Analysis of Protein Function Prediction using DeepGOPlus Model

This report provides an analysis of protein function prediction, a crucial task in bioinformatics, focusing on the DeepGOPlus model developed by Maxat Kulmanove and Robert Hoehndorf. The report begins by highlighting the limitations of the earlier DeepGO model, which primarily included constraints on sequence length and the number of functions it could predict, limiting its applicability to only a fraction of proteins. In response, DeepGOPlus was developed by integrating neural network predictions with sequence similarity methods to overcome these limitations, enhancing the speed and accuracy of protein function prediction. The new model addresses issues like the sequence length limitations of its predecessor and can be applied to genome-scale annotation. The report emphasizes the importance of understanding bioinformatics techniques and associated terminology, such as CCO, CNN, and various statistical tools, for a comprehensive understanding of the research. The study concludes that DeepGOPlus is a faster and more accurate tool, particularly useful for annotating proteins in newly sequenced organisms and in metagenomics projects, due to its efficiency and ability to handle proteins from different kingdoms. Finally, it suggests potential future research directions, including the incorporation of additional features and deep neural network models to broaden the application of this technology.

News and Views
Protein Function Prediction
Prediction of protein function refers to one of the major but important tasks of bioinformatics,
where a number of techniques are used for determining the range of biological problems, in
terms of role of protein in developing disease mechanisms, so that targeted drugs could find. It
includes sequence based features, protein structure, protein–protein interaction networks and
other bioinformatics tools used for analysing the composition of molecules and model of
biological system, by collecting genomic data. Predictions of protein structure helps in producing
automatically the predicted protein structures by using amino acid sequence. However, these
methods may consume a lot of time for predicting the protein functions, therefore, a new model
is developed in this article by Maxat Kulmanove and Robert Hoehndorf as DeepGOPlus, for
improving the function prediction from sequence. But before developing this model, a proper
study is conducted to determine the factors that arise limitations in performance of DeepGO
model. It has been identified by the researchers that DeepGO model only predicts protein

functions having sequence length of less than 1002 and do not contain any ambiguous amino
acids, where more than 90% of protein sequences satisfy such criteria. So, it states that this
model only predicts functions of approximate 10% proteins. Further, lack of computation
technique is another main drawback point of DeepGO model that predicts only 2000 functions of
about 45000 proteins. In addition to these, a number of points have been researched under
present article that articulate that performance of DeepGO model is not so efficient. So,
considering these limitations, a new architectural model DeepGOPlus is developed by combining
the neural network predictions with methods that are based on sequence similarity for capturing
orthology and other interaction information. This would help in overcoming from those
limitations that reduces the performance of DeepGo used in experimental annotation. Therefore,
with aim to identify the protein structure, one can predict the functions of protein on the
similarity structural basis. It would also make able to predict molecules or drug that could
efficiently bind to protein.
Important Concepts
To read this paper, it is essential for readers to have knowledge of Bioinformatics techniques and
its used for predicting the protein functions. Along with this, many abbreviated forms are used in
this article like CCO, CNN, BPO, MFO, CAFA and models like GOLabeler and DeepText2GO,
with some statistical measurement tools like Fmax, so, without knowledge of these forms, it would
be difficult to analyse the methods and results obtain from experiments.
Major Contribution
This research has aid to identify the factors where DeepGO technique has failed to give efficient
output in predicting the protein functions. So, developing a new model - DeepGOPlus and
investigating its performance by comparing it other models, it has been concluded that it is more
fast and accurate tool for predicting protein functions from sequence of protein. DeepGOPlus
novel model has overcome from several limitations of DeepGO model by justifying that it has no
limits in terms of sequence length of amino acid. So, it can be used in applications of genome-
scale annotation of protein functions, especially, within newly sequenced organisms. It doesn’t
make any assumptions based on taxa or kingdom to which a protein belongs, like meta-genomics
protein function in which proteins from different kingdoms could be mixed. Along with this, due
to fast speed, DeepGOPlus can annotate thousands of proteins within few minutes even on single

computer system, which further enables its application in project based on metagenomics or
other for finding unknown functions.
Future Directions
Through this article, readers can evaluate further planning of researchers in terms of how
additional features and different types of test based on deep neural network models will be
incorporated. As pervious related methods could be used only for deriving known proteins, like
information obtained from interaction networks or literature. But with more development,
DeepGOPlus will be relied primarily on features which can be derived from sequence amino
acid, for ensuring that it could be applied in wider manner as possible.