The complexity and variations in LC-MS experimental design has led to the population of databases with MS/MS spectra obtained using diverse assay conditions. Reference MS/MS spectra are biased towards biomedically relevant compounds, and a vast majority of the plant metabolome is not represented therein. Thus, database-dependent identification of LC-MS-derived mass features using commonly used MS/MS spectral matching algorithms leaves >95% of the thousands of metabolites from plant extracts unidentified. However, if the specific metabolite itself cannot be identified, accurate determination of the structural class of the metabolite can also be useful to associate LC-MS peaks to candidate metabolic pathways, obtain a bird’s eye view of differentially accumulating metabolites under stress, and to frame testable functional hypotheses in QTL mapping experiments. Towards this goal, we developed a new machine learning based approach for feature-based classification of LC-MS/MS peaks into pre-existing structural categories. Preliminary analyses were performed on two lipid databases – LIPID MAPS and LipidBlast – with different but well-laid-out ontologies. A random forests model, trained to classify lipids using just the chemical formula, was able to achieve 95-100% accuracy for almost all categories. This single-label classification approach was also successful in categorizing a broader set of compounds from PlantCyc and the ReSpect for Phytochemicals databases using formula and MS/MS fragmentation features, respectively. We implemented multi-label learning to better model structural diversity in the broader ChEBI database of organic compounds, and found ~85% precision, recall and accuracy in predicting multiple labels of a given instance using just the formula. Using in silico modeled MS/MS fragments from lipids, we further demonstrated that adding signature fragments to the model can increase predictive accuracy for structurally related lipid classes. This algorithm, which we call MetClass, is database-independent, scalable, very fast, and can help provide more insights into the vastly understudied plant metabolic diversity.