A knowledge-based potential function predicts the specificity and relative binding energy of RNA-binding proteins Suxin Zheng 1, *, Timothy A. Robertson 2, * and Gabriele Varani 1,2 1 Department of Chemistry, University of Washington, Seattle, WA, USA 2 Department of Biochemistry, University of Washington, Seattle, WA, USA The sequence-specific recognition of RNA by proteins plays a fundamental role in gene expression by direct- ing different cellular RNAs to specific processing path- ways or subcellular locations. Many experimental studies have explored the molecular basis for the sequence dependence of protein–RNA recognition [1– 4]; more recently, a few studies have explored this prob- lem from a computational perspective as well [5–16]. However, these early studies have emphasized qualita- tive descriptions of the recognition process; relatively few attempts have been made to quantify the character- istics of protein–RNA interactions using computational approaches [17]. Here, we present a new approach for predicting the specificity of RNA-binding proteins and to evaluate the contribution of individual amino acids to the energetic of protein–RNA complexes. Knowledge-based potential functions have been employed in protein structure prediction [18–27], as well as in the prediction of protein–protein [25,28–30] and protein–ligand interactions [30–33]. A few studies have explored the use of knowledge-based methods for the prediction of protein–DNA interactions from structure [30,34,35]. More recently, our group [36] and others [37] have independently demonstrated that knowledge-based potentials can provide quantitative descriptions of protein–DNA interfaces comparable to those provided using molecular mechanics force fields [37]. The relative scarcity of high-resolution structures of protein–RNA complexes has represented an under- standable barrier to the quantitative application of computational approaches to the problem of protein– RNA recognition. However, we have previously dem- onstrated that a statistical hydrogen bonding potential can discriminate native structures of protein–RNA complexes from docking decoy sets [17]. As hydrogen Keywords distance-dependent potential; protein–RNA interaction; RRM recognition; statistical potential Correspondence G. Varani, Department of Chemistry and Department of Biochemistry, University of Washington, Seattle, WA 98195, USA Fax: +1 206 685 8665 Tel: +1 206 543 7113 E-mail: varani@chem.washington.edu *These authors contributed equally to this work (Received 25 July 2007, revised 22 Septem- ber 2007, accepted 19 October 2007) doi:10.1111/j.1742-4658.2007.06155.x RNA–protein interactions are fundamental to gene expression. Thus, the molecular basis for the sequence dependence of protein–RNA recognition has been extensively studied experimentally. However, there have been very few computational studies of this problem, and no sustained attempt has been made towards using computational methods to predict or alter the sequence-specificity of these proteins. In the present study, we provide a distance-dependent statistical potential function derived from our previous work on protein–DNA interactions. This potential function discriminates native structures from decoys, successfully predicts the native sequences recognized by sequence-specific RNA-binding proteins, and recapitulates experimentally determined relative changes in binding energy due to muta- tions of individual amino acids at protein–RNA interfaces. Thus, this work demonstrates that statistical models allow the quantitative analysis of protein–RNA recognition based on their structure and can be applied to modeling protein–RNA interfaces for prediction and design purposes. Abbreviations KH, K homology; MD, molecular dynamics; PDB, Protein Data Bank; RRM, RNA recognition motif; SRP, signal recognition particle. 6378 FEBS Journal 274 (2007) 6378–6391 ª 2007 The Authors Journal compilation ª 2007 FEBS bonds represent only approximately 25% of contacts between protein and RNA [12], we reasoned that a more comprehensive approach would describe these interactions more effectively. In the present study, we report the application of an all-atom, distance-dependent statistical potential to the prediction of sequence-specific recognition between proteins and RNA. We demonstrate that this approach can discriminate native structures of complexes from even close docking decoys, recapitulate experimentally determined relative binding energies (DDGs) for several protein–RNA complexes, and predict the RNA sequences recognized by a number of different RNA recognition motif (RRM) and K homology (KH) domains. These results demonstrate that statistical models can be applied to problems requiring the high- resolution modeling of protein–RNA interactions. The anticipated future enrichment of the structural data- base will further improve the predictive performance of the potential. Results The all-atom distance potential is constructed from the distribution of interatomic distances observed in the high resolution (< 2.5 A ˚ ) structures of protein–RNA complexes deposited in the Protein Data Bank (PDB). In this approach, the ‘correctness’ of a protein–RNA structure is assumed to be approximated by the sum of the probabilities of observing the set intermolecular distances defined in the 3D structure, relative to the likelihood of encountering such distance in the dataset of all protein–RNA structures. This kind of method was proposed by Sipple [20], and has been applied to protein structure prediction, protein–protein and pro- tein–ligand interactions [18–33], as well as to protein– DNA recognition [30,34–37]. The distance-dependent statistical potential used here for protein–RNA inter- faces is essentially identical to the score recently described by us for protein–DNA complexes [36]. The primary difference is the introduction of a new pseud- count correction, where an optimized number of pseudocounts are added to the observed counts for each atom pair (for additional details, see Experimen- tal procedures). As a control, we also tested a simple contact-counting method, wherein every contact between protein and RNA (within a given distance cut-off) was assigned the same score of )1. Docking decoy discrimination An important property of any potential function is its ability to discriminate cognate (native crystallographi- cally determined structures) from noncognate (decoy) structures [38]. As a preliminary test of our method, and a direct comparison with previous work, we used our distance-dependent potential to evaluate five sets of docking decoys generated for the application of the rosetta physical potential function to protein– RNA interactions [17]. These decoys were created using a combination of rigid-body docking and pro- tein side-chain repacking, and range in rmsd (relative to the native structure) from 0.2 A ˚ to over 20 A ˚ . Thus, they represent a solid basis for comparison to a much more complex scoring method (the multiterm, hybrid physical ⁄ statistical potential function used by rosetta). When scored with the distance-dependent potential, the native complex can always be identified as the best structure in each of the five decoy sets (Fig. 1), even for decoys that are very close to the native structure. The native structure Z-scores for these decoy sets are shown in Table 1. These values indicate a strong dis- criminatory ability, comparable to that reported by Chen et al. [17] using their significantly more complex scoring method. Overall, the distance potential (using a6A ˚ cut-off) results in a mean native Z-score of )5.45, versus the value of )6.37 obtained by Chen et al. [17] (Table 1); this difference is statistically insig- nificant (P ¼ 0.53, Welch’s two-sided t-test), indicating that the two methods are equivalent. When we investigated protein–DNA complexes using the same approach, we demonstrated that the all-atom potential outperformed a reduced atom description, where relevant groups were grouped according to their chemical similarity (as described in the Experimental procedures) [36]. Given the relative sparsity of the structural database, we investigated whether a reduced-atom representation would not lead to improved performance in the protein–RNA case. The all-atom potential performs better than the reduced atom potential (mean Z-score )5.45 versus )4.66; see also supplementary Table S4), although the difference is not as striking as for protein–DNA com- plexes. We believe this is due to less favorable statistics (fewer structures of protein–RNA complexes). We anticipate that the increasing availability of protein– RNA structures, together with the availability of data on specificity, will further improve the performance of the knowledge-based predictive method presented here. We retained the all atom representation because it is already slightly better than the reduced atom approach. The protein–RNA score has distinctive properties compared to the protein–DNA potential. When we scored the protein–RNA decoy set using the protein– S. Zheng et al. A knowledge-based potential function FEBS Journal 274 (2007) 6378–6391 ª 2007 The Authors Journal compilation ª 2007 FEBS 6379 DNA potential, the average Z-score was approxi- mately half that obtained with the protein–RNA potential ()2.84 versus )5.45; see also supplementary Table S4). Thus, although the chemistry of RNA and DNA are very similar, the structure of RNA allows for different interactions between proteins and the two nucleic acids that are reflected in this result. To investigate whether the statistical potential is not simply reflecting the size of an interface or the number of intermolecular contacts, we also used a very simple contact-counting potential to evaluate the same decoys; in this method, the fitness of an interface is evaluated by counting the number of close approaches between the protein and RNA. Satisfactorily, this method was A B C E D Fig. 1. Score–rmsd plots for the five docking decoy sets generated by Chen et al. [17]; the score generated by the distance-dependent potential (in arbitrary units) is plotted versus the deviation from the native structure (open circle at rmsd ¼ 0). (A) Poly A-binding protein in complex with polyadenylate RNA (PDB code: 1CVJ). (B) Nova-2 KH RNA-binding domain 3 (PDB code: 1EC6). (C) HuD protein in complex with AU-rich RNA (PDB code: 1FXL). (D) Human SRP19 in complex with human SRP RNA (PDB code: 1JID). (E) Human U1A protein in com- plex with U1 snRNA hairpin (PDB code: 1URN). Close-up views of near-native decoys (0–3 A ˚ rmsd) are shown in the insets. A knowledge-based potential function S. Zheng et al. 6380 FEBS Journal 274 (2007) 6378–6391 ª 2007 The Authors Journal compilation ª 2007 FEBS much less effective, providing an average Z-score of )2.64, less than half of the average native Z-score found using the distance potential (Table 1). Interestingly, the magnitude of the observed Z-scores declines significantly as the contact cut-off is increased from 6 A ˚ to 10 A ˚ and then to 12 A ˚ (see sup- plementary Table S5), suggesting that short-range con- tacts provide the bulk of the discriminatory power in this test. This result suggests that protein–RNA recog- nition specificity is primarily determined by short- range intermolecular contacts. Long-range effects (e.g. nonlocal electrostatics) appear to play a more limited role, at least in decoy discrimination. To test the discrimination ability of the potential for near native decoys, we next compared its ability to discriminate near-native protein–RNA structures with that of the force field implemented in the amber 8 molecular simulation