|Home | About | Journals | Submit | Contact Us | Français|
Motivation: Mutating residues into alanine (alanine scanning) is one of the fastest experimental means of probing hypotheses about protein function. Alanine scans can reveal functional hot spots, i.e. residues that alter function upon mutation. In vitro mutagenesis is cumbersome and costly: probing all residues in a protein is typically as impossible as substituting by all non-native amino acids. In contrast, such exhaustive mutagenesis is feasible in silico.
Results: Previously, we developed SNAP to predict functional changes due to non-synonymous single nucleotide polymorphisms. Here, we applied SNAP to all experimental mutations in the ASEdb database of alanine scans; we identified 70% of the hot spots (≥1 kCal/mol change in binding energy); more severe changes were predicted more accurately. Encouraged, we carried out a complete all-against-all in silico mutagenesis for human glucokinase. Many of the residues predicted as functionally important have indeed been confirmed in the literature, others await experimental verification, and our method is ready to aid in the design of in vitro mutagenesis.
Availability: ASEdb and glucokinase scores are available at http://www.rostlab.org/services/SNAP. For submissions of large/whole proteins for processing please contact the author.
The role of a protein in an interaction pathway is arguably its most important function (Eisenberg et al., 2000). Thus, protein–protein and protein–substrate interactions are essential for survival. Typically very few residues are essential for any protein interaction interface in the sense that mutating these significantly impacts the reaction (Bogan and Thorn, 1998; Weiss et al., 2000); these crucial residues are often referred to as protein–protein interaction hot spots. One coarse-grained experimental probe for elucidating the function of a protein is to mutate residues that are hypothesized to be involved in function. Alanine, glycine, proline and cysteine scanning mutagenesis (individual substitutions of residues by any of the said amino acids) are used to identify functionally important sites (Clackson and Wells, 1995; Gardsvoll et al., 2006; Konishi et al., 1999; Kouadio et al., 2005; Qin et al., 2003). Because of a variety of biophysical and technical reasons, alanine scans dominate. Rarely multiple mutations are tested for the same residue (Xiang et al., 2006; Yang et al., 2003). The impact of mutations on function is captured by a variety of probes; one of the more accurate means is the measurement of the change in the binding energy between the wild-type (native sequence) and the mutated protein. Although, large energy changes may result from destabilization of the affected proteins and from deformation of the binding sites, such dramatic alterations often indicate that a hot spot was mutated. To illustrate the relevance of hot spots to research: over 400 PubMed records mention hot spots in 2007 alone. One reasonable definition for a hot spot is that its mutation alters the binding energy by ≥1kcal/mol (Kortemme and Baker, 2002).
Computational methods can identify hot spots for proteins of known three-dimensional (3D) structure (DeLano, 2002; Guerois et al., 2002; Shulman-Peleg et al., 2007), and more recent attempts even spot these crucial sites from sequence (Gonzalez-Ruiz and Gohlke, 2006; Ofran and Rost, 2007b). ISIS (Ofran and Rost, 2007a) was the first tool to specifically predict protein–protein interaction hot spots from sequence, but estimates for the effects of single substitutions have long been around (Epstein, 1966; Vegotsky and Fox, 1962; Zuckerkandl and Pauling, 1965). The most recent methods are tailored to predict the effects of non-synonymous single nucleotide polymorphisms (SNPs), i.e. single nucleotide changes that alter the protein sequence (Bromberg and Rost, 2007; Ng and Henikoff, 2003; Ramensky et al., 2002; Yue et al., 2006). Such methods have not been assessed in light of large-scale alanine scans and hot spots. One reason might be that while function changes are sensed by such methods, the amount or severity of change is not. Thus, the predicted functional change may just as likely be a hot spot as it may not be.
Here, we examined the potential of one particular implementation for in silico mutagenesis, namely SNAP (Bromberg and Rost, 2007), that has been optimized to predict the effect of non-synonymous SNPs on a version of the public database PMD (Kawabata et al., 1999; Nishikawa et al., 1994) curated by us. SNAP evaluates functional effects of single amino acid substitutions using neural networks; its output is a value from –100 (no effect) to +100 (effect). First, we established that SNAP correctly captured the effect of alanine scans extracted from ASEdb (Thorn and Bogan, 2001). Then, we assessed substitutions by amino acids other than alanine. Combining these results, we could analyze in silico to which extent alanine scans correlate with all possible mutations. For technical reasons, we confined this analysis to one particular protein with ample experimental data (hexokinase).
