iPNHOT: A knowledge-based approach for identifying protein-nucleic acid interaction hot spots

doi:10.21203/rs.2.9629/v1

Download PDF

Research article

iPNHOT: A knowledge-based approach for identifying protein-nucleic acid interaction hot spots

https://doi.org/10.21203/rs.2.9629/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 06 Jul, 2020

Read the published version in BMC Bioinformatics →

You are reading this older preprint version

Read the latest preprint version →

Background

The interaction between proteins and nucleic acids plays pivotal roles in various biological processes such as transcription, translation, and gene regulation. Hot spots are a small set of residues that contribute most to the binding affinity of a protein-nucleic acid interaction. Compared to the extensive studies of the hot spots on protein-protein interfaces, the hot spot residues within protein-nucleic acids interfaces remains less well-studied, in part because mutagenesis data for protein-nucleic acids interaction are not as abundant as that for protein-protein interactions.

Results

In this study, we built a new computational model, iPNHOT, to effectively predict hot spot residues on protein-nucleic acids interfaces. One training data set and an independent test set were collected from dbAMEPNI and some recent literature, respectively. To build our model, we generated 97 different sequential and structural features and used a two-step strategy to select the relevant features. The final model was built based only on 9 features using a support vector machine (SVM). The features include three unique features such as ∆SASsa1/2, ∆SASna1/2, and esp3, which are newly proposed in this study. Based on the cross validation results, our model gave F1 score and AUC as 0.720 and 0.807 on the subset collected from ProNIT, respectively, compared to 0.407 and 0.670 of mCSM-NA, a state-of-the art model to predict the thermodynamic effects of protein-nucleic acid interaction. The iPNHOT model was further tested on the independent test set, which showed that our model outperformed other methods.

Conclusion

In this study, by collecting data from a recently published database dbAMEPNI, we proposed a new model, iPNHOT, to predict hotspots on protein-nucleic acid interfaces. The result shows that our model outperformed the existing state-of-art models. Our model is available for users through a webserver: http://zhulab.ahu.edu.cn/iPNHOT/.

Bioinformatics

Protein-nucleic acid interaction

Hot spots

Feature selection

electrostatic potential

Support vector machine

The interaction of proteins with nucleic acids is essential in many different cellular processes, such as translation, RNA-metabolism, gene regulation, DNA replication and repair, and so on [1, 2]. The understanding of the interactions between proteins and nucleic acids sheds light on designing new functions and regulating cellular behaviors.

Mutagenesis studies on protein-protein and protein-nucleic acid interfaces provide important clues in exploring the drivers of the interactions [3-5]. It has been shown that the mutation of a few interface residues to alanine can dramatically decrease the binding affinity [6, 7]. Those residues are called hotspots (HS) [8]， often defined as a residue whose mutation to alanine generates a binding free energy difference over 2 kcal/mol [9].

While hotspots on protein-protein interfaces have been extensively studied by both experimental and computational methods [6, 9-16], the hotspots on protein-nucleic acid interfaces are not as comprehensively investigated. Possibly, the inherent characteristics, such as electrostatic and hydration of the protein-nucleic acid interfaces, make it difficult to characterize the energetics of mutations. In addition, very few of the energetic data about the residues on protein-nucleic acid interfaces were collected in the past decades, which make the development of computational methods at a slow pace.

Some protein-nucleic acid alanine mutagenesis data from the literature were collected in the ProNIT database [17]. Based on these data, several computational methods had been developed to predict the effect of the mutagenesis or hot spots on protein-nucleic acid interfaces [11, 18-20]. Pires et al. used the concept of graph-based signatures to predict the effects of the mutations on protein-nucleic acids interfaces [19]. They built a model called mCSM-NA that can quantitatively predict the effects of the mutations in protein coding regions on nucleic acid binding affinities. Barik et al. developed a method, HotSPRing, to predict the hot spots at protein-RNA recognition sites [18]. The model was built by using random forests based on structural and physicochemical features. Ramos et al. developed a computational alanine scanning mutagenesis methodology to predict the hot spots on protein-DNA interfaces [20]. Munteanu et al. developed a model by the combination of a SVM (support vector machine) with a genetic algorithm (GA) as the wrapper for feature selection to predict the hot spots on protein-nucleic acid interfaces based on solvent accessible surface area and residue conservation [11]. Recently, Pan et al. developed a method, PrabHot, for predicting hot spots on protein-RNA interfaces by collecting data from ProNIT and literature [21].

Although a few computational methods have been developed to predict the mutation effects on protein-nucleic acid interfaces, the generalization of those models were not well validated due to the limited number of data points. In this study, alanine mutagenesis data of protein-nucleic acid interfaces were collected from a comprehensive database dbAMEPNI [22]. Seven kinds of sequential or structural features were generated for the interface residues. Then, we used a two-step strategy to select the relevant features, through which, we were able to develop a knowledge-based model to predict the HS on protein-nucleic acid interfaces.

Benchmark datasets

Training dataset

The training dataset comes from dbAMEPNI database [22] which was built in our group. The database contains alanine mutagenic effects data from ProNIT database [17] and our curated data collected from literature between 2011 and 2017.

