Leveraging a meta-learning approach to advance the accuracy of Nav blocking peptides prediction

The voltage-gated sodium (Nav) channel is a crucial molecular component responsible for initiating and propagating action potentials. While the α subunit, forming the channel pore, plays a central role in this function, the complete physiological function of Nav channels relies on crucial interactions between the α subunit and auxiliary proteins, known as protein–protein interactions (PPI). Nav blocking peptides (NaBPs) have been recognized as a promising and alternative therapeutic agent for pain and itch. Although traditional experimental methods can precisely determine the effect and activity of NaBPs, they remain time-consuming and costly. Hence, machine learning (ML)-based methods that are capable of accurately contributing in silico prediction of NaBPs are highly desirable. In this study, we develop an innovative meta-learning-based NaBP prediction method (MetaNaBP). MetaNaBP generates new feature representations by employing a wide range of sequence-based feature descriptors that cover multiple perspectives, in combination with powerful ML algorithms. Then, these feature representations were optimized to identify informative features using a two-step feature selection method. Finally, the selected informative features were applied to develop the final meta-predictor. To the best of our knowledge, MetaNaBP is the first meta-predictor for NaBP prediction. Experimental results demonstrated that MetaNaBP achieved an accuracy of 0.948 and a Matthews correlation coefficient of 0.898 over the independent test dataset, which were 5.79% and 11.76% higher than the existing method. In addition, the discriminative power of our feature representations surpassed that of conventional feature descriptors over both the training and independent test datasets. We anticipate that MetaNaBP will be exploited for the large-scale prediction and analysis of NaBPs to narrow down the potential NaBPs.


Dataset construction and preprocessing
To ensure a fair comparison between the existing method and our proposed method, we utilized the same benchmark dataset generated by the previous work 3 .This dataset was originally extracted from NTXpred, created by Saha and Raghava 21 .It comprises 244 NaBPs and 244 non-NaBPs, with 240 positive samples and 236 negative samples remaining after excluding samples containing special symbols.From these samples, 80% of the benchmark dataset were randomly selected to construct the training dataset (i.e., 192 positive samples and 188 negative samples), whereas the remaining samples were used to construct the independent test dataset (i.e., 48 positive samples and 48 negative samples).

Feature representation
Here, we employed fifteen types of feature descriptors using eight different feature encoding schemes to fully capture the useful patterns of NaBPs (Table 1).These feature descriptors have previously demonstrated reasonable prediction performance for several protein properties [22][23][24][25][26][27][28][29] .They are derived from three groups of feature descriptors, including sequence composition-based, physicochemical property-based, and pseudo-amino acid composition-based features.A detailed description of three groups is summarized as follows.
For the sequence composition-based group, we applied three well-known feature encodings, including AAC, DPC, and TPC [30][31][32] .AAC encoding calculates the occurrence frequencies of the 20 possible amino acids.The composition of the ith amino acid ( aa(i) ) can be defined as: where AA i is the number of the ith amino acid, and L is the length of the query protein.Meanwhile, DPC and TPC encodings calculate the occurrence frequencies of one and two consecutive amino acids, respectively.The composition of the jth dipeptide ( dp j ) and kth tripeptide ( tp(k) ) can be defined as: (1) where DP j and TP k are the number of the jth dipeptide and kth tripeptide, respectively.As a result, AAC, DPC, and TPC encoding generate 20-D, 400-D, and 8000-D feature vectors, respectively.
CTD encoding involves three feature descriptors (i.e., composition (C), transition (T) and distribution (D)) 33 .This encoding is naturally used to characterize the global distribution patterns and physicochemical properties of amino acids in the protein sequences.In particular, CTD groups 20 amino acids into 3 major groups based on seven physicochemical properties of amino acids, as summarized in Table S1.
For a given the ith physicochemical property ( PCP i ), CTDC represents the global percentage of each group of 20 amino acids, which can be calculated as follows: where N PCP1,G1 is the number of amino acids involved in Group #1 of physicochemical property #1, with phys- icochemical property #1 being hydrophobicity (Table S1).Using CTDC, any protein sequence can be encoded as a 21-D feature vector.CTDT encoding computes the number of amino acids in the current group followed by amino acids of another group, while CTDD encoding calculates the distribution descriptor involving the fractions of the entire protein sequence.Thus, the dimensions of the CTDT and CTDD feature vectors are 21 and 105, respectively.
In case of PAAC encoding 34 , it was initially introduced to address the issue of sequence-order information.In particular, the PAAC feature vector can be represented as: where f c is the percentage composition of the amino acid type c, while θ j represents the jth rank of �(P(S i ), P(S i+j )) and is the maximum correlation length.As a result, the dimension of the PAAC feature vector is 21 + , respectively, where = 1, 2, 3, …, 8.All the fifteen types of feature descriptors were extracted using the protr package in the R programming language 35 . (2)

