SecProCT: In Silico Prediction of Human Secretory Proteins Based on Capsule Network and Transformer

Identifying secretory proteins from blood, saliva or other body fluids has become an effective method of diagnosing diseases. Existing secretory protein prediction methods are mainly based on conventional machine learning algorithms and are highly dependent on the feature set from the protein. In this article, we propose a deep learning model based on the capsule network and transformer architecture, SecProCT, to predict secretory proteins using only amino acid sequences. The proposed model was validated using cross-validation and achieved 0.921 and 0.892 accuracy for predicting blood-secretory proteins and saliva-secretory proteins, respectively. Meanwhile, the proposed model was validated on an independent test set and achieved 0.917 and 0.905 accuracy for predicting blood-secretory proteins and saliva-secretory proteins, respectively, which are better than conventional machine learning methods and other deep learning methods for biological sequence analysis. The main contributions of this article are as follows: (1) a deep learning model based on a capsule network and transformer architecture is proposed for predicting secretory proteins. The results of this model are better than the those of existing conventional machine learning methods and deep learning methods for biological sequence analysis; (2) only amino acid sequences are used in the proposed model, which overcomes the high dependence of existing methods on the annotated protein features; (3) the proposed model can accurately predict most experimentally verified secretory proteins and cancer protein biomarkers in blood and saliva.     

Computational Prediction of O-linked Glycosylation Sites That Preferentially Map on Intrinsically Disordered Regions of Extracellular Proteins

O-glycosylation of mammalian proteins is one of the important posttranslational modifications. We applied a support vector machine (SVM) to predict whether Ser or Thr is glycosylated, in order to elucidate the O-glycosylation mechanism. O-glycosylated sites were often found clustered along the sequence, whereas other sites were located sporadically. Therefore, we developed two types of SVMs for predicting clustered and isolated sites separately. We found that the amino acid composition was effective for predicting the clustered type, whereas the site-specific algorithm was effective for the isolated type. The highest prediction accuracy for the clustered type was 74%, while that for the isolated type was 79%. The existence frequency of amino acids around the O-glycosylation sites was different in the two types: namely, Pro, Val and Ala had high existence probabilities at each specific position relative to a glycosylation site, especially for the isolated type. Independent component analyses for the amino acid sequences around O-glycosylation sites showed the position-specific existences of the identified amino acids as independent components. The O-glycosylation sites were preferentially located within intrinsically disordered regions of extracellular proteins: particularly, more than 90% of the clustered O-GalNAc glycosylation sites were observed in intrinsically disordered regions. This feature could be the key for understanding the non-conservation property of O-glycosylation, and its role in functional diversity and structural stability.     

Physicochemical feature-based classification of amino acid mutations

A huge quantity of gene and protein sequences have become available during the post-genomic era, and information about genetic variations, including amino acid substitutions and SNPs, is accumulating rapidly. To understand the effects of these changes, it is often essential to apply bioinformatics tools. Where there is a lack of homologous sequences or a three-dimensional structure, it becomes essential to predict the effects of mutations based solely on protein sequence information. Several computational methods utilizing machine learning techniques have been developed. These predictions generally use the 20-alphabet amino acid code to train the model. With limited available data, the 20-alphabet amino acid features may introduce so many parameters that the model becomes over-fitted. To decrease the number of parameters, we propose a physicochemical feature-based method to forecast the effects of amino acid substitutions on protein stability. Protein structure alterations caused by mutations can be classified as stabilizing or destabilizing. Based on experimental folding-unfolding free energy (DeltaDeltaG) values, we trained a support vector machine with a cleaned data set. The physicochemical properties of the mutated residues, the number of neighboring residues in the primary sequence and the temperature and pH were used as input attributes. Different kernel functions, attributes and window sizes were optimized. An average accuracy of 80% was obtained in cross-validation experiments.     

Random forest models to predict the densities and surface tensions of deep eutectic solvents

The use of machine learning in physicochemical properties modeling has great potential to accelerate the application of emerging materials. Deep eutectic solvents (DESs), an emerging class of solvents, are promising for applications as inexpensive “designer” solvents. Due to the unique structure of DESs, the hydrogen bond donor and hydrogen bond acceptor can be varied to create a mixture with specific physical properties. In this work, we proposed random forest (RF) models to predict the densities and the surface tensions of DESs, which are essential for the separation process. In the proposed models, the structural information and the calculated critical properties were used as two different types of features, respectively. The results demonstrate that the RF models predict the densities and surface tensions of DESs with high accuracy, with absolute average relative deviation (AARD%) less than 1% in the prediction of density and 3% in the prediction of surface tension.     

