IGPred-HDnet: Prediction of Immunoglobulin Proteins Using Graphical Features and the Hierarchal Deep Learning-Based Approach

Ali, Zakir; Alturise, Fahad; Alkhalifah, Tamim; Khan, Yaser Daanial

doi:https://doi.org/10.1155/2023/2465414

Computational Intelligence and Neuroscience

On this page

Abstract Introduction Materials and Methods Conclusion Data Availability Conflicts of Interest Acknowledgments Supplementary Materials References Copyright Related Articles

Special Issue

Advanced Artificial Intelligence for Computational Healthcare and Bioinformatics

View this Special Issue

Research Article | Open Access

Volume 2023 | Article ID 2465414 | https://doi.org/10.1155/2023/2465414

IGPred-HDnet: Prediction of Immunoglobulin Proteins Using Graphical Features and the Hierarchal Deep Learning-Based Approach

Zakir Ali,¹Fahad Alturise,²Tamim Alkhalifah,²and Yaser Daanial Khan¹

Academic Editor: Xiaolong Zhou

Received08 Jul 2022

Revised16 Sept 2022

Accepted12 Oct 2022

Published25 Jan 2023

Abstract

Motivation. Immunoglobulin proteins (IGP) (also called antibodies) are glycoproteins that act as B-cell receptors against external or internal antigens like viruses and bacteria. IGPs play a significant role in diverse cellular processes ranging from adhesion to cell recognition. IGP identifications via the in-silico approach are faster and more cost-effective than wet-lab technological methods. Methods. In this study, we developed an intelligent theoretical deep learning framework, “IGPred-HDnet” for the discrimination of IGPs and non-IGPs. Three types of promising descriptors are feature extraction based on graphical and statistical features (FEGS), amphiphilic pseudo-amino acid composition (Amp-PseAAC), and dipeptide composition (DPC) to extract the graphical, physicochemical, and sequential features. Next, the extracted attributes are evaluated through machine learning, i.e., decision tree (DT), support vector machine (SVM), k-nearest neighbour (KNN), and hierarchical deep network (HDnet) classifiers. The proposed predictor IGPred-HDnet was trained and tested using a 10-fold cross-validation and independent test. Results and Conclusion. The success rates in terms of accuracy (ACC) and Matthew’s correlation coefficient (MCC) of IGPred-HDnet on training and independent dataset (D_train D_test) are ACC = 98.00%, 99.10%, and MCC = 0.958, and 0.980 points, respectively. The empirical outcomes demonstrate that the IGPred-HDnet model efficacy on both datasets using the novel FEGS feature and HDnet algorithm achieved superior predictions to other existing computational models. We hope this research will provide great insights into the large-scale identification of IGPs and pharmaceutical companies in new drug design.

1. Introduction

Immunoglobulins are serum proteins in the human body. These proteins act as an antibody involved in the various cellular processes such as a decision, binding, or recognition of the cell. Immunoglobulin significantly boosts the immune system by discovering the dangerous macromolecules that entered the body [1]. When unfamiliar elements inject into the body, the immune system has a unique skill to detect the attacker and then activates B lymphocytes to hide the immunoglobulin from invader antigens. For instance, immunoglobulins will deactivate the toxin by altering its chemical structure when averting its appearance. To provide a shield against bacterial infection, stabilin-2 can attach to both Gram-positive and Gram-negative bacterial contagions.

Immunoglobulins are linked/related to various disease treatments [2], such as autoimmune, inflammation in the skin, and Bechet’s diseases [3, 4]. In other words, intravenous immunoglobulin provides a fighting strength to cure such kinds of diseases for people who have suffered from muscle problems and systemic swelling in skin infections. The use of immunoglobulin for lupus erythematosus dermatosis in association with the treatment of Bechet’s infection has a great potential without any harmful impact [3, 4]. In Ref. [5], it is shown that immunoglobulins have a better understanding of immunological processes, permitting the development of an enhanced version of drugs to cure the infection. Considering the medical application of immunoglobulin proteins, in-depth knowledge of their functional level is still under development.

Over the past years, immunoglobulin protein classification and characterization have become a hot topic in bioinformatics and computational biology. Wet-lab approaches such as X-ray crystallography and mass spectrometry are used to discover immunoglobulin proteins. However, such laboratory-based approaches are unfavourable due to their high cost and time consumption. In this regard, researchers have designed various machine learning-based methods to identify immunoglobulin protein sequence analysis. Efficient machine learning-based methods can quickly and accurately predict unannotated proteins from large databases. Machine learning techniques are applied in numerous areas of medicine like diagnostics. Clonal dynamics and relative frequencies are utilized to develop an antibody clonal examining framework to explore certain antigenic human monoclonal antibodies [5–7]. In the various field of the healthcare system, immunological and biological usage, including infection control, immunization diagnostics, and B-cell detection, is of key significance [8, 9]. The research community has reported numerous studies related to antigen range that can be selected by specific antibodies or by a group of antibodies, e.g., antibody stock provided by applying a Rep-Seq in many areas [10]. The said key observation headed to another and well-defined technique for tackling the B-cell epitope detection in which the intellectual purpose of a specific antibody is detected [11, 12]. This study incorporates optical, electrochemical, and piezoelectric biosensors to predict complete immunoglobulin degrees, in which electrochemical is most generally employed. Several immunoglobulin optical biosensors depend on surface plasmon resonance (SPR) prediction present in buffer solutions. For an immunoglobulin study, these available state-of-the-art technologies are useful; however, conducting the biochemical study is very expensive in terms of money and time. For accurate and speedy execution of a huge amount of protein data, it is a need of time to develop a computational framework for immunoglobulins. For example, the first phase declares the purpose of immunoglobulins proteins which design a useful and inexpensive framework to predict them efficiently. The research community has designed various frameworks based on machine learning procedures for protein sequence analysis and classification in the last decades [13–17]. In bioinformatics, predicting immunoglobulins transforms protein sequences into feature metrics to uncover the core formation of proteins. The essential characteristics of protein prediction are itemized as follows: feature representation and key feature selection based on their importance and classification. Amino acid composition (AAC), dipeptides (Dip), and tripeptides are feature extraction techniques to extract n-gram features representation, where the occurrence of n-length peptides are utilized as feature matrices [18–20].

