Using structural bioinformatics to investigate the impact of non synonymous SNPs and disease mutations: scope and limitations

Abstract

Background: Linking structural e ects of mutations to functional outcomes is a major issue in structural bioinformatics, and many tools and studies have shown that speci c structural properties such as stability and residue burial can be used to distinguish neutral variations and disease associated mutations. Results: We have investigated 39 structural properties on a set of SNPs and disease mutations from the Uniprot Knowledge Base that could be mapped on high quality crystal structures and show that none of these properties can be used as a sole classi cation criterion to separate the two data sets. Furthermore, we have reviewed the annotation process from mutation to result and identi ed the liabilities in each step. Conclusions: Although excellent annotation results of various research groups underline the great potential of using structural bioinformatics to investigate the mechanisms underlying disease, the interpretation of such annotations cannot always be extrapolated to proteome wide variation studies. Di culties for large-scale studies can be found both on the technical level, i.e. the scarcity of data and the incompleteness of the structural tool suites, and on the conceptual level, i.e. the correct interpretation of the results in a cellular context.

Corresponding author

Background

The molecular phenotype of a coding non synonymous SNP or disease associated mutation describes the functional and structural properties of a protein that are a ected by a single amino acid substitution [ 25 ]. In this study we want to address whether the concept of the in silico determined molecular phenotype can be employed for large-scale classi cation of SNPs and disease mutations. The attempt to classify a large set of mutations based on an incomplete molecular phenotype may seem naive at rst glance, had it not been suggested that individual properties such as protein stability, the accessibility of the amino acid substitution site, and the location of variants in surface pockets are predictive determinants of the phenotypic e ect of a variation [1{4]. A comparative study of protein stability predictors by Blundell and co-workers demonstrated that although protein stability changes caused by mutation can be relatively accurately estimated in silico, these predictions by themselves do not yield accuracy on large-scale classi cation between benign and disruptive mutations [5{7].

Furthermore, computational analyses rely heavily on the quality of the data under scrutiny and the computational methods used to evaluate these data. Before investigating 39 structural properties of proteins and amino acid substitutions for their predictive power regarding SNP classi cation, we have investigated what major liabilities are encountered when implementing an structural approach to SNP annotation and classi cation. The results are compared with those achieved by the best performers among the state-of-the-art tools.

Results and Discussion

In this study we have identi ed the common issues that are encountered when performing large-scale analyses of structural properties of human coding variation. The rst issue concerns the availability of structural data for nsSNPs and disease mutations, while the second involves the availability of computational tools to predict structural properties. The last issue concerns the quality of classi cation: are the training and evaluation data sets used in the analyses su cient to extrapolate results for larger studies, and do the properties used have su cient predictive power to separate the two data sets? Structural coverage of human genetic variation Despite structural genomics projects, the gap between sequence and structural information is still wide, and the coverage of variation data with structural data is estimated to be as low as 14% [ 4 ]. We have investigated the boundaries of structural coverage by varying the quality requirements on the structural model (Supplementary Figure S1A), the sequence identity between query sequence and modelled structure (Figure S1B), the percentage of the wild type sequence covered by the structural model (Figure S1C), and the length of the alignment between query and target (Figure S1D). Without applying any restrictions, about 12% of all nsSNPs present in the Ensembl Variation Database (release 44) can be mapped on a structural model, in accordance with the estimate cited previously. However, this percentage is valid only when no restrictions regarding sequence identity, sequence coverage or structure quality are applied. Our standard restrictions on building high-con dence structural models using the FoldX force eld are X-ray structures with a resolution lower than 2.5 A and sequence identity higher than 80%. Applying these restrictions to the Ensembl data results in a data set of 5416 nsSNPs (circa 4% of the data, Figure S1B).

Predictability of structural properties The second issue for a large-scale structural bioinformatics approach is the structural properties that are predictable with state of the art tools: how well can we describe the structural behaviour of a protein and its mutants? Previous structural studies have identi ed protein stability, aggregation and misfolding as determinants of correct functioning on the single protein level [ 7,11,12 ]. Mutations a ecting the functional sites of a protein, such as DNA, ligand and protein interaction sites, are not considered within this scope, but the investigation of these sites will most certainly be of great importance to assess the impact of amino acid substitutions.