Machine learning algorithms
In this work, six well-known ML algorithms were applied to develop the proposed model, including random forest (RF), extreme gradient boosting (XGB), support vector machine (SVM), partial least squares (PLS), multilayer perceptron (MLP), and k-nearest neighbor (KNN).The detailed implementations and parameter optimizations of these six ML models are briefly summarized below.RF is considered a powerful ensemble learning model that makes the final prediction output based on the voting strategy of multiple decision trees 36 .Two essential parameters in the development of RF models are the number of randomly selected features (mtry) and decision trees (ntree).The search range for mtry and ntree were {3, 5, 7, 9, 10} and {20, 50, 100, 200, 300}, respectively.We trained and optimized RF model using the ran-domForest package in the R language.XGB has been shown to achieve reasonable accuracy in many machine learning problems [37][38][39] .The basic idea of XGB is to create a strong prediction model by combining several weak prediction models.More detailed information on XGB is provided in the study by Chen and Carlos 40 .Herein, the XGB model was trained and optimized using the xgboost package in R language.The details of search ranges of XGB' parameters are summarized in Table S2.For SVM, it is well-recognized as an efficient ML algorithm for binary classification tasks.The principle idea of SVM is to determine an optimal hyperplane that can maximize the margin between two classes (i.e., positive and negative samples).The SVM model was trained based on the cost parameter ({0.25, 0.5, 1, 2, 4}) using the e1071 package in the R language.Detailed information on MLP, PLS, and KNN is summarized our previous studies 25,26,32 .

Meta-learning framework of MetaNaBP
In this study, MetaNaBP is a meta-learning approach developed to enhance the prediction of NaBPs.To date, the meta-learning approach has been efficiently used for the prediction and analysis of various peptides [22][23][24] and drugs [41][42][43] .As illustrated in Fig. 1, the meta-learning process involves two main steps: (i) base-classifier construction and (ii) meta-classifier optimization 22,23,32,44 .
In the first step, fifteen types of feature descriptors from eight different feature encoding were applied to represent NaBPs.These feature descriptors were individually inputted into six different ML algorithms.As a result, we obtained a total of 90 base-classifiers.During model training, all the base-classifiers were built and optimized using a grid search and the tenfold cross-validation approach.Detailed information regarding the search ranges of ML classifiers' parameters are recorded in Table S2.In the next step, we utilized all the 90 baseclassifiers to generate the estimated probabilities to be NaPBs and then obtained a 90-dimensional (90-D) feature vector containing 90 probabilistic features (PFs).For a given protein sequence P, its PFs are defined as follows: (9) PFs = [P(M 1 , FD 1 ), P(M 1 , FD 2 ), P(M 1 , FD 3 ) . . .., P(M i , FD j ), . . ., P(M 6 , FD 15 )] where P(M i , FD j ) is the estimated probability derived from the i th ML algorithm and the j th feature descriptor.In general, each base-classifier is capable of providing two important pieces of information: estimated probabilities and classes to be NaPBs.In case of class information, it can be can represented by where C(M i , FD j ) is the estimated class label derived from the ith ML algorithm and the jth feature descriptor.
As a result, we obtained a 90-D feature vector containing 90 CFs.In this study, we fused them to construct a multi-view feature vector (PCF) containing 90 PFs and 90 CFs.To enhance the feature ability, a two-step feature selection approach was applied to optimize the PF, CF, and PCF feature vectors.In brief, the first step is to calculate the importance score of each feature using the mean decrease of Gini index (MDGI) and then generate the ranked features in descending order.The second step is to identify the optimal feature set using the sequential forward search (SFS) strategy.In SFS, we construct several feature sets containing m top-ranked importance features, where m = 5, 10, 15, …, n.After that, we sequentially input m top-ranked importance features into RF classifiers.In this study, the feature sets is considered the best one when the corresponding RF classifiers provide the highest MCC.The best feature sets of PF, CF, and PCF are referred as FS-PF, FS-CF, and FS-PCF, respectively.