package. We generated near- native protein–RNA decoys for 21 protein–RNA complexes by conducting molecular dynamics (MD) simulations of the native complexes, and by selecting multiple time-steps from the resulting trajectories for each structure. We then scored these structures using the distance-dependent potential function, and exam- ined the correlations between distance scores and amber energies for each decoy set. This is a difficult test of score performance because the structures are very close to native. Indeed, neither the distance-dependent score, nor the amber potential appears to be able to discriminate native structures from these very near-native, MD-generated decoys (average Z-score of )0.69 versus )0.59; Table 2). Although there is no correlation of the either score with rmsd, the distance-dependent statistical potential is somewhat correlated (average R 2 ¼ 0.41) with the energy values predicted by the amber force field. Thus, it remains very difficult for either approach to discrim- inate the native structure from structures that are close to it in energy. Identifying RNA-binding sequences from structure Having established the performance of the statistical potential function in decoy discrimination, we investi- gated the ability of the potential to perform tasks rele- vant to its intended application. First, we sought to evaluate whether the potential could predict the cog- nate recognition sequences of RNA-binding proteins. This is a particularly important problem because sequence specificity is known for only a fraction of all RNA-binding proteins. The ability to predict (or at least narrow down) the cognate sequence for ‘orphan’ RNA-binding proteins would greatly facilitate the design of biological experiments aimed at dissecting the function of these proteins. It is also a problem that is not well suited for MD approaches because of the demanding computational requirements. This application relies on a specific structural model of RNA recognition by RRM and KH Table 1. Native Z-scores and score–rmsd correlation coefficients for the protein–RNA docking decoy sets prepared by Chen et al. [17] . Z-scores Distance- dependent a Coulomb b ROSETTA + HB c Contact count a 1CVJ )7.02 )1.19 )5.11 )2.44 1EC6 )6.46 )1.09 )6.53 )3.00 1FXL )2.66 )1.55 )2.70 )1.26 1JID )6.29 )1.36 )9.12 )3.09 1URN )4.80 )1.35 )8.39 )3.39 Mean ± SD )5.45 ± 1.76 )1.31 ± 0.18 )6.37 ± 2.58 )2.64 ± 0.84 a Using a 6 A ˚ contact cut-off. b From Chen et al. [17] and referring to a potential lacking the directional component of hydrogen bonding (HB) interactions. c From Chen et al. [17] and referring to the com- plete potential function. Table 2. Z-scores and correlations for near-native decoys generated by MD simulation. Largest rmsd (A ˚ ) Z-scores Distance- dependent versus AMBER (R 2 ) Distance- dependent AMBER 1B7F 2.33 )3.51 )3.01 0.37 1CVJ 2.63 )1.13 )0.44 0.51 1DFU 2.45 )1.82 )1.93 0.39 1E7K 2.58 )0.38 0.83 0.34 1EC6 5.21 )3.02 )2.29 0.71 1FJE 3.54 1.49 2.51 0.42 1FXL 2.22 )1.32 )1.34 0.47 1JBS 2.46 )0.31 0.60 0.38 1JID 2.92 0.26 1.05 0.54 1K1G 3.48 0.52 1.95 0.52 1KNZ 2.44 )1.39 )1.69 0.24 1M8W 2.47 )1.10 )1.21 0.44 1R9F 2.55 )1.24 )0.19 0.11 1RKJ 3.85 0.92 3.00 0.19 1URN 3.06 )1.08 )2.88 0.46 2AD9 3.94 )0.68 )2.09 0.42 2ADB 2.66 0.37 )2.15 0.16 2ADC 3.01 )0.21 )0.65 0.63 2ASB 2.12 )0.94 )2.07 0.48 2ATW 2.19 )1.38 )2.94 0.33 2CJK 2.75 1.42 2.47 0.42 Mean ± SD )0.69 ± 1.28 )0.59 ± 1.94 0.41 ± 0.15 S. Zheng et al. A knowledge-based potential function FEBS Journal 274 (2007) 6378–6391 ª 2007 The Authors Journal compilation ª 2007 FEBS 6381 domains involving four nucleotides, as detailed in the Experimental procedures. This model is strongly sup- ported by previous research on the mechanism of RNA recognition for RRM proteins [6,39,40] and by the structure of existing KH domains bound to RNA [6,41–44]. As a consequence of the assumptions of the model, complexes containing two RNA-binding domains were divided into independent structures (e.g. 1CVJ_1 and 1CVJ_2 represent the first and sec- ond Poly A binding protein domain of structure 1CVJ, respectively), and the two domains were con- sidered structurally and thermodynamically unrelated. Because the model assumes that each RRM and KH domain binds to each of four nucleotides indepen- dently, we generated a set of 4 4 (256) different structures for each protein–RNA complex by compu- tationally ‘threading’ all possible four-nucleotide com- binations onto the RNA bases nearest the center of the b-sheet structure of the RRM. We then scored these sequence-variant structures with the distance- dependent potential function. Figure 2 shows the results of this analysis. If the potential and model of recognition were perfect, and if each structure was sequence-specific and corresponded to the most favorable sequence recognized by a given domain, the cognate sequences of the tested structures would be expected to rank as number 1. Because it is unlikely that the cognate recognition sequences for all domains will be consistently assigned the best score, we expressed sequence-discrimination performance in terms of percentiles (where perfect discrimination of the cognate recognition sequence would result in a percentile score of 100). Remarkably, we found that 18 of the 29 tested RRM and KH domain complexes had their cognate recognition sequence ranked above the 90th percentile (i.e. had better than ten-fold enrich- ment for the correct sequence). Furthermore, the distance-dependent potential ranks the cognate recog- nition sequences of the protein–RNA complexes in our test set above the 90th percentile, on average. By con- trast, when we performed the same test using a simple counting potential as a control (Fig. 2), the average rank was the 41st percentile. Among successful examples of binding-sequence discrimination, the native