To the best of our knowledge this is the first comprehensive study that connects biophysical data from alanine scans with methods optimized to capture the functional effects of SNPs. Making this connection is by itself an important novelty. What makes it even more interesting is that only in silico can we comprehensively address the question as to how representative current alanine scanning is, and only by this means can we comprehensively study the effects of mutagenesis without exorbitant costs. Further large-scale testing of our pilot study is required to establish more clearly that our approach actually captures functionally important residues and hot spots.
Alanine scanning data was extracted from ASEdb database (Thorn and Bogan, 2001). For each complex we recorded the name of the mutated partner, the position of the mutation, and the change in energy (ΔΔGcomplex) of stability of the given complex due to the mutation. If more than one complex was reported for the given mutant, only the complex resulting in the highest energy change was retained. For the purposes of ASEdb, ΔΔGcomplex is computed as the difference in energy of the wild-type complex (ΔGwild) as compared to the energy of the mutated complex (ΔGmut). Thus, a negative ΔΔG represents a more stable complex (ΔGmut >ΔGwild) and a positive ΔΔGcomplex represents a destabilized complex (ΔGmut<ΔGwild). We used a value of 1 kCal/mol change in binding energy as cutoff for determining hot spot residues.
SNAP outputs a score that ranges from −100 (no effect) to+100 (strong effect). A score cutoff is chosen to classify all mutations into neutral and non-neutral. By default, positive scores define non-neutral mutations; scores ≤ 0 identify neutral mutations; higher scores yield stronger predictions. For this work, we recorded SNAP predictions for all 19 non-native substitutions for each mutated residue in the by ASEdb data sets. We also compiled the average over all substitution scores at each position. Accuracy (often also referred to as specificity) and coverage (also referred to as sensitivity) of all performances were computed using Equation (1), where TP is the number of hot spots predicted to be non-neutral, FP is the number of non-hot spots predicted to be non-neutral and FN is the number of hot spots predicted to be neutral.
We assumed that all residues predicted and not observed to be functionally important were incorrect predictions (false positives). In particular, we assumed that for each protein in ASEdb there is only one binding site, namely the one probed in that experiment. This is obviously an extreme position that will considerably underestimate our levels of accuracy.
The correlation between score distributions was computed by:
SNAP networks were trained on data from PMD (Glaser et al., 1998; Kawabata et al., 1999) which slightly overlaps with ASEdb. To avoid over-estimating performance by testing on mutants that were seen in training we aligned all sequences in ASEdb against all proteins in PMD (BLAST at e=0.001). For each of the aligned sequences we collected the mutants found in both databases and recorded their functional effects according to PMD. These were then compared to the corresponding classifications from ASEdb.
We utilized PROFacc (Rost, 2000, 2005; Rost and Sander, 1994) to predict location of affected residues in ASEdb in protein structure. Residues were split into three classes: buried=<9% exposed surface area, intermediate=>9% and <36%, exposed=>36%. SNAP prediction accuracy and coverage [Equation (1)] were computed separately for each accessibility class as well as over all classes.
The sequence of human hexokinase (SWISS-PROT identifier HXK4_HUMAN;P35557;465 amino acids) was taken from SWISS-PROT (Bairoch and Apweiler, 2000; Bairoch et al., 2005) Four evaluations of residue importance were performed using scores from alanine, glycine, cysteine and average substitutions. For residues with the native amino acid non-X, the SNAP score of the by-X substitution was recorded; for residues with acid X, the average SNAP score was taken.
We extracted 1073 mutants from 48 distinct protein chains from ASEdb. Of these 323 were classified as hot spots at the cutoff of >1 kcal/mol change. Using this distribution with a random model (probability of observing a hot spot at any given residue is 0.5) to predict hot spots would result in 30% accuracy at 50% coverage [Equation (1)]. With default parameters, accuracy and coverage of SNAP predictions were 36% and 70%, respectively. When excluding any overlap between ASEdb and PMD (Section 2), these numbers fell to 33% and 67% (Fig. 1). While both of these sets of numbers significantly exceeded random, it is unclear which better estimated the method's performance. Of 174 overlapping mutants 45 (~26%) were annotated differently between PMD and ASEdb (i.e. PMD annotated the mutant as non-neutral when the corresponding ASEdb energy change was <1 kCal/mol, or vice versa). SNAP correctly classified 20 (~44% of 45) of these according to the ASEdb energy change. This implies that SNAP did not ‘memorize’ the training samples, but learned to make decisions based on observed patterns. Arguably, removing the overlapping mutants is therefore unnecessary and artificially reduces performance by decreasing sample diversity in the data set.