Firstly, we identified 335 interface residues from dbAMEPNI database by defining the interface residue as a residue whose buried solvent accessible surface area is larger than 0.0. Then, we detected redundancy among homologous proteins using the PISCES server [23] with a sequence identity cutoff set to 25%. If homologous pairs are included, the sites of recognition differ between the two proteins. Finally, we obtained a dataset containing 299 interface residues, which come from 106 protein-nucleic acid complexes that consist of 74 protein-DNA complexes, 31 protein-RNA complexes and 1 protein-RNA/DNA complexes. The complexes and the 299 interface residues are listed in Table S1 and S2 of the additional file 1, respectively. According to Table S2, 106 interface residues are common to the data in ProNIT. By using as cutoff, 86 of the 299 interface residues are defined as hot spot residues and the remaining 213 residues are defined as non-hot spot residues.

Independent test set

The independent test set comes from three different sources: (i) The independent test set in Pires et al.’s work [19] which comes from Barik et al.’s paper [18]. In their paper, they collected 80 alanine mutations from 14 protein-RNA complexes. Note that these alanine mutations were also included in the dbAMEPNI database later. After checking these protein-RNA complexes, we found one complex (PDBID: 2Y8W) appeared in our training dataset, so we removed this complex. The other complex 2XS2 was also removed because the only one corresponding residue is not on the protein-RNA interface. (ii) The independent test set reported in Pan et al.’s work [21]. In their work, they collected the 58 mutations as the independent test set. After carefully checking the 58 data, we found several problems of the dataset: (1) It includes 13 non-alanine mutations; (2) The multiple mutants were wrongly considered as single alanine mutants. For example, the mutants of 4JVH are all double mutants, however, they were used as single alanine mutants in their dataset; (3) There is redundancy between 5EN1 and 5HO4. According to these problems, we removed part of the data and kept 33 data points. We have summited the 33 data points on the PrabHot server, and 23 PrabHot scores (Table S4 in additional file 1) were available for plotting the ROC curve (Figure S1 in additional file 1). (iii) Literature corresponding to the 3D structures of protein-nucleic acid complexes deposited in PDB database [24] from 2017 to 2018. We identified 51 protein-nucleic acid complexes, and the corresponding references were carefully examined to find the alanine mutation information. From these articles, we obtained 16 alanine mutation data.

For the proteins collected from the three sources mentioned above, we used PISCES server to determine the sequence identity between them and the proteins in the training dataset using the sequence identity cutoff 25%. If homologous pairs are included, the sites of recognition differ between the two proteins. In all, we obtained 99 interface residues which come from 27 protein-nucleic acids complexes that consist of 22 protein-RNA complexes, 4 protein-DNA complexes and 1 protein-RNA/DNA complexes (Table S3 in additional file 1). By using as cutoff, 13 of the 99 interface residues are hotspot residues and 86 of them are non-hot spot residues (Table S4 in additional file 1).

Feature extraction

Due to the special characteristics of the protein-nucleic acid interfaces, we generated 7 different kinds of features to build our model, which were described as below.

Physicochemical characteristics of 20 amino acids.

As the basic unit that comprises a protein, it is the residues that interact with other molecules. Different properties of the 20 residues have been deposited in AAindex [25], 10 of which were used as features to predict the hotspots on protein-protein interfaces in previous studies [26-29]. The 10 physicochemical properties were considered as the first 10 features as showed in Table S5 (see additional file 1). The numerical values of the 10 features are showed in Table S6 (see additional file 1).

Depth index and protrusion index.

The surface shape complementarity between proteins and nucleic acids is an important factor in protein-nucleic acid binding. The surface geometry of residues in the interface is quantified using features in this study. We used the PSAIA [30] program to calculate the depth index (DI) and protrusion index (PI) for each interface residue. The program calculates several different kinds of depth index and protrusion index including the average values of the entire and side chain of the residue, the maximum and minimum values of the residue’s atoms. We used the former 4 values for each residue in both bound and unbound state as features. More specifically, these features contain the average DI of the entire residue, the average DI of the side chain of the residue, the average PI of the entire residue, and the average PI of the side chain of the residue.

In addition, we calculated the differences of these 4 values of each residue between bound and unbound states by using the following equations:

(1) Due to technical limitations, this equation is only available as a download in the supplemental files section.

(2) Due to technical limitations, this equation is only available as a download in the supplemental files section.

(3) Due to technical limitations, this equation is only available as a download in the supplemental files section.

(4) Due to technical limitations, this equation is only available as a download in the supplemental files section.

where, the and mean the average DIs of the total residue in unbound and bound states, respectively. the and mean the average DIs of the side chain of the residue in unbound state and bound states, respectively. We did the same for PI. Furthermore, we calculated the relative DIs and PIs according to the following equations:

(5) Due to technical limitations, this equation is only available as a download in the supplemental files section.

(6) Due to technical limitations, this equation is only available as a download in the supplemental files section.

(7) Due to technical limitations, this equation is only available as a download in the supplemental files section.

(8) Due to technical limitations, this equation is only available as a download in the supplemental files section.

In all, we obtained 16 features related to depth index and protrusion index.

Features related to solvent accessible surface area (SASA).