Tools have been developed that describe the structure and dynamics of a protein: stability, aggregation, amyloidosis, and folding. We have used computational methods that are capable of assessing the e ects of a mutation on protein stability (FoldX), aggregation (Tango) and amyloidosis (Waltz). Although algorithms exist that can predict folding of small single domain proteins (e.g. Rosetta [ 13 ], FoldX [ 14 ], SimFold [ 15 ]), to date no computational method exists that can predict folding events on large multi-domain proteins, or that is applicable in genome wide studies.

Although we have not investigated proteinprotein interactions in this study, we have included an analysis of the binding of proteins to molecular chaperones, as it is directly related to correct folding of the protein. The high abundance of chaperones in the cell emphasises their crucial role in the cell [ 16 ], but this is not re ected in the availability of computational tools for chaperone binding. We have used the only available tool, the Hsp70 binding predictor Limbo [ 17 ], to assess chaperone binding variation caused by amino acid alteration.

The predictive power of structural properties Following the recommendations of Care et al [ 18 ], we have used the SwissProt annotated disease and polymorphism data (SwissProt Variation Index release 52) as the evaluation data for our analyses. Mapping of these variants on high quality structural models (X-ray structures with resolution 2:5A, sequence identity with the model above 80%) yielded a data set of 240 positive (disease-associated) mutations and 400 negative variations (neutral nsSNPs) in 98 proteins. To ensure that the analyses are comparable, we applied the sequence based predictors to the same small data set as the predictors that use 3D structures or structural models.

Before we evaluated the discriminative power of the individual structural parameters, we wanted to assess whether our data showed distinguishable patterns for three important parameters. The rst two criteria, stability di erence and the degree of burial of the mutation site, have previously been identied as providing information about the severity of a mutation [ 4, 19 ]. The third criterion is di erence in aggregation propensity, which has been cited as likely to be an important factor in disease susceptibility [ 12, 20 ] but thus far has not been applied in a proteome wide mutation analysis.

Figure 1 shows the distributions for the stability di erences (A) and di erences in aggregation propensity (B) between wild type and variant proteins, and the burial of the mutation site (C). The rst observation of both the stability and the aggregation analysis is that the observed changes are not discrete but follow a smooth distribution from negative to positive change. Second, there are noticeable di erences between SNPs and disease mutations, but they cannot be distinguished by a simple cut-o value on the output, as there is large overlap between the distributions. This is con rmed by the P-values obtained from paired student t-tests, which are 0.96 for the stability distributions, 0.99 for the aggregation distributions, and 0.99 for the burial distributions, respectively. For the stability distributions, we see that disease mutations are generally more destabilising than SNPs, but their distributions overlap largely. A similar analysis has been performed on SwissProt variants using the Site Directed Mutator stability predictor [ 7 ], and the distributions of stability di erences of disease mutations and neutral variations are similar to our ndings.

In a rst series of properties to test as classi ers, we have investigated 15 properties of the amino acid substitution site that contribute to the assessment of the e ect of the mutation using the FoldX algorithm (Table 3). Cut o values were generated that varied between the minimal and maximal values measure for the speci c property, and the true and false positive rate, and the Matthews correlation coe cient (MCC) were calculated for each cut-o value. Table 3 lists the data for both the best MCC and the MCC90, i.e. the coe cient that is measured at high speci city (true negative rate = 90%). The corresponding ROC curves for these analyses can be found in Supplementary Figure S1.

The same strategy was then applied to predicted values of structural di erences between mutant and wild type proteins (24 properties). Statistics were calculated for stability and entropy parameters, as well as for di erences concerning protein aggregation, amyloidosis and chaperone binding (Table 4, Supplementary Figure S2).

