SÄLEN and the Art of Metal Binding - An Explainable Graph Neural Network for Fragment-Wise Prediction of Metal-Adduction from Mass Spectrometry Data

Abstract

When searching for and analyzing existing pharmaceutical molecules, mass spectrometry is often used. One method utilizes metal-adduct formation, adding metals to the sample to see which compound its rate and location of binding match. Detecting and interpreting metal-adduct formation in mass spectrometry data is challenging, especially when the goal is not only to predict whether a signal will appear, but also to understand which parts of a molecule contribute to that prediction. In this thesis, we develop the Substructure Aware Ligand Explanation Network (SÄLEN), an ion-conditioned, fragment-wise graph neural network for predicting Mass Spectrometry (MS) derived metal-adduct-like signals from molecule–ion pairs. SÄLEN was trained on experimental data from AstraZeneca and designed to provide intrinsic fragment-level explanations by decomposing each prediction into contributions from molecular substructures. The model achieved competitive regression performance compared with established molecular models, with a test Mean Absolute Error (MAE) of 0.81 in log-intensity space. In the multitask setting, SÄLEN reached a positive classification precision of 92% and recall of 70%, making it suitable for high-confidence prioritization of molecule–ion pairs. The fragment-wise explanations were evaluated through counterfactual atom edits, fragment swaps, inter-model comparison, and comparison with Quantum Mechanical (QM) descriptors. These analyses suggest that SÄLEN often highlights chemically plausible regions; although the explanations should be interpreted as model sensitivities rather than proof of physical binding mechanisms. Overall, SÄLEN provides a useful framework for interpretable prediction of metaladduct- like MS signals and may support hypothesis generation and prioritization in experimental workflows.