Consensus Prediction of Protein Conformational Disorder from Amino Acidic Sequence

INTRODUCTION

It was recently shown that several proteins do not assume a well defined and stable three-dimensional (3D) structure but are natively unfolded [1]. This was absolutely surprising since unfolded proteins are known to be less stable and soluble in vitro and protein misfolding is known to be associated with several conformational diseases, including Parkinson and Alzheimer [2]. However, a considerable fraction of the proteome is constituted by natively unfolded proteins and this fraction seems to be larger in higher organisms than in simpler prokaryotes.

Several techniques to predict conformational disorder in proteins have been designed [3-5] and the performance of many of them is periodically checked, within the CASP initiatives [6], where several blinded predictions are made on targets, the conformational status of which is known only by the CASP organizers and is unknown by the various prediction teams that participate to CASP. In general, it appears that (i) the reliability of these predictions is rather modest and that (ii) different predictions are made by different predictors. The first point is per se not surprising, given the intrinsic difficulty of predicting 3D features on the basis of amino acidic sequences. The second point - the inconsistency between different prediction methods - is also not very surprising. In fact, various predictors do not differ only in their algorithms but also in what they define as "conformational disorder" and thus in what they want to predict. For example, in one of the DISEMBL versions [7], all the residues in loops are considered to be conformationally disordered, while in another of the DISEMBL versions, only the residues that were not visible in the crystallographic electron density maps are considered to be disordered. Alternatively, in IUPRED no a priori definition of disorder is used [8]. Despite their limitations, the techniques for predicting conformational disorder are extremely important. Initially, they were designed principally to study the interesting phenome-non of conformation disorder and for large-scale proteome comparisons. Later, it became clear that they have also a series of practical applications, like for example in structural genomics, where they are becoming routine filters in the pipeline of finding suitable targets to be analyzed [3, 4]. In fact, it is obvious that the 3D structure of natively unfolded proteins cannot be determined and that these disordered targets must not be analyzed experimentally by structural biologists.

In this paper we present a consensus method, based on various prediction methods, the performance of which is significantly better than that of each individual predictor. Such a new technique is easily usable with freely available software and is interesting not only for structural genomics initiatives but also for traditional hypothesis-driven structural biology. We also report a statistical survey of the Protein Data Bank that shows the fraction of disordered residues in proteins the crystal structure of which was determined. It appears that a moderate fraction of conformationally disordered residues can be tolerated. About 22% of these crystal structures have more the 5% of the residues that are disordered, though only about 2% of them have more than 20% of the residues in a conformationally disordered status.

METHODS

RESULTS AND DISCUSSION

Besides their basic biological importance, predictions of protein conformational disorder are important in structural biology, where "impossible" targets must be identified before inserting them in the experimental pipe-line that goes from cloning to structural determination. This is particularly important not only in structural genomics initiatives, the success rate of which is still rather modest, but also in traditional hypothesis-driven applications, especially when the protein construct must be designed by the scientists, like for example in multi-domain protein and viral poly-proteins [10].

Predictions of conformational disorder are thus one of the bioinformatics filters that must be used before moving towards experimental analyses. Other filters are focused on the quaternary structural requirements of a protein chain [11], on protein solubility and stability [12, 13], and some web-based servers were created to assist the users in this task [14, 15].

However, before doing predictions of conformational disorder it is necessary to know what level of disorder can be tolerated by well folded proteins. In fact, while it is clear that the 3D structure of a completely disordered protein cannot be determined, it is also clear that many (or, maybe, most) proteins are partially disordered.

For example, many loops at the protein surface are very flexible and tend to adopt more than a single shape. For this reason, we scanned the Protein Data Bank (PDB) [16, 17] looking for regions conformationally disordered.

This information was extracted from the records labeled with "REMARK 465", where the depositors of the crystal structures declare, if necessary, which residues were not observed in the electron density maps. This analysis was limited to the crystal structures, which are nevertheless the large majority of the entries of the PDB, and it was assumed that the location of completely unfolded segments cannot be detected in the electron density maps.

Fig.(1) shows the distribution of the PDB entries according to their fraction of residues not observed which are likely to be conformationally disordered. It appears that a considerable number of structures have conformational disorder. In 22% of them, more than 5% of the residues are disordered. However, only about 2% of the crystal structures contain more than 20% of the residues that lack a well defined structure. The most extreme case is the entry 1VCR, the light-harvesting complex from Pisum sativum thylacoid membrane, where 56% of the residues were not observed, though this crystal structure was determined and refined at very low resolution (9.5 Å) [18].

Fig. (2) shows the relationships between the crystallographic resolution and the percentage of disordered residues. It can be seen that resolution tends to decrease if the amount of disorder increases, though the effect of disorder on resolution is not spectacular. In fact the average resolution decreases only from 2.13 to 2.45 Å if the disorder fraction increases from 2.5 to 32.5%.

This clearly shows that protein 3D structures are often partially disordered and that a moderate fraction of conformationally disordered residues can be tolerated. Keeping this in mind, one can now try to predict if a protein has a reasonable probability to be suitable for a structural biology analysis. We designed a prediction method that is based on several individual prediction algorithms. The only necessary input is the amino acidic sequence of the protein and all the predictors are freely available. Each prediction algorithm must be used separately (Table 1) and its results must be inserted into equation (2), together with the coefficients x_i reported in Table 2. If the value of p_cons is positive, the residue is predicted to be disordered and if it smaller than zero, the residue is predicted to be ordered. This can easily be done for each residue and, as a consequence, it is possible to reach a global picture of the conformational status of the protein.

This new prediction method, which is essentially a weighted consensus approach, performs quite well, better than any individual prediction algorithm. Table 3 shows the values of several figures of merit, obtained with a 20-fold cross validation procedure. It can be seen that predictions are very accurate, with all the figures of merit larger than 80%. This is impossible by using individual predictors, though all of them have very high specificity. The probability excess, which is the best figure of merit because little influenced by the fact that the data are unbalanced, is equal to 80.1%, a value much larger than any other predictor.

It must be observed that the prediction reliability described above is based on the particular set of proteins available at the DISPROT database. Therefore, it would not be surprising to obtain other estimations of reliability by using different data.

As a consequence, the reliability indicators shown in Table 3 cannot be used to rank various prediction methods according to their performances. It is however clear that the consensus approach presented in this manuscript is likely to be superior to all the individual methods on which it is based and it is also reasonable to suppose that an increase of experimental knowledge, which is likely to occur in the future, will allow more accurate predictions.

REFERENCES