sequences of the RRM1 of Sex-lethal protein (1B7F_1) and KH1 domain of Poly C-binding protein-2 were both ranked first out of 256 sequences, whereas KH domain 3 of hnRNP K (1ZZI), RRM of U2B¢ protein (1A9N) and RRM 4 of Polypyrimidine Tract Binding protein (2ADC_1) each had their cognate recognition sequences ranked in the top 3 (Supplemental Table S2). However, prediction was less successful for other RRM domains, such as the U1A complex (the cognate recognition sequence of U1A protein was ranked at 30). This result is none- theless not too surprising due to the noncanonical, seven-nucleotide recognition sequence (AUUGCAC) recognized by U1A that makes an unusually specific and strong interaction with RNA, unparalleled in other known RRMs [45]. Relatively poor results were also obtained for the Poly A binding potein (1CVJ_1, rank 19), and for RRM1 of the HuD protein (1FXL_1, rank 32). Both Pab and HuD utilize two domains to achieve sequence-specific recognition in a cooperative manner and do not discriminate well between sequences that are related to their cognate rec- ognition motif (A-rich and AU-rich sequences, respec- tively) [46]. Notably, however, the nonsequence-specific RNA helicase protein (PDB code: 2DB3, included as a negative control) had an expectedly poor cognate sequence rank of 226 ⁄ 256. Estimating experimentally determined relative RNA-binding affinities A second very important property of any potential function is the ability to recapitulate the sequence dependence of experimental binding energies; this is a prerequisite if the potential is to be applied to prob- lems of protein–RNA interface prediction or design. Fortunately, a few structures have a relatively dense set of experimentally determined binding constants for interface mutations. We used these experimentally characterized mutants to create a set of computation- ally ‘mutated’ structures of the complexes (Table 3), Fig. 2. Structure-based identification of RRM recognition sequen- ces. The cognate sequence is ranked by the distance potential (cut-off ¼ 6A ˚ ) for RRM ⁄ KH domain proteins. The red line repre- sents the rank of cognate recognition sequences using the contact- counting score; the blue line represents the rank of these sequences using the distance-dependent potential. The points in each colored line are sorted independently by rank; the x-axis is the sort order. The dashed line represents the 10th percentile. A knowledge-based potential function S. Zheng et al. 6382 FEBS Journal 274 (2007) 6378–6391 ª 2007 The Authors Journal compilation ª 2007 FEBS and have scored these structures using the distance- dependent statistical potential. A first very instructive example is provided by mutants of bacteriophage MS2 coat protein [47,48]. Starting with the crystal structure of the complex between MS2 coat protein and the cognate RNA hair- pin (PDB code: 1ZDI), a series of structures were gen- erated, representing the RNA and protein mutants for which binding constants are reported in the literature. Then the distance-dependent potential scores for these structures were compared with the known binding con- stants for each mutation. Unfortunately, when all of the MS2 mutations were considered together, a poor correlation was observed between distance score and experimental binding affinities (data not shown). How- ever, excellent correlations were obtained between these values when the binding-affinity data were divided into two subsets (Table 3, Fig. 3). A first set corresponds to complexes where the bound RNA hair- pin contained adenine, guanine or uridine base at posi- tion )5; the second set contains instead protein mutants where the bound RNA contained a cytosine at this position. Within each sets of mutants, the corre- lation between distance score and experimental binding affinity is strong (R 2 ¼ 0.65, Fig. 3A; R 2 ¼ 0.97, Fig. 3B), and statistically significant at the 95% confi- dence level. Figure 3C shows a likely explanation for this result: an intramolecular hydrogen bond formed by the cytosine at position )5 [47]. When this nucleo- tide is mutated to any other base, the intramolecular hydrogen bond is lost, leading to a reorganization of the RNA structure. This result does not provide direct information on the relative contribution of that hydrogen bond to the overall binding energy; it is simply implied that Table 3. Correlations between the distance-dependent score and the experimental free energy of binding for several mutant protein– RNA complexes. Distance-dependent Contact counting 6A ˚ 10 A ˚ 12 A ˚ 6A ˚ 10 A ˚ 12 A ˚ Protein mutations MS2 (no cytosine at position )5) 0.43 0.50 0.65 0.19 0.10 0.08 MS2 mutations (with cytosine at )5) 0.81 0.81 0.97 0.43 0.14 0.09 Fox-1 0.40 0.45 0.47 0.47 0.43 0.42 U1A a 0.27 0.48 0.65 0.29 )0.06 )0.03 U1A b 0.04 0.14 0.39 0.29 )0.06 )0.03 RNA mutations Fox-1 0.20 )0.39 )0.57 0.30 0.33 0.35 SRP; 2¢-OH mutations 0.87 0.56 0.52 0.36 0.30 0.29 SRP; base mutations )0.07 )0.03 )0.07 0.01 0.07 0.05 a The native U1A complex was included in the training set for this experiment. b The U2B¢ complex (U1A homolog) was included in the training set for this experiment. A B C Fig. 3. Correlation between scores generated by the distance- dependent statistical potential and experimental binding free ener- gies (logK d ) for mutants of the MS2 coat protein. (A) Complexes between protein mutants and RNA-containing nucleotides other than cytosine at position )5. (B) Complexes between protein mutants and RNA containing cytosine at position )5. (C) The char- acteristic intramolecular hydrogen bond between the amino group of C5 and the O1P atom of U6 observed in the structure of the MS2–RNA complex containing a cytosine at position )5 that helps organize the RNA structure for protein binding [47]. S. Zheng et al. A knowledge-based potential