher accuracy than those previously published in the literature, showing high promise toward semiquantitative lipidomics.

Figure 1

Collection of training and evaluation data sets for machine learning models. Two lipid standard mixes, the Avanti UltimateSPLASH ONE and Avanti SPLASH LIPIDOMIX, were spiked according to the (a) scheme into cell pellet samples composed of MSCs from two donors. From these samples, several lipid (b) data sets were created for model training, testing, and target evaluation. UI27RMEWO3 Created in BioRender. Fernandez, F. (2025).

Figure 2

Model development workflow depicting (a) generation of lipid calibration curves for determination of experimental sensitivity values (m_R), (b) identifying optimal train/test split percentage, (c) optimization of data transformation and feature selection, d) generation of 1,101 model iterations to select median error model, (e) model training with 5-fold cross validation, (f) model validation on test set lipids, and (g) external validation on separate lipids in separate samples. Full workflow results in two final models: one for each ionization mode, positive and negative. NA27RMF9PW Created in BioRender. Fernandez, F. (2025).

Figure 3

Electrospray ionization efficiency violin plots of observed lipid classes in (a) positive ion mode and (b) negative ion mode. The plot displays the ionization efficiencies of the lipid components of the Avanti Ultimate SplashOne mix spiked into donor 1 and donor 2 MSC samples. For each individual lipid, the ionization efficiencies measured were averaged across donor 1 and donor 2 samples. Ionization efficiency was defined as the slope of the concentration–response curve (m) using non-normalized peak areas.

Figure 4

Model calibration and internal validation performance. Plots depicting measured relative slopes (m_R) for UltimateSplash components in donor 1 samples against their predicted relative slopes for (a) positive ion mode model calibration, (b) negative ion mode calibration, (c) positive model cross-validation, (d) negative mode cross-validation, (e) positive mode internal validation, and (f) negative mode internal validation. For positive and negative ion mode model iterations, the median was chosen for display. All models utilized Box-Cox transformation and recursive feature elimination (RFE). Dashed lines indicate the regression fit line, and solid lines indicate a x = y fit.

Figure 5

Errors for internal and external validation of positive and negative ion mode models. Average percent errors for (a) positive and (b) negative mode concentration predictions during internal validation using 20% of the analytes in UltimateSplash Mix. Errors were averaged at each calibration level. Average percent errors for (c) positive and (d) negative mode concentration predictions during external validation using Splash Lipidomix in donor 1 and donor 2 samples. Errors were averaged across three high samples and three low samples. Errors were calculated by comparing predicted concentrations to spiked concentrations. Note: ceramide names abbreviated for space; full names can be found in Table S3.

Figure 6

Model accuracy compared to traditional nontargeted quantitation approaches. The accuracies of predicted and calculated concentrations for Splash Lipidomix standards in donor 1 and donor 2 samples (averaged across 12 samples). Concentrations were determined using the prediction workflow, one-point calibration, and surrogate lipid class curve calibration and are displayed as (a) a global average for positive and negative modes, (b) averages by lipid component in positive mode, and (c) averages by lipid component in negative mode.