Combining text mining and sequence analysis to discover protein functional regions

Recently presented protein sequence classification models can identify relevant regions of the sequence. This observation has many potential applications to detecting functional regions of proteins. However, identifying such sequence regions automatically is difficult in practice, as relatively few types of information have enough annotated sequences to perform this analysis. Our approach addresses this data scarcity problem by combining text and sequence analysis. First, we train a text classifier over the explicit textual annotations available for some of the sequences in the dataset, and use the trained classifier to predict the class for the rest of the unlabeled sequences. We then train a joint sequence text classifier over the text contained in the functional annotations of the sequences, and the actual sequences in this larger, automatically extended dataset. Finally, we project the classifier onto the original sequences to determine the relevant regions of the sequences. We demonstrate the effectiveness of our approach by predicting protein sub-cellular localization and determining localization specific functional regions of these proteins.