The residue’s solvent accessible surface area has been used as features in previous studies [11, 14, 31-33] in predicting hotspots on protein-protein interfaces. In this study, we used the different representations of SASA as features to build our model. The SASA was calculated by NACCESS program [34], which calculated the SASA of a residue in different scenarios, for example, the absolute SASA and the relative SASA, the SASA of all atoms, side chain atoms, backbone atoms, polar, and nonpolar atoms of the residue. We obtained these SASAs in both bound and unbound states.

In addition, we calculated the buried SASA that is the difference of the SASA between proteins in bound and unbound states. The buried SASA has been thought to correlate with different energy terms such as desolvation energy. In this work, we calculated different kinds of buried absolute SASA and relative SASA mentioned above. Furthermore, we considered different powers of the buried absolute SASA and relative SASA as features. The three powers we tested are 0.5, 1.5, and 2.0.

In all, we obtained 54 features related to SASA. These features can be found in Table S5 (see additional file 1).

Features related to electrostatic potential

Considering the electrostatic characteristics of nucleic acids, we supposed the electrostatic potential could be benefit for predicting hotspots on protein-nucleic acid interfaces. In this study, we used the APBS program [35] to calculate the electrostatic potential around the proteins, and the procedure to calculate the electrostatic potential of a residue has been described in our previous study [36]. The description of the 5 features related to electrostatic potential can be found in Table S5.

Hydrogen bond features.

By using the WHATIF server [37] we obtained the hydrogen bonds [38] on protein-nucleic acid interfaces. The hydrogen bond numbers formed by the entire residue and those of the side chain with nucleic acid were counted as two features.

Secondary structure features

A residue’s secondary structure is assigned by the DSSP program [39, 40], which outputs 8 different kinds of secondary structure that include H (α-helix), B (isolated β-bridge), E (extended strand), G (3-helix), I (5-helix), T (hydrogen bonded turn), S (bend), and blank (loops). We re-categorized them into 5 different types by combining B, T, S as turn, and G, I as helix1. Then the 5 different types of secondary structure were represented as binary vectors by using (1, 0, 0, 0, 0) as H, (0, 1, 0, 0, 0) as E, (0, 0, 1, 0, 0) as turn, (0, 0, 0, 1, 0) as helix1, and (0, 0, 0, 0, 1) as loops.

Sequence conservation features

Based on our previous works[36, 41], we obtained 5 features from the PSSM file generated by PsiBlast. The first one is the information entropy that represents the conservation of the corresponding sequence position. In addition, we defined two kinds of relative conservation based on the weighted observed percentage of each kind of residues for each sequence position as follows:

(9) Due to technical limitations, this equation is only available as a download in the supplemental files section.

(10) Due to technical limitations, this equation is only available as a download in the supplemental files section.

where, , is the weighted observed percentage of residue type at the certain sequence position, with the formulas designed to avoid division by 0. is the weighted observed percentage of the residue type “alanine” at the certain sequence position. Label ‘rm’ means the residue type with the maximum percentage, and ‘ra’ means the actual residue type at that sequence position. And ‘wop’ is the abbreviation of ‘weighted observed percentage’. Similarly, we also defined two kinds of relative conservation based on the position specific scores in the position specific scoring matrix (PSSM) as follow:

(11) Due to technical limitations, this equation is only available as a download in the supplemental files section.

(12) Due to technical limitations, this equation is only available as a download in the supplemental files section.

where, is the position specific score of residue type on the certain sequence position. Labels ‘ra’, ‘rm’, and ‘A’ have the same meaning as above, and ‘pps’ is the abbreviation of ‘position specific score’.

In all, we obtained 97 features in this study, and the z-scores were calculated to standardize all the features.

Feature selection

Feature selection has become an important step for building machine learning models, especially for high-dimensional applications. By feature selection, redundant and irrelevant features can be removed, and we can also avoid over-fitting, improve model performance and provide faster and more cost-effective models.

Previous study [41] shows that a hybrid two-step feature selection strategy is effective to detect relevant feature subset. In this work, we combined decision tree and sequential forward feature selection as a two-step strategy to determine the relevant feature subset. First, we used a MATLAB function FITCTREE to select a feature subset. Then, we used the sequential forward feature selection (SFS) method to determine the final feature subset.

Evaluation with SVM

Support vector machine (SVM) has been used to build models for predicting hotspots on protein-protein interfaces in several previous studies [11, 14, 15, 31], due to its low complexity and robust output. In this study, SVMlight [42] and the radial basis function were used to train our models. The two parameters, G and C, were optimized by a grid search with G values from 0 to 2 and C values from 0 to 40. To avoid over-fitting, we used the leave-one-out cross validation to evaluate the models. Then, the model was further validated on an independent test set.

Due to the imbalance of our data set, the overall accuracy is heavily biased by the accuracy of the negative examples. Therefore, we provide several different metrics, sensitivity (SEN), specificity (SPE), accuracy (ACC), precision (PRE), F1 score and Matthew correlation coefficient (MCC) to evaluate the performances of different models. These metrics are defined as follows:

(13) Due to technical limitations, this equation is only available as a download in the supplemental files section.

(14) Due to technical limitations, this equation is only available as a download in the supplemental files section.