Increasing the SNAP non-neutrality cutoff (to 5 or 10, i.e. fewer residues predicted as hot spots; Fig. 1) reduced coverage without increasing accuracy correspondingly. Slightly increasing the threshold for considering a residue as a hot spot (from 1 to 2 or 2.5 kcal/mol) slightly increased coverage and decreased accuracy. In contrast, significantly increasing this energy threshold (from 1 to 4 or 4.5 kCal/mol) significant raised coverage (80 and 90%, respectively). Overall, more severe (larger) changes in binding energy tended to yield higher SNAP scores. When we considered as neutral only mutations for which the binding energy remained identical between wild-type and mutant, our default method achieved 84% accuracy at 62% coverage.
Extending ‘no change’ to an interval of ±0.2kcal/mol in the change of binding energy (approximation of experimental error in energy change measurement) yielded 68% accuracy at 63% coverage. SNAP predictions were more accurate for residues that were predicted to be buried: 80% buried hot spots were identified, 79% intermediate ones, and only 55% of the exposed hot spots.
SNAP identifies functional effects of single amino acid substitutions. The tool was not explicitly developed to outline residues of functional importance. Surprisingly, it recognized 70% of the hot spots in the ASEdb data set, albeit it did so at very low accuracy. To some extent, low accuracy undoubtedly reflected limitations in our method. However, there are three major problems with the data and the way we used them that also contributed to low accuracy. Firstly, a particular mutation may not destabilize an interaction enough to pass the chosen threshold. For example, the K110A mutant in the basic fibroblast growth factor (bFGF) is part of a second important binding site (Springer et al., 1994). Mutation of this residue by an alanine slightly stabilizes the probed complex (ΔΔGcomplex=−0.33kcal/mol). Secondly, all experiments probe only one particular reaction. A residue not predicted to be a hot spot might be involved in another interaction. For instance, the H114A mutation in angiogenin is known to greatly decrease enzymatic activity of angiogenin with respect to tRNA (Shapiro and Vallee, 1989). However, the change in energy recorded in ASEdb is of the bound angiogenin to ribonuclease inhibitor complex. The mutant described has very little effect on this binding (~0.7kCal/mol). Thirdly, the precise threshold for considering a residue a hot spot is neither well-defined nor reaction-independent. For instance, the mutation of residue D28 in CD2 to alanine changes the binding energy of its complex with CD48 by >1.7kcal/mol although this residue has been shown in a more detailed study to contribute little to the actual binding (Davis et al., 1998). Instead, this particular mutation likely induces local changes in the adjacent binding site. Considering all the possible false assignments of functionality importance using alanine scans, it is not surprising that a fair number of non-hot spot residues are assigned to the non-neutral class by SNAP, and vice versa. Nevertheless, as the severity of change correlated fairly well with the SNAPs scores absolutely crucial hot spots (e.g. >4kCal/mol change) are virtually guaranteed to be included in the prediction at any cutoff.
The observation that buried hot spots are predicted more reliably could be due to the fact that buried residues are, on average, more sequence conserved than exposed residues (Rost and Sander, 1994) and that the success of SNAP is intricately linked to sequence conservation. Another reason might simply be that the experimental results are more reliable for the exceptional buried hot spots.
We used scores for substitutions ‘by alanine’, ‘by cysteine’, ‘by glycine’, and the average over all possible substitutions to highlight residues of importance in the human glucokinase protein. For alanine substitutions, the most conservative of all, a total of 214 of 465 (46%) residues in the human glucokinase (Hexokinase IV or D; HXK4) sequence were predicted to be functionally important at the default SNAP score cutoff. For cysteine and glycine, the functional residue counts were 254 and 275, respectively. The average substitution by all 19 non-native amino acids outlined 232 residues as functionally important (Fig. 2).
We chose this example because HXK4 is experimentally well studied; it is an enzyme that functions in glucose metabolism (Kamata et al., 2004). Variants of the glucokinase encoding gene are implicated in type 2 diabetes (MODY-2 maturity onset diabetes of the young) (Vionnet et al., 1992). The enzyme exists in three forms—super-open, open and closed. It has at least two functional sites: the glucose binding site (including residues E256, E290, T168, K169, N204 and D205) and the allosteric binding site [including V455, A456 and Y214, mutations of which cause a metabolic disease persistent hyperinsulemic hypoglycemia (Christesen et al., 2002; Glaser et al., 1998)]. Kamata et al., (2004) describe a synthetic glucokinase activator which binds the allosteric site and interacts with residues R63, M210, I211, Y215, M235 and V452. Allosteric binding is facilitated by the flexibility of connecting region I (residues 64–72), which, although not responsible for binding itself, is very important to proper function. In the super-open form glucokinase has reduced affinity for glucose and no allosteric binding site. A slow, energetically costly, conformational change transforms the protein into the open form upon glucose binding; this form has higher affinity for glucose binding, and is capable to rapidly transform into the closed form.
