Computational Approaches to Immunogenicity

From the perspective of drug development, avoiding immunogenicity may begin with candidate selection. Accurate prediction of immunogenic peptides early in a project should allow the design of therapeutic proteins with reduced risk of anti-drug antibody responses. Identification of potential T-cell epitopes within therapeutic proteins is therefore an active field of research. Empirical methods to directly measure peptide-MHC binding affinities are now available. Although usually quite powerful, most of these methods have been set up and tested on a limited number of alleles, restricting their sphere of utility. In vitro approaches include stimulating T cells with peptide sequences from the therapeutic protein, using either a panel of normal volunteers or previously treated patients as donors. Again, low-throughput limits the utility of this method for screening potential drug candidates.

Computational predictions, based on the peptide sequences of lead molecules, can be used to broadly and rapidly screen numerous candidates for possible immunogenic effects. Because the TCR is intrinsically variable, computational rules governing the specific interaction between TCR and peptide-MHC class II complexes have not been developed. Thus, in silico identification of potentially immunogenic peptides currently relies on predicting which pep-tides bind to MHC class II molecules (126-129). Only a subset of the peptide-MHC class II complexes will also bind and activate a TCR. As antigen presentation is necessary, but not sufficient, for T-cell activation, such prediction methods will systematically produce many false positives.

Qualitative prediction methods calculate an immunogenicity score which, combined to a threshold, provide a list of binder and nonbinder peptides (130). The sequence-based approaches often have difficulty in detecting binding motifs because of the highly variable amino acid composition of binding peptides. However, some clear rules have been established, such as the presence of specific anchor residues required for binding to the MHC groove, at well-defined positions in the peptide (131).

A very popular and relatively easy method based on the concept of "virtual matrices" was introduced in 1999 by Sturniolo et al. (132). The interaction between a peptide in an extended conformation and the MHC class II molecule involves several pockets along the groove of MHC, one for each amino acid side chain of the peptide. The method assumes that each pocket binds independently to an amino acid of the peptide. The strength of the method results from treating MHC pockets as independent units that are shared between HLA alleles. By generating pocket profiles for a limited set of pockets, the authors estimated quantitative matrices for a range of HLA alleles, using sequence comparison of the residues that make up each of the variable pockets. Affinity of a peptide for a particular HLA allele was predicted by summing the individual matrix-derived scores of the amino acids at each position along the peptide. The calculated affinity score was shown to be in good agreement with experimentally determined affinity. The virtual matrices have been implemented in the program TEPITOPE developed by the same group (133).

Experimental confirmation of in silico immunogenicity predictions can be labor intensive. A recent study compared serum anti-drug antibody levels with ex vivo responses of patient T cells stimulated with drug peptide sequences and correlated the magnitude of these responses with the patients' MHC allele carriage. Good correlation was found between the predictions made by in silico screening of eight common MHC alleles against overlapping peptide fragments and the in vivo as well as ex vivo antibody responses to drug (134). Another approach to validate an in silico screening method is to screen proteins with well-known immunogenic epitopes and observe the predictive value of the algorithm to identify those peptides. In some cases, additional empirical evidence, such as co-crystallization of peptides with MHC class II molecules, may be used to qualify the algorithm and to benchmark the score thresholds that are used to identify peptides most likely to bind.

0 0

Post a comment