ProLanGO2

Abstract

Predicting protein function from protein sequence is a main challenge in the computational biology field. Traditional methods that search protein sequences against existing databases may not work well in practice, particularly when little or no homology exists in the database. We introduce the ProLanGO2 method which utilizes the natural language processing and machine learning techniques to tackle the protein function prediction problem with protein sequence as input. Our method has been benchmarked blindly in the latest Critical Assessment of protein Function Annotation algorithms (CAFA 4) experiment. There are a few changes compared to the old version of ProLanGO. First of all, the latest version of the UniProt database is used. Second, the Uniprot database is filtered by the newly created fragment sequence database FSD to prepare for the protein sequence language. Third, the Encoder-Decoder network, a model consisting of two RNNs (encoder and decoder), is used to train models on the dataset. Fourth, if no k-mers of a protein sequence exist in the FSD, we select the top ten GO terms with the highest probability in all sequences from the Uniprot database that didn't contain any k-mers in FSD, and use those ten GO terms as back up for the prediction of new protein sequence. Finally, we selected the 100 best performing models and explored all combinations of those models to select the best performance ensemble model. We benchmark those different combinations of models on CAFA 3 dataset and select three top performance ensemble models for prediction in the latest CAFA 4 experiment as CaoLab. We have also evaluated the performance of our ProLanGO2 method on 253 unseen sequences taken from the UniProt database and compared with several other protein function prediction methods, the results show that our method achieves great performance among sequence-based protein function prediction methods. Our method is available in GitHub: https://github.com/caorenzhi/ProLanGO2.git.