HomPPI: a class of sequence homology based protein-protein interface prediction methods

Computational Prediction of Protein-Protein Interfaces

Several different genetic, biochemical, and biophysical methods have been used to identify and characterize protein interfaces [1]. These experiments are very valuable and have contributed greatly to our knowledge of protein-protein interfaces. However, the high cost in time and resources required for these experiments call for reliable computational approaches to identify interface residues. In addition to providing important clues to biological function of novel proteins, computational predictions can reduce the searching space required for docking two polypeptides [3].

To distinguish interface residues from non-interface surface residues, a wide range of sequence, physicochemical and structural features have been investigated [3–18], and many in silico approaches to protein-protein interface prediction have been explored in the literature (reviewed in [19–21]). Protein-protein interface prediction algorithms can be classified into three categories: (i) sequence-based methods, which use only the primary amino acid sequence of the query protein as input [3, 22–28]; (ii) structure-based methods, which make use of information derived from the structure of the query protein [5, 18, 29–31]; and (iii) methods that use both sequence and structure derived information in making predictions [32, 33].

Several sequence-based protein-protein interface prediction methods have been explored in the literature [3, 22–28]. Most, if not all, of these methods, extract for each residue in the query protein, a fixed length window that includes the target residue and a fixed number of its sequence neighbours. Each residue is classified as an interface residue or a non-interface residue based on features of the amino acids in the corresponding window. Various methods differ both in the specific machine learning algorithms or statistical methods employed and in terms of the specific features of the amino acids used. Commonly used features include the identity of the amino acids in the window [27], the amino acid composition of interfaces [34], the physicochemical properties of the amino acids [35], and the degree of conservation of the amino acids (obtained by aligning the query sequence with homologous sequences) [3]. Some studies report substantial improvements in interface residue prediction when predicted structural properties, e.g., solvent surface accessibility and secondary structure of the residues are utilized [21].

A number of structure-based methods [5, 18, 29–31] or hybrid methods that combine both sequence and structure-derived information [32, 33] have been proposed for predicting protein interfaces. The performance of the best-performing sequence-based methods is generally lower than that of structure-based methods (see [21] for a comparison). A possible explanation for the difference in the performance of sequence-based and structure-based protein interface residue predictors is that the latter can trivially eliminate non-surface residues from the set of candidate interface residues and potentially exploit a rich set of features derived from the 3D structures.

The use of structure-based methods, however, is limited to proteins for which the structure of the query protein is available, and the number of solved structures significantly lags behind the number of protein sequences [35]. Even when the structure of a query protein is available, the application of structure-based prediction methods is complicated by conformational changes that take place when some proteins bind to their partners. Structure-based methods rely on structural features extracted from the structure in the unbound state or from a bound complex that has been separated into constituent proteins. It is unclear whether such structural features are indeed reliable predictors of interfaces for proteins that undergo significant conformational changes upon binding [20, 36]. Moreover, higher organisms have a large number of intrinsically disordered proteins/regions (IDPs/IDRs) that undergo induced folding only after binding to their partners [37]. Such disordered regions - for which experimental structure information is, by definition, lacking - participate in many important cellular recognition events, and are believed to contribute to the ability of some hub proteins to interact with multiple partners in protein-protein interaction networks [38]. Hence, there is an urgent need for sequence-based methods for reliable prediction of protein-protein interfaces.

Analysis of Interface Residue Conservation

The relationship between sequence conservation and various aspects of protein structure, interaction, expression, and function has been the focus of many studies over the past decades [39, 40], and sequence homology-based methods have been used for predicting both protein structure and protein function [41–53]. Thus, it is natural to ask whether protein-protein interface residues can be reliably identified using sequence homology-based methods. Published studies disagree on whether protein-protein interfaces are more conserved than the rest of the protein sequences. Grishin and Phillips [54], after examining five enzyme families, concluded that the degree of conservation of interfaces is same as that of protein sequences as a whole. The studies by Caffrey et al. [55] as well as Reddy and Kaznessis [56], found that the interacting surface-patches are not significantly more conserved than other surface-patches. Caffrey et al. [55], based on their study of 64 protein-protein interacting chains, found that interface residues are slightly more conserved than the rest of the protein surface residues. Reddy and Kaznessis [56], based on their study of 28 hetero transient and non-transient complexes, found that the fraction of highly conserved interface residues is greater than that of highly conserved non-interface surface residues. They suggested that the number of conserved residue positions is more predictive of protein-protein binding sites than the average conservation index of residues in the target patch. Choi et al. [57] analyzed 2,646 protein interfaces based on a conservation score that measures the position-specific evolutionary rate estimated using a phylogenetic tree [58], and concluded that protein interface residues are more conserved than non-interface surface residues.

Despite the disagreement regarding whether interface residues are conserved or not, several researchers have used conservation of residues to predict protein-protein interfaces with varying degrees of success. For example, the Evolutionary Trace (ET) method [59, 60] and its variants [58, 61–64] calculate conservation score for each residue using a phylogenetic tree built from a multiple sequence alignment. Residues with conservation scores above a certain threshold are mapped onto the 3D structure of the protein to identify putative binding sites. Carl et al. [65] used a dataset of sixteen transient protein chains to explore the feasibility of predicting protein-protein binding sites based on their membership in structurally conserved surface patches (where conserved patches are identified using structural alignment of a query protein with one or more of its structural homologs). Bordner and Abagyan [66] and Wang et al. [67] calculated evolution rate for each amino acid of protein sequences using phylogenetic trees, and used evolution rate as an attribute along with other physicochemical and sequential attributes to train a SVM classifier for interface residue prediction. Panchenko et al. [68] predicted functional sites of proteins using spatial averages of sequence conservation scores. Shoemaker et al. [69] have recently developed a web server for predicting protein binding sites by inspecting homologous proteins with similar structures. Based on a statistical analysis of target-template sequence alignments on a benchmark dataset of 329 two-chain complexes, Kundrotas and Vakser [70] have shown that it is possible to obtain high quality alignment of interface residues even when the overall alignment quality is rather poor. Specifically, they concluded that in approximately 50% of the complexes considered, the overall accuracy of the modelled interfaces was good enough for guiding docking.

