DeepGO: predicting protein functions from sequence and interactions using a deep ontology-aware classifier.

Kulmanov M, Khan MA, Hoehndorf R, Wren J. · 2018

Read DeepGO: predicting protein functions from sequence and interactions using a deep ontology-aware classifier. by Kulmanov M, Khan MA, Hoehndorf R, Wren J. (2018) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc5860606.txt.

Introduction

Kulmanov and co-authors present DeepGO, a tool that predicts protein functions. Their article appeared in Bioinformatics. New sequencing yields protein sequences far faster than labs can test them. DeepGO uses deep learning to guess what a protein does. It learns from the amino acid sequence itself. It also uses a network of protein interactions across species.

Predictions follow the structure of the Gene Ontology. That system holds over 40,000 function classes with set links. The model uses those links as background knowledge. Tests used the standards of a community assessment exercise. The authors report clear gains over BLAST baselines. Gains were strongest for cell location predictions. A web server and source code are both provided.

Catalog record

CREATORKulmanov M, Khan MA, Hoehndorf R, Wren J.
YEAR2018
COLLECTIONPAPERS
FORMATTXT
FILE SIZE43 KB
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC5860606

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

DeepGO: predicting protein functions from sequence and interactions using a deep ontology-aware classifier By Maxat Kulmanov, Mohammed Asif Khan, Robert Hoehndorf, Jonathan Wren Published in: Bioinformatics DOI: 10.1093/bioinformatics/btx624 ABSTRACT Abstract Motivation A large number of protein sequences are becoming available through the application of novel high-throughput sequencing technologies. Experimental functional characterization of these proteins is time-consuming and expensive, and is often only done rigorously for few selected model organisms. Computational function prediction approaches have been suggested to fill this gap. The functions of proteins are classified using the Gene Ontology (GO), which contains over 40 000 classes. Additionally, proteins have multiple functions, making function prediction a large-scale, multi-class, multi-label problem. Results We have developed a novel method to predict protein function from sequence. We use deep learning to learn features from protein sequences as well as a cross-species protein–protein interaction network. Our approach specifically outputs information in the structure of the GO and utilizes the dependencies between GO classes as background information to construct a deep learning model. We evaluate our method using the standards established by the Computational Assessment of Function Annotation (CAFA) and demonstrate a significant improvement over baseline methods such as BLAST, in particular for predicting cellular locations. Availability and implementation Web server: http://deepgo.bio2vec.net, Source code: https://github.com/bio-ontology-research-group/deepgo Supplementary information Supplementary data are available at Bioinformatics online. 1 INTRODUCTION Advances in sequencing technology have led to a large and rapidly increasing amount of genetic and protein sequences, and the amount is expected to increase further through sequencing of additional organisms as well as metagenomics. Although knowledge of protein sequences is useful for many applications, such as phylogenetics and evolutionary biology, understanding the behavior of biological systems additionally requires knowledge of the proteins’ functions. Identifying protein functions is challenging and commonly requires in vitro or in vivo experiments (Costanzo et al., 2016), and it is obvious that experimental functional annotation of proteins will not scale with the amount of novel protein sequences becoming available. One approach to address the challenge of identifying proteins’ functions is the computational prediction of protein functions (Radivojac et al., 2013). Function prediction can use several sources of information, including protein–protein interactions (Hou, 2017; Jiang and McQuay, 2012; Kirac and Ozsoyoglu, 2008; Nguyen et al., 2011; Sharan et al., 2007), genetic interactions (Costanzo et al., 2016), evolutionary relations (Gaudet et al., 2011), protein structures and structure prediction methods (Konc et al., 2013), literature (Verspoor, 2014) or combinations of these (Sokolov and Ben-Hur, 2010). These methods have been developed for many years, and their predictive performance is improving steadily (Radivojac et al., 2013). There are several key challenges for protein function prediction methods. One of these is the complex relation between protein sequence, structure and function (Alberts et al., 2002); despite