PredNTS: Improved and Robust Prediction of Nitrotyrosine Sites by Integrating Multiple Sequence Features

Nitrotyrosine, which is generated by numerous reactive nitrogen species, is a type of protein post-translational modification. Identification of site-specific nitration modification on tyrosine is a prerequisite to understanding the molecular function of nitrated proteins. Thanks to the progress of machine learning, computational prediction can play a vital role before the biological experimentation. Herein, we developed a computational predictor PredNTS by integrating multiple sequence features including K-mer, composition of k-spaced amino acid pairs (CKSAAP), AAindex, and binary encoding schemes. The important features were selected by the recursive feature elimination approach using a random forest classifier. Finally, we linearly combined the successive random forest (RF) probability scores generated by the different, single encoding-employing RF models. The resultant PredNTS predictor achieved an area under a curve (AUC) of 0.910 using five-fold cross validation. It outperformed the existing predictors on a comprehensive and independent dataset. Furthermore, we investigated several machine learning algorithms to demonstrate the superiority of the employed RF algorithm. The PredNTS is a useful computational resource for the prediction of nitrotyrosine sites. The web-application with the curated datasets of the PredNTS is publicly available.     

Prediction of protein-protein interaction sites using support vector machines

The identification of protein-protein interaction sites is essential for the mutant design and prediction of protein-protein networks. The interaction sites of residue units were predicted using support vector machines (SVM) and the profiles of sequentially/spatially neighboring residues, plus additional information. When only sequence information was used, prediction performance was highest using the feature vectors, sequentially neighboring profiles and predicted interaction site ratios, which were calculated by SVM regression using amino acid compositions. When structural information was also used, prediction performance was highest using the feature vectors, spatially neighboring residue profiles, accessible surface areas, and the with/without protein interaction sites ratios predicted by SVM regression and amino acid compositions. In the latter case, the precision at recall = 50% was 54-56% for a homo-hetero mixed test set and >20% higher than for random prediction. Approximately 30% of the residues wrongly predicted as interaction sites were the closest sequentially/spatially neighboring on the interaction site residues. The predicted residues covered 86-87% of the actual interfaces (96-97% of interfaces with over 20 residues). This prediction performance appeared to be slightly higher than a previously reported study. Comparing the prediction accuracy of each molecule, it seems to be easier to predict interaction sites for stable complexes.     

PYTHIA: Deep Learning Approach for Local Protein Conformation Prediction

Protein Blocks (PBs) are a widely used structural alphabet describing local protein backbone conformation in terms of 16 possible conformational states, adopted by five consecutive amino acids. The representation of complex protein 3D structures as 1D PB sequences was previously successfully applied to protein structure alignment and protein structure prediction. In the current study, we present a new model, PYTHIA (predicting any conformation at high accuracy), for the prediction of the protein local conformations in terms of PBs directly from the amino acid sequence. PYTHIA is based on a deep residual inception-inside-inception neural network with convolutional block attention modules, predicting 1 of 16 PB classes from evolutionary information combined to physicochemical properties of individual amino acids. PYTHIA clearly outperforms the LOCUSTRA reference method for all PB classes and demonstrates great performance for PB prediction on particularly challenging proteins from the CASP14 free modelling category.     

PD-BertEDL: An Ensemble Deep Learning Method Using BERT and Multivariate Representation to Predict Peptide Detectability

Peptide detectability is defined as the probability of identifying a peptide from a mixture of standard samples, which is a key step in protein identification and analysis. Exploring effective methods for predicting peptide detectability is helpful for disease treatment and clinical research. However, most existing computational methods for predicting peptide detectability rely on a single information. With the increasing complexity of feature representation, it is necessary to explore the influence of multivariate information on peptide detectability. Thus, we propose an ensemble deep learning method, PD-BertEDL. Bidirectional encoder representations from transformers (BERT) is introduced to capture the context information of peptides. Context information, sequence information, and physicochemical information of peptides were combined to construct the multivariate feature space of peptides. We use different deep learning methods to capture the high-quality features of different categories of peptides information and use the average fusion strategy to integrate three model prediction results to solve the heterogeneity problem and to enhance the robustness and adaptability of the model. The experimental results show that PD-BertEDL is superior to the existing prediction methods, which can effectively predict peptide detectability and provide strong support for protein identification and quantitative analysis, as well as disease treatment.     