Overview of the Paper

Against this background, we study a class of sequence homology-based methods for protein-protein interface prediction. We introduce a novel measure of interface conservation that captures the degree to which interface residues in each protein are conserved among its sequence homologs. First, we describe the results of our analysis of the interface conservation among homologous sequences using several large non-redundant datasets of protein-protein interfaces extracted from the Protein Data Bank (PDB) [71], including datasets that allow us to compare "obligate" versus "transient" interfaces. To explore the extent to which interface conservation can be exploited in the prediction of interface residues, we systematically examined the relationship between interface conservation and six sequence-based variables. In one set of experiments, we examined binding interfaces in homologous proteins without specifying a specific interaction partner (i.e., non-partner specific, NPS-interfaces). The results of this analysis indicated that interfaces in obligate complexes are, in general, more highly conserved than those in transient complexes. In a complementary set of experiments, we examined interfaces in complexes between specific pairs of proteins (i.e., partner-specific, PS-interfaces). In contrast to the results for NPS-interfaces, by focusing on the interface of each query protein with a specific binding partner, we discovered a high degree of sequence conservation in transient PS-interfaces. This analysis revealed that transient interfaces tend to be highly partner-specific.

Second, based on the results of protein interface conservation analysis we propose HomPPI, a class of sequence homology-based approaches to protein interface prediction. We present two variants of HomPPI: (i) NPS-HomPPI (non partner-specific HomPPI), which can be used to predict interface residues of a query protein in the absence of knowledge of the interaction partner; and (ii) PS-HomPPI (partner-specific HomPPI), which can be used to predict the interface residues of a query protein with a specific target protein. The performance of both HomPPI methods was evaluated on several benchmark datasets, including a large non-redundant set of transient complexes. Due to the increasing importance of intrinsically disordered proteins in understanding molecular recognition mechanics and in rational drug design and discovery [72–75], we also tested NPS-HomPPI on two datasets of intrinsically disordered proteins.

Finally, we compare the performance of HomPPI with that of other web-based servers for interface residue prediction, using several performance measures that assess the reliability of correctly predicting, on average, interface and non-interface residues in a given protein. We discuss the relative advantages and limitations of homology-based methods for interface residue prediction.

Conservation of PPIs in Non-Partner Specific (NPS) Interfaces

First, we examined the conservation of PPI residues in the absence of knowledge of interaction partners. For this study, we analyzed interfaces in putative homologs (hereafter, we refer to putative homologs as "homologs" for simplicity) of each protein in a large non-redundant dataset, Nr6505. After removing chains with interfaces containing fewer than 3 amino acids, we were left with 5853 chains. For each of the 5853 remaining proteins, we extracted homologs from the nr_pdbaa_s2c database using BLASTP [77] with expectation value (EVal) ≤ 10 from the resulting set of homologs, we eliminated those that were nearly identical to the query sequence (to ensure an accurate estimate of conservation). To ensure that the interface residues of the homologs could be reliably determined, we retained only those homologs that were part of complexes with resolution 3.5 Å or better. For each query-homolog pair in sequence alignments generated by BLASTP, we used the interface residues of the homolog(s) to predict the interface residues of the query protein. We calculated the correlation coefficient (CC) between the predicted and actual interface residues of the query protein, and refer to this value as the interface conservation (IC) score, i.e., the degree of conservation of interface residues between the query protein and its homologs (see Methods for details).

We examined the dependence of the interface conservation score on six NCBI BLAST alignment statistics: Expectation value (EVal), Identity Score, Positive Score, Local Alignment Length (LAL) and two Alignment Length Fractions (LAL/Query Length) and (LAL/Homolog Length). The EVal is a statistic that estimates the number of hits expected by chance when searching database of a particular size; the lower the EVal value, the more significant the score. The Identity Score is a measure of the degree of sequence identity between two amino acid sequences. The Positive Score returned by BLASTP is the number of positive-scoring matches in an alignment. It takes into account observed substitutions that preserve the physicochemical properties of the original residue. The LAL is the length of the local alignment; Alignment Length Fractions are LAL normalized by the length of the query or the length of the identified homologous sequence. We represent each query-homolog pair as a six dimensional vector defined by these six variables.

Principal Components Analysis of NPS-interface Conservation Space

As a first exploratory step, PCA (Principal Component Analysis) was applied to visualize the relationships between the interface conservation (IC) scores and the six BLAST alignment statistics. PCA, which is a dimensionality reduction technique, is typically used to represent dimensions that explain maximum variability and provide a simple and parsimonious description of the covariance structure [78].

Figure 1 shows a PCA biplot in which each data point, representing a query-homolog pair, is projected from the original 6-dimensional space to a 2-dimensional space defined by the first and second principal components (PC1 and PC2). A large fraction (88.58%) of the variance is explained by the first two principal components (48.75% + 39.83%). Based on IC scores, the PCA biplot can be subdivided into three regions that correspond to: (i) Dark Zone: containing query-homolog pairs with poorly conserved interface residues (blue and green data points), corresponding to low values of the CC between predicted and actual interfaces and thus low IC scores; (ii) Twilight Zone: containing pairs with moderately conserved interfaces (yellow and orange data points); and (iii) Safe Zone: containing pairs with highly conserved interfaces (red data points).

The PCA analysis allows us to identify highly correlated explanatory variables. In Figure 1, the axes of the original 6 dimensional space are represented as blue vectors with red circles at their tips in the 2-dimensional space defined by PC1 and PC2. Highly correlated vectors (variables) have small angles between them. This type of analysis reveals, for example, that the two Alignment Length Fractions are highly correlated with each other, as are the Positive Score and Identity Score. Explanatory variables that are highly correlated with each other make similar contributions to the IC score.