Performance measurement
We used five well-known performance metrics to assess the prediction performance of the prediction models, including ACC, MCC, SN, F1, specificity (SP), the area under the receiver operating characteristics (ROC) curve (AUC), and precision (PRE) [45][46][47] .ACC, MCC, SN, and SP are defined as: In this study, the numbers of correctly predicted NaBPs and non-NaBPs are denoted as TP and TN, respectively.Additionally, the number of non-NaBPs predicted as NaBPs is denoted as FP, while the number of NaBPs predicted as non-NaBPs is denoted as FN [48][49][50] .

Analysis of amino acid preference of Na v s blocking peptides
We employed a Student's t-test to assess the statistically significant difference in amino acid preference between NaBPs and non-NaBPs, as summarized in Table 2. Interestingly, hydrophobic amino acid residues (i.e., Tyr, Ala, Trp, and Gly) and negatively charge amino acids (Asp and Glu) significantly dominated in NaBPs with a p-value of less than 0.05 (Fig. 2).Among these nonpolar side-chain residues, Try and Gly showed the highest difference scores of 0.026 and 0.025, respectively.On the other hand, hydrophilic or polar uncharged residues (Cys, Ser, Gln, and Thr) and positively charged side-chain amino acids (Arg and Lys) were more preferable in non-NaBPs.
A similar pattern of the hydrophobic and negatively charged portions of Na v inhibitors were also observed in domain IV voltage sensor-targeting peptides 51 .For instance, the conserved hydrophobic patch surrounded by a charged ring was typically observed in GMTs isolated from spiders.According to experiments with hydrophobic substitution analogs of tarantula peptide GpTx-1, the lower hydrophobicity of NaBPs could lead to a significant decrease in potency on Na v voltage sensor binding ability 5,51 .This structural feature has been proposed to promote not only NaBPs-voltage-gated ion channel interactions but also NaBPs-lipid membrane interactions 15 .Typically, increasing the hydrophobicity at specific amino acid positions within the peptide, presumed to directly interact with the channel, generally resulted in the maintenance or augmentation of the potency of domain IV voltage sensors in Na v 1.4 and Na v 1.7 6,52 .Moreover, only van der Waals force and hydrophobic interactions to the compound was provided by the lipid-exposed pocket.Dealing with the effects of drug binding, deglycosylation, and the role of hydrophobic residues in the voltage sensors, it could be implied that side-chain hydrophobicity 1 , P(M i , FD j ) ≥ 0.5 0 , P(M i , FD j ) < 0.5  53 .This basic feature of the region makes the fragments of structure with an acidic binding property easily attach to it.Thus, negatively charged side chains (Asp and Glu) might be required for the interaction of NaBPs with the S4s segment.
Focusing on the aromatic side-chain residues, both Trp and Tyr were significantly dominated in NaBPs with a p-value of less than 0.05.These pieces of evidence might be correlated with the fact that the conformations of the aromatic groups of Trp and Tyr are prone to forming π-π stacking interaction, binding to the selectivity pocket of the voltage sensor domains.Similar to the conserved aromatic residue in NaBPs isolated from scorpion, Trp is crucial for the interaction with sodium channels Na v 1.7, Na v 1.4 and Na v 1.5 54 .

Performance evaluation of different feature encoding methods and ML algorithms
In this section, we used five performance metrics to compare the impact of different feature encoding methods and ML algorithms on the prediction of NaBPs.Each developed ML classifier was evaluated using both the  www.nature.com/scientificreports/tenfold cross-validation and independent tests.After that, we selected the top-five ML classifiers with the highest cross-validation MCC for conducting the comparative analysis.The results of all the ML classifiers over the tenfold cross-validation and independent tests are recorded in Fig. 3 along with Fig. S1 and Tables S3-S4, while the results of the top-ten ML classifiers are documented in Table S5.Based on the cross-validation MCC, the top-ten ML classifiers consisted of MLP-TPC, RF-DPC, PLS-TPC, SVM-TPC, MLP-DPC, SVM-DPC, PLS-DPC, RF-CTDD, RF-CTD, and SVM-PAAC( = 4), with corresponding MCCs of 0.902, 0.886, 0.883, 0.882, 0.858, 0.851, 0.847, 0.835, 0.829, and 0.828.Interestingly, seven out of the top-ten ML classifiers were developed based on DPC and TPC encoding methods.This indicates that DPC and TPC might be important for effectively capturing information about NaBPs.To demonstrate this point, we calculate average five performance metrics for fifteen types of feature descriptors over six ML algorithms (Table S6).The top-five feature descriptors, having the highest average MCC of 0.830, 0.795, 0.751, 0.750, and 0.747, were DPC, TPC, PAAC( = 7), PAAC( = 6), and AAC, respectively.As shown in Fig. 3 and Table S5, MLP-TPC is considered the best among ML classifiers developed in this study, achieving ACC of 0.950, SN of 0.917, SP of 0.984, and AUC of 0.988 on the training dataset.On the other hand, for the independent test dataset, the best ML classifier was SVM-DPC, achieving ACC, SN, SP, MCC, and AUC of 0.958, 0.938, 0.979, 0.917, and 0.988, respectively.In brief, the ranks of MLP-TPC were 1 and 6 as evidenced by both the tenfold cross-validation and independent test, respectively.This confirms that using single ML-based approaches could not provide robust and stable prediction performance.

