Metabolomics has emerged as a promising discipline in pharmaceuticals and preventive healthcare. However, analysing large metabolomics datasets remains challenging due to limited and incompletely annotated biological pathways. To address this limitation, we recently proposed training machine learning classifiers on molecular fingerprints of metabolites to predict their responses under specific conditions and analysing feature importance to identify key chemical configurations, providing insights into the affected biological processes. This study extends our previous research by evaluating various metabolite structural representations, including Morgan fingerprint and its variants, graph-based structural encodings and proposing novel representations to improve resolution and interpretability of the state-of-the-art approaches. These structural encodings were evaluated on mass spectrometry metabolomic data for a cellular model of the genetic disease Ataxia Telangiectasia. The study found that machine learning classifiers trained on the new representations improved in classification accuracy and interpretability. Notably, models trained on graph-based encoding do not exhibit performance gains, not even with pre-training on a larger metabolite dataset, underlining the efficacy of our proposed representations. Finally, feature importance analysis across different encoding methods consistently identifies similar structures as relevant for classification, underscoring the robustness of our approach across diverse structural representations.