Binding of the allosteric regulator prevents glucokinase from going into its super-open form and thus contributes to continuous glucose metabolism (Kamata et al., 2004). The crystal structure of HXK4 was solved by Kamata et al., (2004) (PDB: 1v4s,Fig. 2); it captures the closed (glucose bound) conformation of HXK4. The synthetic activator loosely bound to the allosteric site is also seen. In all SNAP evaluations, the glucose binding site is very well highlighted with red (implying sites predicted to be functionally important). Neighboring internal regions also shown in red somewhat correspond to the stretches of sequence involved in facilitating transformation changes. Some of the residues interacting with the synthetic compound are also lit up. Quantitative predictions for the binding residues discussed here are given in Table 1.
When considering the four images of glucokinase (Fig. 2), it is intuitively clear that for this example by-alanine substitutions appear to be best in identifying functionally important residues (red predictions limited to potential functional sites and there is a higher resolution of color; i.e. very few residues for which prediction is made with low confidence). However, a more detailed study/comparison is required to determine which, if any one (as opposed to a few), substitution scoring is best at finding all functionally important residues.
Because in silico mutagenesis is so much cheaper than its experimental sister, we could comprehensively analyze the degree to which alanine scans are representative of all possible mutations. We found that SNAP prediction scores for by-alanine substitutions correlated strongly with the average SNAP scores over all possible substitutions [for both, reported ASEdb mutant locations (Fig. 3) and over all glucokinase residues (data not shown)]. This suggested that using alanine scans aimed at estimating functional importance of residues may likely be just as informative as sequentially substituting each of the other 18 amino acids. For ASEdb mutagenesis sites, the average correlated also significantly with by-cysteine and by-glycine substitutions.
ASEdb data is likely skewed with regard to interface residues; i.e. most alanine scanning mutagenesis experiments are performed on suspected sets of binders. When considering entire protein sequences, however, other notions become important. For instance, core residues may be predicted as functionally important due to their utter necessity for maintenance of protein stability. There currently is no simple automated way to separate out the reasons behind functional importance annotations. However, as the example with HXK4 shows, there is validity in filtering entire sequences.
First, the ability to consider all possible substitutions at each residue may aid experimentalists in choosing the optimal site for mutagenesis. Second, in this particular sequence, and likely in many others, over half of the residues are excluded from functional considerations by almost any measure. This significantly narrows down the number of suspects. Third, SNAP scores have a scale meaning; i.e. substitutions that have severe effects are more likely to have higher scores. This suggests priorities for processing mutations of interest. While in silico mutagenesis may not yet be good enough to do the experiment, we challenge that tools of the type we used have finally come sufficiently of age to aid experimental mutagenesis in their design and prioritization. In other words, comprehensive in silico mutagenesis is not ready to be an end, but certainly it is ready to make for a good beginning.
Alanine scans aid the experimental elucidation of protein function. We demonstrated that SNAP, a method developed for a very different purpose, namely to predict the effects of non-synonymous SNPs, correctly identified over 70% of the functionally important sites in ASEdb. As an example for a comprehensive in silico mutagenesis, we presented a demi-formal, graphical and intuitive evaluation of predictions made for all possible substitutions in the human glucokinase. This exercise highlighted the potential value in using SNAP predictions to guide experiments. Our work also suggested that alanine scans may be surprisingly representative of what could be found if we had the means to experimentally test the mutation of all residues by all non-native amino acids in say all human proteins.
Thanks to Rudolph L. Leibel, Marco Punta, Ta-tsen Soong and Chani Weinreb (all Columbia) for helpful discussions. Particular thanks to Guy Yachdav (Columbia) for all his help with setting up and maintaining the SNAP web-server. Last not least, thanks to all those who deposit experimental data into databases and to all of those who make their carefully evaluated tools available.
Funding: The work of Y.B. and B.R. was supported by the grant RO1-LM07329-01 from the National Library of Medicine (NLM).
Conflict of Interest: none declared.