significant progress in the past years in protein structure prediction (Moult et al., 2014), it still requires large efforts to predict protein structure with sufficient quality to be useful in function prediction. Another challenge is the large and complex output space for any classification method. Protein functions are classified using the Gene Ontology (GO) (Ashburner et al., 2000) which contains over 40 000 functions and cellular locations. Additionally, the GO contains strong, formally defined relations between functions that need to be taken into account during function prediction to ensure that these predictions are consistent (Radivojac et al., 2013; Sokolov and Ben-Hur, 2010). The formal dependencies between classes in GO also lead to the situation where proteins are assigned to multiple function classes in GO, for different levels of abstraction. Furthermore, several proteins do not only have a single function but may be peiotropic and have multiple different functions, making function prediction inherently a multi-label, multi-class problem. A final challenge is that proteins do not function in isolation. In particular higher-level physiological functions that go beyond simple molecular interactions, such as apoptosis or regulation of heart rate, will require other proteins and cannot usually be predicted by considering a single protein in isolation. Due to these challenges, it is also not obvious what kind of features should be used to predict the functions of a protein, and whether they can be generated efficiently for a large number of proteins. Here, we present a novel method for predicting protein functions from protein sequence and known interactions. We combine two forms of representation learning based on multiple layers of neural networks to learn features that are useful for predicting protein functions, one method that learns features from protein sequence and another that learns representations of proteins based on their location in an interaction network. We then utilize these features in a novel deep neuro-symbolic model that is built to resemble the structure and dependencies between classes that exist within the GO, refine predictions and features on each level of GO, and ultimately optimize the performance of function prediction based on the performance over the whole ontology hierarchy. We demonstrate that our model improves performance of function prediction over a BLAST baseline, and performs particularly well in predicting cellular locations of proteins. The main advantage of our approach is that it does not rely on manually crafted features but is entirely data-driven. 2 MATERIALS AND METHODS 2.1 Datasets For our experiments, we use the Gene Ontology (GO) (Ashburner et al., 2000), downloaded on 05 January 2016 from http://geneontology.org/page/download-ontology in OBO format. The version of GO has 44 683 classes of which 1968 are obsolete. GO has three major branches, one for biological processes (BP), molecular functions (MF) and cellular components (CC), each containing 28 647, 10 161 and 3907 classes, respectively. We use SwissProt’s (Boutet et al., 2016) reviewed and manually annotated protein sequences with GO annotations downloaded on 05 January 2016 from http://www.uniprot.org/uniprot/. The dataset contains 553 232 proteins, and 525 931 proteins have function annotations. Furthermore, we select proteins with annotations with experimental evidence code (EXP, IDA, IPI, IMP, IGI, IEP, TAS and IC) and filter the proteins by maximum length of 1002 ignoring proteins with ambiguous amino acid codes (B, O, J, U, X, Z) in their sequence. Our final dataset contains 60 710 proteins annotated with 27 760 classes (19 181 in BP, 6221 in MF and 2358 in CC). The dataset covers more than 90% of all proteins with experimental annotations in SwissProt. Supplementary Figure S1 shows the sequence length distribution. 2.2 Training We trained three models, one for each sub-ontology in GO. First, we propagate annotations using the GO ontology structure and randomly split proteins into a training set (80%) and testing set (20%). Due to computational limitations and the small number of annotations for very specific GO classes, we ranked GO classes by their number of annotations and selected the top 932 terms for BP, 589 terms for MF and 436 terms for the CC ontology. These cutoff values correspond to selecting only classes with the minimum number of annotations 250, 50 and 50, for BP, MF and CC, respectively. We create three binary label vectors for each protein sequence, one for each of the GO hierarchies. If a protein sequence is annotated with a GO class from our lists of selected classes, then we assign 1 to the term’s position in the binary label vector and use it as positive sample

Read the complete text