Furthermore, another feature extraction method pseudo-amino acid composition (PseAAC), is commonly implemented, considering physicochemical properties among residues [15, 17, 21–23]. The pseudotype protein structure led to a protein density drop in dyscalculia; for this purpose, the notion of pseudo-K-tuple is combined with the idea of PseAAC [24, 25] to design a framework of AAC minimized with pseudo-K-tuples amino acid composition (PseKRAAC) [26]. They developed a classifier IGPred by considering nine (9) physicochemical properties of amino acid-generated proteins with replica ACC [27, 28]. In Ref. [29], a predictor was developed via a support vector machine (SVM) to predict immunoglobulins and nonimmunoglobulins. They used PseAAC with nine physical and chemical characteristics of amino acids; A cross-validation technique was used to train a model, and they got 96.3% accuracy. However, the performance is good but still needs an efficient bioinformatics tool to predict immunoglobulin with a less error rate.

Various feature representations and multifaced prediction methods may produce unnecessary knowledge representation [30, 31]. However, to deal with this problem, many studies suggested feature selection algorithms for eliminating unnecessary information to enhance the performance of the prediction methods. The first one is PCC, which stands for Pearson’s correlation coefficient, used to measure the significance of feature representation in a subgroup. In contrast, the second part is related to computing the repetition among features representation by using Euclidean distance (ED), cosine distance (CD), and Tanimoto (TO). Maximum-Relevance-Maximum-Distance in [32, 33] and Analysis of Variance (ANOVA) in [34] are typical feature selection approaches. For optimum feature representation, [35–37] used the principal component analysis (PCA) and misclassification error (MCE) to extract optimal feature representation for pentatricopeptide-repeat proteins prediction and got 97.9% accuracy. Li et al. in [33] used the above method to design a model for the prediction of anticancer peptide sequences with 19-dimensional attributes.

Although significant contribution has been devoted to the prediction of IGPs, some shortcomings should be acknowledged in terms of feature-encoding schemes and learning models. One major limitation of the existing methods is the lack of feature learning algorithms to extract the structured pattern information from protein sequences properly. Secondly, only machine learning classifiers are not accurate enough to discriminate IGPs from non-IGPs. Thirdly, the developed immunoglobulin predictors only showed the training dataset results using a cross-validation test while ignoring the external/independent test results. Independent test results are significant as they show the trained model’s generalization power.

To our best knowledge, IGPred-HDnet is the first deep learning-based predictor for identifying IGPs. IGPred-HDnet extracts the nominal feature vectors using novel feature descriptors such as FEGS (extract the graphical features), AAPse (extracting physicochemical features), and DPC (sequential features) from the given protein sequence and fed to the hierarchical deep net model (HDnet) as the base classifier for constructing the model. The model opts for deep representations instead of manually extracted handcrafted features and aims to perform the classification of IGPs. We have validated the model through exhaustive methods which shows that the overall prediction on both training and testing datasets outperformed the existing state-of-the-art methods. The study provides great insights into the large-scale identification of IGPs which pharmaceutical companies can opt for in novel drug design.

2. Materials and Methods

In the subsequent subsections, we will describe the stepwise approach to the classification of IGPs. Figure 1 shows these stepwise approaches. Firstly, the dataset collection and preprocessing method will be discussed. The feature representation method will be presented in the next section; the classification framework and model evaluation will be disused in the third stage of the methodology.

2.1. Dataset Construction and Preprocessing

This portion will discuss dataset collection for experimenting, i.e., training and evaluating the designed framework. The dataset contained the immunoglobulins sequences downloaded from the UniProt database present in or outside the cell membrane. There are some standard techniques to assure the quality of the baseline dataset; in the first stage, we eliminated the ambiguous residues, i.e., “B,” “J,” “O,” “X,” “U,” and “Z” from the protein sequences to obtained typical amino acid sequences [38]. We also eliminate the sequence if it is the portion of other proteins. We picked the protein sequences from the human, mouse, and rat categories in the second stage. We used CD-HIT software to diminish hugely indistinguishable bias in the last stage, which caused overfitting predicted results, and the cutoff value is set at 60%:

Our dataset consists of 302 samples, with 110 positive and 192 negative samples of immunoglobulins for training the model:

Our independent dataset contains 112 samples to evaluate our trained model, of which 40 are positive and 72 are negative samples. Overall, 150 positive and 264 negative samples are provided in Supplementary File S1 and Supplementary File S2, respectively.

2.2. Existing Feature Extraction Schemes

In designing a computerized framework, a series of steps are carried out to predict immunoglobulins. Among them, the feature extraction scheme is a challenging and essential step in formulating a biological sequence into some numerical values [39]. Conventional classification learning models, including K-nearest neighbour (KNN), random forest (RF) [40, 41], and support vector machine (SVM) [42], are based on fixed-length statistical values and are unable to handle the variable-length protein sequence; hence, the features representation algorithm can tackle this problem by extracting the fixed-length feature vector form the variable-length sequences [43–45]. Several researchers have used different feature encoding schemes [46] as shown in Figure 2; however, none of them used the proposed method for extracting vital pattern information from the immunoglobulins. A detailed description is given in Section 2.3.

