Viruses have a long history of posing threat to living organisms (1) as they have caused more than 300 million deaths worldwide (2). A recent emanation Coronavirus) is an example of an acute virus that caused a global pandemic (3). According to World Health Organization, Coronavirus has caused approximately more than 400 million infections and 6 million deaths across the globe (4). Similarly, the Ebola virus was also responsible for an epidemic that caused more than 11 thousand deaths in Africa (5).

Viruses are small microscopic particles that contain genetic material (DNA or RNA) surrounded by a protein coat (1). These particles are considered non-living because of their inability to reproduce or perform any other biological function since they lack specific proteins (6). However, once they get a chance to enter inside the host cell, they make interactions with available proteins in the cell and become capable to reproduce themselves (7). Initially, to enter inside a host cell, the viruses interact with the host cell receptor proteins (8) and replicate themselves by injecting their genetic material into the cell's genome (9). After the entrance into the cell, the aim of viruses is to interact with diverse types of proteins through which they can control the process of the cell cycle, particle assembly, apoptosis, and cell metabolism (7, 10). The relationships between host and virus proteins are termed virus-host protein-protein interactions (11).

To prevent viruses from interacting with host proteins, hosts have sophisticated mechanisms to recognize and confine the viruses, such as the dendritic and β- cells, T-cells, and major histocompatibility complex (12). Therefore, viruses tend to adapt in an efficient manner by interacting with specific host proteins and cellular pathways that prove to be substantial for evading or inactivating factors that are detrimental to viral growth (7). Meanwhile, to enhance immunity against viruses, it is difficult to develop efficient vaccines/drugs because of the poor understanding of different mechanisms that have been adapted by the viruses, and their frequent transmissibility from cell-to-cell or species-to-species (13). Consequently, analyzes of virus-host PPIs are essential to explore their effects on diverse types of biological functions and to design antiviral strategies (14). Furthermore, through such analyzes essential viral proteins and viral dependencies on host proteins can be identified as drug targets to halt the replication process of viruses by pharmacological inhibition (15).

Multiple experimental techniques have been utilized to identify virus-host protein-protein interactions (VHPPIs) such as protease assay (16), surface plasmon resonance (SPR) (17), Förster resonance energy (FRET) (18), Yeast two hybrid screening (Y2H) (19) and affinity purification mass spectrometry (AP-MS) (20). Such conventional wet lab methods are expensive, time-consuming, and error-prone, which impede inter and intra species large scale proteomics sequence analyzes. To overcome the shortcomings of experimental approaches, the development of machine learning applications for efficient proteomics sequence analyzes across different species (e.g., humans, viruses) is an active area of research (21–23). Researchers have developed a machine learning based clustering applications to distinguish several microbial pathogens (24, 25) and classification applications to categorize the genes associated with the survival of pathogens under certain environmental conditions, antibiotics, or other disturbances (26). Similarly, researchers have developed classification applications to determine VHPPIs that play a key role in understanding the functional paradigms of viruses as well as host responses (27, 28). With an aim to provide cheap, fast, and accurate virus-host protein-protein analyzes, to date, around 13 AI-based predictors (21–23, 27–36) have been proposed.

Recently, Yang et al. (36) proposed a VHPPI predictor by utilizing position-specific scoring matrices to statistically represent virus and host protein sequences that were further passed to Siamese convolutional neural network (CNN) for VHPPI prediction. The predictor was evaluated on VHPPI data of human proteins and 8 different viruses. Another similar predictor namely, Deep Viral (35) used one hot vector encoding (OHE) for the discretization of sequences and convolutional neural network architecture for VHPPI prediction. Deep Viral was evaluated on VHPPIs of humans and 12 different viruses. Deep-VHPPI (27) predictor also used OHE and attention mechanism along with CNN for VHPPI prediction. The predictor was evaluated on VHPPI data related to humans and 4 different viruses.

