In information retrieval, query performance prediction aims to predict whether a search engine is likely to succeed in retrieving potentially relevant documents to a user's query. This problem is usually cast into a regression problem where a machine should predict the effectiveness (in terms of an information retrieval measure) of the search engine on a given query. The solutions range from simple unsupervised approaches where a single source of information (e.g., the variance of the retrieval similarity scores in NQC), predicts the search engine effectiveness for a given query, to more involved ones that rely on supervised machine learning making use of several sources of information, e.g., the learning to rank (LETOR) features, word embedding similarities etc. In this paper, we investigate the combination of two different types of evidences into a single neural network model. While our first source of information corresponds to the semantic interaction between the terms in queries and their top-retrieved documents, our second source of information corresponds to that of LETOR features.