(a)

(b)

(c)

2.3. Feature Extraction Based on Graphical and Statistical Features (FEGS)

Herein, we have opted for a novel feature representation method named Feature Extraction based on Graphical and Statistical features (FEGS) [47] for immunoglobulins sequences, as shown in Figure 3. The proposed deep neural network is not novel; however, the extraction of features through this method is novel. Extracting the hidden pattern information through graphs is different from other sequence-based feature descriptors. The main shortcoming of traditional methods is the loss of sequence order information. For example, amino acid composition and reduced amino acid alphabet cannot retain the protein’s global correlated properties. Furthermore, the manual extraction of features requires extensive approaches which can be somehow not sufficient. These handcrafted features are not that much powerful to discriminate biological sequences as compared to the deep representations, as shown in [15]. The FEGS algorithm was proposed to tackle this issue by formulating the biological proteins using a three-dimensional curve. The working principle of the FEGS algorithm is that initially, FEGS employs the graphical depiction of primary proteins using circular cones in 3D space by extending the notion of 3D protein paths. Secondly, using the physicochemical properties of amino acids that efficiently extract the statistical attributes of protein pairs, FEGS seeks to form many circular cones in 3D space. Finally, the 578-dimensional vector is generated by combining mono-amino acid and dipeptide compositions for each protein sequence.

Initially, the protein sequences are provided in the FASTA format as input, and then FEGS starts eliminating unnecessary indices with identical values and generating 158 space curves for the subsequent protein sequence.

2.3.1. Generation of 3D Graphical Curves for Immunoglobulins Sequences

In this method, the protein sequences are provided in the FASTA format as input; then, according to their physicochemical indices, 20 amino acids are first linked with 20 points in the 3D area. In the second step, the graphical curve of an immunoglobulins sequence can be generated by enlarging a 3D protein track centred on a right circular cone.

(1) Preparation of the 20 Amino Acids and the 400 Amino Acid Sets. Physicochemical properties (PCP) of amino acids (AAs) play a vital role in analyzing and characterizing protein function. We arranged the 20 AAs with respect to their PCP from lower to higher order. Then, we organized them on the circumference of the bottommost of a right circular cone with a height of 1 by the following formula:

The above equation denoted 20 amino acids, whereas all 400 amino acid pairs are linked to the bottom of the right circular cone via the formula below:

represents each of the 400 amino acid pairs.

(2) Building 3D Graphical Curves for Protein Sequences. Consider that we have a protein sequence S having N AA residues . Constructing the 3D graph for the protein sequence is quite challenging. The 3D graphical curve is generated by enlarging a 3D protein track centred on a right circular cone as follows. Initiating from the origin point broadens it to the subsequent point in the 3D area, conforming to the first AA and the second point related to the second AA and so on till the 3D track is accomplished at the last AA , and via this process, the P path is obtained, coordinating with a 3D graphical curve of the immunoglobulins sequence S, whereas is the amino acid , and the point coordinates and are described in the following formulas:

In the above equation, , and is the number of amino acid sets determined. The selected 158 physicochemical properties are linked with the exclusive right circular cone; in this way, we got 158 various 3-dimensional graphical curves for every immunoglobulin sequence related to the 158 physicochemical properties of amino acids.

2.3.2. Numerical Features of Protein Sequences

Another challenging job is to transform the generated graphical curves into numerical feature vectors for the similarity analysis of immunoglobulins samples. Here, for each curve, the L/L matrix denoted by M is calculated, and off-diagonal values are well-defined as a measure of the Euclidean distance and the sum of geometric lengths of boundaries between and of the curve. At the same time, on-diagonal elements are equal to zero. Subsequently, all 158 curves are converted into 158-dimensional feature representation matrices as a graphical features representation described below:

There are many other feature extraction techniques in which AAC and DPC are commonly utilized in protein sequence analyses. To count the frequency of AA in a given sequence, normalized by sequence length, AAC is widely used for this process to extract 20 fixed-length features as formulated below:

The above equation f represents the number of AA occurrences in the protein sequence. DPC also counts the number of occurrences of the 400 AA sets of the given protein sequence; and it extracts 400 fixed-length features below:where f represents the number of occurrences of AA sets, i.e., in the protein sequence. The statistical features, i.e., AA and DPC are merged with graphical features represented to get a 578-dimensional feature vector for the protein sequence S. In general, a dataset that contains N number of immunoglobulins sequences is given to FEGS, then we can get the N × 578 feature representation matrix, in which every row represents a feature representation vector of immunoglobulins sequences.

3. The Proposed Model Workflow

We developed a robust immunoglobulins predictor called Immunoglobulin Proteins Prediction Hierarchical Deep net (IGPred-HDnet). Figure 4 illustrates the flow of the proposed framework, in which the main stages of the IGPred-HDnet framework are shown such as data collection, data distribution, feature representation computation through FEGS, and classification through HDNet and evaluation. In feature representation, a novel feature encoding method is proposed to extract valuable feature representation from immunoglobulin sequences.

3.1. Hierarchical Deep Net Model (HDnet)

The hierarchical deep net (HDnet) model is an ensemble-based model inspired by [48], which is a substitute for a deep neural network (DNN) to learn hyperlevel feature representation using various resources and efforts. In contrast, DNN used complex architecture, i.e., forward and backward propagation algorithms, to learn hidden information. In developing an HDnet classifier, it is crucial to determine the learning algorithms employed in each layer. In our proposed model, we set the combination of Extreme Gradient Boost (XGBoost) [49, 50], random forest (RF) [51–53], and extremely randomized trees (ERT) [54, 55] classifiers which achieved outstanding performance and feed it with the previously computed 578-dimensional vector. HDnet is based on the deep ensemble method that cascades conventional classifiers, for example, RF, ERT, and XGBoost. Compared to DNN, HDnet uses decision trees instead of various neural network (NN) models for feature representation learning in each layer. Figure 5 shows the generic representation of HDNet, elaborating that if there are multiple feature vectors from multiple encoding schemes, they are concatenated at the level-N. These feature vectors are actually deep representations learnt at different layers, similar to other deep neural networks. Due to the hierarchical type nature, the HDnet model allows the training process to be more robust, and it will be more appropriate for training a limited amount of protein samples. DNN involves various parameters that need tunes during training a model, while our proposed model easily tunes the hyperparameter.

