A GRAPHICAL INTERFACE FOR CORRELATED MUTATIONS AND OTHER PROTEIN STRUCTURE PREDICTION METHODS.
Florencio Pazos, Osvaldo Olmea1, Alfonso Valencia*.
Protein Design Group. CNB-CSIC. Madrid.
1 Permanent address: Physical Chemistry Division, CIGB, Havana, Cuba.
* corresponding author.
Alfonso Valencia. Protein Design Group. CNB-CSIC. Campus U. Autonoma. Cantoblanco. Madrid 28049.
Tfn. +34-1-585 45 70, Fax. +34-1-585 45 06.
e-mail "valencia@samba.cnb.uam.es"
url http://gredos.cnb.uam.es
Keywords: Graphical interface, Structure Prediction, Correlated Mutations, Sequence Conservation, Protein Contacts.
Introduction
We present here a graphical interface for the representation of different types of structural predictions derived from sequence information.
The interface (Fig. 1) is able to represent : a) secondary structure, b) accessibility predictions, both of them derived from the Predict-Protein server(Rost, et al., 1994) (Rostet al., 1994), c) sequence conserved residues, as calculated by (Sander et al., 1993) also obtained from the Predict-protein server and c) prediction of contacts by correlated mutations. The degree of conservation, level of correlation and confidence of the secondary structure predictions can be set interactively by the user. This is especially useful since the quality of the different predictions depends on their reliability values.
The need to integrate all this information in a single representation becomes apparent in recent protein structure prediction exercise (Hubbard et al. IRBM Workshop team., 1996) and in the context of the recent contest on structure prediction (CASP2 http://www.mrc-cpe.cam.ac.uk/casp2/). We have found the combined representation useful. Its use will certainly contribute to new improvements and hopefully will facilitate the access to otherwise unconnected prediction methods.
The newest information displayed is the contact prediction based on correlated mutations. Correlated mutations are those changes occurred during evolution to compensate for structural drift and/or to maintain stability. They can be found in multiple sequence alignments as groups of positions with similar patterns of variation. Our initial approach to the calculation of correlated mutations was described in (Göbel et al., 1994) . The main difference with other published methods (see Rost et al. 1994 for a review) is that in our method the positions in the alignments are not compared directly. Instead, the relative variation of the sequences at each position is compared. It is important to notice that our approach is also completely different from mutual information calculated as conditional probabilities for each type of residue pair at two different positions (Clarke, 1995) . Our procedure can be viewed as the comparison of two grey-colour matrices by the relative distribution of the tones in each rather than by comparing the actual intensities between corresponding pixels.
We have recently made some technical improvements to the published method (Göbel, et al., 1994). First we use the McLachlan scoring matrix directly, avoiding a previous normalization step that was found to be responsible for numerical rounding errors. Also we are treating gaps as dummy observations, so that gaps are set to an arbitrary value of 0 in the scoring matrix. This value has no effect on the calculation as assessed independently. In this way the number of sequences at each position (N) is kept constant along the whole alignment, regardless of the presence of gaps. This new treatment of gaps only affects positions with less than 10% gaps, since positions with more gaps as well as completely conserved positions are excluded from the calculation.
These rather technical changes have a quite dramatic effect in the predictive power of the method, with a two-fold improvement in the prediction of three-dimensional contacts over our previous implementation. In figure 2 the ACCURACY of the current implementation is represented for a large set of non-homologous proteins of known three dimensional structure. ACCURACY, also called precision by others, is defined as the fraction of correct predictions over the total number of predictions. A typical example could be Gamma-b Crystallin (170 residues) for which 9 contacts are predicted correctly out of 71 predictions (ACCURACY of 0.12 when contacts between sequence neighbours closer than 6 residues are excluded). ACCURACY depends on protein size, since it is easier to predict contacts in small proteins that have higher densities in their contact maps.
The graphical interface itself is independent of the program used to predict the contacts. Any method can be used provided that its output fits the simple format described in our web -page (http://gredos.cnb.uam.es/pazos/plotcorr.html). Indeed, a more comprehensive approach that integrates correlated mutations with other sources of information derived from multiple sequence alignments is described elsewhere (Olmea et al., submitted). The three main sources of information are: a) sequence conservation, b) contact density per residue type and environment and c) stability of correlated mutations to changes in the input sequence alignment. The new program is already available at the same site and the results can be examined with the graphical interface. The ACCURACY of this combined approach is better: in the example of Gamma-b Crystallin mentioned above, 15% contact predictions are correct.
