Predicting protein function is a difficult problem in bioinformatics. Many recent techniques employ embeddings to learn representations of protein sequences and infer function from these; however there have been no studies that have utilized protein function text embeddings to forecast protein function. Here, we propose to learn and explore text-driven embedding representations of protein function comment sections kept as part of the Swiss-Prot entries and understand how the resulting data can be used to enhance protein function annotations. The comparative study is based on protein function text embeddings derived from two approaches which include a combination of natural language processing frameworks such as Word2Vec, Doc2Vec and dictionary-based Named Entity Recognition and acts as a preliminary assessment based on direct propagation techniques such as sequence similarity and by-similarity prediction.
Understanding the role of proteins is crucial to life. However, there exists only a small subset of
proteins whose function is well predicted thereby making protein function prediction a fundamental
task in the field of Bioinformatics. Numerous techniques have been developed for protein function
prediction using sequence embeddings, protein structures or protein-protein interactions [
Our work is based on the hypothesis that states a direct correlation between sequence similarity (corresponding to the BLAST identity score) and similar biological function (as expressed in the protein function comment). The idea here is to capture this correlation with the help of the corresponding protein embeddings. Here we consider the text-based embeddings that are derived from the protein function comment sections of the UniProtKB reviewed entries: SwissProt entries. Specifically, we aim to learn and compare two embedding models that map functions of protein to sequences of vector representations such that two proteins having similar function as stated in the function comment section appear closer in the embedding vector space. A secondary objective is to analyze the vocabulary that emerges from the text-based embeddings.
Methods of representing textual data range from traditional term-frequency based methods to
embeddings. In our study, we experiment and generate embeddings using two methods namely:
word2doc2vec and hybrid-word2doc2vec. word2doc2vec is an approach that makes use of the
Word2Vec framework [
Our second method employs a hybrid approach exploring a combination of a dictionary-based
Named Entity Recognition using Whatizit tool [
In order to analyze the potential offered by these embeddings we intend to perform a visualization based on clustering between both our approaches, judge clusters from protein embeddings against UniRef clusters accounting for the direct propagation technique based on sequence similarity, perform a basic analysis on the emerging protein text vocabulary as well as test prediction potential. To test the prediction potential, we intend to narrow down the emerging results to two criterias: (i) high sequence similarity but low embedding similarity and (ii) low sequence similarity but high embedding similarity. The purpose of doing so is to ultimately define and implement a strategy on how embeddings could propagate function annotations as well as estimate their scope for curation work.
This work was partially supported by the BMBF-funded de.NBI Cloud within the German Network for Bioinformatics Infrastructure (de.NBI) (031A532B, 031A533A, 031A533B, 031A534A, 031A535A, 031A537A, 031A537B, 031A537C, 031A537D, 031A538A).