The results obtained from these detailed analyses are unanimous: none of the parameters evaluated can be used to separate the data. All MCC values are close to zero, and thus the predictions are no better than a random predictor would perform on the data. The high accuracy of FoldX for stability estimation has been proven in various studies [ 6,9,10 ], so we have high con dence in our stability estimations. In accordance with the analyses of [ 7 ], we nd that high stability di erences alone are no su cient criterion to distinguish deleterious mutations and neutral variation. These results show that the dominant effect of for instance stability that was proposed in earlier large-scale studies [ 4, 22 ] can not be always generalised for other data.

The fact that none of the properties representing conformational di erences between wild type and variant protein contain enough information to distinguish neutral and deleterious variation implies that large-scale classi cation based on singular structural properties is not feasible and requires a better understanding of how the complex interplay between biophysical and biochemical properties of a protein conspire to di erent tolerance for mutations in different proteins.

Recent studies that combine structural and evolutionary information using machine learning techniques are able to classify relatively large data sets obtained for the SwissProt database successfully (summarised in Table S2). Machine learning approaches suggest that data integration is indeed the way forward, but the creation of this black box style of classi er does not o er insight into the biological processes. In the same way that using evolutionary information to classify SNPs obscures the how and why a speci c mutation is deleterious, using black box machine learning methods will not teach us what the underlying reason of disease is. Although knowing that an amino acid is critical for correct function is of course useful, in a structural bioinformatics approach the focus is more on the molecular mechanism underlying disease.

A simple combination of the SNPe ect structural bioinformatics toolsuite on our evaluation data set showed that in our case, at least a linear combination of these methods is not su cient to classify the data (TPR = 0.73, TNR= 0.27, MCC=0). A large part of the polymorphism data is predicted to have deleterious e ect. To assess the \predictiveness" of our data set, we applied the well-established evolutionary method SIFT [ 24 ] to our data and found that SIFT was also not able to classify e ectively. In fact the results were even worse than our naive classi er (TPR=0.69, TNR=0.21, MCC=-0.12).

As an illustration of the in uence of the data set used for evaluation on the performance of a predictor, we list the results for the variation in performance of SNP classi cation of SIFT, that uses evolutionary information to label SNPs (Supplementary Table S3). The Matthews correlation coe cient varies between -0.12 on our data set over 0.25 on human mutagenesis data, up to 0.59 on the HIV-1 protease mutagenesis set in the original SIFT paper [ 24 ]. This is yet another informative example on how crucial the choice of training and test data are to build and evaluate predictors: generalisation of results is only possible when the training data are expressive enough to represent the entire feature space.

Conclusions

The concept of using the molecular phenotypic effect of a nsSNP to assess its e ect on the structure and function of the protein it alters was rst introduced by Bork and co-workers [ 25 ]. The question has been raised to how much of this molecular phenotype is necessary to evaluate the contribution of a SNP to a disease phenotype: are there singular dominant properties that determine the impairment of structure and function, or do we need to consider the full ensemble of molecular properties to interpret the impact of the SNP? Other research groups have proposed that single properties such as stability [ 4 ] and solvent accessibility [ 1 ] can be used to classify SNPs. We have examined all the individual structural bioinformatics tools that were proposed in the SNPe ect toolsuite [ 26 ] for their ability to act as a binary classi er for deleterious and neutral SNPs. Neither of the individual properties that were examined could serve this purpose. Because several approaches were able to classify similar data sets as the one we have used, we applied the most used evolutionary method, SIFT [ 23 ], to our data set. As it was not able to classify our data set accurately, we argued that generalisation of the results presented by the state of the art classi ers might be an important issue. We illustrated this problem with the variability of performance of SIFT on 8 di erent data sets used in various analyses.

From these analyses we concluded that strict classi cation of SNPs is not feasible at the time, both because there are still many technical di culties to overcome, and because the biological interpretation of the molecular phenotype in relation to a disease phenotype is a complex matter. Even at the single molecule level, we cannot assess how tolerant a speci c protein is to structural variation. The inherent rigidity of a protein might in uence the change in stability that is allowed before severe conformational changes are introduced. Furthermore, on the cellular level biological interpretation is even harder: we can not predict the role of the protein quality control system plays in this tolerance level, not all interactions are described at the molecular level, and much more. Even if we can predict the molecular e ect accurately, this might not necessarily result in a disease phenotype because of functional redundancy of the protein.

However, not being able to classify human variation into disease mutations and neutral or bene cial variation does not mean that this approach or the methods developed are useless. By using high quality bioinformatics tools, we can select from a large pool of variations the candidates that are interesting for detailed investigation. This in itself is a valuable contribution, because the amount of variation data available is too massive to be investigated experimentally. In silico analyses can and will be used successfully as an addition to in vitro and in vivo studies.

Assembly of data sets

Statistics on the structural coverage and validation status of human non synonymous coding SNPs were performed on data from the Ensembl human variation database release 44, containing 12.2 million SNPs, of which 133698 cause an amino acid variation in a known transcript. The mapping of SNPs on protein structures was evaluated using the \ensppdbmapping" DAS service provided by the SPICE server [ 27 ]. Positive and negative data sets for the evaluation of SNP classi cation were designed with data from the SwissProt variation index [ 28 ] in the UniProt knowledge base (version 52.0, March 2007, [ 29 ]) that were mapped onto known PDB structures and high quality homologs thereof. The quality criteria described in the results section (models with resolution of 3 Aor higher, sequence identity of 80% or more) lead to structural models of 400 SNPs (negative) and 240 disease associated mutations (positive).

Structural bioinformatics tools

We have used the FoldX force eld [33] for all mutant properties regarding structural location, protein stability and its various components, the Tango [34] and Waltz [35, submitted] algorithms to assess the propensity for aggregation of wild type and variant proteins, and the Limbo algorithm [17, submitted] to evaluate the chaperone-binding properties of amino acid sequences. A novel tool developed by Lenaerts et al (unpublished) was used to estimate the entropy of a speci c amino acid site in a high-resolution structure. Detailed descriptions of these ve tools can be found in the Supplementary Material.

Authors contributions

Conceived and designed the experiments: JR JS FR. Performed the experiments: JR. Analysed the data: JR JS FR. Wrote the paper: JR.

Acknowledgements

Joke Reumers was supported by a grant from the Federal Research O ce (FWO, IUAP P6/43), Belgium, and the Institute for the encouragement of Scienti c Research and Innovation of Brussels (ISRIB), Belgium. 29. UniProt Consortium:

Resource (UniProt). 35(Database issue):D193{7.

The Nucleic

Universal Protein

Acids Res 2007, 30. Zweig MH, Campbell G: Receiver-operating characteristic (ROC) plots: a fundamental evaluation tool in clinical medicine. Clin Chem 1993, 39(4):561{ 577. 31. Matthews BW: Comparison of the predicted and observed secondary structure of T4 phage lysozyme. Biochim Biophys Acta 1975, 405(2):442{451. Best MCC

Threshold

MCC90 1.61 -1.05 -1.76 -0.10 0.32 1.96 -0.98 -0.6 1.5 0.22 0.43 0.73 0.93 0.19 0.22 <0 -0.01 0.05 0.10 0.15 0.16 0.06 0.15 -0.1 0.05 0

Additional Files

Figure 1 { gure1.pdf Additional le 2 | supplementary.pdf Several of the less critical gures and tables are added as supplementary material, together with detailed descriptions of the structural bioinformatics tools used. Using structural bioinformatics to investigate the impact of non synonymous SNPs and disease mutations: scope and limitations Supplementary Material Joke Reumers1, Joost Schymkowitz1 and Frederic Rousseau 1 1Switch Laboratory, VIB, Vrije Universiteit Brussel, Pleinlaan 2, 1050 Brussels, Belgium Email: Joke Reumers - [e-mail]; Joost Schymkowitz - [e-mail]; Frederic Rousseau [e-mail];

Corresponding author

Methods Structural bioinformatics tools FoldX The FoldX force eld was developed for the fast and accurate estimation of the free change upon mutation on the stability of a protein or a protein complex [23{26]. It uses an all-atom representation of these macromolecules, and has been validated on a test database of more than 1000 mutants from more than 20 di erent proteins. It currently yields a correlation of 0.78 with a standard deviation of 0.41 kcal/mol.

