Prediction of Residue Status to Be Protected or Not Protected From Hydrogen Exchange Using Amino Acid Sequence Only

Search for Structural Parameters with the Best Correlation with Experimental Data on Hydrogen Exchange

We try to find what structural parameter will be the best for the prediction of hydrogen exchange protection factors. We consider seven proteins for which we have experimental data: alpha-lactalbumin, equine lysozyme [18], bovine pancreatic trypsin inhibitor [19], staphylococcal nuclease [20], horse heart cytochrome c [21], staphylococcal nuclease [22] and chymotripsin inhibitor 2 (CI2 [23]. We extracted B-factor data from the corresponding PDB files from our database. We operated with two types of B-factor values. The first type of values is an average of B-factor values over all atoms of the given amino acid residue in the studied protein (. The second type of values is the B-factor value for the C_α atom of the given amino acid residue in the studied protein (B(C_α.

It turns out that the maximal correlation (47% is obtained if to use the number of atom-atom contacts calculated from the known 3D structure, but not the B-factor (33% and entropy scale, as the number of degrees of freedom for angles φ, ψ, and χ for each amino acid [24] (about 1% (see Table 1. It has been shown previously in the work of Vendruscolo [15] that the correlations between experimental protection factors and B factors are relatively weak, consistent with the view that protection factors mainly probe larger-amplitude fluctuations than B factors [25]. The CamP method gives not a very high correlation for our database of proteins (correlation coefficient is 40%. This result confirms the suggestion that the result of the program trained on the database will depend on the considered database. Vendruscolo and the authors also compared the result of disordered predictions from different programs with experimental data [15]. The authors explained weak correlations found in this case by the fact that the intrinsic propensity for being unfolded is strongly modulated by the interactions in the folded state to define the local fluctuations probed by hydrogen exchange measurements [15].

We suppose that using only amino acid sequence we can predict the status of the residue to be protected or not in 3D structure (but not absolute values using statistics of atom-atom contacts for each 20 amino-acids sequence in the globular state (see Table 2 [16].

Using Fold/Unfold Method for the Prediction of Residue Status to be Protected or not from Hydrogen Exchange

In our studies, to search for status of the residue to be protected or not from hydrogen exchange using amino acid sequence only, we have employed the FoldUnfold program, which permits choosing different widths of the averaged window [16, 17]. As seen from our studies, the choice of the window width depends on the length of the expected loop and, consequently, on the task posed by researchers. Thus, the window width of 41 amino acid residues is optimal for the search of long unstructured regions in proteins that are considered to be completely disordered [16, 17]. The window width of 11 amino acid residues is optimal for the search of unstructured region in a polypeptide chain of 10-20 amino acid residues long.

To predict the status of the residue from the amino acid sequence, we used the scale based on the average number of contacts observed for 20 amino acids in globular proteins [16].

The procedure that was used for construction of a contact profile is as follows. First of all, contact indices are assigned to the given amino acids sequence. The second step is smoothing of the obtained profile by averaging with some sliding window. After such manipulations we translated the given body of contacts into a binary system with the following rules: if the difference between indices of the given amino acid and some threshold (the value of threshold, which is 20.4, discriminates regular and loop regions is less than 0, it should be an unprotected amino acid. Otherwise we name this amino acid as protected. The value of the threshold and rules were taken from previous works where a contact scale was used [16, 17]. In those works it was shown experimentally that there exists correlation between contacts profile (obtained using a contact scale and structured/unstructured regions [16, 17]. For estimation of quality of our prediction we used sensitivity and specificity. The specificity is defined as the rate of true negatives to the sum of true negatives and false positives, while sensitivity is defined as the rate of true positives to the sum of true positives and false negatives. In other words, the sensitivity defines the percentage of correct predictions of protected amino acids which are experimentally proved to be protected, while the specificity defines the percentage of correct predicted unprotected amino acids.

The FoldUnfold method with a different size of windows has been applied to seven proteins for which we have experimental data: alpha-lactalbumin, equine lysozyme [18], bovine pancreatic trypsin inhibitor [19], staphylococcal nuclease [20], horse heart cytochrome c [21], staphylococcal nuclease [22] and chymotripsin inhibitor 2 (CI2 [23].

The results of calculations are summarized in Table 3. As shown from the Table, the best scores belong to the size of 11 amino acids. We also tried the sliding window size of 13 amino acids, but this leads to lower results as compared to those of 11 amino acids (data not shown. One can see that for protein alpha-lactalbumin (pdb code 1hml we have got the maximum accuracy – 97%. A little bit lower results were obtained for other proteins, thus proteins 1hrc, 2eql, 2rn2 and 6pti have about 80% of accuracy (sensitivity on average (see Fig. 1. The next important question that arises is whether the accuracy (sensitivity is true in terms of specificity. We have calculated the amount of false positive predictions for every protein (specificity as well. It can be seen from the table that the results vary greatly, from 35% to 60 %, which is about 40% on average. It should be underlined that the majority of mispredictions for proteins represent cases in which protection was predicted but not observed.

We have considered the results of predictions by the CamP method for our database of proteins to compare with our results (see Table 3.

In spite of the fact that the CamP method is based on a neural net and gives almost 97% of accuracy, its specificity is very low (13%. There is an explanation to this fact: the method simply overpredicts the number of protected residues. According to the definition [15] the residue is protected if lnp>5, and not protected if lnp<5. Summarizing this information we can resume that in spite of its simplicity our simple approach keeps abreast with the CamP method.