BLAST EVal is a strong indicator of NPS-interface conservation

We studied the relationship of each individual variable with interface conservation. A scatter plot in which the IC score for each query-homolog pair is plotted against log(EVal) is shown in Figure 2. One can see that log(EVal) is a good indicator of protein interface conservation. When log(EVal) > -50 the median values of IC scores cluster around 0 (low conservation). In the region of log(EVal) ≤-50 (that is, EVal ≤ 1.9287E-022) the medians of IC scores increase as the log (EVal) decreases. When log(EVal) < -100 the medians of IC scores tend to be greater than 0.5 (strong conservation).

NPS-interface conservation in Twilight/Safe Zone is strongly positively correlated with log(LAL)

Figure 3 is a scatter plot showing the IC score for each query-homolog pair plotted against the log of its LAL value. We can clearly see that when log(LAL) is larger than 4, the medians of IC score show a strong positive correlation with log(LAL). When the LAL is shorter than 55 residues (log(LAL) <4), the probability that interface is conserved in these homologs is low (the medians of the IC scores are ~0). We define this region as the Dark Zone. When the LAL is longer than 700 residues (log(LAL)>6.55), interface conservation is high (the medians of IC scores are usually > 0.7). We define this region as the Safe Zone.

A high BLAST Positive Score reflects NPS-interface conservation

The relationship between IC scores and the Positive Scores of query-homolog alignments is shown in Figure 4. The median values of the IC scores begin to increase at a BLAST Positive Score of ~ 90%.

We also studied the relationship of IC score with the Identity Score, and the Local Alignment Length Fractions (LAL/Query Length) and (LAL/Homolog Length). As expected, the Identity Score results were similar to those for the Positive Score. The IC score was not as strongly linearly related to LAL fraction as it was to the log(LAL) (data not shown). Taken together, these results provide guidelines for choosing sequence similarity thresholds that reflect the degree of conservation in NPS interfaces.

NPS-Interface Conservation in Transient versus Obligate Binding Proteins

In light of reports that protein interfaces in transient complexes are not as conserved as those in obligate (permanent) complexes [57], it is interesting to ask whether the query-homolog pairs with near-zero IC scores (Figure 2 and Figure 3) tend to involve proteins that participate in transient interactions. To address this question, we further studied the differences in protein interface conservation among proteins that participate in transient versus obligate interactions.

To compare protein interfaces in transient and obligate complexes, we used the Trans135 and Oblig94 dataset obtained from [76], which includes a total of 270 chains from transient and 188 chains from obligate complexes. We extracted the homologs of each chain from nr_pdbaa_s2c using BLASTP with EVal ≤ 10 Query and homolog proteins with interfaces containing fewer than 3 amino acids were removed, as were homologs that were nearly identical to the query proteins. We extracted 43,115 query-homolog pairs containing chains that participate in transient interactions and 24,212 pairs containing chains that participate in obligate interactions.

In agreement with previous studies [57], our analyses showed that PPIs are conserved in both obligate and transient binding proteins. As before, we performed PCA to examine the conservation of interfaces as a function of log(EVal), Identity Score, Positive Score, log(LAL), and alignment length fractions (LAL/Query Length) and (LAL/Homolog Length). The PCA biplots in Figure 5 show that data points corresponding to different IC scores (different colors) are partially segregated, indicating that the six alignment statistics can distinguish query-homolog pairs with highly conserved interface residues (red) from those in which interface residues are not conserved (blue or green).

The results in Figure 5 also reveal that interface residues in proteins from obligate complexes (left panel) are more conserved among their sequence homologs than those from transient complexes (right panel). Figure 6 further illustrates differences in interface conservation in obligate (left) versus transient complexes (right). The median values of IC scores plotted as a function of log (LAL) are more frequently above 0 for pairs that involve obligate binding proteins (Figure 6a) than for those that involve transient binding proteins (Figure 6b). Regression analysis of these data confirms that log(LAL) for the obligate dataset has a larger coefficient (0.095) than that for the transient dataset (0.052), which confirms that protein interfaces are more conserved in the obligate complexes than in transient complexes analyzed in this study.

Figure 6c reveals an obvious pattern of interface conservation in obligate binding proteins: a strong trend of increasing median IC score with decreasing log(EVal). In contrast, Figure 6d shows that for transient binding proteins, more of the median values of IC scores cluster around 0, indicating that log(EVal) has little relation to interface conservation in transient complexes.

Also, comparison of Figure 6e and 6f reveals that the Positive Score is a good indicator of interface conservation in the case of proteins from obligate complexes; however, this is not the case for proteins from transient complexes. For obligate binding proteins, when the Positive Score exceeds 45%, the medians of IC scores begin to show an increasing trend (Figure 6e). In contrast, in the case of transient binding proteins, medians of IC scores do not begin to increase until the Positive Score approaches 70% (Figure 6f).

It is important to emphasize that all of the interfaces analyzed above are what we refer to as "non partner-specific" (NPS). That is, the interface residues of a query protein represent the complete set of its interface residues with all of its partners. However, a given query protein can interact with different binding partners through different interfaces. A possible explanation for the low IC scores for NPS-transient interfaces is that the union of all interface residues of a transient binding protein are not highly conserved across its homologs. This does not preclude the possibility that such interfaces are conserved in the context of partner-specific interactions. We investigate this possibility in the following section.