Development of MetaNaBP
To overcome the limitation of single ML-based approaches, we employed several types of feature representations (i.e.PF, CF, PCF, FS-PF, FS-CF, and FS-PCF) to develop different meta-classifiers.Table 3 and Fig. 4 shows that PF and PCF attain similar and better performance compared with CF in terms of MCC.Specifically, PF and PCF exhibited ACC, SN, and MCC, and AUC with ranges of 0.966-0.982,0.943-0.964,and 0.933-0.964,respectively, on the training dataset.To improve the prediction performance of the model and reduce computational time, we employed a two-step feature selection strategy.As seen in Table 3, the MCC of FS-PF and FS-PCF is better than their controls in terms of the tenfold cross-validation test.Among several types of our feature representations, it can be stated that FS-PCF outperformed other fusion features in terms of cross-validation MCC.Regarding the results of the independent test, FS-PCF secures the best prediction performance in terms of ACC (0.948), SN (0.979), and MCC (0.898) (Fig. 4).Altogether, we selected FS-PCF as the best feature representation and inputted it into the RF classifier to develop the final meta-classifier, namely MetaNaBP.

Meta-learning can improve the effectiveness and generalization ability
To reveal the effectiveness of the meta-learning approach, assessed and compared the performance of MetaNaBP and other base-classifiers using five performance metrics, on both the training and independent datasets.For convenient of comparative analysis, the performance of MetaNaBP was compared against the top-five baseclassifiers, including MLP-TPC, RF-DPC, PLS-TPC, SVM-TPC, and MLP-DPC (Table 4).Impressively, for both the cross-validation and independent test, MetaNaBP demonstrated better performance in terms of ACC, MCC, AUC, and SP compared to the top-five base-classifiers.Moreover, MCC values of MetaNaBP were 7.19% and 6.13% higher than the best base-classifier (i.e., MLP-TPC) on the training and independent dataset, respectively.In addition, we investigated the feature ability of FS-PCF by comparing its performance against conventional feature descriptors.To make a fair test, we trained 15 individual RF models in conjunction with each type of feature descriptors and evaluate their performance.Summarized in Fig. 5 and Table 5, along with Table S7, are the detailed comparative results in terms of the tenfold cross-validation and independent tests.As visualized in Fig. 5, FS-PCF attained outstanding performance on four metrics (i.e., AUC, SN, ACC, and MCC) as observed in both the tenfold cross-validation and independent tests.Intriguingly, when compared with top-three conventional feature descriptors (i.e., DPC, CTDD, and CTD) over the independent test, FS-PCF was more effective and demonstrated an excellent ability in identifying NaBPs, achieving ACC of 0.948, SN of 0.979, MCC of 0.898, and AUC of 0.991 (Table 5).The improved performance of FS-PCF on the training dataset suggested that our feature

Comparison of MetaNaBP with the existing method
To assess the predictive efficacy of our proposed method, we compared the results of tenfold cross-validation and independent tests for MetaNaBP with those of PEP-PREDNa + 3 .As seen in Table 6, it's worth noting that there are five different prediction models (i.e., ET, RF, XGB classifier (XGBC), XGBRF classifier (XGBRFC), LightGBM classifier (LGBMC)) found in PEP-PRED Na+ .Please note that the experimental results of these five different prediction models were directly obtained from the previous work 3 to ensure a fair comparison.As can be seen, amongst the five different prediction models found in PEP-PRED Na+ , LGBMC performed better that the others, as indicated by the tenfold cross-validation results, achieving an ACC of 0.840, SN of 0.840, SP of 0.850, F1 of 0.840, and MCC of 0.680.Compared with PEP-PRED Na+ (LGBMC), the ACC, SN, SP, F1, and MCC of our proposed method significantly improved by 14.42, 12.88, 15.00, 14.43, and 28.69%, respectively.The improved performance of MetaNaBP on the training dataset suggests that the multi-view feature fusion strategy used herein can advance the efficiency of model training.Furthermore, on the independent test dataset, MetaNaBP also achieved the best ACC, SN, F1, and MCC, which were 5.79, 11.92, 6.62, and 11.76% higher than those of PEP-PRED Na+ (LGBMC).Overall, MetaNaBP surpassed the existing method and provided a stable prediction performance, indicating its effectiveness and generalization ability.

Interpretability of the MetaNaBP model
In the feature selection process, we calculated the discriminative ability of each feature using MDGI.Figure 6 displays the ranks of feature importance for FS-CF, FS-PF, and FS-PCF.As mentioned above, because FS-PF and FS-PCF displayed better performance compared to FS-CF, we analyzed their impact to understand the prediction outputs of MetaNaBP.As seen in Fig. 6C, the top-ten informative features of FS-PCF include PLS-TPC_PF, MLP-TPC_PF, PLS-TPC_CF, MLP-TPC_CF, RF-DPC_PF, RF-CTD_PF, RF-DPC_CF, RF-AAC_PF, RF-TPC_PF, and RF-CTD_CF, with corresponding MDGI values of 15.64, 14.03, 9.22, 8.35, 7.30, 6.18, 5.54,  5.53, 5.08, and 4.88.In the meanwhile, PLS-TPC_PF, MLP-TPC_PF, RF-DPC_PF, RF-CTD_PF, and RF-AAC_PF were identified among the top-ten informative features for both FS-PF and FS-PCF (Fig. 6B,C).To reveal the directionality of these features, we employed the Shapley Additive explanation (SHAP) approach.Figure 7 shows that the values for these five highlighted features tend to be relatively low for most non-NaBPs and high for most NaBPs.This observation is further confirmed by the five boxplots illustrated in Fig. 8.It suggests that these five important features are beneficial for discriminating NaBPs from non-NaBPs and play a pivitol role in achieving the improved performance of MetaNaBP.

Future perspective and direction of NaBP prediction models
Rapid and accurate identification of NaBPs based solely on sequence information is capable of enhancing the comprehension of the structure and function of targeted NaBPs at the molecular level.This understanding www.nature.com/scientificreports/proves valuable for the screening and design of therapeutic drugs for pain.Specific NaBPs targeting the Nav channel have been proposed through computational design, followed by experimental validation.For instance, the computationally designed NaBPs (PTx2-3127 and PTx2-3258), inspired by ProTx-II from tarantula venom, effectively inhibit Nav1.7 activation in mouse and human sensory neurons 55,56 .At the cellular level, validation through a tetrazolium-based mouse neuroblastoma cell assay revealed that the NaBP candidate 'CcNT, ' derived from the tentacle venom of the Scyphozoa C. capillata, specifically inhibits Na v channels at receptor site 1 57 .In tissue and animal model experiments, rodents are commonly used for physiological and systemic validations of potential NaBP candidates from the computational design pipeline.For example, PTx2-3127 inhibits Na v 1.7 and demonstrates efficacy in rat models of chronic and thermal pain when administered intrathecally 55 .Notably, cytotoxicity and cardiotoxicity assays are also required to avoid the undesirable side effects of NaBPs 58 .In addition to chronic and thermal pain experiments, inflammatory and neuropathic pain in rodent models can also be considered for the functional validation of NaBPs, similar to the case of μ-EPTX-Na1a from the venom of the Chinese cobra (Naja atra) 58 .
In the future work, we will incorporate our proposed ML framework with ordinary differential equation (ODE)-based theoretical modelling to provide insights into the dynamics, interactions, and effects of NaBPs on specific Nav targets, revealing regulatory mechanisms similar to studies on gene/protein signaling networks aimed at identifying therapeutic targets in diseases 59,60 .Additionally, advanced interaction predictive models based on deep learning methods could predict associations between NaBP candidates and specific Nav targets, mirroring the development of computational models identifying relationships between genetic markers and diseases [61][62][63] .Beyond bioactive peptides, non-coding RNAs (ncRNAs), such as MicroRNAs (miRNAs) and long non-coding RNAs (lncRNAs), also have potential as diverse channel blockers in neurons [64][65][66] .In this context, a meta-learning approach could be applied alongside interaction predictive models for NaBPs and ncRNAs, aiming to understand their functions and identify related protein targets.This approach aligns with various deep learning methods that have utilized similar algorithms in molecular marker studies for other diseases 63,[67][68][69][70][71][72] .