We set the boosting parameter value k = 20 for the XGBoost classifier. For RF and ERT, the number of decision trees is also set at 20, and the node values are picked by randomly picking features. In our model, every layer is an ensemble of diverse learners (e.g., six XGBoost, six RF, and six ERT) who accept the feature representation processed by previous layer classification models. The outcome of the previous layer is the input for the subsequent layer for processing. To produce the enhanced feature representation related to the multivariate class vectors, we have integrated, stacked, and summed output as a supreme probability score. The process of training is terminated if enhancement is not observed in performance. Figure 5 reveals the layer-by-layer framework of the HDnet.

4. Performance Evaluation

In this research, we utilized four performance evaluation measures, e.g., accuracy (ACC), specificity (SP), sensitivity (SN), and Matthew correlation coefficient (MCC), to figure out the achievement rate of our proposed prediction models described as

In the above equations, TP represents True-IGPs, which are correctly predicted as positive instances, whereas TN corresponds to true non-IGPs, which are correctly classified as negative samples. FN indicates non-IGPs, which the model incorrectly predicts as immunoglobulins.

The performance above measures containing the MCC is dependent on the threshold, which delivers the comprehensive evaluation for the binary class classification. Furthermore, to describe the model performance on a large scale, we utilized the Area Under the ROC (Receiver Operating Characteristic) Curve (AUC), which is in the shape of an independent threshold analysis like a further essential assessment of the model.

5. Proposed Framework Evaluation

In machine learning (ML), the model performance is naturally assessed via cross-validation (CV). There are three tests in the research community to determine the discriminatory power of the designed framework: K-fold also called subsampling, Jackknife, i.e., leave-one-out and independent tests [56, 57]. The Jackknife test provides exceptional and encouraging results to train a model [58]; however, the main cons are computational cast due to a large number of calculations [59]. To overcome the weakness of the Jackknife and improve the simplification power, we implemented the K-Fold CV test to train our model and test the performance [60]. In this method, we randomly divided the train data into K-folds (subsets), in which K − 1 is utilized to train the proposed model, and the leftover is utilized to test the model [61]. Subsequently, for the particular approximation, the obtained results are averaged. We set the value of K to 10 after conducting various experiments.

5.1. Predictive Performance of Hypothesis Learners Using Various Feature Encoding Schemes on Training Dataset D_train

In this section, we experimentally determine the prediction performance of various classifiers, i.e., KNN [62], DT [63], SVM [46, 64], and HDnet using various descriptors, i.e., APAAC (physicochemical features), DPC (sequential features), and FEGS (graphical features), as shown in Figure 6. Each learning engine is computed by conducting a ten-fold CV test on the training dataset D_train with four evaluation measures ACC, SN, SP, and MCC. In the case of APAAC feature vectors, the SVM classifier secured the worst AAC = 89.72% and MCC = 0.786, while HDnet achieved a higher ACC of 95.69% and MCC of 0.909 points. Similarly, in the case of the DPC method, again the HDnet classifier produced 0.33% high ACC and 0.007 points MCC, respectively. Furthermore, in the case of the FEGS feature method, the highest performance is obtained by the HDnet classifier, which is ACC = 98.00%, SN = 94.55%, SP = 100%, and MCC = 0.958. The second-best predictor is KNN which achieved 90.41% ACC and 0.809 points MCC, while SVM comparatively produce good predictions on all feature methods.

(a)

(b)

(c)

Several judgements are made on the reported results of all classifiers in Table 1. First, the HDnet model consistently produced the best outcomes among the classification algorithms compared to other machine-learning classifiers for all feature encoding schemes. The main reason is due to the high learning potential of a deep neural network as compared to the conventional classifiers. The internal structure of the HDnet classifier is based on decision trees that enable the model to predict the extracted features better [65]. Further, it is evident in the literature that deeper networks have more learning potential as compared to conventional neural networks [15, 66, 67].

Secondly, among the feature representation approaches, FEGS (graphical features) produced the best results for overall hypothesis learners (classifiers) than other feature vectors such as DPC and APAAC. The underlyingreason for the high prediction rate of FEGS methods is that FEGS extracts the conserved local and global graphical, physicochemical and statistical attributes from a protein sequence. As in Figure 1, the visualization influence of the extracted features through t-distributed stochastic neighbour embedding (t-SNE) can be seen. The red colour represents the IGPs class, and the green colour represents the non-IGPs class. The features with a high correlation, like DPC and APAAC, cannot incorporate the correct predictions of immunoglobulins. In contrast, the novel features of FEGS are less correlated enabling the classifiers to produce high performance.

5.2. Predictive Performance of Hypothesis Learners Using Various Feature Encoding Schemes on the Testing Dataset D_test

In this subsection, we examine the success rates of our model via an independent test to show its generalization power. It was ensured that the samples in the independent test D_test were unseen, and none of the immunoglobulin samples was used in training the model. Table 1 depicts the prediction outcomes of all classifiers using the APAAC, DPC, and FEGS feature methods. Comparative analysis reveals that our proposed learning model HDnet using novel feature FEGS achieved outstanding results in terms of all performance metrics, likewise ACC = 99.10%, SN = 97.50%, SP = 100%, and MCC = 0.980 points, respectively. In contrast, the same learning engine using the APAAC feature produced the worst results as shown in Figure 7.

(a)

(b)

(c)

5.3. Predictive Performance of the Proposed Predictor with Existing Methods on Training and Testing Datasets

In this section, we theoretically compare the efficacy of our proposed model with the three developed approaches such as CC-PSSM [39], IGPred [19], and Ghulam et al.’s approach [68] on training and testing datasets. The results in Table 2 are extracted from the previous literature [69]. It is worth noting that none of the existing predictors generated the prediction outcomes on independent tests to show the generalization power of their model. Driven by the novel feature descriptor FEGS with the intelligent deep learning-based algorithm HDnet, IGPred-HDnet outperformed the existing methods for IGs identification in terms of all performance metrics, i.e., ACC, SN, SP, MCC, and AUC. On the training benchmark dataset, our method notable increased ACC by 1.9%, SN by 1%, SP by 1.5%, and MCC by 0.026 points over the second-best performer XGBoost. An independent test was performed to investigate further the IGPred-HDnet model’s predictive capability on unseen data. Both the ACC and AUC results are 0.99 and 1.00, as shown in Table 2 and Figure 8.