Modelling and evaluation of mutations in FoldX is performed with the BuildModel command. It is used rst to model a homologous sequence on a structural model and to optimise the side chains to

t the new sequence, and then to evaluate the e ect of a single amino acid variation. The Gibbs free energy of a protein is calculated with the Stability command. The various structural parameters used in the classi cation tests (backbone clash, backbone H bond formation , sidechain H bond formation, electrostatics , solvation of hydrophobic residues, solvation of polar residues, torsion clash, Van der Waals contribution,Van der Waals clash)

Entropy calculations based on side chain sampling In addition to the entropy calculations intrinsic to the FoldX force eld, we use a novel method based on extensive sampling of side chain conformations as developed by Lenaerts et al. (unpublished). The sampling method produces for each side chain the probability (P (X)) of nding the residue's side chain in a particular conformational state. From these probabilities entropy can easily be derived:

H(X) =

X iP (xi)log2P (xi)

The method uses a rotamer database based on conditional statistics of dihedral angles derived from the WHAT IF data set [ 27 ]. All amino acids from this data and their corresponding dihedral angles (10 bin) were used to derive the following probabilities: P ( i); P ( ij i 1) and P ( ij i 1; i 2), except for 1(P ( 1) and P ( 1j ; )). A set of n random rotamers can be derived from the probability distribution thus calculated. This will allow sampling of rotamers with greater resolution than classical rotamer libraries.

The sampling itself is performed by Monte Carlo based sampling method with Metropolis criterion (at 298K). The Metropolis criterion states that a certain conformational change is accepted with a probability p that depends on the free energy change G associated with the conformational change as given by the following formula:

Tango The -aggregation prediction algorithm Tango [ 28 ] uses a statistical mechanics approach to represent a competition between major conformational states: the random coil and the native conformations, as well as -turn, -helix and -aggregate. Two windows of variable length slide over the sequence, and each such window can populate these conformational states according to a Boltzmann distribution. The frequency of population of each structural state for a given segment will be relative to its energy, which is derived from statistical and empirical parameters.

To predict the -aggregating segments of a peptide, Tango calculates the partition function of the phase space involving these conformational states. In our analysis we have used Tango to calculate the di erence in aggregation tendency that results from an single amino acid variation.

Waltz Current methods for the prediction of the sequence determinants of amyloidosis su er from two major problems: overpredicting amorphous cross aggregates and missing amylogenic sequences that are enriched in the polar Q and N residues, such as the prion protein. The Waltz algorithm [29, submitted] tackles these problems by taking into account amyloid hexapeptides from 48 new amyloid forming sequences, derived from 31 proteins. About half the proteins in this extended data set were not previously known to contain amyloidogenic sequences such as presenilin-2, titin and myosin. Waltz combines terms from amino acid sequence scoring in the learning set, physical property analysis and homology modelling. The method shows 84% sensitivity at 92% speci city on the AmylHex data set [ 30 ], and correctly identi es mutations in human proteins known to be associated with amyloid deposition.

Limbo Limbo is a Hsp70 binding site predictor that was built using a dual method combining sequence and structural information [31, submitted]. Experimental DnaK binding data of 53 non-redundant peptide sequences was used to generate a sequencebased position-speci c scoring matrix (PSSM) based on logarithm of the odds scores. Following an in silico alanine scan of the substrate peptide in the crystal structure of a DnaK-substrate complex (PDBID 1DKX Zhu1996) using FoldX, a structurebased PSSM that re ects the individual contribution of certain substrate residue types for DnaK binding was generated. The Limbo DnaK binding site predictor was obtained by combining the structurebased PSSM with a normalisation factor of 0.2 with the sequence-based PSSM. Limbo is able to correctly predict 89% of the true positives in a tested peptide set (high sensitivity), with a concurrent amount of only 5.9% false positives for a speci c score threshold (high speci city). The robustness of the predictor was evaluated with a cross-validation test, resulting in a true positive rate of 72% true positives and a false positive rate of 5.9%. The predictor was able to identify an entire known DnaK binding site in the heat-shock promoter 32 [ 32 ]. We have used Limbo to rank mutated proteins according to their DnaK binding a nity. Supplementary Table S1 - Types of data sets used to train and test SNP classi ers.