function FEBS Journal 274 (2007) 6378–6391 ª 2007 The Authors Journal compilation ª 2007 FEBS 6383 mutations must be segregated into two groups to obtain a clear correlation between experimental and predicted relative affinities. The most likely explana- tion for this result is that, at present, the statistical potential does not consider RNA intramolecular con- tacts; therefore, contributions to binding energy due to changes in RNA structure (i.e. that occur when that hydrogen bond is lost) cannot be captured by our cur- rent approach. A second example that reinforces our interpretation of the results obtained with MS2 is provided by Fox-1 protein, which regulates alternative splicing of tissue- specific exons by binding to the GCAUG sequence [49]. The structure of the complex (PDB code: 2ERR) and the experimental binding constants for two sets of related mutations have been reported [49]: one set for mutations on the Fox-1 protein and a second set for mutations to its target RNA molecule. A moderately strong correlation was observed between the distance score and the protein mutation data (R 2 ¼ 0.46, Fig. 4), but an anticorrelation was observed for the set of RNA mutations (R 2 ¼ )0.57; Table 3). As in the previous case, this result reflects the failure of the current statistical potential to capture the energetic contribution associated with the disruption of RNA intramolecular interactions that are a characteristic of this complex [49]. A third example is human U1A protein (PDB code: 1URN), a great model for the RRM superfamily because of the availability of NMR and crystallo- graphic structures [50,51], as well as binding data. In this case, we observed poor correlations between the distance-dependent score and the experimentally determined dissociation constants (K d ) [52] when we conducted a test using a training set of strictly non- homologous protein–RNA structures. Initially, we assumed that this observation would reflect the very large and energetically significant conformational changes that have been observed in the RNA and protein upon complex formation [53]. However, when the U1A complex itself was included in the training set, we obtained moderate to strong correlations (R 2 values between 0.27 and 0.65, depending on the choice of distance cut-off). This suggests that U1A binds to RNA by forming intermolecular interactions that are not commonly observed in the database of training structures. This hypothesis is supported by the observation that the inclusion of a close U1A homolog (the U2B¢–U2A¢ complex) in the training set improves the results of this test as well (R 2 increases from 0.04 to 0.39; Table 3). Thus, it appears that the structure of the U1A or of its homologous complex contains a set of protein–RNA atomic contacts (i.e. interatomic distances) that are not well represented in the 71 other protein–RNA complexes in our training set. Figure 5 shows the final example, a universally con- served component of the core of the signal recognition particle (SRP). The structure of the complex (PDB code: 1HQ1) and the binding affinity of a series of RNA mutants have been determined [54]. The distance potential results in scores that correlate significantly (R 2 ¼ 0.52, P £ 0.05) with experimental binding affini- ties for mutations involving substitutions of deoxy- nucleotides for their corresponding ribonucleotides. However, as observed for Fox-1, no significant Ade-4 Cyt-3 Ura-1 Gua-2 A B Fig. 4. (A) Correlation between scores generated by the distance- dependent statistical potential and experimental binding free ener- gies (logK d ) for mutants of the Fox-1 protein. (B) The intramolecular hydrogen bond between uracil 1 and cytosine 3, and the non-Wat- son–Crick base pair between guanine 2 and adenine 4 for the RNA in complex with Fox-1 protein (PDB code: 2ERR). The protein is represented in yellow; the RNA structure is colored by atom type. A knowledge-based potential function S. Zheng et al. 6384 FEBS Journal 274 (2007) 6378–6391 ª 2007 The Authors Journal compilation ª 2007 FEBS correlation was found for mutations of nucleotides that disrupt critical RNA intramolecular interactions. In this final case, these mutations involve the disrup- tion of base pairs near the binding interface that define the secondary structure of the RNA, which is obvi- ously important for recognition, but do not contribute directly to the formation of intermolecular contacts [54]. Disscussion The central role of protein–RNA interactions in the regulation of gene expression has led to consider- able interest in the biochemical processes underlying these interactions [55–57]. However, much of this research has been devoted to the study of the struc- ture ⁄ function relationship for individual protein– RNA complexes, and little effort has been made to develop quantitative models that might describe these interactions more comprehensively. Thus, our understanding of the mechanisms driving protein– RNA recognition is still largely descriptive [11]. Recent work on protein–DNA interactions has shown that quantitative models of protein–nucleic acid recognition can provide insight into the mecha- nisms of gene regulation [58,59], and, in the not too distant future, promise to allow the rational design of DNA-binding proteins with altered specificity [60]. The development of computational tools capable of predicting the specificity of RNA-binding proteins across entire families (such as the RRM superfam- ily), or of redesigning the specificity of these pro- teins, would be of equal importance in dissecting post-transcriptional regulatory mechanisms, and in providing new tools to interrogate gene expression pathways. In a previous study, our group demonstrated that a statistical potential function could be surprisingly accu- rate when used to predict protein–DNA