Dataset Information

Harnessing the potential of artificial neural networks for predicting protein glycosylation.

ABSTRACT: Kinetic models offer incomparable insight on cellular mechanisms controlling protein glycosylation. However, their ability to reproduce site-specific glycoform distributions depends on accurate estimation of a large number of protein-specific kinetic parameters and prior knowledge of enzyme and transport protein levels in the Golgi membrane. Herein we propose an artificial neural network (ANN) for protein glycosylation and apply this to four recombinant glycoproteins produced in Chinese hamster ovary (CHO) cells, two monoclonal antibodies and two fusion proteins. We demonstrate that the ANN model accurately predicts site-specific glycoform distributions of up to eighteen glycan species with an average absolute error of 1.1%, correctly reproducing the effect of metabolic perturbations as part of a hybrid, kinetic/ANN, glycosylation model (HyGlycoM), as well as the impact of manganese supplementation and glycosyltransferase knock out experiments as a stand-alone machine learning algorithm. These results showcase the potential of machine learning and hybrid approaches for rapidly developing performance-driven models of protein glycosylation.

SUBMITTER: Kotidis P

PROVIDER: S-EPMC7256630 | biostudies-literature | 2020 Jun

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

Harnessing the potential of artificial neural networks for predicting protein glycosylation.

Kotidis Pavlos P Kontoravdi Cleo C

Metabolic engineering communications 20200515

Kinetic models offer incomparable insight on cellular mechanisms controlling protein glycosylation. However, their ability to reproduce site-specific glycoform distributions depends on accurate estimation of a large number of protein-specific kinetic parameters and prior knowledge of enzyme and transport protein levels in the Golgi membrane. Herein we propose an artificial neural network (ANN) for protein glycosylation and apply this to four recombinant glycoproteins produced in Chinese hamster ...[more]

PMID: 32489858

Similar Datasets

Project description:Post-translational modifications (PTMs) provide an extensible framework for regulation of protein behavior beyond the diversity represented within the genome alone. While the rate of identification of PTMs has rapidly increased in recent years, our knowledge of PTM functionality encompasses less than 5% of this data. We previously developed SAPH-ire (Structural Analysis of PTM Hotspots) for the prioritization of eukaryotic PTMs based on function potential of discrete modified alignment positions (MAPs) in a set of 8 protein families. A proteome-wide expansion of the dataset to all families of PTM-bearing, eukaryotic proteins with a representational crystal structure and the application of artificial neural network (ANN) models demonstrated the broader applicability of this approach. Although structural features of proteins have been repeatedly demonstrated to be predictive of PTM functionality, the availability of adequately resolved 3D structures in the Protein Data Bank (PDB) limits the scope of these methods. In order to bridge this gap and capture the larger set of PTM-bearing proteins without an available, homologous structure, we explored all available MAP features as ANN inputs to identify predictive models that do not rely on 3D protein structural data. This systematic, algorithmic approach explores 8 available input features in exhaustive combinations (247 models; size 2-8). To control for potential bias in random sampling for holdback in training sets, we iterated each model across 100 randomized, sample training and testing sets-yielding 24,700 individual ANNs. The size of the analyzed dataset and iterative generation of ANNs represents the largest and most thorough investigation of predictive models for PTM functionality to date. Comparison of input layer combinations allows us to quantify ANN performance with a high degree of confidence and subsequently select a top-ranked, robust fit model which highlights 3,687 MAPs, including 10,933 PTMs with a high probability of biological impact but without a currently known functional role.

Project description:Machine learning methods for protein function prediction are urgently needed, especially now that a substantial fraction of known sequences remains unannotated despite the extensive use of functional assignments based on sequence similarity. One major bottleneck supervised learning faces in protein function prediction is the structured, multi-label nature of the problem, because biological roles are represented by lists of terms from hierarchically organised controlled vocabularies such as the Gene Ontology. In this work, we build on recent developments in the area of deep learning and investigate the usefulness of multi-task deep neural networks (MTDNN), which consist of upstream shared layers upon which are stacked in parallel as many independent modules (additional hidden layers with their own output units) as the number of output GO terms (the tasks). MTDNN learns individual tasks partially using shared representations and partially from task-specific characteristics. When no close homologues with experimentally validated functions can be identified, MTDNN gives more accurate predictions than baseline methods based on annotation frequencies in public databases or homology transfers. More importantly, the results show that MTDNN binary classification accuracy is higher than alternative machine learning-based methods that do not exploit commonalities and differences among prediction tasks. Interestingly, compared with a single-task predictor, the performance improvement is not linearly correlated with the number of tasks in MTDNN, but medium size models provide more improvement in our case. One of advantages of MTDNN is that given a set of features, there is no requirement for MTDNN to have a bootstrap feature selection procedure as what traditional machine learning algorithms do. Overall, the results indicate that the proposed MTDNN algorithm improves the performance of protein function prediction. On the other hand, there is still large room for deep learning techniques to further enhance prediction ability.

Dataset Information

Harnessing the potential of artificial neural networks for predicting protein glycosylation.

Publications

Harnessing the potential of artificial neural networks for predicting protein glycosylation.

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets