https://doi.org/10.6084/m9.figshare.8135393.v3. In another study, deepNF (Gligorijevi, Barot & Bonneau, 2018) was constructed by multimodal deep AE to capture hidden information in proteins from different types of interaction networks. Iranian journal of pharmaceutical research: IJPR, Proteins: Structure, Function, and Bioinformatics, Frontiers in Bioengineering and Biotechnology, Department of Computer Science and Engineering, University of Minnesota, IEEE/ACM Transactions on Computational Biology and Bioinformatics, http://bioinf.cs.ucl.ac.uk/downloads/mtdnn, https://github.com/duongvtt96/Comparison-GO-annotation-systems, https://doi.org/10.6084/m9.figshare.8135393.v3, Database resources of the national center for biotechnology information, Gapped blast and psi-blast: a new generation of protein database search programs, Proteomics applications in health: biomarker and drug discovery and food industry, Gene ontology: tool for the unification of biology, Autoencoders, unsupervised learning, and deep architectures, Proceedings of ICML workshop on unsupervised and transfer learning, Machine learning techniques for protein function prediction, Sdn2go: an integrated deep learning model for protein function prediction, Prolango: protein function prediction using neural machine translation based on a recurrent neural network, TALE: transformer-based protein function Annotation with joint sequenceLabel Embedding, Deep autoencoder neural networks for gene ontology annotation predictions, Proceedings of the 5th ACM conference on bioinformatics, computational biology, and health informatics, Computational methods for annotation transfers from sequence, Ffpred 3: feature-based function prediction for all gene ontology domains, An integrated probabilistic model for functional prediction of proteins, Mapping gene ontology to proteins based on proteinprotein interaction data, Deepadd: protein function prediction from k-mer embedding and additional features, Predicting human protein function with multi-task deep neural networks, A decision-theoretic generalization of on-line learning and an application to boosting, Automated protein function predictionthe genomic challenge, deepnf: deep network fusion for protein function prediction, Gofdr: a sequence alignment based method for predicting protein functions, Pfp: automated prediction of gene ontology functional annotations with confidence scores using protein sequence data, Automated gene ontology annotation for anonymous sequence data, The genomematrix information retrieval system, Poster Abstracts of HGM2002 Human Genome Meeting (HGM2002), An expanded evaluation of protein function prediction methods shows an improvement in accuracy, Automatic annotation of protein functional class from sparse and imbalanced data sets, VLDB workshop on data mining and bioinformatics, Pogo: prediction of gene ontology terms for fungal proteins, The kegg resource for deciphering the genome, Gofigure: automated gene ontology annotation, Bayesian markov random field analysis for protein function prediction based on network data, Deepgoplus: improved protein function prediction from sequence, Deepgo: predicting protein functions from sequence and interactions using a deep ontology-aware classifier, Ms-k nn: protein function prediction by integrating multiple data sources, Handwritten digit recognition with a back-propagation network, Advances in neural information processing systems, Predicting protein function from protein/protein interaction data: a probabilistic approach, Deep learning in bioinformatics: introduction, application, and perspective in the big data era, Gonet: a deep network to annotate proteins via recurrent convolution networks, 2020 IEEE international conference on bioinformatics and biomedicine (BIBM), volume 2, Ffpred: an integrated feature-based function prediction server for vertebrate proteomes, Inferring function using patterns of native disorder in proteins, A combined algorithm for genome-wide prediction of protein function, Gotcha: a new method for prediction of protein function assessed by the annotation of seven genomes, Probabilistic protein function prediction from heterogeneous genome-wide data, Beyond homology transfer: deep learning for automated annotation of proteins, Computational approaches for protein function prediction: a survey, Integrating multi-network topology for gene function prediction using deep neural networks, Improved biomolecular annotation prediction through weighting scheme methods, International meeting on computational intelligence methods for bioinformatics and biostatistics, Tenth edition, CIBB 2013, Computational algorithms to predict gene ontology annotations, Inga: protein function prediction combining interaction networks, domain assignments and sequence similarity, A large-scale evaluation of computational protein function prediction, Protein function predictionthe power of multiplicity, Deepred: automated protein function prediction with multi-task feed-forward deep neural networks, The