(15) Due to technical limitations, this equation is only available as a download in the supplemental files section.

(16) Due to technical limitations, this equation is only available as a download in the supplemental files section.

(17) Due to technical limitations, this equation is only available as a download in the supplemental files section.

(18) Due to technical limitations, this equation is only available as a download in the supplemental files section.

where TP, FP, TN, and FN represent the numbers of true positive (predicted hot spot residues are actual hot spots), false positive (predicted hot spot residues are actual non-hot spots), true negative (predicted non-hot spot residues are actual non-hot spots) and false negative (predicted non-hot spot residues are actual hot spots), respectively. In addition to these 6 parameters, the Area under the Curve (AUC) of the Receiver Operating Characteristic (ROC) curve was also used as a metric to evaluate our model. The ROC curve shows the relationship between true positive rate and false positive rate, and the area under the curve (AUC) of ROC curve indicates how strongly the model separates the positive and negative examples.

Statistical analysis to detect the relationship between features and hotspots (Wilcoxon Rank Sum Test)

Statistical analysis is useful to reveal the role of each feature on differentiating hot spots from non-hot spots. Because the normal distribution of our data was not guaranteed, the t-test could not be used to analyze the selected features. Instead, the Wilcoxon Rank Sum test was used in the statistical analysis. The Wilcoxon Rank Sum test is a nonparametric test to assess whether two samples of observations come from the same distribution. The RANKSUM function in MATLAB was used in this statistical analysis.

Feature Selection

Features selected by decision tree

As showed in List 1, the decision tree selected 21 features from 97 original features, which include 3 physicochemical features of amino acids, 7 features related to depth index and protrusion index, 7 features related to solvent accessible surface area, 2 features related to electrostatic potential, 1 feature related to secondary structure, and 1 feature related to conservation.

List 1. The features selected by Decision tree. The number in the parenthesis is corresponding to the feature number in Table S5

Due to technical limitations, List 1 is only available as a download in the supplemental files section.

The final feature subset selected by SFS

From the preliminary subset of features selected by decision tree, we further used a sequential forward feature selection (SFS) process to determine the final subset of features as input of the final model. In each round of the SFS process, different feature combinations were used to train models by SVM, and the cross-validation results (F1 score) were used to evaluate these feature combinations. We selected the top three feature combinations in each round for the next round. Table S7 in the additional file 1 shows the features selected in each round and the corresponding cross validation F1 scores. The results show that the predictive performance is convergent at the 9th round. The best cross-validated F1 score is 0.698. The corresponding feature combination contains 9 features, which are Nphb, Hdrpo, , , SASbau,, , , and Helix. Nphb is the number of potential hydrogen bonds of the residue. Hdrpo, is the hydrophobicity of the residue. is the difference of the side chain depth indexes between bound and unbound states. is the difference of the side chain protrusion indexes between bound and unbound states. SASbau is the backbone absolute SASA of the residue in the unbound state. and are the square roots of the differences of the absolute SASAs of residue side chain and nonpolar part between unbound and bound states. is the electrostatic potential of the neighbor residue and the target residue. Helix is if the residue lies in a helix secondary structure.

Based on these 9 features, we built our final model, iPNHOT (identification of protein-nucleic acid interaction hot spots), using SVM. The parameters of G and C for radial basis function used in the final model are 0.1 and 20.0, respectively. The cross-validation results show that our model achieved 0.686, 0.711, 0.696, and 0.829 for recall, precision, F1 score, and accuracy, respectively.

In addition, we plot the ROC curve based on the cross-validation results as shown in Figure 1. The AUC of the curve is 0.862.

Our previous work [14] in predicting hotspots on protein-protein interfaces showed that differences between the leave one-residue out cross validation and the leave one-protein our cross validation is small. In this work, we also did a leave one-protein out cross validation based on the final feature subset. The results indicated that the leave one-protein out cross validation achieved the sensitivity, specificity and F1 score of 0.605, 0.878 and 0.634, respectively, which is a little bit worse than that of the leave one out cross validation.

Models based on all features or the features selected only by decision tree or SFS

To validate the effectiveness of our two-step feature selection process, we also built models based on all 97 features (AFmodel), the 21 features selected only by decision trees (DTmodel), and the features selected only by SFS (SFSmodel), respectively. As shown in Table 1, the AFmodel gives the lowest predictive accuracies compared with other models. The iPNHOT model is superior to DTmodel on all the six evaluation metrics. We inferred that using all features or the 21 features selected by decision tree may have over-fitted the models. In addition, we did the SFS feature selection based on the original 97 features, although it is five more times time-consuming than our two-step feature selection process. Table S8 in the additional file 1 shows the SFS process, it was convergent at the 6th round. Table 1 shows that the iPNHOT model is also superior to SFSmodel on all the six evaluation metrics.

To further demonstrate the effectiveness of our two-step feature selection strategy, we also combined NSGA-II and SVM to select the relevant feature subset and optimize the G and C parameters of SVM. NSGA-II (Non-dominated Sorting Genetic Algorithm II) is a popular method for multiple objective optimization [43]. We tried different populations (50-250) and different generations (50-250) of NSGA, as shown in Figure 2, the best F1 score is 0.662 which was obtained when population and generation were set to 50 and 250, respectively. Thus, we show that our two-step feature selection strategy is superior to GA in this study.

