advertisement: compare things at compare-stuff.com!

Predicting antigen contacts

Having investigated relationships between antibody sequences, antigen contacts, antigen type and combining site topography, we applied these results by predicting the antigen-binding pockets of antibody structures. This is particularly relevant for antibodies whose sequences are known and from which a model can be constructed. From the model we aim to predict those residues most likely to be involved in antigen contact, which can then be subjected to site-directed mutagenesis to test both theory and model.

We have developed a rapid and simple method which uses only average contact and surface shape information and test it by application to the complexed crystal structures (all antigen-contact data from the antibody being predicted are excluded). Firstly the mean burial (by antigen) data, $\overline{\rho}$, for each residue is projected onto the molecular surface of the antibody (Figure 2.7(a)). Figure 2.7(b) shows the same antibody with the surface coloured according to its curvature, $C$, as calculated by the program GRASP[Nicholls et al., 1991]. This alternative description of curvature was adopted for purely practical reasons and allows easy application of the technique by anyone with access to GRASP. To find the probable contact surface we calculate $\overline{\rho}(30-C)$, which combines the measure of concavity with the probability of antigen contact. An arbitrary cutoff of $\overline{\rho}(30-C) > 35$ was chosen by visual inspection (using just one complex), and surface points satisfying this condition are coloured red (Figure 2.7(c)). This red surface is patchy and discontinuous, so neighbouring patches are merged and disconnected patches are eliminated using a dilation/erosion procedure[Delaney, 1992] developed within GRASP. The red surfaces are dilated five times, eroded eight times and dilated again three times (each by 1Å). The binding pocket prediction is the resultant smooth red patch (Figure 2.7(d)).

Figure 2.7: Antigen binding pocket prediction using mean fractional burial by antigen calculated from the set of 26 antibody-antigen complexes. Sequence (a) to (d) Antibody CHA255[Love et al., 1993] molecular surface: (a) coloured by predictive mean fractional burial by antigen, $\overline{\rho}$ (see Section 2.2.2) (b) coloured by GRASP curvature, $C$ (c) red patches indicate where $\overline{\rho}(30-C) > 35$ (d) final binding pocket prediction (red patches in (c) have been smoothed) and bound antigen (black wireframe) (e) Antibody 40-50[Jeffrey et al., 1995], prediction as in (d). Figures produced using the program GRASP[Nicholls et al., 1991].

The `worst case' accuracy, $Q_w$, and `fraction correct' accuracy, $Q_c$, for predicting antigen contacting residues by this protocol with each of the 26 complexed crystal structures are calculated as follows:

$$Q_w = \frac{N_o + N_u} {N_c}$$

$$Q_c = \frac{N_c} {N_i}$$

where $N_c$ is the number of correctly predicted antigen contacting residues, $N_i$ is the total number of residues in the antibody-antigen interface and $N_o$ and $N_u$ are the numbers of over- and under-predicted contact residues respectively. The mean $Q_w$ is 1.0 (best 0.25, worst 2.5) whilst the mean $Q_c$ is 0.79 (best 1.0, worst 0.53). As expected (since contact data are derived from all antigen classes), the method tends to over-predict interactions with smaller antigens, and under-predict interactions with larger antigens. Improvements can be made by using separate average contact data for different sizes of antigen. Human judgement can also play an important role in using these predictions. For example, Figure 2.7(e) shows the prediction for antibody 40-50 (1ibg, anti-ouabain[Jeffrey et al., 1995]); the choice between the two putative binding regions is simple if it is known that the antigen is small and could not bind across both predicted surface patches.