funcat, a functional annotation scheme for systematic classification of proteins from whole genomes, Proceedings of the Twelfth International Conference on Artificial Intelligence and Statistics, Functional annotation prediction: all for one and one for all, Pfp-wgan: protein function prediction by discovering gene ontology term correlations with generative adversarial networks, Network-based prediction of protein function, A survey of computational methods for protein function prediction, An overview of in silico protein function prediction, Hierachial protein function prediction with tails-gnns, Genome annotation: from sequence to biology, Near perfect protein multi-label classification with deep neural networks, Seclaf: a webserver and deep neural network design tool for hierarchical biological sequence classification, The string database in 2017: quality-controlled proteinprotein association networks, made broadly accessible, Towards recognition of protein function based on its structure using deep convolutional networks, 2016 IEEE international conference on bioinformatics and biomedicine (BIBM), Pannzer2: a rapid functional annotation web server, Gopet: a tool for automated predictions of gene ontology terms, Protein function prediction is improved by creating synthetic feature samples with generative adversarial networks, Predicting protein function from sequence and structural data, Enzyme nomenclature 1992. Novel FFPred that uses training sample augmented by a deep learning architecture. Probabilistic framework based on PPI, predicting GO assignments for 10% unannotated proteins in yeast. They automatically extract high-level features from raw data and provide predictions in an end-to-end manner. Consequently, assigning protein function from scratch using machine learning, i.e., directly inferring the annotation from the amino acid sequence, without access to any additional references or databases, is an ongoing task. Note: You are now also subscribed to the subject areas of this publication The second one is structure-based or otherwise exploits big data from several available resources. Meanwhile, PFP-WGAN (Seyyedsalehi et al., 2021) is one of two latest ideas that use GAN to infer the functionalities of proteins. Working on the same model organism, S. cerevisiae, as used in the methods described above, Kourmpetis et al. and will receive updates in the daily or weekly email digests if turned on. We used the finalized benchmark set provided in the CAFA3 report (Zhou et al., 2019). In these methods, the unknown sequence is searched in a database that curates well-annotated proteins. Further, part-of implies that the child node is necessarily part of the parent. The current rate at which new DNA and protein sequences are being generated is too fast to experimentally discover the functions of those sequences, emphasizing the need for accurate Automatic Function Prediction (AFP) methods. Alternatively, low-dimensional feature representations in the hidden layer are extracted and fed into an SVM classifier for final classification (Gligorijevi, Barot & Bonneau, 2018) or a CNN model (Peng et al., 2020) for final classification. Bonetta & Valentino (2020) demonstrate protein function prediction in the machine learning workflow. The challenges we consider are how to: (1) include condition-specific functional annotation, (2) predict functions for non-model species, (3) include new informative data sources, (4) deal with the biases of Gene Ontology (GO) annotations, and (5) maximally exploit the GO to obtain performance gains. Because of the variability in the vocabulary used to define protein function, which makes the annotation process confusing to both humans and machines, various databases have been proposed to provide a standardized scheme, such as the Enzyme Commission (EC)(Webb, 1992), Functional Catalogue (FunCat)(Ruepp et al., 2004), and Kyoto Encyclopedia of Genes and Genomes (KEGG)(Kanehisa et al., 2004). We also provide recommendations for addressing those challenges, by adapting (1) the way we represent proteins and genes, (2) the way we represent gene functions, and (3) the algorithms that perform the prediction from gene to function. Letovsky & Kasif (2003) employed a functional linkage graph (Marcotte et al., 1999) constructed based on a PPI network of the yeast Saccharomyces cerevisiae. The GO database has facilitated a comprehensive vocabulary for functional annotation because it presents structured GO functions in three domains (MF, BP, and CC), thereby effectively supporting in silico protein function assignment. Unlike supervised learning, unsupervised learning independently mines hidden patterns in input distribution. no more than one email per day or week based on your preferences. These tools relied on local alignment search tools, such as the Basic Local Alignment Search Tool (BLAST)(Altschul et al., 1997). GO annotation predictors using the traditional approach. Please note that many of the page functionalities won't work as expected