The underlying reason for achieving high predictions is to extract the graphical-based, physicochemical-based, and sequence-based attributes. Also, the hierarchical type structure of the HDnet classifier enables a better forecast of the IGs samples from the extracted attributes [65].

6. Conclusion and Future Work

IGPs are a crucial constituent of the immune system. Understanding deep insight IGPs can provide useful hints in drug discovery for disease treatment. Thus, the objective of this research was to construct a novel sequence-based computational method for predicting and analyzing IGPs. The proposed theoretical model “IGPred-HDnet” is superior to other advance immunoglobulin-based predictors due to several reasons. Firstly, we designed an innovative graphical algorithm FEGS to capture structured information buried in the protein sample. The structure features produced better results than the other feature schemes. Secondly, we implemented a deep learning model called HDnet for the first time as a learning model for recognizing IGPs.

Despite enhancing the model’s overall performance, further gaps still exist for future, such as several previous publications like Tang et al. [27] established public webservers that can enrich the applicability of the anticipated model. Also, using novel feature selection algorithms is vital to avoid overfitting and improve the generalization power of the trained model. We hope that the proposed IGPred-HDnet will become a potential tool for large-scale IGPs characterization in particular and other protein problems in general.

Data Availability

The dataset analyzed in this study can be found in the supplementary files.

Conflicts of Interest

The authors declare that they have no conflicts of interest.

Acknowledgments

The researchers would like to thank the Deanship of Scientific Research, Qassim University for funding the publication of this project.

Supplementary Materials

Supplementary File S1 contains the positive samples (immunoglobulins sequences). Supplementary File S2 contains the negative samples (nonimmunoglobulins sequences). (Supplementary Materials)