Origin data set Neutral variations Mutagenesis studies Orthologs SwissProt SNP OMIM dbSNP Disease mutations Mutagenesis studies COSMIC database HGMD OMIM SwissProt Disease Data [ 21 ] Data [ 22 ]

Size of data set

Number of studies [1{9] [ 3, 9, 10 ] [3, 8, 11{14] [ 15 ] [ 16, 17 ] [1{9] [ 18 ] [ 9 ] [ 3, 8, 13, 15, 18 ] [3, 8, 10{14, 19, 20] [ 20 ] [ 19, 20 ] Supplementary Table S2 - Performance of state-of-the-art predictors on representative data sets. The performance of a few selected tools on SwissProt disease associated mutations and SNP data are shown.

Study Bao et al [ 11 ] Capriotti et al [ 13 ] Karchin et al [ 14 ] Ng & Heniko [ 19 ] Wang & Moult [ 20 ] Worth et al [ 16 ] Yue & Moult [ 9 ]

Method

FPR

FNR

TPR

TNR

MCC Random Forest HybridMeth SVM SIFT Stability Combined SVM Supplementary Table S3 - Variation of the performance of SIFT on di erent data sets.

Dataset Test set Human lac I repressor HIV 1-protease T4 lysozyme SwissProt disease SwissProt + dbSNP SwissProt

A

2 104 C 0 0

2 104 1.5 104 s P N S ro 1 104 f e b m u N

5000 B 0

0 100 100 80 80 TPR4600 TPR4600 20 20 0 0 20 40 FPR 60 80 100 0 0 20 40 FPR 60 80 100 Figure S2. ROC curves for classi cation of disease mutations and neutral variation by using structural properties of the amino acid substitution site. 100 80 R60 TP40 20 0 0

Solvation hydrophobic 20

40 FPR 60 Main chain burial 80 100 20

40 FPR 60 Side chain burial 80 100 20

40 FPR 60 Solvation polar 80 100 20

40 FPR 60 Van der Waals clash 80 100 100 100 80 80 TPR4600 TPR4600 20 20 0 0 20 40 FPR 60 80 100 0 0 20 40 FPR 60 80 100 Figure S2 (continued). ROC curves for classi cation of disease mutations and neutral variation by using structural properties of the amino acid substitution site. 100 80 60 40 20 0 0 20 80 100 0 20 40 60 80

100

FPR 40

60

FPR properties of the amino acid substitution site. Overall stability difference

Overall stability difference

(surface residues)

40 FPR 60 Backbone H bond 80 100 20

40 FPR 60 Sidechain H bond 80 100 100 100 80 80 TPR4600 TPR4600 20 20 0 0 20 40 FPR 60 80 100 0 0 20 40 FPR 60 80 100 Figure S3. ROC curves for classi cation of disease mutations and neutral variation by using structural di erences between the wild type and variant protein. Electrostatics

Entropy main chain

40 FPR 60 Entropy side chain 80 100 20 40 FPR 60 80 Solvation hydrophobic

100 80 100 20 80

100 40 FPR 60

Torsion 20 100 100 80 80 TPR4600 TPR4600 20 20 0 0 20 40 FPR 60 80 100 0 0 20 40 FPR 60 80 100 Figure S3 (continued). ROC curves for classi cation of disease mutations and neutral variation by using structural di erences between the wild type and variant protein.