Ding et al. (34) proposed a VHPPI predictor based on long short-term memory (LSTM) neural network. At preprocessing stage, they generated statistical representations of viral and host proteins by reaping the benefits of 3 different encoders namely, the relative frequency of amino acid triplets (RFAT), frequency difference of amino acid triplets (FDAT), and amino acid composition (AC). The predictor (34) was evaluated on VHPPIs across proteins belonging to 137 different viruses and 13 hosts. Denovo (29) used amino acid properties such as dipoles and volumes of side chains to represent 20 amino acids (AAs) with only 7 cluster numbers to reduce the diversity of amino acids. The sequences were then encoded based on the normalized kmer frequencies of 7 unique clusters. Denovo predictor used SVM classifier and was evaluated on the dataset of 10 viruses and human proteins. HOPITOR (37) used a similar encoding method as Denovo (29). HOPITOR used an SVM classifier and was evaluated on 10 different viruses and human proteins. Yang et al. (31) proposed InterSPPI-HVPPI which utilized Doc2vec embeddings and random forest (RF) classifier for VHPPI prediction. The predictor (31) was evaluated on data related to 12 viruses, and human proteins. Karabulut et al. (28) proposed meta predictor (ML-AdVInfect) that reaped the benefits of 4 existing predictors namely HOPITOR (37), InterSPPI-HVPPI (31), VHPPI, and Denovo (29). Specifically, the authors passed the predictions of existing predictors to the SVM classifier for the final VHPPI prediction.

Barman et al. (22) proposed a VHPPI predictor that utilized an RF classifier and statistical vectors generated through 4 different encoding methods namely, average domain-domain association score, virus methionine, virus seline, and virus valine. The predictor was evaluated on VHPPI data related to human proteins and 5 different viruses. Zhou et al. (30) used 7 sequence encoding methods i.e., RFAT, FDAT, AC, composition, transition, and distribution of amino acid groups. The approach (30) used an SVM classifier for VHPPI predictions across the proteins of 332 viruses and 29 hosts. Alguwaizani et al. (32) combined statistical vectors of 4 different encoders namely, amino acid repeats, the sum of the squared length of single amino acid repeats (SARs), maximum of the sum of the squared length of SARs in a window of 6 residues, and composition of amino acids in 5 partitions of the protein sequence. The predictor used an SVM classifier and experimentation was performed on VHPPI data related to 6 hosts and 5 viruses. Recently, Asim et al. proposed an LCGA-VHPPI predictor (38), that made use of a local-global residue context aware sequence encoding scheme and a deep forest model. The authors evaluated their predictor on data related to 23 viruses and human proteins.

Following the success of neural word embedding approaches in natural language processing and bioinformatics, Tsukiyama et al. proposed LSTM-PHV (21) that transformed viral host protein sequences to statistical vectors by learning statistical representation of k-mers in an unsupervised manner using Word2vec approach. The study (21) used bidirectional LSTM for VHPPI prediction and data of proteins belonging to 332 viruses and 29 hosts. Similarly, MTT (23) predictor utilized randomly initialized embeddings and LSTM based classifier. MTT predictor was evaluated on data related to 16 viruses and human proteins. Hangyu et al. (33) developed a VHPPI predictor based on Node2vec and Word2vec embeddings methods and a multilayer perceptron (MLP) classifier. Authors performed experimentation over 7 variants of the SARS virus and 16 different host proteins.

The working paradigm of existing VHPPI predictors can be broadly categorized into two different stages. In the first stage, raw sequences are transformed into statistical vectors where the aim is to capture distributional information of 21 unique amino acids. In the second stage, a machine or deep learning classifier is utilized to discriminate interactive viral-host protein pairs from non-interactive ones.

In the first stage, while transforming raw sequences to statistical vectors, 2 predictors (27, 35), make use of one hot vector encoding method which lacks information related to correlations of amino acids. Moreover, 3 predictors use word embedding generation approaches (21, 23, 35), that capture kmer-kmer associations but lack information related to the distribution of amino acids. To capture distribution and various patterns of amino acids, other predictors utilized 10 different mathematical encoders (22, 29, 31, 34, 36) however, these methods do not capture sequence order or amino acids correlation information. Such information is crucial for the analyzes of protein sequences as reported in the existing studies (39–42) which include sequence encoders such as APAAC and QS order. Despite the promising performance shown by APAAC and QS order encoders for subcellular location prediction (39), Cyclin protein classification (40), and protein-protein interaction prediction (41, 42) tasks, no researcher has explored their potential to effectively generate numerical representations of viral-host protein sequences.