without javascript enabled. For INGA, we obtained predictions for the CAFA3 benchmark based on the predictions of CAFA3 targets, which were provided on their website. Machine learning approaches are highly advantageous and considered the future direction for AFP. Another type of neural network, fully connected deep network (FCDN), is a series of fully connected layers. Protein sequences are experimentally represented by three types of descriptors(subsequence profile map, pseudo-amino acid composition, and conjoint triad feature), with subsequence profile map performing best in the analysis. The very first solution of similarity-based methods was homology-based. CNN-based framework, predicting protein function from sequence and additional information (PPI or SSP). The authors converted amino acid sequences and GO terms into ProLan and GOLan languages, respectively. In addition to proposing a new approach, accuracy assessment is important for demonstrating the improvements of a novel methodology. DeepGOPlus is a recently developed representative annotation system that is based on deep learning. Working on a dataset identical to the one collated by FFPred3, Fa et al. 
https://doi.org/10.3390/genes11111264, Makrodimitris, Stavros, Roeland C. H. J. van Ham, and Marcel J. T. Reinders. Papers are submitted upon individual invitation or recommendation by the scientific editors and undergo peer review We review the representative solutions below, focusing on three categories: similarity-based methods, probabilistic methods, and machine learning methods. Feature Papers represent the most advanced research with significant potential for high impact in the field. As it is difficult to introduce all the available methodologies and comprehensively compare them, we tested the prominent predictors in this study. This model was utilized for converting input vectors from InterPro and predicting GO terms (Zhang et al., 2019). NetGO improves the performance of a large scale AFP by accessing the enormous proteinprotein network of over 2000 species in the STRING database (Szklarczyk et al., 2016). One of the earliest deep learning-based methods for GO annotation predictions is the one proposed by Chicco, Sadowski & Baldi (2014). Deep learning-based methods for assigning GO terms to proteins. We used Google Scholar https://scholar.google.com/ as the literature database to retrieve relevant publications, without applying any restrictions as to the publishing data, journal, or publisher. 
automatic function prediction; Gene Ontology; protein representation; machine learning, Help us to further improve by taking part in this short 5 minute survey, Genetic-Based Hypertension Subtype Identification Using Informative SNPs, Expression of Genes Encoding Manganese Peroxidase and Laccase of. Finally, we discussed the remaining major challenges in the field, and emphasized the future directions for protein function prediction with GO. Branch (a) of Fig. The statistics are presented in Table 3. https://doi.org/10.3390/genes11111264, Makrodimitris S, van Ham RCHJ, Reinders MJT. One is a sequence-only approach, which is useful for predicting the functions of novel proteins in the absence of homologous information or other references. The source code and data are available at, Advanced method of DeepGO, relying on a structure of several CNN layers. NetGO (You et al., 2019) is an extension of GoLabeler (You et al., 2018b), which employs the learning-to-rank(LTR) model to integrate sequence-based evidence. Lately, TALE (Cao & Shen, 2021) has been developed to generate GO predictions by integrating sequence patterns based on transformer encoder and the joint similarity of sequence-term. On the one hand, understanding protein function is essential for deciphering biological evolution and for countless applications, such as drug design and disease treatment. First, protein domain, family, and motif information is queried from InterPro and encoded before passing through fully connected layers. With respect to the specific approaches, machine learning models may be limited by the heterogeneity of genomic data when mining different sources of information, while parameter and hyper-parameter tuning is the challenging step of the deep learning approach. Author to whom correspondence should be addressed. Please let us know what you think of our products and services. This research was supported by the National Research Foundation of Korea (NRF) funded by the Ministry of Science, ICT and Future Planning (NRF-2019R1A2C1084308). CNNs can be applied in AFP either alone (Kulmanov, Khan & Hoehndorf, 2018; Cai, Wang & Deng, 2020; Du et al., 2020) or in combination with other architectures (Zhang et al., 2020; Spalevi et al., 2020). In addition to information generated by InterPro and PPI, global and local semantic features of amino acid sequences are extracted by Bi-LSTM and a convolutional layer, respectively. Conversely, scores in the full mode were computed on all protein sequences in the benchmark set. Protein function