The step of predicting small-molecule drug transformation by hepatic enzymes is still a time-consuming and congested preclinical screening. In this framework, each molecule under consideration is represented as a graph, and the molecular information is represented by graph spectral invariants that are based on its adjacency and Laplacian spectra. These spectral molecular descriptors are employed to build an ensemble learning-based classifier which discriminates metabolically stable and labile molecules and also to predict intrinsic clearance. A set of 2400 drug-like molecular graphs was compiled, and the latent metabolic response was designed to be nonlinear and depend on the shape of the spectra via interactions. The ensemble proposed achieved an accuracy of 82.8 % and an area under the curve of 0.917, outperforming the logistic baseline, kernel baseline, tree baseline and neural baseline. This accuracy was raised to 81.2 % when the plain topological counts were supplemented by spectral invariants, as confirmed by ablation. The coefficient of determination of the clearance regression was 0.72. Results have shown that spectral characterisation of drug metabolism gives a discriminative signal which is not captured by size-based descriptors.