Conservation of PPIs in Partner-Specific (PS) Interfaces

To examine the conservation of partner-specific (PS) interfaces in transient protein complexes, we again used the Trans135 dataset of protein pairs that participate in transient interactions [76]. For each of the proteins in an interacting pair, we separately extracted the corresponding homologs, using BLASTP with expectation value EVal≤10 against the nr_pdbaa_s2c database. We removed homologs that are part of complexes with resolution worse than 3.5 Å. If query proteins A and B form a complex A-B, and have homologs A' and B' that interact in a complex A'-B', we consider A'-B' as a homo-interolog of A-B. To ensure an accurate estimate of conservation, from the resulting set of homo-interologs, we eliminated those that were within the same PDB complex as the query proteins, and those that were nearly identical to the query pairs (see Methods for additional details). For each protein chain in a query pair, we use the interface residues of its homolog in a homo-interolog to infer the PS interface residues of the query protein chain. Thus, we use the interface residues of A' in the homo-interolog (A'-B') of query pair A-B to infer the interfaces of A with B, based on the sequence alignment between A and A' obtained using BLASTP. We measure the similarity between a pair of interacting proteins A-B and its homo-interolog A'-B', in terms of the metrics for the quality of sequence alignment between A and A' and between B and B', using the six BLAST alignment statistics described above.

We used PCA of 3, 456 candidate homo-interologs to explore the relationship between interface conservation (IC score) and the six alignment statistics computed from the predicted PS interfaces, e.g., of chain A when it interacts with B, using known interfaces of A' with B'. This analysis revealed that much of the observed variance in IC scores is explained by three factors: (i) the average log (EVal); (ii) the average Positive Score of the homo-interolog and (iii) the alignment fractions Frac _A , Frac _{A '}, Frac _B , and Frac _{B '} computed from the alignments of constituent chains (A with A' and B with B') (see Methods for additional details).

The results in Figure 7 show that transient interfaces are highly conserved in homo-interologs. The trend of increasing median IC scores, as a function of decreasing logEval (Figure 7a) or increasing Positive Score (Figure 7b) or the combination of Positive Score and Frac _A × Frac _{A '} is clear (Figure 7d). The trend of increasing IC scores as a function of Frac _B × Frac _{B '} is similar to that as a function of Frac _A × Frac _{A '} (data not shown). In contrast, the, logLAL, which is the average of alignment length between A and A', and between B and B', is not strongly correlated with interface conservation for PS-interfaces (Figure 7c).

A comparison of the results for PS-interface conservation in transient complexes here (Figure 7a and 7b) with those obtained for NPS-interface conservation in transient complexes above (Figure 6d and 6f), reveals that the conservation of transient interfaces can be detected easily when the binding partner sequence information is utilized. The seemingly weak conservation of interfaces in transient complexes shown in Figure 6 is thus a consequence of the specificity of transient interfaces for different partners. Therefore, we conclude that interfaces in transient complexes are both highly partner-specific and highly conserved, when their partner-specificity is taken into account.

HomPPI - Homologous Sequence-Based Protein-Protein Interface Prediction

Performance Evaluation of HomPPI Methods

We report several performance measures that provide estimates of the reliability of interface (and non-interface) residue predictions obtained using the HomPPI family of predictors. We compare the performance of HomPPI predictors with several state-of-the-art interface prediction methods on a benchmark dataset. We evaluate the effectiveness of HomPPI in predicting the interface residues of disordered proteins. Finally, we compare the partner-specific and non-partner-specific versions of HomPPI.

We focus our discussion on results using several performance measures that assess the effectiveness of the methods in reliably predicting, on average, the interface and non-interface residues of any given protein (See Methods for details). However, because several of the published studies report performance measures that assess the effectiveness of the methods in reliably assigning interface versus non-interface labels, on average, to any given protein residue, we also include results using "residue-based" performance measures in Supplementary Materials (See http://homppi.cs.iastate.edu/supplementaryData.html).

(ii) Comparison of NPS-HomPPI with other PPI Prediction Servers

Direct comparison of NPS-HomPPI with other methods described in the literature is complicated by the limited availability of implementations of the underlying methods (many of which are available only in the form of servers), and differences in the choice of training and evaluation datasets, evaluation procedures and evaluation measures [80]. Hence, we limit our comparisons of HomPPI with five state-of-the-art methods available as web-based servers: Promate [18], Cons-PPISP [33, 81], meta-PPISP [82], PIER [83] and PSIVER[22]. All of these methods except PSIVER take advantage of both sequence and experimentally determined protein structure of the query proteins. They have been reported to be among the best performing methods currently available for predicting PPIs (see [20, 21] for reviews). PSIVER is one of the most recently published methods for interface residue prediction that only uses protein sequence-derived information. Although direct comparisons of the data representation and the algorithms used by PSIVER with those used by other sequence-based interface residue predictors are currently not available, PSIVER has been reported to outperform two other sequence-based servers: ISIS [3] and the sequence-based variant (made available as an experimental version in 2008) of SPPIDER [84].

Promate samples the protein surface using circular patches around a set of anchoring dots and estimates the probability that each surface dot belongs to an interface, based on the distribution of various physicochemical properties within interface and non-interface patches. Cons-PPISP is a consensus method that combines six neural networks trained on six datasets. Meta-PPISP is a consensus method that combines the output from cons-PPISP, Promate, and PINUP [85]. PIER relies on partial least squares (PLS) regression of surface patch properties of the query protein. PSIVER uses PSSM profiles and predicted solvent accessibility as input features, and uses a Naïve Bayes classifier with parameters obtained using kernel density estimation. Because NPS-HomPPI does not take structural information into account, to compare its performance with the structure-based servers, we mapped the interfaces predicted by each server onto the full sequence of each query protein in order to evaluate prediction performance on the entire protein sequence.

We compared the performance of NPS-HomPPI with all five PPI servers on a subset of the Benchmark180 dataset [79], specifically, 125 out of 180 proteins for which NPS-HomPPI was able to identify homologs in the Safe or Twilight zones. The sensitivity-specificity plots (also called precision-recall plots) are shown in Figure 8. Each data point corresponds to a different classification threshold value. The prediction score of NPS-HomPPI is simply the normalized vote (for each residue total votes for interfaces from homologs are normalized by the number of homologs) from 10 (or fewer available) homologs. Thus, NPS-HomPPI produces a limited number of distinct prediction scores.

