Nowadays, spam detection has been one of the foremost machine learning -oriented applications in the context of security in computer networks. In this work, we propose to learn intrinsic properties of e-mail messages by means of Restricted Boltzmann Machines (RBM s) in order to identity whether such messages contain relevant (ham) or non-relevant (spam) content. The main idea contribution of this work is to employ Harmony Search-based optimization techniques to fine-tune RBM parameters, as well as to evaluate their robustness in the context spam detection. The unsupervised learned features are then used to feed the Optimum-Path Forest classifier, being the original features extracted from e-mail content and compared against the new ones. The results have shown RBMs are suitable to learn features from e-mail data, since they obtained favorable results in the datasets considered in this work.