Availability
The graphical interface is written in C using X graphical libraries. The program that calculates correlated mutations is coded in C++. The typical execution time for the whole procedure is 20 s on a R4400 MIPS processor for a 170 residue long alignment with 50 sequences. User instructions, detailed description of the methods, and the binary code for SGI IRIX 5.2 and Linux Operating Systems are available at http://gredos.cnb.uam.es/pazos/plotcorr.html. The programs are free for academic. Commercial users are requested to contact valencia@samba.cnb.uam.es.
Acknowledgements.
Comments and suggestions from U. Göbel, P. Cronet, J. Lozano, L. Segovia, N. Brown and members of the Protein Design Group are acknowledged. comments This work was supported by Grant BIO94-1067. O. O. is a recipient of an I.C.I. fellowship.
Figures
Fig. 1. Graphical interface for the representation of correlated positions.
In the X-window the protein sequence is represented as a circle with the N-and C-termini on the right of the figure and the sequence running counter-clock wise. Starting from the outside, the following information is displayed in circles: a) Position of the conserved residues. Conservation is taken from Predict-Protein server (Rost, et al., 1994) as calculated by the program MAXHOM (Sander et al., 1991). The level of variability, starting with 0 (conserved), can be chosen from the dials. b) Predicted accessibility, and c) Secondary structure schema are given, as obtained from the Predict-Protein server (Rost, et al., 1994). One of the dial buttons controls the reliability level of the predictions. A colour code indicates the different structural elements. d) Selected residues: it is possible to pick up residues from the screen to get residue type and number labelled. e) Correlated pairs of residues are represented by lines connecting the corresponding residues. The level of correlation, between -1 and 1, can be set by the user.
In many cases it is interesting to visualise the results excluding some residues. Residues to be excluded can be read from an external file. This option is particularly useful for excluding tree-determinant residues as calculated by the program SEQUENCESPACE (Casari, et al., 1994).
Fig. 2 Performances of Correlated mutations in contact prediction.
X-axis, protein families sorted by number of aligned positions. Proteins are labelled by their PDB code. The list include all proteins in the PDB-select list of non sequence-redundant protein structures (Hobohm et al., 1994) for which there were more than alignment of 15 sequences and at least 40 valid positions. in the HSSP data base (Sander et al., 1993). Valid positions are those positions with less than 10% gaps and not completely conserved. Y-axis. Log scale of ACCURACY (number of true-predictions over true plus false-predictions. True predictions are those that correspond to pairs of residues with their beta carbons closer than 8 Å (Alpha carbons are used for Gly). The number of predicted pairs for each protein is 1/2 of protein length. These pairs are chosen from the most correlated positions in the sorted list of correlation values.
References.
Casari, G., Sander, C., & Valencia, A. (1994). Functional residues in protein sequence space. Nature Structural Biology. 2, 171-178.
Clarke, N. (1995). Covariation of residues in the homeodomain sequence family. Protein Science. 4, 2269-2278.
Göbel, U., Sander, C., Schneider, R., & Valencia, A. (1994). Correlated mutations and residue contacts in proteins. Proteins. 18, 309-317.
Hobohm, U., & Sander, C. (1994). Enlarged representative set of protein structures. Protein Sci 3, 522-524.
Hubbard, T., Tramontano, A., IRBM Workshop team. (1996). Update on protein structure prediction: results of the 1995 IRBM Workshop. Review. Folding &Design. 1, R55-R63.
Rost, B., Sander, C., & Schneider, R. (1994). PHD - A mail server for protein secondary structure prediction. CABIOS. 10, 53-60.
Sander, C., & Schneider, R. (1991). Database of Homology-Derived Structures and the Structural Meaning of Sequence Alignment. Proteins. 9, 56-68.
Sander, C., & Schneider, R. (1993). The HSSP data base of protein structure-sequence alignments. Nucl. Acids Res. 21, 3105-3109.