prediction is a crucial part of genome annotation. We also referred to references cited in the downloaded papers, to capture significant studies in the field. Finally, successful solutions for GO term prediction can be expanded to include other functional resources (EC, pathways, etc. RNNs have a DNN backbone, and the units of the hidden layer are connected. Further, we selected some suitable predictors from among the reviewed tools and conducted a mini comparison of their performance using a worldwide challenge dataset. Generally, protein function identification is accomplished through manual or computational annotation. Software package allowing the user to define the. Each individual network is built for a specific GO term level on DAG, which allows the hierarchical post-processing of predictions. Researchers are employing various approaches to efficiently predict the GO terms. Autoencoder (AE) (Baldi, 2012) is an unsupervised learning model developed for assigning GO terms to amino acid sequences. Combination of a text-based method and a sequence-based method to improve large-scale protein function prediction. Leverages heterogeneous data to increase the probabilistic based model for the functional prediction of. (2010) suggested applying a Bayesian approach to the MRF model. Similar to DeepGOPlus, TALE is also combined with sequence similarity as TALE+ model to enhance its performance. Meanwhile, the deep learning approach had demonstrable advantages in this area and yielded highly competitive predictions. Although methods based on a local alignment search are straightforward and perform well to some extent, they also have some drawbacks, including database annotation errors or excessive function transfer, threshold relativity, and low sensitivity or specificity (Sasson, Kaplan & Linial, 2006). Restricted Boltzmann machine (RBM) (Salakhutdinov & Hinton, 2009) has one hidden layer for representing latent features and an input layer encoding the observed data. The two latest reviews are prepared byBonetta & Valentino (2020) andZhao et al. Instead of only using InterPro terms as features, PoGO integrates three more sources (sequence similarity, biochemical properties, and protein tertiary structure). This type of Ensemble model with InterPro sequence similarity, biochemical property, and protein tertiary structure to predict the GO of fungal proteins. The computational GO annotation of proteins has been an actively pursued and challenging task in bioinformatics since around the 2000s; this is a response to the need to bridge the gap between the known and unknown, newly discovered amino acid sequences. DeepGOPlus (Kulmanov & Hoehndorf, 2020) has been developed to overcome the existing limitations of DeepGO (Kulmanov, Khan & Hoehndorf, 2018), such as sequence length, unavailable PPI features, and the number of GO labels. "Automatic Gene Function Prediction in the 2020s" Genes 11, no. See further details. Effective LTR-based web server, combining both sequence and massive network information of proteins to annotate gene products. For the CC class, two baselines (Naive and BLAST) were at the top for Fmax and recall, followed by FFPred3. Models using the traditional approach and deep learning approach are depicted in blue and orange tones, respectively. Supervised learning is an essential approach for in silico protein function prediction. interesting to authors, or important in this field. This was complemented by the analysis of related publications, with the assessment outcomes computed by CAFA, a worldwide venue for comparing computational protein function predictors. (This article belongs to the Special Issue. The user can implement their self-configured models in a downloadable program or run two pre-trained models (Szalkai & Grolmusz, 2018a) on a web server. The statements, opinions and data contained in the journal, 1996-2022 MDPI (Basel, Switzerland) unless otherwise stated. Next, topological features of PPI are obtained by the Deepwalk algorithm. Based on a similar concept of combining primary protein structure and PPI, GONET (Li et al., 2020), a novel model, was built by employing CNN, RNN, and Attention layer for human and mouse sequences. They play a role in numerous processes, including biochemical reactions, transmission of signals, nutrient transport, immune system boosting, etc. In that part, we summarize three main sub-categories of the traditional approach, and mention prominent or most recent corresponding studies. Then, experimental annotations for a subset of target sequences are accumulated until t1, to complete a benchmark dataset for the performance evaluation. Produces output as a clickable graph in four steps, including homologous sequence search, minimum cover graph construction, and assigning ontologies after scoring them. Another tool, Prediction of Gene Ontology terms (PoGO) (Jung et al., 2010), has been developed from Automatic Annotation of Protein Functional Class (AAPFC) (Jung & Thon, 2006). Previous reviews focused on AFP(Rost et al., 2003) in terms of the data type used(Watson, Laskowski & Thornton, 2005; Pandey, Kumar & Steinbach, 2006; Sleator & Walsh, 2010; Shehu, Barbar & Molloy, 2016), drawbacks and corresponding solutions(Friedberg, 2006), protein interaction networks(Sharan, Ulitsky & Shamir, 2007), types of classified function(Rentzsch & Orengo, 2009), and GO assignment based on sequence information(Cozzetto & Jones, 2017). Deep model for protein function prediction using representation learning to embed protein sequences and networks. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. You can add specific subject areas through your profile settings. Typically, unsupervised learning models are employed for clustering, reducing dimensions, and transforming data. Some approaches have been suggested and developed to cope with imbalanced data; these include working on the subgroups in which the classes are more balanced, data augmentation using GAN, and considering evaluation metrics specified for imbalanced data, such as AUPR. The servers of several tools have been updated recently, but we need to control their training data or databases used, in which do not include sequences in the CAFA3 benchmark. The former approach is the gold standard for functional annotation, because it is implemented by expert annotators and yields high quality curated results. Automated function prediction (AFP) based on the GO system is a challenging problem in bioinformatics. Multiple alignment-based method utilizing FDRs and PSSM to rank GO terms for amino acid sequences. The following information was supplied regarding data availability: Supplemental documents are available at GitHub: https://github.com/duongvtt96/Comparison-GO-annotation-systems. For DeepMNE-CNN (Peng et al., 2020) has a superior performance than deepNF in the human data by utilizing CNN layers instead of SVM for the classification model. You are accessing a machine-readable page. Prediction methods have recently witnessed rapid development, owing to the emergence of high-throughput sequencing technologies. The statements, opinions and data contained in the journals are solely Our promise DeepAdd (Du et al., 2020) was inspired by DeepGO server and provides a solution for AFP, utilizing a CNN framework to learn vector representations from sequences and additional information. The scores of Naive, BLAST, FFPred3 and DEEPred are only available for the full mode. Based on the timeline of the challenge, t0 is the deadline for the prediction submission of the target sequences. Therefore, we chose INGA, which was one of the top models in CAFA3, as a representative conventional tool. You seem to have javascript disabled. For instance, if term A is denoted as is-a of term B, it means that A is a sub-type of B. Accordingly, the critical assessment of functional annotation (CAFA) (Radivojac et al., 2013; Jiang et al., 2016; Zhou et al., 2019) is a community-based experiment that provides a large-scale evaluation of computational protein function prediction methods in a time-delayed manner. TypoMissing or incorrect metadataQuality: PDF, figure, table, or data qualityDownload issuesAbusive behaviorResearch misconductOther issue not listed above. (2020). Find support for a specific problem in the support section of our website. After the learning stage, a model that captures the relationship between feature and function is produced, and used to predict the GO terms for novel amino acid sequences. Further, we highlighted the challenges and future prospects of the field. 
These updates will appear in your home dashboard each time you visit PeerJ. Four main competitions (CAFA1CAFA4) have been held every 3 years since 2010. Herein, we reviewed the currently available computational GO annotation methods for proteins, ranging from conventional to deep learning approach. Based on previous studies(Jung & Thon, 2006; Jung et al., 2010) and the aforementioned surveys, our review is divided in two main parts offering an overview of the field (from its inception to its present state). No special The GO consortium created a database for a controlled vocabulary describing the functional properties of genomic products (e.g.,genes, proteins, and RNA). Dataset used in this comparison (final benchmark CAFA3) is available at Figshare: Zhou, Naihui (2019): Supplementary_data. Recurrent neural network (RNN) is a deep learning architecture developed especially for sequential data. published in the various research areas of the journal. In terms of output, the GO database has been updating because GO annotations are still imbalanced and not complete for all species. Each ontology (vocabulary) belongs to one of three categories: molecular function (MF), biological process (BP), and cellular component (CC). ProLanGo (Cao et al., 2017) is the first tool that applies Neural Machine Translation (NMT) developed by Google in AFP. DeepFunc (Zhang et al., 2019) is a novel predictor that surpasses DeepGO, FFPred3, and BLAST. The components presented in the dashed box are optional, depending on each method. Nonetheless, this approach is expensive and laborious, and thus, it is difficult to scale. Currently, the amount of generated genomic data, and the numbers of sophisticated algorithms and computational resources are rapidly growing. The overall outcome is presented similarly in Fig. Integrates architectures with three sub-models and a weighting classifier to achieve GO term predictions. The multi-label problem can be addressed via the advancement of computational resources and well-defined solutions, for example, stacking many individual solutions together.