Van der Waals clash 0 20 40 60 80 100 0 20 40 60 80 100

FPR 100 80 60 40 20 0 80 60 40 20 0 80 60 40 20 0 100 R P T 0 20 40 60 80

100

FPR 60 40 20 0 80 60 40 20 0 100 100 R P T 80 60 40 20 0

FPR di erences between the wild type and variant protein. 40 FPR 60

Waltz 80 100 20 40 FPR 60 80 Waltz (positive scores)

100 100 20 80

100 40 FPR 60

Limbo 20 40 FPR 60 80

Waltz (negative scores) 100 100 80 80 TPR4600 TPR4600 20 20 0 0 20 40 FPR 60 80 100 0 0 20 40 FPR 60 80 100 Figure S3 (continued). ROC curves for classi cation of disease mutations and neutral variation by using structural di erences between the wild type and variant protein. 100 80 R60 TP40 20 0 0

References

1. Chasman D , Adams RM : Predicting the functional consequences of non-synonymous single nucleotide polymorphisms: Structure-based assessment of amino acid variation . J Mol Biol 2001 , 307 ( 2 ): 683 { 706 .

2. Cli ord RJ , Edmonson MN , Nguyen C , Buetow KH : Large-scale analysis of non-synonymous coding region single nucleotide polymorphisms . Bioinformatics 2004 , 20 ( 7 ): 1006 { 1014 .

3. Ferrer-Costa C , Orozco M , de la Cruz X : Sequencebased prediction of pathological mutations . Proteins 2004 , 57 ( 4 ): 811 { 819 .

4. Jiang R , Yang H , Sun F , Chen T : Searching for interpretable rules for disease mutations: a simulated annealing bump hunting strategy . BMC Bioinformatics 2006 , 7: 417 .

5. Krishnan VG , Westhead DR : A comparative study of machine-learning methods to predict the e ects of single nucleotide polymorphisms on protein function . Bioinformatics 2003 , 19 ( 17 ): 2199 { 2209 .

6. Needham CJ , Bradford JR , Bulpitt AJ , Care MA , Westhead DR : Predicting the e ect of missense mutations on protein function: analysis with Bayesian networks . BMC Bioinformatics 2006 , 7: 405 .

7. Ng PC , Heniko S : Predicting deleterious amino acid substitutions . Genome Res 2001 , 11 ( 5 ): 863 { 874 .

8. Saunders CT , Baker D : Evaluation of structural and evolutionary contributions to deleterious mutation prediction . J Mol Biol 2002 , 322 ( 4 ): 891 { 901 .

9. Yue P , Li Z , Moult J : Loss of protein structure stability as a major causative factor in monogenic disease . J Mol Biol 2005 , 353 ( 2 ): 459 { 473 .

10. Ferrer-Costa C , Orozco M , de la Cruz X : Characterization of disease-associated single amino acid polymorphisms in terms of sequence and structure properties . J Mol Biol 2002 , 315 ( 4 ): 771 { 786 .

11. Bao L , Cui Y : Prediction of the phenotypic effects of non-synonymous single nucleotide polymorphisms using structural and evolutionary information . Bioinformatics 2005 , 21 ( 10 ): 2185 { 2190 .

12. Bao L , Cui Y : Functional impacts of nonsynonymous single nucleotide polymorphisms: Selective constraint and structural environments . FEBS Lett 2006 , 580 ( 5 ): 1231 { 4 .

13. Capriotti E , Calabrese R , Casadio R : Predicting the insurgence of human genetic diseases associated to single point protein mutations with support vector machines and evolutionary information . Bioinformatics 2006 , 22 ( 22 ): 2729 { 2734 .

14. Karchin R , Diekhans M , Kelly L , Thomas DJ , Pieper U , Eswar N , Haussler D , Sali A : LS-SNP: largescale annotation of coding non-synonymous SNPs based on multiple information sources . Bioinformatics 2005 , 21 ( 12 ): 2814 { 2820 .

15. Stitziel NO , Tseng YY , Pervouchine D , Goddeau D , Kasif S , Liang J : Structural location of disease-associated single-nucleotide polymorphisms . J Mol Biol 2003 , 327 ( 5 ): 1021 { 1030 .

16. Worth CL , Bickerton GRJ , Schreyer A , Forman JR , Cheng TMK , Lee S , Gong S , Burke DF , Blundell TL : A structural bioinformatics approach to the analysis of nonsynonymous single nucleotide polymorphisms (nsSNPs) and their relation to disease . J Bioinform Comput Biol 2007 , 5 ( 6 ): 1297 { 1318 .