For the two transient complex types, enzyme-inhibitors (Figure 8a) and transient non-enzyme-inhibitors, transient (Figure 8b), NPS-HomPPI consistently outperforms Promate, PIER, meta-PPISP, cons-PPISP, and PSIVER except for sensitivity values lower than 0.2 (which is very low to be useful in practice). On both obligate heterodimers (Figure 8c) and homodimers (Figure 8d), NPS-HomPPI outperforms all five servers across the full range of sensitivity and specificity values for which it can generate homology-based predictions. It should be noted that structure-based methods predict which surface residues are interface residues. In contrast, sequence-based methods have the more challenging task of identifying interface residues from the set of all residues. In other words, structure-based methods can trivially eliminate all non-surface residues from the set of candidate interface residues. Viewed in this light, the observed predictive performance of NPS-HomPPI, a purely sequence-based method, suggests that it is possible to make reliable non-partner-specific interface residue predictions using only the sequences of a protein by taking advantage of the conservation of interfaces in the context of non-partner-specific interactions.

(iii) Performance of NPS-HomPPI on Intrinsically Disordered Proteins

Intrinsically disordered proteins (IDPs) and proteins containing intrinsically disordered regions (IDRs) are attractive targets for drug discovery [73]. The lack of defined tertiary structure in IDPs/IDRs poses a major challenge to structure-based interface prediction methods. Hence, we compared the performance of NPS-HomPPI with ANCHOR [86], a recently published method for the prediction of binding regions in disordered proteins. For this comparison, we used two non-redundant disordered protein datasets, S1 and S2, recently collected by Meszaros et al. [87]. Some of the test proteins are based on data from NMR structures. In order to compare NPS-HomPPI with ANCHOR on the largest possible number of cases available to us, we extracted interface residues from these NMR cases; however, we used only sequence homologs with interface residues determined from X-ray structures to make predictions.

Figure 9 shows the performance comparison of NPS-HomPPI with ANCHOR on the prediction of interface residues in disordered proteins. NPS-HomPPI significantly outperforms ANCHOR over a broad range of sensitivity and specificity for both short as well as long disordered proteins for which sequence homologs are available in Safe, Twilight or Dark Zones (Figure 9a and 9b respectively). For example, as shown in Figure 9b, on the S2 dataset, at a prediction sensitivity value of 0.70, ANCHOR achieves a specificity of ~0.40, whereas NPS-HomPPI achieves a specificity of ~0.64.

At present, NPS-HomPPI has relatively high prediction coverage for long disordered proteins (78%; 31 out of 40 interfaces of disordered proteins), but lower coverage for short disordered proteins (50%; 28 out of 56 interfaces of disordered proteins). This is in part due to that fact that many disordered proteins available in the PDB have only NMR structures, which were excluded from the current study. Incorporation of data from NMR structures in the future can be expected to increase the coverage of NPS-HomPPI for disordered proteins.

(iv) Performance of NPS-HomPPI versus PS-HomPPI

Our analysis of the conservation of PS-transient interfaces described earlier suggests that many interfaces in transient protein complexes are highly partner-specific. Thus, we implemented a variant of HomPPI, designated PS-HomPPI, to evaluate the possibility that prediction of interface residues, especially in transient complexes, can be improved by using sequence information about specific binding partners, when available.

We first evaluated the performance of PS-HomPPI on a transient complex dataset, Trans135 (dimers from the dataset in [76]). PS-HomPPI found at least one homo-interolog that meets the Safe or Twilight similarity thresholds for 60% (162/270) proteins in the Trans135 dataset. Overall, PS-HomPPI had an average CC of 0.65, sensitivity of 0.69, specificity of 0.70 and accuracy of 0.92.

To investigate whether the partner information is, in fact, helpful in predicting interfaces we directly compared the performance of PS-HomPPI with NPS-HomPPI on the Trans135 dataset. In Trans135, there were 139 out of 270 chains that for which predictions could be generated by both NPS-HomPPI (using homologs) and PS-HomPPI (using homo-interologs) from the Safe or Twilight zones (see Methods for details).

The results shown in Figure 10 indicate that, at least for transient interfaces in the Trans135 dataset, PS-HomPPI outperforms NPS-HomPPI. Although the average values over proteins (green dots) for CC, sensitivity and specificity are similar, the median values (the red bar in the box) for PS predictions (left panel) are much higher than that for NPS predictions (right panel). Also, the observed variance (length of the box) of PS predictions (left panel) is much smaller than that of NPS predictions (right panel). These results suggest that the reliability of interface residue predictions can be improved by exploiting the knowledge of the binding partner of a query protein.

Protein Interface Conservation across Structure Space

The study of protein interface conservation among proteins with similar structures has received considerable attention in recent years. By analyzing the structural similarity of representative protein-protein interfaces in dimeric proteins, Gao and Skolnick [88] showed that the vast majority of native interfaces have a close structural neighbor with similar backbone Cα geometry and interface contact pattern.

In a related study, Zhang et al. [89] explored the conservation of interface residues among structural neighbors of a query protein (i.e., proteins that share the same SCOP family, superfamily or fold, or a high degree of structural similarity regardless of their SCOP classification). They showed that: (i) interfaces are indeed conserved among structural neighbors; (ii) the degree of interface conservation is most significant among proteins that have a clear evolutionary relationship. They further showed that conservation of interface residues among structural neighbors can be successfully exploited to predict protein-protein interfaces based on protein structure information.