In the second stage, 4 predictors (27, 34–36) utilize CNNs, 2 predictors (21, 34) make use of LSTM architecture and 8 predictors (22, 23, 28–33) use traditional classifiers. As such predictors have shown better performances across limited hosts and viruses, therefore these predictors cannot be generalized across multiple hosts and viruses. For instance, LSTM-PHV is the most recent predictor which managed to produce better performance for human and Coronavirus related VHPPI, but failed to produce similar performance over Zhou et al. (43) datasets that contain multiple hosts and viruses. To make a generic predictor capable of accurately predicting interactions across multiple hosts and viruses, only one meta predictor (28) has been developed. However, this meta-predictor relies on the predictions of 4 existing VHPPI predictors that have their own drawbacks at the sequence encoding and classification level.

With an aim to develop a more accurate and generic meta predictor, the contributions of this paper are manifold, i) It makes use of two different physicochemical properties-based sequence encoding methods namely, APAAC and QS order. In addition, unlike other protein sequence analysis tasks where numerical representations of complete raw protein sequences have been generated through these encoders by utilizing a combination of different physicochemical properties, the paper in hand proposes an effective way to generate numerical representations by using a precise subset of physicochemical properties. ii) Considering different physicochemical properties in both encoders extract some irrelevant and redundant features, to remove such features, it transforms the original feature space into a reduced and more discriminative feature space by utilizing a dimensionality reduction method named feature agglomeration. iii) Using separate and combined statistical vectors generated through APAAC and Qsorder, it generates more effective and discriminative probabilistic feature space by fusing the predictions of two different classifiers. Optimized probabilistic feature space is used to feed the SVM classifier which makes final predictions. iv) Large-scale experimentation over 7 public benchmark datasets and performance comparison of the proposed meta predictor with existing predictors is performed. v) To facilitate researchers and practitioners, a web application based on the proposed meta predictor is developed.

This section briefly describes the working paradigm of the proposed predictor, benchmark datasets, and diverse types of evaluation measures.

To evaluate the integrity, effectiveness, and prediction performance of the proposed virus-host PPIs meta predictor in a reliable manner, following the evaluation criteria of existing studies (21, 30, 36), we utilize 8 different evaluation measures, i.e., accuracy (ACC), specificity (SP), sensitivity (SN), precision (PR), F1 score, area under the receiver operating characteristic (AUROC), area under the precision-recall curve (AUPRC) and Matthews correlation coefficient (MCC).

Accuracy (ACC) (30) measures the proportion of correct predictions with respect to total predictions. Specificity (SP) or True Negative Rate (TNR) is the ratio between true negative class predictions and overall predictions of negative class. Similarly, Recall/Sensitivity (21) computes the score by taking the ratio of correct predictions made on positive class samples to the sum of correct and false predictions of positive class samples. Precision (PR) computes performance score by taking the ratio between correct predictions of positive class samples and all samples which predictor labeled as the positive class. The area under receiver operating characteristics (AUROC) (32) calculates performance score by using true positive rate (TPR) and false positive rate (FPR) at different thresholds. Whereas, the area under the precision-recall curve (AUPRC) (32) calculates the performance scores using precision (P) and recall (FPR) at different thresholds. F1 score combines the precision (P) and recall (R) into a single measure by taking their harmonic mean. MCC (21) measures the correlation of the true classes with the predicted classes by considering all predictions related to positive and negative class samples. Mathematical equations of the aforementioned measures are given as,

In the above numerical cases, T_P and T_N denote true predictions on the positive and negative classes. While F_P and F_N refer to the false predictions related to the positive and negative classes, respectively.