17. Burke DF , Worth CL , Priego EM , Cheng T , Smink LJ , Todd JA , Blundell TL : Genome bioinformatic analysis of nonsynonymous SNPs . BMC Bioinformatics 2007 , 8: 301 .

18. Worth CL , Burke DF , Blundell TL : Estimating the effects of single nucleotide polymorphisms on protein structure: how good are we at identifying likely disease associated mutations ? In Proceedings of Molecular Interactions - Bringing Chemistry to Life 2006 .

19. Ng PC , Heniko S : Accounting for human polymorphisms predicted to a ect protein function . Genome Res 2002 , 12 ( 3 ): 436 { 446 .

20. Wang Z , Moult J : SNPs, protein structure, and disease . Hum Mutat 2001 , 17 ( 4 ): 263 { 270 .

21. Halushka MK , Fan JB , Bentley K , Hsie L , Shen N , Weder A , Cooper R , Lipshutz R , Chakravarti A : Patterns of single-nucleotide polymorphisms in candidate genes for blood-pressure homeostasis . Nat Genet 1999 , 22 ( 3 ): 239 { 247 .

22. Cargill M , Altshuler D , Ireland J , Sklar P , Ardlie K , Patil N , Shaw N , Lane CR , Lim EP , Kalyanaraman N , Nemesh J , Ziaugra L , Friedland L , Rolfe A , Warrington J , Lipshutz R , Daley GQ , Lander ES : Characterization of single-nucleotide polymorphisms in coding regions of human genes (vol 22 , pg 231, 1999 ). Nat Genet 1999 , 23 ( 3 ): 373 { 373 .

23. Serrano L , Guerois R : Fold-X: An algorithm to predict and engineer folding pathways . Abstr Pap Am Chem Soc 2001 , 221 :U395{ U395 .

24. Guerois R , Nielsen JE , Serrano L : Predicting changes in the stability of proteins and protein complexes: A study of more than 1000 mutations . J Mol Biol 2002 , 320 ( 2 ): 369 { 387 .

25. Schymkowitz J , Borg J , Stricher F , Nys R , Rousseau F , Serrano L : The FoldX web server: an online force eld . Nucleic Acid Res 2005 , 33 :W382{ W388 .

26. Schymkowitz JWH , Rousseau F , Martins IC , Ferkingho - Borg J , Stricher F , Serrano L : Prediction of water and metal binding sites and their a nities by using the Fold-X force eld . Proc Natl Acad Sci USA 2005 , 102 ( 29 ): 10147 { 10152 .

27. Vriend G : What If - a molecular modeling and drug design program . J Mol Graph 1990 , 8 : 52 {.

28. Fernandez-Escamilla AM , Rousseau F , Schymkowitz J , Serrano L : Prediction of sequence-dependent and mutational e ects on the aggregation of peptides and proteins . Nat Biotechnol 2004 , 22 ( 10 ): 1302 { 1306 .

29. Maurer-Stroh S , Kuemmerer N , Lopez de la Paz M , Martins I , Reumers J , Serrano L , Rousseau F , Schymkowitz J : Accurate prediction of sequence determinants of amyloid formation using the Waltz algorithm . Submitted 2007 .

30. Thompson MJ , Sievers SA , Karanicolas J , Ivanova MI , Baker D , Eisenberg D : The 3D pro le method for identifying bril-forming segments of proteins . Proc Natl Acad Sci U S A 2006 , 103 ( 11 ): 4074 { 4078 .

31. Van Durme J , Maurer-Stroh S , Wilkinson H , Rousseau F , Schymkowitz J : Accurate prediction of the sequence determinants of DnaK-peptide binding via a method that integrates homology modelling and experimental data . Submitted 2007 .

32. McCarty JS , Rudiger S , Schonfeld HJ , SchneiderMergener J , Nakahigashi K , Yura T , Bukau B : Regulatory region C of the E. coli heat shock transcription factor, sigma32, constitutes a DnaK binding site and is conserved among eubacteria . J Mol Biol 1996 , 256 ( 5 ): 829 { 37 .