Comparison of different classifiers on the selected 9 features

To evaluate the effectiveness of the SVM learning method in predicting the hot spots within protein-nucleic acid interfaces, we compared the performance of models built by different machine learning algorithms (KNN, naïve Bayesian (NB) and Logistic Regression (LR)) based on the selected 9 features. Table 2 shows that the model built based on SVM (iPNHOT) achieved the highest recall (0.686), the highest precision (0.711), the highest accuracy (0.829), the highest F1 score (0.698), and the highest MCC (0.580) compared with other models. These results indicated that the SVM model outperformed the models built by KNN, naïve Bayesian, logistic regression based on the selected 9 features.

In addition, to further evaluate the effectiveness of our feature selection process and the SVM learning method in predicting the hot spots on protein-nucleic acid interfaces, we also compared our iPNHOT model with the random forest model built using all the 97 features. We used all the 97 features because random forest classifier is an ensemble learning method and the diversity of trees is important for the algorithm. One of the important steps of the random forest algorithm is to select a feature subset randomly, then to determine an optimal feature from the feature subset to divide the examples. Thus, the diversity of the trees can be enhanced by using all features. We tried different tree numbers and selected the one which gives the best predictive accuracy. The optimal tree number is 54. Table 2 shows that iPNHOT achieved higher values than the random forest model for all the six evaluation metrics except specificity, demonstrating that our two-step feature selection strategy and the SVM learning method are effective in predicting hot spot on protein-nucleic acid interfaces.

Evaluation of our model on the independent test set

The generalization of the iPNHOT model was evaluated on the independent test set. Table 3 shows that the recall, the specificity, the accuracy on the independent test set is 0.538, 0.872, 0.828 that is close to the cross-validation recall, specificity, accuracy of 0.686, 0.887, 0.829, respectively, which shows the good generalization of the iPNHOT model.

Comparison with other methods

Our iPNHOT model was built to predict the interface hot spot residues on both protein-RNA and protein-DNA interfaces. The SBHD server [11] is also for predicting hotspot residues on both protein-RNA and protein-DNA interfaces, however, it is not available now. mCSM-NA server [19] contains modules to predict mutagenic effect of residues on both protein-RNA or protein-DNA interfaces, and it is available to the community. HotSPRing [18] and ProbHot [21] are two models for predicting hot spots on protein-RNA interfaces. However, it turned out HotSPRing server does not work well after giving it several tries. PrabHot server only outputs the predicted scores for predicted hotspot residues. In addition, PrabHot defined the hotspot residues by using a cutoff value 1.0 kcal/mol, which is different from the cutoff value 2.0 kcal/mol used in this study.

Thus, the AUC is the only metric that can be compared between PrabHot and iPNHOT.

First, we compared our model to mCSM-NA on both the training data set and the independent test set. The prediction results for all examples in the training data set and the independent test set are shown in Table S2 and Table S4, respectively. As shown in Table 3, the cross-validation results of iPNHOT outperform the predictive results of mCSM-NA according to all the 6 evaluation metrics. However, only part of the training data set, collected from ProNIT, has been used to train the mCSM-NA model. To fairly compare with mCSM-NA, we extracted the 106 interface residues obtained from ProNIT, and compared the predictive results between iPNHOT and mCSM-NA on these 106 data points. Table 3 indicates that iPNHOT outperforms mCSM-NA on all the metrics except specificity.

In addition, we also compared iPNHOT with mCSM-NA on the independent test set. Table 3 shows that iPNHOT outperforms mCSM-NA on all the metrics except recall.

In addition to the 6 performance metrics, we also plotted the ROC curves to compare different methods. Figure 1 shows the ROC curves based on the predictive results of mCSM-NA vs. iPNHOT on the training data set. For the 106 data collected from ProNIT, the area under the curve (AUC) of mCSM-NA is 0.670 that is substantially lower than the AUC of iPNHOT (0.807). Figure 3 shows that the AUC of mCSM-NA is 0.734 which is lower than the AUC of iPNHOT (0.832) on the independent test set.

Furthermore, we compared the AUCs between iPNHOT and the PrabHot. Because part of the data in the independent test set have been used to train the PrabHot model, the AUCs were calculated based on 23 samples which were not used to train the PrabHot model and whose predicted scores of PrabHot are available. As shown in Fig. S1, the AUC of PrabHot is 0.525 which is lower than the AUC of iPNHOT (0.658).

Thus, we demonstrated that our model outperforms other state-of-art methods for predicting hotspots on protein-nucleic acid interfaces.

Statistical analysis of the selected 9 features

To further evaluate the ability of the 9 selected features to distinguish hot spot from non-hot spots, we used the Wilcoxon rank sum analysis. Figure 4 shows that five of the 9 selected features can significantly differentiate hot spots from non-hot spots with p-values less than 0.05, which are , , , , and esp3. The former two features, and , reflect the shape complementarity between protein residues and nucleic acid upon binding. As we proposed in the “Feature extraction” section, and may related to the desolvation energy upon binding. As for esp3, it is the electrostatic potential of protein surface patch around the target residue. For hot spots, the average value of the feature is 11.6 compared to 2.53 for non-hot spots. Because of the negative electrostatic potential of nucleic acid surface, this feature may partially reflect the electrostatic potential complementarity between the protein surface patch and the nucleic acid surface patch around the target residue. Thus, these 5 features combined the effects of shape complementarity, electrostatic potential complementarity, and the desolvation energy.