Conclusions
In this study, we have developed a novel and high-accuracy computational model to accurately identify NaBPs, named MetaNaBP.To the best of our knowledge, MetaNaBP is the first meta-learning model designed for NaBPs prediction.MetaNaBP utilizes a wide range of sequence-based feature descriptors that cover multiple perspectives, inputting them to six powerful ML algorithms to construct several base-classifiers.These base-classifiers,  in turn, provide useful feature representations.Simultaneously, these feature representations were optimized using a two-step feature selection method to identify the most effective feature set, which was then applied to develop the final meta-predictor.To validate the performance of MetaNaBP, we conducted a comparative analysis of our meta-learning model and its base-classifiers, along with an existing method.The results from both crossvalidation and independent tests indicated that MetaNaBP outperformed its base-classifiers, underscoring the effectiveness of the meta-learning approach.Impressively, on the independent test dataset, MetaNaBP was capable of increasing MCC of 0.820 to 0.898, ACC of 0.890 to 0.948, SN of 0.900 to 0.979.This indicates that MetaNaBP could be a useful tool to precisely identify NaBPs.Anticipating MetaNaBP to be a useful tool, we envision it playing a crucial role in drug discovery and development by screening and identifying potential NaBPs.Although MetaNaBP achieved high-accuracy performance in NaBP prediction, we still have a few aims to enhance its overall performance in our future work.Firstly, we aim to employ other types of feature encoding schemes, such as Word2vec and one-hot encoding methods 73 .Secondly, we plan to integrate our proposed model with powerful deep learning methods, such as convolutional neural networks and long short-term memory networks 74 .Lastly, we intend to establish a web server to directly perform NaBP prediction based solely on the primary sequence.

Figure 1 .
Figure 1.The framework of the proposed MetaNaBP based on a meta-learning approach.There exist four main steps, including (A) dataset preparation.(B) base-classifier construction.(C) meta-classifier optimization.(D) performance evaluation.

Figure 2 .
Figure 2. Boxplots of amino acid compositions of 20 amino acids for NaBPs and non-NaBPs.X-and Y-axes represent 20 amino acids along with their p-value.

Figure 4 .
Figure 4. Performance comparison of different feature representations over the independent test dataset.(A-C) The performances of different feature representations in terms of ACC, SN, SP, MCC and AUC.(D) The feature number of different feature representations.

Figure 5 .
Figure 5. ACC, SN, SP, MCC, and AUC values of FS-PCF and conventional feature descriptors as evaluated by the tenfold cross-validation (A) and independent tests (B).

Figure 6 .
Figure 6.Feature importance of our feature representations (i.e., FS-CF (A), FS-PF (B), FS-PCF (C)).The feature having the largest value of mean decrease of Gini index (MDGI) is the most important one.

Figure 7 .
Figure 7. SHAP values of the 20 based-classifiers selected by the MetaNaBP.

Figure 8 .
Figure 8. Boxplots of five important PFs selected by MDGI values on the training dataset for NaBPs and non-NaBPs.

Table 2 .
Average amino acid compositions of Na v blocking (Positive) and non-Na v (Negative) blocking peptides along with difference values and p-value.

Table 3 .
Cross-validation and independent test results of different feature representations.The best performance value for each performance metrics across different feature representations is highlighted in bold.

Table 4 .
Performance comparison of MetaNaBP and top-five powerful prediction models on the training and independent test datasets.The best performance value for each performance metrics across different methods is highlighted in bold.

Table 5 .
Performance comparison of our feature (FS-PCF) and conventional feature descriptors on the independent test dataset.The best performance value for each performance metrics across different methods is highlighted in bold.

Table 6 .
Performance comparison of MetaNaBP and the existing methods on the training and independent test datasets.The best performance value for each performance metrics across different methods is highlighted in bold.