interactions from structure [36]; this result was corroborated by a similar study published concurrently by another group [37]. Given these results, we hypothesized that the same approach would be equally successful with pro- tein–RNA interfaces. Indeed, although various statisti- cal techniques have been used by a number of groups for the prediction of protein structures, protein–DNA and protein–ligand interactions [18–35], such an approach has never been applied to protein–RNA interactions. In the present study, we describe the successful application of the distance-dependent, all-atom statis- tical potential function to the prediction of the ener- getics and recognition specificity of protein–RNA interactions. We demonstrate that the statistical potential can recapitulate experimentally determined relative binding constants for a number of protein– RNA complexes (with the caveat that it cannot yet capture the effect of mutations on RNA–RNA inter- actions). We also demonstrate that this simple tech- nique is remarkably successful at predicting the cognate recognition sequences of a wide variety of RNA-binding proteins. The challenge of near native decoy discrimination The statistical potential performs very well in classi- cal decoy discriminations tests. It is quite remarkable that similar Z-scores in tests of decoy discrimination are obtained for the statistical score and the rosetta-derived score because this second method contains many more adjustable parameters that are optimized to reproduce the average composition of these interfaces as observed in nature. By contrast, the current statistical potential was generated ‘as is’ from the observed frequency of intermolecular con- tacts in the database of protein–RNA structures. Thus, it appears that the distance-dependent statisti- cal potential implicitly captures at least some of the complexities of these intermolecular interactions that are explicitly enumerated in physical energy functions. The question of how to generate and discriminate near-native decoys is still an open challenge for many areas of computational structural biology [61,62]. The docking decoy set used here contains many near-native decoys (e.g. < 1 A ˚ rmsd) that can be discriminated by the distance-dependent potential (Fig. 1). However, when testing against the exceptionally near-native Fig. 5. Correlation between scores generated by the distance- dependent statistical potential and experimental binding free energies (logK d ) for ribose-to-deoxyribose mutants of a universally conserved protein component of the SRP. S. Zheng et al. A knowledge-based potential function FEBS Journal 274 (2007) 6378–6391 ª 2007 The Authors Journal compilation ª 2007 FEBS 6385 decoys generated by extracting snapshots from MD simulations (Table 2), we found that near non-native decoys could not be reliably discriminated from native structures, not even by amber, which was used to con- duct the MD simulations. Thus, the question of how to create a potential that is sensitive to the extremely subtle structural variations present in very near-native decoys remains a challenging and important area of research. We are hopeful that the incorporation of terms describing the higher-order geometric preferences of protein–RNA interfaces (e.g. the incorporation of a directional hydrogen-bonding potential) [17] may enhance the discriminatory power of our method, as will the inevitable increase in high-resolution structural data available for training. Nevertheless, the distance- dependent potential function already performs on par with the amber and rosetta force fields in decoy dis- crimination tests. The impact of contact distance cut-off on discriminatory power The contact distance cut-offs used in the present study were varied to determine the value that maxi- mizes decoy discrimination performance for protein– RNA complexes. Previously, Robertson et al. [36] showed that shorter contact cut-offs result in optimal discrimination ability in protein–DNA complexes, whereas Samudrala et al. [21] found that a longer cut-off (> 10 A ˚ ) was better able to discriminate cor- rect structures during protein structure prediction experiments. Finally, Lu et al. [23] demonstrated that the first coordination shell (i.e. a cut-off between 3.5 A ˚ and 6.5 A ˚ ) achieves the greatest selectivity for protein decoys created using gapless threading pro- cedures; thus, the question remains as to the best choice of contact cut-off. To evaluate the influence of different cut-off values in our study, replicate experiments were conducted using 6 A ˚ ,10A ˚ and 12 A ˚ distance cut-offs. In nearly all of our tests, the use of a shorter contact cut-off (6 A ˚ ) results in greater selectivity for structural details of the interface (Table 1). For the prediction of mutation energies, however, a longer cut-off appears to outperform shorter cut-off values for some sets of mutation data (Table 3). Some of these mutations are not near the protein–RNA interface (e.g. one of the U1A mutations, D79V, is 9 A ˚ from the RNA mole- cule), and only the use of a longer cut-off value can capture these effects. In light of the differing conclu- sions of previous research [21,23,36], these results imply that a ‘one size fits all’ approach to energy function design may be limiting. In other words, it may be possible to significantly improve potential functions by customizing their parameterization to particular problems. Prediction of RNA recognition sequences from protein–RNA complex structures An obvious but yet to be attempted application of any potential function for protein–RNA interactions is the prediction of cognate binding sequences. In a test of sequence recognition for 29 unique KH and RRM domains, we found that the potential is able to identify (within the 10th percentile) the cognate RNA