To investigate the extent to which conservation of interface residues can be used to improve the prediction of protein-protein interfaces based on protein sequence information, we systematically studied interface conservation across sequence space. Our results demonstrate that protein interfaces from different binding types are conserved among proteins with homologous sequences. We further showed that the degree of conservation of interfaces is even greater when putative interaction partners are taken into account. The IC score, our measure of interface conservation, unlike those used in previous studies [57] (e.g., residue conservation in sequence alignments), makes direct use of experimentally determined interface residues to measure the degree of interface conservation. Specifically, the IC score directly measures the extent to which the interface residues of sequence homologs of a query protein are predictive of the interface residues of a query protein. Hence, the IC score provides the basis for setting the parameters of our sequence homology-based interface prediction methods.

Distance Functions for Identifying Putative Homologs with Conserved Interfaces

Because we do not know the IC score for a query sequence with unknown interface residues, we identified several statistics associated with the BLASTP alignment of a query sequence with its homologs that are correlated with the IC score. We found that interface residues of a query protein can be reliably predicted from the known interfaces of its homologs (and in the case of partner-specific predictions, the homologs of its interaction partner as well) when the homologs are selected taking into account measures of quality of sequence alignment, specifically NCBI BLAST sequence alignment statistics. The HomPPI methods presented here use simple linear combinations of BLAST sequence alignment statistics, determined using PCA analysis of the relationship between the statistics and the IC score. It would be interesting to explore optimal, perhaps non-linear, combinations of parameters to maximize the desired performance criteria (e.g., sensitivity, specificity, or some combination thereof).

Conservation of Interfaces in Obligate and Transient Complexes

Our results are consistent with previous studies [57], in that we found interface residues to be more highly conserved than non-interface residues, in both obligate and transient complexes. We also found that when information regarding the specific binding partner of a query protein is not taken into account in estimating the conservation score, interfaces in transient complexes appear to be less highly conserved than those in obligate complexes. Our results further show that transient interfaces are highly partner-specific, and that the partner-specific interfaces in transient complexes are, in fact, highly conserved. Interfaces of intrinsically disordered proteins that nevertheless form ordered complexes with globular proteins are also highly conserved (see below).

Interfaces of Disordered Proteins Are Highly Conserved and Non Partner-Specific

Compared with its performance on transient binding proteins in the Benchmark180 dataset, NPS-HomPPI performs much better on interfaces of disordered proteins in the S1 and S2 datasets. This is consistent with the conclusion of Meszaros et al. [90] that interfaces of intrinsically disordered proteins are evolutionarily conserved. The high degree of conservation of interface (binding) regions in IDPs also reflects the important biological functions in which many disordered proteins participate. It is believed that the flexibility of disordered binding regions may facilitate the binding of IDPs using the same set of binding residues to different binding partners (at different times) [91]. Our results suggest that this specialized disorder-to-order transition as a result of binding may be associated with a high degree of interface conservation.

The conservation of interfaces in IDPs may contribute to the generally successful application of interface residue predictors to interfaces in IDPs. Several groups have developed methods for predicting disordered binding regions, including PONDR VL-XT [92, 93], ANCHOR [86], and other examples reviewed in [94], that have produced encouraging results. The success of these predictors suggests that at least some sequence features are likely to be conserved within binding regions of different IDPs.

The fact that disordered interfaces can be reliably inferred by NPS-HomPPI indicates that disordered interfaces are non-partner-specific, which is consistent with findings that these proteins are able to bind a broad range of ligands through common binding regions [95, 96].

Performance of HomPPI Compared with Published Methods

Our results show that whenever the interfaces of the close sequence homologs of a query protein are available, NPS-HomPPI outperforms several state-of-the-art protein interface prediction servers (many of which take advantage of the structure of the query protein), over a broad range of sensitivity and specificity values. In the case of transient complexes (Figure 8a and 8b), NPS-HomPPI consistently outperforms Promate, PIER, meta-PPISP, cons-PPISP, and PSIVER except for sensitivity values lower than 0.2. On obligate dimers (Figure 8c and 8d), NPS-HomPPI significantly outperforms all five servers across the full range of sensitivity and specificity values for which it can generate homology-based predictions. These results strongly suggest that it is possible to reliably predict protein interface residues using only sequence information whenever the interface residues of sequence homologs of the query protein are known. Each of the webbased PPI servers with which we compared our NPS-HomPPI server, except PSIVER, take advantage of the structure of the query proteins to determine surface residues, and restrict the predicted interface residues to a subset of the surface residues. This trivially reduces the number of false positive interface residue predictions (relative to the total number of residues in the query protein) which, in turn, yields a substantial increase in the specificity of interface predictions produced by structure-based servers. Consequently, purely sequence-based protein interface prediction servers have a handicap relative to structure-based prediction servers. When viewed in this light, performance of NPS-HomPPI relative to the state-of-the-art protein interface prediction methods is especially impressive.

The HomPPI methods for interface residue prediction do have an important limitation, however, in that they rely on the availability putative homologs for which experimentally-determined structures of bound complexes are available in the PDB. One may ask whether the coverage of the HomPPI family of protein-protein interface prediction methods is broad enough to be sufficiently useful in practice. We address this question below.

Prediction Coverage of HomPPI Methods

The current coverage of HomPPI protein interface prediction methods can be assessed from our results as follows:

Parameters for HomPPI Can Be Relaxed for Obligate Interactions

The current default parameters for HomPPI are intentionally rather stringently set based on the results of our statistical analysis of interface conservation using Trans135, which is a dataset of transient binding proteins. Our analyses suggest that NPS-HomPPI has wider Safe and Twilight Zones for obligate binding proteins than for transient binding proteins. Furthermore, even Dark Zone homologs yield interface predictions that are accurate enough to be useful in practice, with average specificity of 0.47 and sensitivity of 0.66 for hetero-obligate dimers, average specificity of 0.44 and sensitivity of 0.47 for homo-obligate dimers (see Table 3). Therefore, for obligate interactions, if a query protein has little sequence similarity with proteins in the PDB, the thresholds of NPS-HomPPI can be relaxed to allow identification of more distant homologs with potentially conserved interfaces that still provide reliable interface predictions.