To compute the performance of the proposed predictor in terms of the aforementioned evaluation measures, the evaluation of the predictive pipeline can be performed under the hood of two different settings, k-fold cross-validation and independent test set based evaluation. In k-fold cross-validation, corpus sequences are split into k-folds where iteratively k-1 folds are used for training and the remaining fold is used for testing the predictive pipeline. In this setting, train test split biaseness does not exist as each sequence participates in training and evaluation. Whereas in the independent test set based evaluation, standard train and test splits of sequences are already available, hence training sequences are used to train the predictive pipeline and test sequences are used to test the predictive pipeline. There is a possibility that authors of benchmark datasets may partition hard sequences in the training set and simple sequences in the test set or vice versa. However, we believe that authors of benchmark datasets have carefully partitioned the sequences into train and test sets to avoid any biaseness and perform a fair evaluation. In our study, following previous work (22, 29–31) which has performed independent test set based evaluation on all datasets except Barman et al. (22) dataset, we also perform independent test set based evaluation to make a fair performance comparison. Contrarily, we evaluate our proposed predictor on Barman et al. (22) dataset using 5-fold cross-validation as done by the previous benchmark study.

The proposed meta-predictor is implemented in Python and protein sequence encoders i.e., APAAC and Qsorder are taken from iLearnPlus (76). The classifiers are implemented using Scikit-Learn (77). In order to determine the optimal hyperparameters of encoding methods and classifiers, we have utilized a grid search strategy (78). A search in the grid finds the most optimal values of hyperparameters by evaluating all possible parameter combinations.

In order to determine the optimal parameters λ for the APAAC encoder and lag for the Qsorder encoder, en_{λ, lag} = [1, ⋯ , 5] is used as the grid search space with a stride size of 1. For the Qsorder encoder, lag=1 is chosen for Denovo, TR3-TS1, and Coronavirus datasets, lag=2 is selected for Barman, TR1-TS1 datasets, lag=3 for TR2-TS2, and lag=4 for TR4-TS4. In addition for the APAAC encoder, λ = (4, 4, 4, 4, 3, 3, 1), are chosen for Barman, Denovo, TR2-TS2, TR4-TS2, Coronavirus, TR1-TS1, and TR3-TS1 datasets.

In terms of ML classifiers, the performance of the SVM classifier is greatly influenced by the base kernel, along with the regularization parameter penalty C that controls the margin among hyperplanes. Whereas the parameter d represents the degree of the polynomial kernel, which affects the flexibility of the decision boundary in SVM. Similarly, the performance of tree-based classifiers (RF, ET) relies on the number of base estimators, maximum features for the best split, and splitting criteria i.e., Gini impurity and entropy. To make sure the reproducibility of results, the grid search space, and the selected hyperparameters are shown in Table 1.

This section briefly describes the performance of the proposed meta predictor at different levels of ensembling. Furthermore, it compares the performance of the proposed meta predictor with existing predictors (11, 21–23, 29, 30, 32, 36) over 7 different benchmark datasets (22, 29, 36, 43).

Since the last decade, the development of machine and deep learning-based computational approaches for virus-host protein-protein interaction prediction has been an active area of research (21, 22). In the marathon of developing robust computational VHPPI predictors, the aim of each newly developed predictor has been to utilize raw virus-host protein sequences and precisely discriminate interactive viral-host protein sequences from non-interactive ones. However, most predictors have been evaluated on a limited type of viruses and hosts, such as 6 different predictors have been evaluated on the Barman dataset that contains 5 different viruses and human proteins as hosts. Seven predictors are evaluated on the Denovo dataset which is comprised of 10 different viruses and human proteins as hosts, and 2 predictors are evaluated on the Coronavirus virus. Only 4 predictors are evaluated on the Zhou et al. (30) dataset, which consists of 332 viruses and 29 host proteins. These datasets are more suitable to evaluate the robustness, generalizability, and predictive performance of a computational predictor. These datasets were developed with the objective to train models on different types of viruses and evaluate them on the particular viruses which were not part of the training set.

Over unseen virus-host protein-protein interaction prediction datasets, the performance of existing predictors is comparably low, as compared to their performance on Barman and Denovo datasets. Recently, we developed a machine learning-based predictor namely LGCA-VHPPI (38), which produced state-of-the-art performance on both Barman and Denovo datasets. We evaluated our predictor on Zhou et al. (30) datasets, where it showed a relatively lower performance as compared to its performance on Barman and Denovo datasets. This motivated us to develop an improved predictor that makes the best use of raw viral host protein sequences to perform better not only on Barman and Denovo datasets but also produce a similar performance for unseen viral-host protein-protein interaction predictions.

