This thesis addresses the core challenges of using amino acid sequence features for predicting the function of unknown proteins using machine learning without relying on homology information. The ultimate aim is to estimate the possibility of correctly predicting the function of unknown proteins using these features. We begin by addressing the challenge of extracting as many sequence and structural features as possible from each bacterial protein by introducing a web tool for protein sequence information, and address the challenge of separating homologs for machine learning classification by implementing the Protein Homology Separator algorithm. Having addressed these initial challenges, we assess the importance of feature selection in protein function prediction and the imbalanced data problem. We show that classifiers generated from feature-selected and balanced data significantly outperform other classifiers. This is clearly evident when the data is used for Support Vector Machine classifiers. We further improved feature selection for protein function prediction by developing a new feature selection method (FrankSum). Using our bacterial datasets, we show that classifiers generated from features selected by FrankSum, outperform classifiers generated from full feature sets, randomly selected features, and features selected by classical Wrapper methods. FrankSum was hence used for addressing the next major challenge. The existing classical approach generates classifiers trained on known proteins and then predicts the function of unknown proteins with no evidence for the correctness and reliability of this transfer. We have introduced a refined approach that provides the likelihood of correctly predicting the function of unknown proteins using machine learning. We show that SVM-based machine learning is able to distinguish proteins from different source organisms with greater reliability and accuracy than proteins with known or unknown function. This means that 'known' and 'unknown' proteins are more similar to each other than proteins from different species. We then show that classifiers transfer successfully between bacterial species. This means that functional class classifiers will be able to transfer from known to unknown proteins with even greater success than they do across species boundaries. Hence we conclude that there is a distinct possibility of being able to correctly predict the biological function of unknown proteins using classifiers trained on known sequences. This is the first critical estimate of predictive performance on proteins of unknown function.