In addition to the features that were statistically important on an individual basis, the other 4 of the 9 selected features were also kept in the final model. This suggests the possibility of coordinated effects between different features. In particular, a feature that is not individually significant can gain significance when combined with other information gleaned from other features.

The analysis of the 21 selected features by decision tree can be found in the additional file 1.

Case Study

To visualized show the hotspot residues on the protein-nucleic acid interfaces, we plotted two cases by using PyMol. The first one is the complex of U1 small nuclear ribonucleoprotein A (U1A) and an RNA, for which the PDB ID is 1AUD. As showed in Figure 5A, 4 hotspot residues and 2 non-hotspot residues at the interface had been recorded in the training dataset. Our model identified all the 4 hot spot residues as hotspots and the 2 non-hotspot residues as non-hotspot residues when doing both leave one out cross validation and leave one-protein out cross validation. On the contrary, mCSM-NA did not assign any of the 4 residue as hot spot residue. The second case is the complex of Pot1(protection of telomere) and a DNA, for which the PDB ID is 1QZG. As showed in Figure 5B, 2 hotspot residues and 3 non-hotspot residues at the interface had been recorded in the training data set. Our model identified all of the 2 hotspot residues as hotspots and all the 3 non-hotspot residues as non-hotspot residues when doing both leave one out cross validation and leave one-protein out cross validation. However, mCSM-NA did not detect any of the 2 hot spot residues. Note that both 1AUD and 1QZG were collected from ProNIT, which had been used to train the mCSM-NA model.

The interface hot spot residues provide clues to understand the principles driving the interaction between protein and nucleic acids. In this study, we collected a non-redundant training dataset with 299 alanine-mutated residues on protein-nucleic acid interfaces from dbAMEPNI database. Based on this data set, we developed a knowledge-based method to predict hot spot residues on protein-nucleic acids interfaces. Using the two-step feature selection strategy, we selected 9 features from the original 97 features, which include some unique feature such as , and esp3. Our model shows better performances compared with mCSM-NA on both the training data set and the independent test set.

The selected features were further analyzed to reveal the relationship between features and hot spots. Among the selected 9 features, the differences of 5 features for hot spot and non-hot spot residues are statistically significant and the 5 features are , , , and esp3. The two features, , and , reflect the shape complementarity or the buried condition of the target residues. The and may reflect the desolvation energy of residues. The esp3 reflect the patch electrostatic potential complementarity around the residue. The differences of the other 4 features are not significant and the 4 features are Nphb, Hdrpo, SASbau, and Helix. Our results show both predictive ability of single feature and the complementarity between features are important for building our model.

SVM: Support Vector Machine HS: Hotspots GA: Genetic Algorithm

DI: Depth Index PI: Protrusion Index SASA: Solvent Accessible Surface Area

PSSM: Position Specific Scoring Matrix SEN: Sensitivity SPE: Specificity

ACC: Accuracy PRE: Precision MCC: Matthew Correlation Coefficient

AUC: Area Under the Curve ROC: Receiver Operating Characteristic

SFS: Sequential Forward feature Selection

Acknowledgements

The authors thank Tianle Wu for assistance with manuscript preparation and revision.

Funding

This work has been supported by the National Natural Science Foundation of China under grants No. 21403002 and 31601074. The fundings had no role in the design of the study and collection, analysis, and interpretation of data and writing the manuscript.

Availability of data and materials

All data generated or analyzed during this study are included in this published article or the Additional files.

Authors’ contributions

Conceived the study: XZ, JCM. Designed the study: XZ, JCM. Participate designed the study: LL, JH, TF. Analyzed the data: XZ, LL, JH, TF, YX. Wrote the paper: XZ, LL, JCM, YX. All authors read and approved the manuscript.

Competing interests

The authors declare that they have no competing interests.

Consent for publication

Not applicable.

Ethics approval and consent to participate

Not applicable.

1. Bowater RP, Cobb AM, Pivonkova H, Havran L, Fojta M: Biophysical and electrochemical studies of protein–nucleic acid interactions. Monatshefte für Chemie - Chemical Monthly 2015, 146(5):723-739.

2. Ollis DL, White SW: Structural basis of protein-nucleic acid interactions Chemical Reviews 1987, 87(5):981-995.

3. Janin J: Principles of protein-protein recognition from structure to thermodynamics. Biochimie 1995, 77(7-8):497-505.

4. Kucukkal TG, Petukh M, Li L, Alexov E: Structural and physico-chemical effects of disease and non-disease nsSNPs on proteins. Current opinion in structural biology 2015, 32:18-24.

5. Keskin O, Gursoy A, Ma B, Nussinov R: Principles of protein-protein interactions: what are the preferred ways for proteins to interact? Chem Rev 2008, 108(4):1225-1244.

6. Kortemme T, Baker D: A simple physical model for binding energy hot spots in protein-protein complexes. Proceedings of the National Academy of Sciences of the United States of America 2002, 99(22):14116-14121.

