Learning new representations from data has been effective in predicting protein functions. However, common techniques tend to extract features irrelevant to the classification task. We propose an autoencoder network that selectively extracts features to produce meaningful representations. By increasing the activation of neurons kept by a winner-take-all operation, hidden units compete to form a subset that encodes relevant features, a process dubbed as mutual competition. We test this method on protein benchmarks, evaluating feature score distribution and classification performance. Results show that the autoencoder extracted features relevant to the classification task, and significantly outperformed other techniques in literature based on non-parameteric statistical tests. This demonstrates that adding competition between neurons encodes meaningful features, further improving the prediction of protein functions.