recognition motifs of these domains approximately 70% of the time. As not all RRM ⁄ KH domains (for example, U1A) obey the simple four-nucleotide recog- nition model that we have introduced (where each nucleotide makes independent interaction with the protein) [6], and the specificity of some proteins is limited (i.e. they bind nearly equally well to a set of related sequences), this is a remarkably strong result. Despite the simple form of the statistical potential, and the over-simplifications of the four-nucleotide recognition model, this method is surprisingly robust over the diverse set of RNA-binding domains that we have considered. Prediction of relative protein–RNA binding energies When we evaluated the relative free energy of a set of mutations for several protein–RNA complexes of known structure, the distance-dependent potential was successful within defined structural classes. We observed strong, statistically significant (P £ 0.05) score–energy correlations for several sets of mutations that we tested; however, to achieve these results, it was necessary to subdivide several of the mutation data sets. For example, for the MS2 complex, the mutation data had to be divided into two classes based on the presence or absence of a cytosine at position )5 in the RNA. A likely explanation for the importance of the )5 cytosine mutation is offered by the observation that the amino group of the cytosine at position )5 makes an intramolecular hydrogen bond that increases the propensity of the free RNA to adopt the structure seen in the complex [48] (Fig. 3C). Because the dis- tance potential currently measures only intermolecular interactions, it is unable to capture the thermodynamic effect of interactions within the RNA or protein, and of mutation-induced changes in RNA (or protein) structure. The good correlations of distance potential with experimental binding energies (i.e. when sequence A knowledge-based potential function S. Zheng et al. 6386 FEBS Journal 274 (2007) 6378–6391 ª 2007 The Authors Journal compilation ª 2007 FEBS mutations are grouped according to the base identity at position )5) strongly suggests that the potential cap- tures the energetic contributions of intermolecular interactions well. The same limitations observed in the MS2 mutation data led to the failures in prediction for RNA mutations in the Fox-1 and SRP complexes. In the structure of the Fox-1 complex, nucleotide U1 interacts with C3 by forming an intramolecular hydrogen bond, whereas G2 and A4 form a non-Watson–Crick base pair [49] (Fig. 4). Four out of seven Fox-1 RNA mutations that were tested directly affect these intramolecular interac- tions, which are not evaluated by the statistical potential used in the present study. In the case of the RNA muta- tions to the SRP complex, the mutated RNA residues are located in a double-stranded region of RNA, and do not interact with the protein [54], yet the disruption of the helix clearly affects the binding energy. The effect of these changes in RNA conformation cannot be captured by the intermolecular potential function used here. Given these observations, it is reasonable to con- clude that the omission of protein intramolecular con- tacts might also limit the predictive power of the method. However, additional examples will need to be examined before definite conclusions can be made con- cerning the applications of statistical potentials to pre- diction of relative binding energies. The effect of training set composition on potential function performance All knowledge-based potentials face the possibility of unintentional bias or over-training because their train- ing depends upon the selection of a representative sam- ple of structures. If great care is not exercised to ensure that this training set is unbiased (i.e. structur- ally heterogeneous), it is possible to create a statistical potential that unfairly scores certain structures more favorably than others simply because they are over- represented in the training set. The challenge of over-fitting is particularly acute for protein–RNA interactions because there are relatively few high-resolution structures of protein–RNA com- plexes. Because of this limitation, a combined train- ing ⁄ test set was used in the present study. To avoid bias, a ‘leave one out’ cross-validation strategy was employed: the tested structure was always excluded from the training set. Thus, every test in the present study was conducted with a different score, and trained using only those structures that were not homologous to the tested protein–RNA complex. This strategy cannot be avoided at the present time, yet it leads to situations where the training data does not contain enough information to capture particular structural phenomena. For example, we observed vir- tually no correlation between the distance-dependent score and the experimental binding affinity for muta- tions of U1A protein until the U1A complex structure was added to the training set (Table 3). Addition of the homologous U2B¢ complex structure (PDB code: 1A9N) to the training set improved these results con- siderably, indicating that the training set was missing critical structural information that would help to dis- criminate native-like contacts unique to the U1A com- plex (an unusually high-affinity RRM, with a long, seven-nucleotide recognition sequence) [52]. We antici- pate that the performance of the method will improve with the size of the structural database, as more high- resolution protein–RNA structures become available. Conclusions We have introduced a statistical potential function that discriminates the structures of native protein–RNA complexes from decoys, reproduces experimentally determined relative binding affinities for a number of RNA-binding proteins, and predicts cognate binding sequences for a