7. Bogan AA, Thorn KS: Anatomy of hot spots in protein interfaces. Journal of molecular biology 1998, 280(1):1-9.

8. DeLano WL: Unraveling hot spots in binding interfaces: progress and challenges. Current opinion in structural biology 2002, 12(1):14-20.

9. Clackson T, Wells JA: A hot spot of binding energy in a hormone-receptor interface. Science 1995, 267(5196):383-386.

10. Lise S, Archambeau C, Pontil M, Jones DT: Prediction of hot spot residues at protein-protein interfaces by combining machine learning and energy-based methods. BMC bioinformatics 2009, 10:365.

11. Munteanu CR, Pimenta AC, Fernandez-Lozano C, Melo A, Cordeiro MN, Moreira IS: Solvent accessible surface area-based hot-spot detection methods for protein-protein and protein-nucleic acid interfaces. Journal of chemical information and modeling 2015, 55(5):1077-1086.

12. Ofran Y, Rost B: Protein-protein interaction hotspots carved into sequences. PLoS computational biology 2007, 3(7):e119.

13. Tuncbag N, Keskin O, Gursoy A: HotPoint: hot spot prediction server for protein interfaces. Nucleic acids research 2010, 38(Web Server issue):W402-406.

14. Zhu X, Mitchell JC: KFC2: a knowledge-based hot spot prediction method based on interface solvation, atomic density, and plasticity features. Proteins 2011, 79(9):2671-2683.

15. Xia JF, Zhao XM, Song J, Huang DS: APIS: accurate prediction of hot spots in protein interfaces by combining protrusion index with solvent accessibility. BMC bioinformatics 2010, 11:174.

16. Moreira IS, Koukos PI, Melo R, Almeida JG, Preto AJ, Schaarschmidt J, Trellet M, Gumus ZH, Costa J, Bonvin A: SpotOn: High Accuracy Identification of Protein-Protein Interface Hot-Spots. Scientific reports 2017, 7(1):8007.

17. Kumar MD, Bava KA, Gromiha MM, Prabakaran P, Kitajima K, Uedaira H, Sarai A: ProTherm and ProNIT: thermodynamic databases for proteins and protein-nucleic acid interactions. Nucleic acids research 2006, 34(Database issue):D204-206.

18. Barik A, Nithin C, Karampudi NB, Mukherjee S, Bahadur RP: Probing binding hot spots at protein-RNA recognition sites. Nucleic acids research 2016, 44(2):e9.

19. Pires DE, Ascher DB: mCSM-NA: predicting the effects of mutations on protein-nucleic acids interactions. Nucleic acids research 2017.

20. Ramos RM, Moreira IS: Computational Alanine Scanning Mutagenesis-An Improved Methodological Approach for Protein-DNA Complexes. Journal of chemical theory and computation 2013, 9(9):4243-4256.

21. Pan Y, Wang Z, Zhan W, Deng L: Computational identification of binding energy hot spots in protein-RNA complexes using an ensemble approach. Bioinformatics 2018, 34(9):1473-1480.

22. Liu L, Xiong Y, Gao H, Wei DQ, Mitchell JC, Zhu X: dbAMEPNI: a database of alanine mutagenic effects for protein-nucleic acid interactions. Database (Oxford) 2018, 2018.

23. Wang G, Dunbrack RL, Jr.: PISCES: a protein sequence culling server. Bioinformatics 2003, 19(12):1589-1591.

24. Rose PW, Prlic A, Altunkaya A, Bi C, Bradley AR, Christie CH, Costanzo LD, Duarte JM, Dutta S, Feng Z et al: The RCSB protein data bank: integrative view of protein, gene and 3D structural information. Nucleic acids research 2017, 45(D1):D271-D281.

25. Kawashima S, Pokarowski P, Pokarowska M, Kolinski A, Katayama T, Kanehisa M: AAindex: amino acid index database, progress report 2008. Nucleic acids research 2008, 36(Database issue):D202-205.

26. Porollo A, Meller J: Prediction-based fingerprints of protein-protein interactions. Proteins 2007, 66(3):630-645.

27. Chen XW, Jeong JC: Sequence-based prediction of protein interaction sites with an integrative method. Bioinformatics 2009, 25(5):585-591.

28. Li N, Sun Z, Jiang F: Prediction of protein-protein binding site by using core interface residue and support vector machine. BMC bioinformatics 2008, 9:553.

29. Ramachandran P, Antoniou A: Identification of Hot-Spot Locations in Proteins Using Digital Filters. IEEE Journal of Selected Topics in Signal Processing 2008, 2(3):378-389.

30. Mihel J, Sikic M, Tomic S, Jeren B, Vlahovicek K: PSAIA - protein structure and interaction analyzer. BMC structural biology 2008, 8:21.

31. Cho KI, Kim D, Lee D: A feature-based approach to modeling protein-protein interaction hot spots. Nucleic acids research 2009, 37(8):2672-2687.

32. Tuncbag N, Gursoy A, Keskin O: Identification of computational hot spots in protein interfaces: combining solvent accessibility and inter-residue potentials improves the accuracy. Bioinformatics 2009, 25(12):1513-1520.