https://doi.org/10.3390/genes11111264, Makrodimitris, Stavros, Roeland C. H. J. van Ham, and Marcel J. T. Reinders. Papers are submitted upon individual invitation or recommendation by the scientific editors and undergo peer review We review the representative solutions below, focusing on three categories: similarity-based methods, probabilistic methods, and machine learning methods. Feature Papers represent the most advanced research with significant potential for high impact in the field. As it is difficult to introduce all the available methodologies and comprehensively compare them, we tested the prominent predictors in this study. This model was utilized for converting input vectors from InterPro and predicting GO terms (Zhang et al., 2019). NetGO improves the performance of a large scale AFP by accessing the enormous proteinprotein network of over 2000 species in the STRING database (Szklarczyk et al., 2016). One of the earliest deep learning-based methods for GO annotation predictions is the one proposed by Chicco, Sadowski & Baldi (2014). Deep learning-based methods for assigning GO terms to proteins. We used Google Scholar https://scholar.google.com/ as the literature database to retrieve relevant publications, without applying any restrictions as to the publishing data, journal, or publisher. automatic function prediction; Gene Ontology; protein representation; machine learning, Help us to further improve by taking part in this short 5 minute survey, Genetic-Based Hypertension Subtype Identification Using Informative SNPs, Expression of Genes Encoding Manganese Peroxidase and Laccase of. Finally, we discussed the remaining major challenges in the field, and emphasized the future directions for protein function prediction with GO. Branch (a) of Fig. The statistics are presented in Table 3. https://doi.org/10.3390/genes11111264, Makrodimitris S, van Ham RCHJ, Reinders MJT. One is a sequence-only approach, which is useful for predicting the functions of novel proteins in the absence of homologous information or other references. The source code and data are available at, Advanced method of DeepGO, relying on a structure of several CNN layers. NetGO (You et al., 2019) is an extension of GoLabeler (You et al., 2018b), which employs the learning-to-rank(LTR) model to integrate sequence-based evidence. Lately, TALE (Cao & Shen, 2021) has been developed to generate GO predictions by integrating sequence patterns based on transformer encoder and the joint similarity of sequence-term. On the one hand, understanding protein function is essential for deciphering biological evolution and for countless applications, such as drug design and disease treatment. First, protein domain, family, and motif information is queried from InterPro and encoded before passing through fully connected layers. With respect to the specific approaches, machine learning models may be limited by the heterogeneity of genomic data when mining different sources of information, while parameter and hyper-parameter tuning is the challenging step of the deep learning approach. Author to whom correspondence should be addressed. Please let us know what you think of our products and services. This research was supported by the National Research Foundation of Korea (NRF) funded by the Ministry of Science, ICT and Future Planning (NRF-2019R1A2C1084308). CNNs can be applied in AFP either alone (Kulmanov, Khan & Hoehndorf, 2018; Cai, Wang & Deng, 2020; Du et al., 2020) or in combination with other architectures (Zhang et al., 2020; Spalevi et al., 2020). In addition to information generated by InterPro and PPI, global and local semantic features of amino acid sequences are extracted by Bi-LSTM and a convolutional layer, respectively. Conversely, scores in the full mode were computed on all protein sequences in the benchmark set. Protein function prediction is a crucial part of genome annotation. We also referred to references cited in the downloaded papers, to capture significant studies in the field. Finally, successful solutions for GO term prediction can be expanded to include other functional resources (EC, pathways, etc. RNNs have a DNN backbone, and the units of the hidden layer are connected. Further, we selected some suitable predictors from among the reviewed tools and conducted a mini comparison of their performance using a worldwide challenge dataset. Generally, protein function identification is accomplished