In viral and host protein sequences, the distribution of amino acids is almost similar across interactive and non-interactive classes. However, an amino acid occurrence at the same position across the sequences of interactive and non-interactive classes varies. The prime reason behind the biaseness of existing viral-host protein-protein interaction predictors toward type I or type II errors is their inability to capture position specific discriminative distribution of amino acids across both classes. To better illustrate this phenomenon, we perform amino acid distribution analysis across both classes with the help of Two Sample Logo (79). Virus and host protein sequences are huge in length and visualizing the amino acid distribution across entire sequence lengths is not feasible at all. Hence, for the purpose of visualization, we take 20 amino acids from the start of virus protein sequences and 20 amino acids from the start of host protein sequences. Using reduced 40 amino acids based sub-sequences, Figure 3 illustrates the amino acid distribution across interactive and non-interactive classes for 7 different datasets. It can be seen that the distribution of amino acids is approximately similar in interactive and non-interactive classes. Considering Barman's dataset as an example (Figure 3A), in interactive and non-interactive samples, there are overlapping amino acids at every position i.e., for position 2, interactive samples contain one of the following amino acid, H, A, E, G, S whereas, non-interactive samples also contain one of the following amino acid, A, E, G, S. In both classes occurrence of 4 amino acids is the same while few samples of the interactive class contain amino acid H, a similar trend exists at other locations as well. Furthermore, other datasets also contain a similar distribution of amino acids as in the Barman dataset. It can be concluded that, across all 7 datasets, we observe the limited discriminative distribution of amino acids and because of that existing predictors lack in performance due to the utilization of sub-optimal sequence encoding methods that generate statistical vectors by neglecting most of the discriminative features about the distribution of amino acids in interactive and non-interactive classes. It is noteworthy to mention that the prime goal of visualizing the amino acid distribution across both classes is to demonstrate the significance of most effectively characterizing viral host protein sequences. However, we have taken entire viral-host protein sequences in our study to generate statistical vectors using physico-chemical properties based sequence encoders, informative feature space using dimensionality reduction algorithm, discriminative feature space using tree based classifiers, and final prediction using SVM classifier. A comprehensive performance comparison of the proposed predictor with VHPPI predictors proves that the proposed predictor manages to capture position specific discriminative distribution of amino acids across both classes.

Furthermore, it is important to mention that all the amino acids are either polar or non-polar in nature and can carry charges, such as out of 21 unique amino acids, 11 amino acids are polar in nature, 4 AAs carry a positive charge (R, D, H, K), 2 AAs carry a negative charge (D, E), and 5 AAs are neutral (C, Q, S, T, Y). Whereas, 10 amino acids are non-polar in nature (A, G, I, L, M, F, P, W, Y, V). Irrespective of positions aware occurrences, considering the overall distribution of amino acids in the protein sequence, charges can be computed by utilizing the physicochemical properties. Overall charge information of amino acids along with their distribution information can extract and encode more discriminative patterns.