Prediction of Binding Partners vs. Prediction of Interface Residues

Protein interface (binding site) predictions and protein interaction (partner) predictions answer closely related, but different questions. Non partner-specific protein interface predictors are designed to identify the residues in a query protein that are likely to make contact with the residues of one or more unspecified interaction partner proteins. Partner-specific protein interface predictors are designed to identify the residues in a query protein that are likely to make contact with residues of a putative interaction partner protein. In contrast, protein interaction predictors are designed to predict whether or not a given pair of proteins is likely to interact [98–101]. Although our study does not directly address the latter question, it is possible to use PS-HomPPI predictions to determine whether or not two query proteins interact: Given a pair of protein sequences, say A and B, we can first use PS-HomPPI to predict the interface residues of A with its putative partner B; and the interface residues of B with its putative partner A. If, in both cases, some number of interface residues are predicted, we can infer that proteins A and B are likely to interact with each other. Conversely, it is possible to use information from predicted protein-protein interactions to refine interface predictions. Yip et al. [2] have proposed an approach to utilize residue level information to improve the accuracy of protein level predictions, and vice versa. They have shown that a two-level machine learning framework that allows information flow between the two levels through shared features yields predictions that are more accurate than those obtained independently at each of the levels.

Using Interface Predictions to Steer Docking and to Rank Docked Conformations

Reliable partner-specific interface predictions can be used to restrict the search space for protein-protein docking by specifying the contacts that need to be preserved in the docked conformation. It is also possible to rank the conformations produced by docking, based on the degree of overlap between the interface of a query protein and its binding partner in the docked conformation with the interface generated by a partner-specific interface prediction method, e.g. PS-HomPPI. In related work[102], we have shown that PS-HomPPI provides reliable interface predictions on a large subset of a Docking Benchmark Dataset, and is both fast and robust in the face of conformational changes induced by complex formation. The quality of the ranking of docked conformations by PS-HomPPI interface prediction is consistently superior to that produced using ClusPro cluster-size-based and energy-based criteria for 61 out of 64 docking complexes for which PS-HomPPI produces interface predictions[102].

Datasets

Five datasets were used in this paper:

Interface Definition

Mapping Interfaces in Structures to Sequences

We label the protein sequences as interface or non-interface residues (according to the definition of interface residues given above) as follows: We first calculate the relevant distances between atoms using the atom coordinates in ATOM section in PDB files. Then, by associating the ATOM section to residues in the SEQRES section, we can map the corresponding residues to protein sequences. However, various errors in PDB files make this a non-trivial task. Hence, we used the mapping files from S2C database, which offers corrected mapping information from ATOM section to residues in the SEQRES section of PDB files, to map interfaces determined in structures to full sequences.

NCBI BLAST Parameters

Performance Evaluation

To evaluate the extent to which protein interfaces are conserved in query-homolog pairs and to estimate the performance of HomPPI and other predictors that we compare with in predicting the interface residues of a novel protein (i.e., one not used to train the predictor), we consider several standard performance measures including sensitivity (recall), specificity (precision), accuracy and Matthews correlation coefficient (CC) [106]. Specifically, for each test protein i, we calculate the corresponding performance measures for each protein i as follows:

where TP _i , FP _i , TN _i and FN _i are respectively the number of interface residues of protein i that are correctly predicted to be interface residues, the number of residues of protein i that are incorrectly predicted to be interface residues, the number of residues of protein i that are correctly predicted to be non-interface residues, and the number of residues of protein i that are incorrectly predicted to be non-interface residues.

We calculate the protein-based overall performance measures as follows:

where N is the total number of test proteins.

These measures describe different aspects of predictor performance. The overall sensitivity is the probability, on average, of correctly predicting the interface residues of a given protein. The overall specificity is the probability, on average, that a predicted interface residue in any given protein is in fact an interface residue. The overall accuracy corresponds to the fraction of residues in any given protein, on average, that are correctly predicted. The overall Matthews correlation coefficient measures of how predictions correlate, on average, with true interfaces and non-interfaces.

Often it is possible to trade off one performance measure (e.g., specificity) against another (e.g., sensitivity) by varying the threshold that is applied to the prediction score to generate the binary (interface versus non-interface) predictions. Hence, we include of the overall sensitivity against overall specificity for different choices of the threshold. The resulting specificity-sensitivity plots or precision-recall plots show the trade-off between sensitivity and specificity and hence provide a much more complete picture of predictive performance.

The performance measures described above provide an estimate of the reliability of the predictor in predicting interface residues of a novel protein. It is worth noting that most of the papers in the literature on interface residue prediction report performance measures by averaging over residues (as opposed to proteins). The residue-based overall performance measures are calculated as follows:

Residue-based specificity-sensitivity plots in this case show how the trade-off between specificity^R and specificity^R is obtained by varying the threshold applied to the prediction score. The residue-based performance measures provide an estimate of the reliability of the predictor in correctly labelling a given residue. However, in practice, it is useful to know how well a predictor can be expected to perform on a given protein sequence as opposed to a residue. sensitivity^P, specificity^P, accuracy^R, and CC^P are more informative than their residue-based counterparts. Hence, in this paper, we report results based on the protein-based measures although, for the purpose of comparison with other published methods, we include the results based on the residue-based measures in Supplementary Materials in HomPPI website.

Interface Conservation (IC) Scores

In protein interface conservation analysis, we used the CC (defined above) as a measure of the extent to which the interface residues in query protein are similar to those in a putative homolog. For clarity, we refer this measure as the Interface Conservation (IC) score.

NPS-HomPPI

NPS-HomPPI is a Non-Partner-Specific Homologous Sequence-Based Protein-Protein Interface Prediction algorithm. NPS-HomPPI is based on the conclusion from statistical analysis of protein interface conservation on Nr6505, Trans135 and Oblig94, i.e., that protein interfaces are conserved across close sequence homologs.