through manual or computational annotation. Software package allowing the user to define the. Each individual network is built for a specific GO term level on DAG, which allows the hierarchical post-processing of predictions. Researchers are employing various approaches to efficiently predict the GO terms. Autoencoder (AE) (Baldi, 2012) is an unsupervised learning model developed for assigning GO terms to amino acid sequences. Combination of a text-based method and a sequence-based method to improve large-scale protein function prediction. Leverages heterogeneous data to increase the probabilistic based model for the functional prediction of. (2010) suggested applying a Bayesian approach to the MRF model. Similar to DeepGOPlus, TALE is also combined with sequence similarity as TALE+ model to enhance its performance. Meanwhile, the deep learning approach had demonstrable advantages in this area and yielded highly competitive predictions. Although methods based on a local alignment search are straightforward and perform well to some extent, they also have some drawbacks, including database annotation errors or excessive function transfer, threshold relativity, and low sensitivity or specificity (Sasson, Kaplan & Linial, 2006). Restricted Boltzmann machine (RBM) (Salakhutdinov & Hinton, 2009) has one hidden layer for representing latent features and an input layer encoding the observed data. The two latest reviews are prepared byBonetta & Valentino (2020) andZhao et al. Instead of only using InterPro terms as features, PoGO integrates three more sources (sequence similarity, biochemical properties, and protein tertiary structure). This type of Ensemble model with InterPro sequence similarity, biochemical property, and protein tertiary structure to predict the GO of fungal proteins. The computational GO annotation of proteins has been an actively pursued and challenging task in bioinformatics since around the 2000s; this is a response to the need to bridge the gap between the known and unknown, newly discovered amino acid sequences. DeepGOPlus (Kulmanov & Hoehndorf, 2020) has been developed to overcome the existing limitations of DeepGO (Kulmanov, Khan & Hoehndorf, 2018), such as sequence length, unavailable PPI features, and the number of GO labels. "Automatic Gene Function Prediction in the 2020s" Genes 11, no. See further details. Effective LTR-based web server, combining both sequence and massive network information of proteins to annotate gene products. For the CC class, two baselines (Naive and BLAST) were at the top for Fmax and recall, followed by FFPred3. Models using the traditional approach and deep learning approach are depicted in blue and orange tones, respectively. Supervised learning is an essential approach for in silico protein function prediction. interesting to authors, or important in this field. This was complemented by the analysis of related publications, with the assessment outcomes computed by CAFA, a worldwide venue for comparing computational protein function predictors. (This article belongs to the Special Issue. The user can implement their self-configured models in a downloadable program or run two pre-trained models (Szalkai & Grolmusz, 2018a) on a web server. The statements, opinions and data contained in the journal, 1996-2022 MDPI (Basel, Switzerland) unless otherwise stated. Next, topological features of PPI are obtained by the Deepwalk algorithm. Based on a similar concept of combining primary protein structure and PPI, GONET (Li et al., 2020), a novel model, was built by employing CNN, RNN, and Attention layer for human and mouse sequences. They play a role in numerous processes, including biochemical reactions, transmission of signals, nutrient transport, immune system boosting, etc. In that part, we summarize three main sub-categories of the traditional approach, and mention prominent or most recent corresponding studies. Then, experimental annotations for a subset of target sequences are accumulated until t1, to complete a benchmark dataset for the performance evaluation. Produces output as a clickable graph in four steps, including homologous sequence search, minimum cover graph construction, and assigning ontologies after scoring them. Another tool, Prediction of Gene Ontology terms (PoGO) (Jung et al., 2010), has been developed from Automatic Annotation of Protein Functional Class (AAPFC) (Jung & Thon, 2006). Previous reviews focused on AFP(Rost et al., 2003) in terms of the data type used(Watson, Laskowski & Thornton, 2005; Pandey, Kumar & Steinbach, 2006; Sleator & Walsh, 2010; Shehu, Barbar & Molloy, 2016), drawbacks and corresponding solutions(Friedberg, 2006), protein interaction networks(Sharan, Ulitsky & Shamir, 2007), types of classified function(Rentzsch & Orengo, 2009), and GO assignment based on sequence information(Cozzetto & Jones, 2017). Deep model for protein function prediction using representation learning to embed protein sequences and networks. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. You can add specific subject areas through your profile settings. Typically, unsupervised learning models are employed for clustering, reducing dimensions, and transforming data. Some approaches have been suggested and developed to cope with imbalanced data; these include working on the subgroups in which the classes are more balanced, data augmentation using GAN, and considering evaluation metrics specified for imbalanced data, such as AUPR. The servers of several tools have been updated recently, but we need to control their training data or databases used, in which do not include sequences in the CAFA3 benchmark. The former approach is the gold standard for functional annotation, because it is implemented by expert annotators and yields high quality curated results. Automated function prediction (AFP) based on the GO system is a challenging problem in bioinformatics. Multiple alignment-based method utilizing FDRs and PSSM to rank GO terms for amino acid sequences. The following information was supplied regarding data availability: Supplemental documents are available at GitHub: https://github.com/duongvtt96/Comparison-GO-annotation-systems. For DeepMNE-CNN (Peng et al., 2020) has a superior performance than deepNF in the human data by utilizing CNN layers instead of SVM for the classification model. You are accessing a machine-readable page. Prediction methods have recently witnessed rapid development, owing to the emergence of high-throughput sequencing technologies. The statements, opinions and data contained in the journals are solely Our promise DeepAdd (Du et al., 2020) was inspired by DeepGO server and provides a solution for AFP, utilizing a CNN framework to learn vector representations from sequences and additional information. The scores of Naive, BLAST, FFPred3 and DEEPred are only available for the full mode. Based on the timeline of the challenge, t0 is the deadline for the prediction submission of the target sequences. Therefore, we chose INGA, which was one of the top models in CAFA3, as a representative conventional tool. You seem to have javascript disabled. For instance, if term A is denoted as is-a of term B, it means that A is a sub-type of B. Accordingly, the critical assessment of functional annotation (CAFA) (Radivojac et al., 2013; Jiang et al., 2016; Zhou et al., 2019) is a community-based experiment that provides a large-scale evaluation of computational protein function prediction methods in a time-delayed manner. TypoMissing or incorrect metadataQuality: PDF, figure, table, or data qualityDownload issuesAbusive behaviorResearch misconductOther issue not listed above. (2020). Find support for a specific problem in the support section of our website. After the learning stage, a model that captures the relationship between feature and function is produced, and used to predict the GO terms for novel amino acid sequences. Further, we highlighted the challenges and future prospects of the field. These updates will appear in your home dashboard each time you visit PeerJ. Four main competitions (CAFA1CAFA4) have been held every 3 years since 2010. Herein, we reviewed the currently available computational GO annotation methods for proteins, ranging from conventional to deep learning approach. Based on previous studies(Jung & Thon, 2006; Jung et al., 2010) and the aforementioned surveys, our review is divided in two main parts offering an overview of the field (from its inception to its present state). No special The GO consortium created a database for a controlled vocabulary describing the functional properties of genomic products (e.g.,genes, proteins, and RNA). Dataset used in this comparison (final benchmark CAFA3) is available at Figshare: Zhou, Naihui (2019): Supplementary_data. Recurrent neural network (RNN) is a deep learning architecture developed especially for sequential data. published in the various research areas of the journal. In terms of output, the GO database has been updating because GO annotations are still imbalanced and not complete for all species. Each ontology (vocabulary) belongs to one of three categories: molecular function (MF), biological process (BP), and cellular component (CC). ProLanGo (Cao et al., 2017) is the first tool that applies Neural Machine Translation (NMT) developed by Google in AFP. DeepFunc (Zhang et al., 2019) is a novel predictor that surpasses DeepGO, FFPred3, and BLAST. The components presented in the dashed box are optional, depending on each method. Nonetheless, this approach is expensive and laborious, and thus, it is difficult to scale. Currently, the amount of generated genomic data, and the numbers of sophisticated algorithms and computational resources are rapidly growing. The overall outcome is presented similarly in Fig. Integrates architectures with three sub-models and a weighting classifier to achieve GO term predictions. The multi-label problem can be addressed via the advancement of computational resources and well-defined solutions, for example, stacking many individual solutions together.