PON-Sol2: Prediction of Effects of Variants on Protein Solubility

Genetic variations have a multitude of effects on proteins. A substantial number of variations affect protein–solvent interactions, either aggregation or solubility. Aggregation is often related to structural alterations, whereas solubilizable proteins in the solid phase can be made again soluble by dilution. Solubility is a central protein property and when reduced can lead to diseases. We developed a prediction method, PON-Sol2, to identify amino acid substitutions that increase, decrease, or have no effect on the protein solubility. The method is a machine learning tool utilizing gradient boosting algorithm and was trained on a large dataset of variants with different outcomes after the selection of features among a large number of tested properties. The method is fast and has high performance. The normalized correct prediction rate for three states is 0.656, and the normalized GC2 score is 0.312 in 10-fold cross-validation. The corresponding numbers in the blind test were 0.545 and 0.157. The performance was superior in comparison to previous methods. The PON-Sol2 predictor is freely available. It can be used to predict the solubility effects of variants for any organism, even in large-scale projects.     

VHH Structural Modelling Approaches: A Critical Review

V_HH, i.e., VH domains of camelid single-chain antibodies, are very promising therapeutic agents due to their significant physicochemical advantages compared to classical mammalian antibodies. The number of experimentally solved V_HH structures has significantly improved recently, which is of great help, because it offers the ability to directly work on 3D structures to humanise or improve them. Unfortunately, most V_HHs do not have 3D structures. Thus, it is essential to find alternative ways to get structural information. The methods of structure prediction from the primary amino acid sequence appear essential to bypass this limitation. This review presents the most extensive overview of structure prediction methods applied for the 3D modelling of a given V_HH sequence (a total of 21). Besides the historical overview, it aims at showing how model software programs have been shaping the structural predictions of V_HHs. A brief explanation of each methodology is supplied, and pertinent examples of their usage are provided. Finally, we present a structure prediction case study of a recently solved V_HH structure. According to some recent studies and the present analysis, AlphaFold 2 and NanoNet appear to be the best tools to predict a structural model of V_HH from its sequence.     

UMPred-FRL: A New Approach for Accurate Prediction of Umami Peptides Using Feature Representation Learning

Umami ingredients have been identified as important factors in food seasoning and production. Traditional experimental methods for characterizing peptides exhibiting umami sensory properties (umami peptides) are time-consuming, laborious, and costly. As a result, it is preferable to develop computational tools for the large-scale identification of available sequences in order to identify novel peptides with umami sensory properties. Although a computational tool has been developed for this purpose, its predictive performance is still insufficient. In this study, we use a feature representation learning approach to create a novel machine-learning meta-predictor called UMPred-FRL for improved umami peptide identification. We combined six well-known machine learning algorithms (extremely randomized trees, k-nearest neighbor, logistic regression, partial least squares, random forest, and support vector machine) with seven different feature encodings (amino acid composition, amphiphilic pseudo-amino acid composition, dipeptide composition, composition-transition-distribution, and pseudo-amino acid composition) to develop the final meta-predictor. Extensive experimental results demonstrated that UMPred-FRL was effective and achieved more accurate performance on the benchmark dataset compared to its baseline models, and consistently outperformed the existing method on the independent test dataset. Finally, to aid in the high-throughput identification of umami peptides, the UMPred-FRL web server was established and made freely available online. It is expected that UMPred-FRL will be a powerful tool for the cost-effective large-scale screening of candidate peptides with potential umami sensory properties.     

Evolving an Accurate Decision Tree-Based Model for Predicting Carbon Dioxide Solubility in Polymers

Solubility is one of the most indispensable physicochemical properties determining the compatibility of components of a blending system. Research has been focused on the solubility of carbon dioxide in polymers as a significant application of green chemistry. To replace costly and time-consuming experiments, a novel solubility prediction model based on a decision tree, called the stochastic gradient boosting algorithm, was proposed to predict CO₂ solubility in 13 different polymers, based on 515 published experimental data lines. The results indicate that the proposed ensemble model is an effective method for predicting the CO₂ solubility in various polymers, with highly satisfactory performance and high efficiency. It produces more accurate outputs than other methods such as machine learning schemes and an equation of state approach.