Database search is a widely used approach for identifying metabolites from tandem mass spectra (MS/MS). In this strategy, an experimental spectrum is matched against a user-specified database of candidate metabolites, and candidates are ranked such that true metabolite-spectrum matches receive the highest scores. Machine-learning methods have been widely incorporated into database-search-based identification tools and have substantially improved performance. To further improve identification accuracy, we propose a new framework for generating negative training samples. The framework first uses autoencoders to learn latent representations of metabolite structures and MS/MS spectra, thereby recasting metabolite-spectrum matching as matching between latent vectors. It then uses a GAN to generate latent vectors of decoy metabolites and constructs decoy metabolite-spectrum matches as negative samples for training. Experimental results show that our tool, MS2MetGAN, achieves better overall performance than existing metabolite identification methods.