Figure 4 shows different clusters of 7 benchmark datasets for the intrinsic analyzes of the statistical vectors generated through APAAC and Qsorder sequence encoders. These clusters are computed by first reducing the dimensions of statistical vectors through principal component analysis (PCA) and then by t-distributed stochastic neighbor embedding (TSNE). In Figures 4A,B, rows represent clusters of interactive and non-interactive classes based on statistical vectors generated through individual encoders (APAAC, Qsorder), and a combination of both encoders. Whereas, the columns represent 7 different benchmark datasets namely, Barman, Denovo, Coronavirus. TR1-TS1, TR2-TS2, TR3-TS1, and TR4-TS2. Overall, statistical vectors from APAAC and Qsorder without dimensionality reduction lead to the formation of overlapping clusters for interactive and non-interactive classes. This overlapping reveals that generated statistical vectors are almost similar and contain less discriminative information about interactive and non-interactive classes, as shown in Figure 4A. Furthermore, this overlapping behavior among clusters exists due to the extraction of some irrelevant and redundant features by different physicochemical properties. To eradicate such type of information, we utilize the feature agglomeration method with the objective to transform generated statistical vectors into a more informative and discriminative feature space. Comparatively, statistical representations of APAAC and Qsorder with dimensionality reduction lead to the formation of slightly unique yet heavily dependent clusters as shown in Figure 4B. Though these encodings could be used for classification purposes, however, still the performance would not be very promising. In addition, the clusters do not seem independent because a single human protein that interacts with some viral proteins, might not interact with some other viral proteins. This means that positive and negative class samples can have very similar representations due to the presence of such proteins. Although dimensionality reduction produces better feature space, however still clusters are not very much separable. To further improve the performance of the predictor, we perform iterative representation learning, where we pass 3 different statistical representations separately to RF and ET classifiers and take their predicted class probabilities to develop a new feature space. The generated feature space leads to the formation of unique and independent clusters as shown in Figure 4C, which suggests the presence of comprehensive discriminatory features for interactive and non-interactive VHPPI pairs. Due to the discriminative and informative nature of the newly generated feature space, we utilize this feature space to train the SVM classifier for virus-host protein-protein interaction prediction.

Overall, as compared to state-of-the-art predictors, the proposed predictor has shown a slight performance improvement on Barman and Denovo datasets, and significant performance improvements on SARS-CoV-2 datasets and the other 4 datasets namely, TR1-TS1, TR2-TS2, TR3-TS1, and TR4-TS2. We believe that the performance of the proposed predictor can be further improved by incorporating representations learned through diverse types of language models such as BERT, and XLNET.

To facilitate the biological community, we provide an interactive, and user-friendly web server for the proposed meta-predictor which is available at https://sds_genetic_analysis.opendfki.de/MP-VHPPI/. This web server can be used to predict virus-host protein-protein interactions across 7 different datasets by using raw human and virus protein sequences. In addition, the online web server allows the user to train the proposed meta-predictor from scratch on all of the datastets. Moreover, it also provides 7 different benchmark datasets that are used in this study.

The prime objective of this research is the development of a robust machine learning-based computational framework capable of precisely predicting viral-host protein-protein interactions across a wide range of hosts and viruses. The proposed meta predictor makes use of APAAC and QS order sequence encoders for statistical representation generation and feature agglomeration method to refine feature space. Furthermore, the meta predictor utilizes the predictions of random forest and extra tree classifiers to feed the SVM classifier that makes final predictions. Experimental results reveal the competence of APAAC and QS order encoders for most effectively generating numerical representations of sequences by capturing amino acids sequence order and distributional information. We have observed dimensionality reduction method removes irrelevant and redundant information which slightly improves the performance of classifiers. The process of iterative representation learning in which predictions of RF and ET classifiers are passed to the SVM classifier significantly improves the accuracy of interaction predictions. The proposed meta predictor has evaluated over 7 benchmark datasets where it outperforms existing predictors with a significant margin of 3.07, 6.07, 2.95, and 2.85%, in terms of accuracy, MCC, precision, and sensitivity, respectively. We believe that the deployment of the proposed meta-predictor as a web interface will assist researchers and practitioners in analyzing the complex phenomenon of VHPPIs at a larger scale to unravel substantial drug targets and optimize antiviral strategies.

Comprehensive performance analysis reveals that the proposed MP-VHPPI predictor manages to outperform existing viral-host protein-protein interaction predictors across 7 benchmark datasets by a decent margin. Although the use of different strategies at the level of representation learning reduces the prediction error decently, however, the proposed model lacks robustness as it is biased toward type II error. In the future, we will optimize the predictive pipeline of the proposed MP-VHPPI predictor with an aim to enhance robustness.

The benchmark datasets for this study can be found at: https://sds_genetic_analysis.opendfki.de/MP-VHPPI/Download/.

MA conceptualized the presented idea and prepared graphics and developed web server. MA and MI performed data curation, formal analysis, validation, and investigation. MA and AF prepared the original draft and final manuscript under the supervision of AD and SA. All authors contributed to the article and approved the submitted version.

This study was supported by Sartorius Artificial Intelligence Lab.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.