LC-MS Ion Suppression Prediction

What this research found

Ion suppression — the signal loss that occurs when co-eluting compounds compete for ionization in liquid chromatography–mass spectrometry (LC-MS) — is usually attributed to a molecule's own chemistry. A machine-learning pipeline was built to test whether chromatographic context matters as well, and context features carried 32.06% of total feature importance, with co-elution count ranking second of ten predictors. Because the raw spectra could not be obtained, the model was trained on a 100-sample synthetic dataset, so the result demonstrates the pipeline rather than confirming the underlying physics.

• Chromatographic context features together carried 32.06% of total SHAP feature importance, clearing the 30% threshold set in advance as the criterion for supporting the droplet surface area competition hypothesis.
• Co-elution count ranked second of ten predictors at 19.99%, ahead of LogP at 14.47% — the property most commonly used to anticipate suppression. Polarity, measured as topological polar surface area, led the ranking at 24.43%.
• Two further context features contributed: the m/z distance to the nearest neighbouring signal at 6.06% and total ion current within the retention window at 6.01%.
• Accuracy on held-out data was weak. The gradient-boosted model reached a test R² of 0.0911 and RMSE of 0.1240 against a training R² of 0.9996, a gap that indicates heavy overfitting on 100 samples.
• The neural network alternative failed outright at a test R² of -2.0040, worse than predicting the mean, and the gradient-boosted model was selected as the better of the two on that basis.

How it was done

Fifteen LC-MS metabolomics studies and their metabolite annotations were curated from the Metabolomics Workbench repository. Ten predictors were defined: seven molecular descriptors, including LogP, molecular weight, and topological polar surface area, plus three chromatographic context features — the number of compounds co-eluting within ±0.1 minutes, total ion current in the retention window, and the m/z distance to the nearest neighbouring signal. Because the raw spectra were never downloaded, the pipeline fell back on a 100-sample synthetic dataset whose ion suppression factors were computed from a known formula. An XGBoost regressor and a three-layer neural network were trained on an 80/20 split, then SHAP (Shapley additive explanations) was applied to the tree model to rank features and check that effect directions matched ionization physics. Deliverables included a suppression risk calculator and a 24-page manuscript of 15,238 words with 21 references.

Data sources

• Metabolomics Workbench — 15 curated LC-MS studies with metabolite annotations
• Synthetic training set of 100 samples, used as a fallback because raw mzML spectra were not available
• 21 peer-reviewed references, including Matuszewski et al. 2003 on matrix effects and Lundberg & Lee 2017 on SHAP

Figures from this analysis

Outputs produced

• selected_studies.json

  Selected studies

• candidate_studies_v5.json

  Candidate studies v5

• download_log.json

  Download log

• ST000004_metabolites.txt

  ST000004 metabolites

• ST000009_metabolites.txt

  ST000009 metabolites

• ST000010_metabolites.txt

  ST000010 metabolites

• ST000011_metabolites.txt

  ST000011 metabolites

• ST000076_metabolites.txt

  ST000076 metabolites