As illustrated in Figure 11, NPS-HomPPI predicts interface residues in a query protein based on the known interface residues of a selected subset of homologs in a sequence alignment. Homologs of the query protein sequence are identified by searching the nr_pdb_s2c database using BLASTP. Note that, in our experiments, in order to allow unbiased evaluation of the performance of NPS-HomPPI, the query sequence itself and sequences that share a high degree (≥95%) of amino acid sequence identity with, and are from the same species as the query sequence are deleted from the set of putative homologs.

If at least one homolog in the Safe Zone is found by the BLASTP search, NPS-HomPPI uses the Safe Zone homolog(s) to infer the interfaces of the query protein. Otherwise, the search is repeated for homologs in the Twilight and Dark Zones. If NPS-HomPPI cannot find homologs in any of the three zones, it does not provide any predictions. The default zone boundaries used by NPS-HomPPI (and hence the parameters used in NPS-HomPPI search for homologs of a query sequence) is based on our interface conservation analysis on the dataset of transient dimers Trans135 (Table 1). The choice of these default parameter thresholds for NPS-HomPPI is intentionally rather conservative; the thresholds can be relaxed if additional information is available (e.g., if we know that the query protein is an obligate binding protein). The IC score of each of the homologs of a query sequence in the alignment returned by BLASTP is predicted using the regression model for the IC score (see eq. 1) from the BLASTP statistics for the alignment of each homolog with the query sequence. For a given query sequence, at most K closest (Safe, Twilight, or Dark Zone homologs, as the case may be, in that order) are selected from the alignment of the query sequence with its homologs to be used to infer the interface residues of the query sequence. In our experiments, K, the maximum number of homologs used in the prediction was set equal to 10. At most K homologs of the query sequence are determined by ranking the homologs in the alignment in decreasing order of their predicted IC scores and choosing (at most) K Safe zone homologs (or Twilight zone homologs if no Safe zone homologs exist or Dark zone homologs if neither Safe nor Twilight zone homologs exist). Once the (at most) K closest homologs to be used for predicting the interface residues of the query sequence are chosen, each residue in the query sequence is labelled as an interface or non-interface residue based on the majority (over the set of at most K closest homologs of the query sequence) of the labels associated with the corresponding position in the alignment. More specifically, each of the at most K homologs provides a positive vote for a given position in the query sequence if the corresponding residue of the homolog is an interface residue; and a negative vote if it is a non-interface residue. The prediction score of NPS-HomPPI for that position in the query sequence is simply the number of positive votes divided by the total number of votes. A query sequence residue with a HomPPI score ≥0.5 is predicted to be an interface residue (See Figure 11 for an example); otherwise, it is predicted to be a non-interface residue. This procedure can be seen as an application of the (at most) K nearest neighbor classifier at each residue of the query sequence.

NPS-Interface Conservation As a Function of Sequence Alignment

We built a linear model for NPS-interface conservation based on the most important sequence alignment statistics identified in the PCA analysis: logEVal, Positive Score, logLAL.

The model is

(1)

Variables, parameter estimates and coefficients are shown in Table 6. All the coefficients are significant.

Variable	Parameter Estimate	Standard Error	t Value	Pr > |t|
β 0	-0.5655	0.0080	-70.66	<.0001
β 1	-0.0004	0.0000	-23.30	<.0001
β 2	0.0037	0.0000	54.44	<.0001
β 3	0.1057	0.0011	94.62	<.0001

PS-HomPPI

PS-HomPPI predicts the interface residues in a protein chain based on the known interface residues of its closest homo-interologs. Given a query protein A and its interaction partner B, PS-HomPPI first identifies the set homo-interologs of A-B using BLASTP to identify the homologs of A and homologs of B. From the BLASTP results, we identify a set of homo-interologs that meet sequence similarity thresholds (determined based on the results of our partner-specific interface conservation analysis, as described in the Results Section). We discard the whole PDB complex that contains A-B, to ensure an objective assessment of the reliability of our prediction procedure. For query A-B and its homologous interacting pair A'-B', we also discard the interacting protein pair A'-B' if A and A' or B and B' share ≥95% sequence identity and belong to the same species.

PS-HomPPI predicts whether an amino acid in query sequence A is an interface residue or not based on the corresponding position in its alignment with (at most) K of the closest homo-interologs of A-B (based on their predicted IC scores). In our experiments, K was set equal to 10. Given a query-partner pair A-B, we label each position in the amino acid sequence of protein A as an interface or non-interface based on whether or not a majority of the corresponding positions of the homologs of A within the homo-interologs of A-B are interface residues. More specifically, each of the at most K homo-interologs provides a positive vote for a given position in the query protein sequence A if the corresponding residue of its homolog A' in its homo-interolog is an interface residue; and a negative vote if it is a non-interface residue. The prediction score of PS-HomPPI for that position in the query sequence is simply the number of positive votes divided by the total number of votes. A residue in the query protein A with a prediction score ≥0.5, is predicted as interface, otherwise, it is predicted as non-interface.

PS-Interface Conservation As a Function of Sequence Alignment

We built a linear model for PS-interface conservation based on the important sequence alignment statistics identified in the PCA analysis: logEVal, Positive Score, Frac _{AA '} and Frac _{BB '}, where

A-B is query protein pair and A'-B' is the homo-interolog of A-B. EVal _{AA '} and EVal _{BB '} are the EVal between A and A', and between B and B'. positiveS _{AA '} and positiveS _{BB '} are the BLAST Positive Score between A and A', between B and B'. The model is

(2)

Variables, parameter estimates and coefficients are shown in Table 8. All the coefficients are significant.

Coefficient	Parameter Estimate	Standard Error	t Value	Pr > |t|
β 0	-0.505	0.040	-12.62	<.0001
β 1	0.001	0.000	6.16	<.0001
β 2	0.009	0.001	14.6	<.0001
β 3	0.341	0.027	12.54	<.0001
β 4	0.205	0.028	7.4	<.0001