References

A. N. Barclay, “Membrane proteins with immunoglobulin-like domains—a master superfamily of interaction molecules,” in Seminars in Immunology, Elsevier, Amsterdam The Netherlands, 2003.
View at: Google Scholar
A. H. A. Latiff and M. A. Kerr, “The clinical significance of immunoglobulin A deficiency,” Annals of Clinical Biochemistry: International Journal of Laboratory Medicine, vol. 44, no. 2, pp. 131–139, 2007.
View at: Publisher Site | Google Scholar
T. Généreau, O. Chosidow, C. Danel, P. Chérin, and S. Herson, “High-dose intravenous immunoglobulin in cutaneous lupus erythematosus,” Archives of Dermatology, vol. 135, no. 9, pp. 1124-1125, 1999.
View at: Publisher Site | Google Scholar
J. K. Presto, E. Z. Hejazi, and V. P. Werth, “Biological therapies in the treatment of cutaneous lupus erythematosus,” Lupus, vol. 26, no. 2, pp. 115–118, 2017.
View at: Publisher Site | Google Scholar
A. Kadurin, S. Nikolenko, K. Khrabrov, A. Aliper, and A. Zhavoronkov, “druGAN: an advanced generative adversarial autoencoder model for de novo generation of new molecules with desired molecular properties in silico,” Molecular Pharmaceutics, vol. 14, no. 9, pp. 3098–3104, 2017.
View at: Publisher Site | Google Scholar
M. Ali, S. A. Khan, K. Wennerberg, and T. Aittokallio, “Global proteomics profiling improves drug sensitivity prediction: results from a multi-omics, pan-cancer modeling approach,” Bioinformatics, vol. 34, no. 8, pp. 1353–1362, 2018.
View at: Publisher Site | Google Scholar
A. A. Shah, H. A. M. Malik, A. Mohammad, Y. D. Khan, and A. Alourani, “Machine learning techniques for identification of carcinogenic mutations, which cause breast adenocarcinoma,” Scientific Reports, vol. 12, no. 1, pp. 11738–11815, 2022.
View at: Publisher Site | Google Scholar
H. Shirai, C. Prades, R. Vita et al., “Antibody informatics for drug discovery,” Biochimica et Biophysica Acta (BBA) - Proteins & Proteomics, vol. 1844, no. 11, pp. 2002–2015, 2014.
View at: Publisher Site | Google Scholar
W. Alghamdi, M. Attique, E. Alzahrani, M. Z. Ullah, and Y. D. Khan, “LBCEPred: a machine learning model to predict linear B-cell epitopes,” Briefings in Bioinformatics, vol. 23, no. 3, Article ID bbac035, 2022.
View at: Publisher Site | Google Scholar
J. Benichou, R. Ben-Hamo, Y. Louzoun, and S. Efroni, “Rep‐Seq: uncovering the immunological repertoire through next‐generation sequencing,” Immunology, vol. 135, no. 3, pp. 183–191, 2012.
View at: Publisher Site | Google Scholar
I. Sela-Culang, S. Ashkenazi, B. Peters, and Y. Ofran, “PEASE: predicting B-cell epitopes utilizing antibody sequence,” Bioinformatics, vol. 31, no. 8, pp. 1313–1315, 2015.
View at: Publisher Site | Google Scholar
C. K. Hua, A. T. Gacerez, C. L. Sentman, M. E. Ackerman, Y. Choi, and C. Bailey-Kellogg, “Computationally-driven identification of antibody epitopes,” Elife, vol. 6, Article ID e29023, 2017.
View at: Publisher Site | Google Scholar
Y. Zhai, Y. Chen, Z. Teng, and Y. Zhao, “Identifying antioxidant proteins by using amino acid composition and protein-protein interactions,” Frontiers in Cell and Developmental Biology, vol. 8, Article ID 591487, 2020.
View at: Publisher Site | Google Scholar
J. Zhang, Z. Zhang, L. Pu, J. Tang, and F. Guo, “AIEpred: an ensemble predictive model of classifier chain to identify anti-inflammatory peptides,” IEEE/ACM Transactions on Computational Biology and Bioinformatics, vol. 18, no. 5, pp. 1831–1840, 2021.
View at: Publisher Site | Google Scholar
S. Naseer, W. Hussain, Y. D. Khan, and N. Rasool, “Optimization of serine phosphorylation prediction in proteins by comparing human engineered features and deep representations,” Analytical Biochemistry, vol. 615, Article ID 114069, 2021.
View at: Publisher Site | Google Scholar
A. H. Butt, S. A. Khan, H. Jamil, N. Rasool, and Y. D. Khan, “A prediction model for membrane proteins using moments based features,” BioMed Research International, vol. 2016, pp. 1–7, 2016.
View at: Publisher Site | Google Scholar
Y. D. Khan, E. Alzahrani, W. Alghamdi, and M. Zaka Ullah, “Sequence-based identification of allergen proteins developed by integration of PseAAC and statistical moments via 5-step rule,” Current Bioinformatics, vol. 15, no. 9, pp. 1046–1055, 2020.
View at: Google Scholar
C. Yang, Y. Ding, Q. Meng, J. Tang, and F. Guo, “Granular multiple kernel learning for identifying RNA-binding protein residues via integrating sequence and structure information,” Neural Computing & Applications, vol. 33, no. 17, pp. 11387–11399, 2021.
View at: Publisher Site | Google Scholar
J. Shao and B. Liu, “ProtFold-DFG: protein fold recognition by combining Directed Fusion Graph and PageRank algorithm,” Briefings in Bioinformatics, vol. 22, no. 3, Article ID bbaa192, 2021.
View at: Publisher Site | Google Scholar
X. Zhao, H. Wang, H. Li, Y. Wu, and G. Wang, “Identifying plant pentatricopeptide repeat proteins using a variable selection method,” Frontiers of Plant Science, vol. 12, Article ID 506681, 2021.
View at: Publisher Site | Google Scholar
M. Awais, W. Hussain, N. Rasool, and Y. Daanial Khan, “ITSP-PseAAC: identifying tumor suppressor proteins by using fully connected neural network and PseAAC,” Current Bioinformatics, vol. 16, no. 5, pp. 700–709, 2021.
View at: Publisher Site | Google Scholar
A. Ehsan, M. K. Mahmood, Y. D. Khan, O. M. Barukab, S. A. Khan, and K. C. Chou, “iHyd-PseAAC (EPSV): identifying hydroxylation sites in proteins by extracting enhanced position and sequence variant feature via chou’s 5-step rule and general pseudo amino acid composition,” Current Genomics, vol. 20, no. 2, pp. 124–133, 2019.
View at: Publisher Site | Google Scholar
S. Ilyas, W. Hussain, A. Ashraf, Y. D. Khan, S. A. Khan, and K. C. Chou, “iMethylK-PseAAC: improving accuracy of lysine methylation sites identification by incorporating statistical moments and position relative features into general PseAAC via Chou’s 5-steps rule,” Current Genomics, vol. 20, no. 4, pp. 275–292, 2019.
View at: Publisher Site | Google Scholar
W. Hussain, N. Rasool, and Y. D. Khan, “A sequence-based predictor of Zika virus proteins developed by integration of PseAAC and statistical moments,” Combinatorial Chemistry & High Throughput Screening, vol. 23, no. 8, pp. 797–804, 2020.
View at: Publisher Site | Google Scholar
S. Naseer, W. Hussain, Y. D. Khan, and N. Rasool, “IPhosS (Deep)-PseAAC: identify phosphoserine sites in proteins using deep learning on general pseudo amino acid compositions via modified 5-Steps rule,” IEEE/ACM Transactions on Computational Biology and Bioinformatics, vol. 19, p. 1, 2020.
View at: Publisher Site | Google Scholar
L.-M. Liu, Y. Xu, and K.-C. Chou, “IPGK-PseAAC: identify lysine phosphoglycerylation sites in proteins by incorporating four different tiers of amino acid pairwise coupling information into the general PseAAC,” Medicinal Chemistry, vol. 13, no. 6, pp. 552–559, 2017.
View at: Publisher Site | Google Scholar
H. Tang, W. Chen, and H. Lin, “Identification of immunoglobulins using Chou’s pseudo amino acid composition with feature selection technique,” Molecular BioSystems, vol. 12, no. 4, pp. 1269–1275, 2016.
View at: Publisher Site | Google Scholar
S. Naseer, H. Waqar, D. Khan Yaser, and R. Nouman, “NPalmitoylDeep-PseAAC: a predictor of N-palmitoylation sites in proteins using deep representations of proteins and PseAAC via modified 5-steps rule,” Current Bioinformatics, vol. 16, no. 2, pp. 294–305, 2021.
View at: Publisher Site | Google Scholar
Y. Ding, J. Tang, and F. Guo, “Identification of drug-target interactions via multiple information integration,” Information Sciences, vol. 418-419, pp. 546–560, 2017.
View at: Publisher Site | Google Scholar
B. Song, F. Li, Y. Liu, and X. Zeng, “Deep learning methods for biomedical named entity recognition: a survey and qualitative comparison,” Briefings in Bioinformatics, vol. 22, no. 6, Article ID bbab282, 2021.
View at: Publisher Site | Google Scholar
S. Ahmed, M. Kabir, M. Arif, Z. Ali, and Z. N. Khan Swati, “Prediction of human phosphorylated proteins by extracting multi-perspective discriminative features from the evolutionary profile and physicochemical properties through LFDA,” Chemometrics and Intelligent Laboratory Systems, vol. 203, Article ID 104066, 2020.
View at: Publisher Site | Google Scholar
C. Meng, J. Wu, F. Guo, B. Dong, and L. Xu, “CWLy-pred: a novel cell wall lytic enzyme identifier based on an improved MRMD feature selection method,” Genomics, vol. 112, no. 6, pp. 4715–4721, 2020.
View at: Publisher Site | Google Scholar
Q. Li, W. Zhou, D. Wang, S. Wang, and Q. Li, “Prediction of anticancer peptides using a low-dimensional feature model,” Frontiers in Bioengineering and Biotechnology, vol. 8, p. 892, 2020.
View at: Publisher Site | Google Scholar