33. Darnell SJ, Page D, Mitchell JC: An automated decision-tree approach to predicting protein interaction hot spots. Proteins 2007, 68(4):813-823.

34. Hubbard SJ, Thornton JM: Naccess. Computer Program, Department of Biochemistry and Molecular Biology. University College London 1993, 2(1).

35. Baker NA, Sept D, Joseph S, Holst MJ, McCammon JA: Electrostatics of nanosystems: application to microtubules and the ribosome. Proceedings of the National Academy of Sciences of the United States of America 2001, 98(18):10037-10041.

36. Zhu X, Ericksen SS, Mitchell JC: DBSI: DNA-binding site identifier. Nucleic acids research 2013, 41(16):e160.

37. Vriend G: WHAT IF: a molecular modeling and drug design program. Journal of molecular graphics 1990, 8(1):52-56, 29.

38. Hooft RW, Sander C, Vriend G: Positioning hydrogen atoms by optimizing hydrogen-bond networks in protein structures. Proteins 1996, 26(4):363-376.

39. Kabsch W, Sander C: Dictionary of protein secondary structure: pattern recognition of hydrogen-bonded and geometrical features. Biopolymers 1983, 22(12):2577-2637.

40. Touw WG, Baakman C, Black J, te Beek TA, Krieger E, Joosten RP, Vriend G: A series of PDB-related databanks for everyday needs. Nucleic acids research 2015, 43(Database issue):D364-368.

41. Qiao Y, Xiong Y, Gao H, Zhu X, Chen P: Protein-protein interface hot spots prediction based on a hybrid feature selection strategy. BMC bioinformatics 2018, 19(1):14.

42. Joachims T: Learning to classify text using support vector machines. Dissertation, Kluwer 2002.

43. Deb K, Pratap A, Agarwal S, Meyarivan T: A fast and elitist multiobjective genetic algorithm: NSGA-II. IEEE Transactions on Evolutionary Computation 2002, 6(2):182-197.

Table 1. Cross-validation results of models based on all features and the features selected by only decision tree, only sequential forward selection, and our two-step feature selection process

Models	REC	PRE	SPE	ACC	F1 score	MCC
iPNHOT	0.686	0.711	0.887	0.829	0.698	0.580
DTmodel	0.547	0.644	0.878	0.783	0.591	0.447
SFSmodel	0.663	0.606	0.826	0.779	0.633	0.477
AFmodel	0.417	0.563	0.869	0.739	0.480	0.317

a The Random forest model is based on the all 97 features generated in this study, and the corresponding tree number is 54

Table 2. Cross validation results of different classifiers based on the selected 9 features.

Learning algorithms	REC	PRE	SPE	ACC	F1 score	MCC
KNN1	0.512	0.518	0.808	0.722	0.515	0.320
KNN3	0.512	0.603	0.864	0.763	0.554	0.396
KNN5	0.442	0.603	0.883	0.756	0.510	0.360
NB	0.477	0.540	0.836	0.732	0.506	0.325
LR	0.302	0.531	0.892	0.722	0.385	0.238
Random Foresta	0.384	0.647	0.916	0.763	0.482	0.360
SVM(iPNHOT)	0.686	0.711	0.887	0.829	0.698	0.580

Table 3. Comparison with mCSM-NA on both the training data set and the independent test set.

Datasets	Methods	REC	PRE	SPE	ACC	F1 score	MCC
Training dataset	mCSM-NA	0.419	0.350	0.685	0.609	0.381	0.0991
	iPNHOT	0.686	0.711	0.887	0.829	0.698	0.580
Training dataset(ProNIT)a	mCSM-NA	0.297	0.647	0.913	0.698	0.407	0.273
	iPNHOT	0.730	0.711	0.841	0.802	0.720	0.567
Independent test set	mCSM-NA	0.615	0.174	0.558	0.566	0.271	0.118
	iPNHOT	0.538	0.389	0.872	0.828	0.452	0.360

a The subset of the training data set which includes 106 residues collected from ProNIT.

Download PDF

Journal Publication

published 06 Jul, 2020

Read the published version in BMC Bioinformatics →

Editorial decision: Major revision
22 Jun, 2019
Review #2 received at journal
16 Jun, 2019
Review #4 received at journal
13 Jun, 2019
Review #3 received at journal
13 Jun, 2019
Review #1 received at journal
09 Jun, 2019
Reviewer #4 agreed at journal
05 Jun, 2019
Reviewer #2 agreed at journal
03 Jun, 2019
Reviewer #3 agreed at journal
03 Jun, 2019
Reviewers invited by journal
26 May, 2019
Reviewer #1 agreed at journal
26 May, 2019
Submission checks completed at journal
17 May, 2019
Editor assigned by journal
17 May, 2019
Editor invited by journal
13 May, 2019

You are reading this older preprint version

Read the latest preprint version →

iPNHOT: A knowledge-based approach for identifying protein-nucleic acid interaction hot spots

Status:

Journal Publication

Version 1

Abstract

Figures

Background

Methods

Results and Discussion

Conclusion

Abbreviations

Declarations

References

Tables

Supplementary Files

Status:

Journal Publication

Version 1