large set of protein–RNA complexes. The statistical potential performs as well as highly optimized physical potential functions in tests of docking decoy discrimination. We anticipate that the performance of the potential will only increase with the size of the structural database and as terms are added to the model to account for protein and RNA intra-molecular interactions that are currently ignored. Nevertheless, even in its current implementation, this statistical model achieves a high degree of sensitivity to subtle changes in protein–RNA interface structure. We are optimistic that this knowledge-based potential function will find broad application to problems requiring the high-resolution modeling of protein– RNA interfaces, such as structure-based genome anno- tation, or the rational design of novel RNA-binding proteins. Experimental procedures All-atom distance potential The potential function used here is identical to a previ- ously described method [36] (a more complete description of the method is provided in supplementary Doc. S1), with the exception of a modified low-count correction. In the present study, the correction described by Sippl [20] is replaced with a weighted pseudocount method, where a constant number of pseudocounts (P) are added to the S. Zheng et al. A knowledge-based potential function FEBS Journal 274 (2007) 6378–6391 ª 2007 The Authors Journal compilation ª 2007 FEBS 6387 [...]... Wang B, Pearlman DA et al (2004) Amber8 University of California, San Francisco, CA Izaguirre JA, Catarello DP, Wozniak JM & Skeel RD (2001) Langevin stabilization of molecular dynamics J Chem Phys 114, 2090–2098 Pastor R, Brooks B & Szabo A (1988) An analysis of the accuracy of Langevin and molecular dynamics algorithms Mol Physics 65, 1409–1419 Supplementary material The following supplementary material... discrimination Table S4 Comparison of the Z-scores obtained on the same set of protein–RNA decoys for the RNA, DNA and reduced atom distance potentials in decoy discrimination Table S5 Comparison of the performance of the potential with different upper cut-off values in decoy discrimination This material is available as part of the online article from http://www.blackwell-synergy.com Please note: Blackwell... we also tested a simple, contact-counting method, wherein every contact between protein and RNA (within a given distance cut-off) was assigned a same score of )1 Atom type selection Atom score types were assigned using the method of Robertson and Varani [36] Briefly, the all-atom potential treats every atom, in every residue, as a unique type (e.g alanine Cb and arginine Cb are considered as unique atom... free energy was calculated using the mm_gbsa module of amber 8 as: DGbind ¼ Gcomplex À ðGprotein þ GRNA Þ where Gcomplex, Gprotein and GRNA represent the mm_gbsacalculated free energies of the protein–RNA complex, the free protein and the free RNA, respectively Prediction of sequence -specificity As many RNA -binding domains of the RRM superfamily interact in a conserved fashion with four nucleotides across... Black DL, Pitsch S & Allain FH-T (2006) Molecular basis of RNA recognition by the human alternative splicing factor Fox-1 EMBO J 25, 163–173 Oubridge C, Ito N, Evans PR, Teo CH & Nagai K (1994) Crystal structure at 1.92 A resolution of the RNA -binding domain of the U 1A spliceosomal protein complexed with an RNA hairpin Nature 372, 432–438 Allain FHT, Gubser CC, Howe PWA, Nagai K, Neuhaus D & Varani.. .A knowledge-based potential function S Zheng et al observed counts for each atom pair A pseudocount correction value of 75 ensured the greatest performance of the potential, as indicated by the average Z-scores [63] (see supplementary Table S3) These pseudocounts are allocated over distance bins in proportion to the background frequency f(dij) values, as calculated using Eqn (4) from a previous... folding An approach to the computational determination of protein structures J Comput Aided Mol Des 7, 473–501 20 Sippl MJ (1990) Calculation of conformational ensembles from potentials of mean force: an approach to the knowledge-based prediction of local structures in globular proteins J Mol Biol 213, 859–883 21 Samudrala R & Moult J (1998) An all-atom distancedependent conditional probability discriminatory... single strand DNA (PDB codes: 2UP1, 1WTB, 1X0F, 1ZZI and 1ZZJ) were also included in the test set because recognition of single stranded RNA and DNA are mechanistically similar Protein mutations were modeled using moe (Chemical Computing Group, Montreal, Canada), followed by energy minimization with amber; the conformation of the mutated residue with side chain conformation most similar to the native residue... was retained Acknowledgements We wish to thank Dr Yu Chen for providing the protein–RNA decoy sets and Mr Daniel Bjerre for many valuable discussions The study was supported by grants from NIH References 1 Amosova O, Broitman SL & Fresco JR (2003) Alaninescanning mutagenesis of the predicted rRNA -binding domain of ErmC¢ redefines the substrate -binding site and suggests a model for protein–RNA interactions... material is available online: Doc S1 Detailed description of the construction and derivation of the distance-dependent potential Table S1 The complex structure (PDB code) in protein–RNA training set Table S2 The cognate sequence rank by distance ˚ potential and contact score (cut-off ¼ 6 A) in RRM ⁄ KH domain sequence decoy (256) sets Table S3 Optimization of the pseudocounts in decoy sets discrimination . A knowledge-based potential function predicts the specificity and relative binding energy of RNA -binding proteins Suxin Zheng 1, *, Timothy A. Robertson 2, *. We anticipate that the increasing availability of protein– RNA structures, together with the availability of data on specificity, will further improve the