Z. Lv, S. Jin, H. Ding, and Q. Zou, “A random forest sub-Golgi protein classifier optimized via dipeptide and amino acid composition features,” Frontiers in Bioengineering and Biotechnology, vol. 7, p. 215, 2019.
View at: Publisher Site | Google Scholar
C. Feng, Q. Zou, and D. Wang, “Using a low correlation high orthogonality feature set and machine learning methods to identify plant pentatricopeptide repeat coding gene/protein,” Neurocomputing, vol. 424, pp. 246–254, 2021.
View at: Publisher Site | Google Scholar
Y. D. Khan, N. S. Khan, S. Farooq et al., “An efficient algorithm for recognition of human actions,” The Scientific World Journal, vol. 2014, pp. 1–11, 2014.
View at: Publisher Site | Google Scholar
S. Saeed, M. K. Mahmood, and Y. D. Khan, “An exposition of facial expression recognition techniques,” Neural Computing & Applications, vol. 29, no. 9, pp. 425–443, 2018.
View at: Publisher Site | Google Scholar
A. A. Shah and Y. D. Khan, “Identification of 4-carboxyglutamate residue sites based on position based statistical feature and multiple classification,” Scientific Reports, vol. 10, no. 1, pp. 16913–17010, 2020.
View at: Publisher Site | Google Scholar
M. Arif, M. Hayat, and Z. Jan, “IMem-2LSAAC: a two-level model for discrimination of membrane proteins and their types by extending the notion of SAAC into chou’s pseudo amino acid composition,” Journal of Theoretical Biology, vol. 442, pp. 11–21, 2018.
View at: Publisher Site | Google Scholar
E. Alzahrani, W. Alghamdi, M. Z. Ullah, and Y. D. Khan, “Identification of stress response proteins through fusion of machine learning models and statistical paradigms,” Scientific Reports, vol. 11, no. 1, pp. 21767–21815, 2021.
View at: Publisher Site | Google Scholar
S. J. Malebary, R. Khan, and Y. D. Khan, “ProtoPred: advancing oncological research through identification of proto-oncogene proteins,” IEEE Access, vol. 9, pp. 68788–68797, 2021.
View at: Publisher Site | Google Scholar
M. K. Mahmood, A. Ehsan, Y. D. Khan, and K. C. Chou, “IHyd-LysSite (EPSV): identifying hydroxylysine sites in protein using statistical formulation by extracting enhanced position and sequence variant feature technique,” Current Genomics, vol. 21, no. 7, pp. 536–545, 2020.
View at: Publisher Site | Google Scholar
K.-C. Chou, “Impacts of bioinformatics to medicinal chemistry,” Medicinal Chemistry, vol. 11, no. 3, pp. 218–234, 2015.
View at: Publisher Site | Google Scholar
F. Ge, Y. Zhang, J. Xu, A. Muhammad, J. Song, and D. J. Yu, “Prediction of disease-associated nsSNPs by integrating multi-scale ResNet models with deep feature fusion,” Briefings in Bioinformatics, vol. 23, no. 1, Article ID bbab530, 2022.
View at: Publisher Site | Google Scholar
F. Ge, A. Muhammad, and D.-J. Yu, “DeepnsSNPs: accurate prediction of non-synonymous single-nucleotide polymorphisms by combining multi-scale convolutional neural network and residue environment information,” Chemometrics and Intelligent Laboratory Systems, vol. 215, Article ID 104326, 2021.
View at: Publisher Site | Google Scholar
W. Hussain, N. Rasool, and Y. D. Khan, “Insights into machine learning-based approaches for virtual screening in drug discovery: existing strategies and streamlining through FP-cadd,” Current Drug Discovery Technologies, vol. 18, no. 4, pp. 463–472, 2021.
View at: Publisher Site | Google Scholar
Z. Mu, T. Yu, X. Liu, H. Zheng, L. Wei, and J. Liu, “FEGS: a novel feature extraction model for protein sequences and its applications,” BMC Bioinformatics, vol. 22, no. 1, pp. 297–315, 2021.
View at: Publisher Site | Google Scholar
Y. Guo, S. Liu, Z. Li, and X. Shang, “BCDForest: a boosting cascade deep forest model towards the classification of cancer subtypes based on gene expression data,” BMC Bioinformatics, vol. 19, no. S5, pp. 118–213, 2018.
View at: Publisher Site | Google Scholar
T. Chen and C. Guestrin, “Xgboost: a scalable tree boosting system,” in Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, San Francisco, CA, USA, August 2016.
View at: Google Scholar
F. Ge, Y. H. Zhu, J. Xu, A. Muhammad, J. Song, and D. J. Yu, “MutTMPredictor: robust and accurate cascade XGBoost classifier for prediction of mutations in transmembrane proteins,” Computational and Structural Biotechnology Journal, vol. 19, pp. 6400–6416, 2021.
View at: Publisher Site | Google Scholar
L. Breiman, “Random forests,” Machine Learning, vol. 45, no. 1, pp. 5–32, 2001.
View at: Publisher Site | Google Scholar
A. H. Butt and Y. D. Khan, “CanLect-Pred: a cancer therapeutics tool for prediction of target cancerlectins using experiential annotated proteomic sequences,” IEEE Access, vol. 8, pp. 9520–9531, 2020.
View at: Publisher Site | Google Scholar
O. Barukab, Y. D. Khan, S. A. Khan, and K. C. Chou, “ISulfoTyr-PseAAC: identify tyrosine sulfation sites by incorporating statistical moments via Chou’s 5-steps rule and pseudo components,” Current Genomics, vol. 20, no. 4, pp. 306–320, 2019.
View at: Publisher Site | Google Scholar
P. Geurts, D. Ernst, and L. Wehenkel, “Extremely randomized trees,” Machine Learning, vol. 63, no. 1, pp. 3–42, 2006.
View at: Publisher Site | Google Scholar
F. Ge, J. Hu, Y. H. Zhu, M. Arif, and D. J. Yu, “TargetMM: accurate missense mutation prediction by utilizing local and global sequence information with classifier ensemble,” Combinatorial Chemistry & High Throughput Screening, vol. 25, no. 1, pp. 38–52, 2021.
View at: Publisher Site | Google Scholar
W. Lee and K. Han, “Constructive prediction of potential RNA aptamers for a protein target,” IEEE/ACM Transactions on Computational Biology and Bioinformatics, vol. 17, no. 5, pp. 1476–1482, 2020.
View at: Publisher Site | Google Scholar
M. Shahid, M. Ilyas, W. Hussain, and Y. D. Khan, “ORI-Deep: improving the accuracy for predicting origin of replication sites by using a blend of features and long short-term memory network,” Briefings in Bioinformatics, vol. 23, no. 2, p. bbac001, 2022.
View at: Publisher Site | Google Scholar
Y.-C. Wang, X. B. Wang, Z. X. Yang, and N. Y. Deng, “Prediction of enzyme subfamily class via pseudo amino acid composition by incorporating the conjoint triad feature,” Protein and Peptide Letters, vol. 17, no. 11, pp. 1441–1449, 2010.
View at: Publisher Site | Google Scholar
X.-X. Chen, H. Tang, W. C. Li et al., “Identification of bacterial cell wall lyases via pseudo amino acid composition,” BioMed Research International, vol. 2016, Article ID 1654623, 8 pages, 2016.
View at: Publisher Site | Google Scholar
S. J. Malebary and Y. D. Khan, “Evaluating machine learning methodologies for identification of cancer driver genes,” Scientific Reports, vol. 11, no. 1, pp. 12281–12313, 2021.
View at: Publisher Site | Google Scholar
L. Wei, Y. Ding, R. Su, J. Tang, and Q. Zou, “Prediction of human protein subcellular localization using deep learning,” Journal of Parallel and Distributed Computing, vol. 117, pp. 212–217, 2018.
View at: Publisher Site | Google Scholar
Z. Ali, A. W. Abbas, T. M. Thasleema, B. Uddin, T. Raaz, and S. A. R. Abid, “Database development and automatic speech recognition of isolated Pashto spoken digits using MFCC and K-NN,” International Journal of Speech Technology, vol. 18, no. 2, pp. 271–275, 2015.
View at: Publisher Site | Google Scholar
X. Chen, M. Wang, and H. Zhang, “The use of classification trees for bioinformatics,” WIREs Data Mining and Knowledge Discovery, vol. 1, no. 1, pp. 55–63, 2011.
View at: Publisher Site | Google Scholar
Y. Gong, B. Liao, D. Peng, and Q. Zou, “Accurate prediction and key feature recognition of immunoglobulin,” Applied Sciences, vol. 11, no. 15, p. 6894, 2021.
View at: Publisher Site | Google Scholar
Z.-H. Zhou and J. Feng, “Deep forest,” 2017, https://arxiv.org/abs/1702.08835.
View at: Google Scholar
W. Hussain, “SAMP-PFPDeep: improving accuracy of short antimicrobial peptides prediction using three different sequence encodings and deep neural networks,” Briefings in Bioinformatics, vol. 23, no. 1, Article ID bbab487, 2022.
View at: Publisher Site | Google Scholar
C.-S. Wang, P. J. Lin, C. L. Cheng, S. H. Tai, Y. H. Kao Yang, and J. H. Chiang, “Detecting potential adverse drug reactions using a deep neural network model,” Journal of Medical Internet Research, vol. 21, no. 2, Article ID e11016, 2019.
View at: Publisher Site | Google Scholar
A. Ghulam, R. Sikander, F. Ali, Z. N. K. Swati, A. Unar, and D. B. Talpur, “Accurate prediction of immunoglobulin proteins using machine learning model,” Informatics in Medicine Unlocked, vol. 29, p. 100885, 2022.
View at: Google Scholar
A. Ghulam, R. Sikander, F. Ali, Z. N. Khan Swati, A. Unar, and D. B. Talpur, “Accurate prediction of immunoglobulin proteins using machine learning model,” Informatics in Medicine Unlocked, vol. 29, Article ID 100885, 2022.
View at: Publisher Site | Google Scholar

Copyright

Copyright © 2023 Zakir Ali et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

PDF Download Citation

Download other formats

Order printed copies

Views

574

Downloads

326

Citations

Computational Intelligence and Neuroscience

Advanced Artificial Intelligence for Computational Healthcare and Bioinformatics

IGPred-HDnet: Prediction of Immunoglobulin Proteins Using Graphical Features and the Hierarchal Deep Learning-Based Approach

Abstract

1. Introduction

2. Materials and Methods

2.1. Dataset Construction and Preprocessing

2.2. Existing Feature Extraction Schemes

2.3. Feature Extraction Based on Graphical and Statistical Features (FEGS)

2.3.1. Generation of 3D Graphical Curves for Immunoglobulins Sequences

2.3.2. Numerical Features of Protein Sequences

3. The Proposed Model Workflow

3.1. Hierarchical Deep Net Model (HDnet)

4. Performance Evaluation

5. Proposed Framework Evaluation

5.1. Predictive Performance of Hypothesis Learners Using Various Feature Encoding Schemes on Training Dataset Dtrain

5.2. Predictive Performance of Hypothesis Learners Using Various Feature Encoding Schemes on the Testing Dataset Dtest

5.3. Predictive Performance of the Proposed Predictor with Existing Methods on Training and Testing Datasets

6. Conclusion and Future Work

Data Availability

Conflicts of Interest

Acknowledgments

Supplementary Materials

References

Copyright

5.1. Predictive Performance of Hypothesis Learners Using Various Feature Encoding Schemes on Training Dataset D_train

5.2. Predictive Performance of Hypothesis Learners Using Various Feature Encoding Schemes on the Testing Dataset D_test