Development of a strategy and computational application to select candidate protein analogues with reduced HLA binding and immunogenicity

Abstract

SummaryUnwanted immune responses against protein therapeutics can reduce efficacy or lead to adverse reactions. T-cell responses are key in the development of such responses, and are directed against immunodominant regions within the protein sequence, often associated with binding to several allelic variants of HLA class II molecules (promiscuous binders). Herein, we report a novel computational strategy to predict 'de-immunized' peptides, based on previous studies of erythropoietin protein immunogenicity. This algorithm (or method) first predicts promiscuous binding regions within the target protein sequence and then identifies residue substitutions predicted to reduce HLA binding. Further, this method anticipates the effect of any given substitution on flanking peptides, thereby circumventing the creation of nascent HLAbinding regions. As a proof-of-principle, the algorithm was applied to Vatreptacog a, an engineered Factor VII molecule associated with unintended immunogenicity. The algorithm correctly predicted the two immunogenic peptides containing the engineered residues. As a further validation, we selected and evaluated the immunogenicity of seven substitutions predicted to simultaneously reduce HLA binding for both peptides, five control substitutions with no predicted reduction in HLAbinding capacity, and additional flanking region controls. In vitro immunogenicity was detected in 21Á4% of the cultures of peptides predicted to have reduced HLA binding and 11Á4% of the flanking regions, compared with 46% for the cultures of the peptides predicted to be immunogenic. This method has been implemented as an interactive application, freely available online at http://tools.iedb.org/deimmu nization/.

Positive selection pressure on E2 protein of classical swine fever virus drives variations in virulence, pathogenesis and antigenicity: Implication for epidemiological surveillance in endemic areas

“…The deimmunization prediction tool has been recently created to identify immunodominant regions and predict those amino acid substitutions that create non‐immunogenic versions of the proteins (Dhanda et al, ). The tool also predicts whether the action of substitution on the immunogenic peptides could create alteration in immunogenic sites in the neighbouring peptides (Dhanda et al, ).…”

“…The deimmunization prediction tool has been recently created to identify immunodominant regions and predict those amino acid substitutions that create non‐immunogenic versions of the proteins (Dhanda et al, ). The tool also predicts whether the action of substitution on the immunogenic peptides could create alteration in immunogenic sites in the neighbouring peptides (Dhanda et al, ). Thus, to assess the role of mutations on different sites of E2 protein of the positively selected strains in comparison with the ancestral CSFV strain, we performed a search for deimmunization for the E2 protein using http://tools.iedb.org/deimmunization/ an default parameters.…”

Validation of a de‐immunization strategy for monoclonal antibodies using cynomolgus macaque as a surrogate for human

“…Aside from EpiVax, predicting T cell epitopes for specific HLA‐II alleles, the other algorithms make predictions based on antibody sequence and amino acid properties of individual peptides, hence these tools are largely species‐agnostic. The areas of the antibodies where all three algorithms overlapped (“hot‐spots”) were then de‐immunized using the IEDB de‐immunization tool (Dhanda et al, ). This algorithm predicts promiscuous binding regions within the protein sequence and identifies amino acid substitutions that are predicted to reduce HLA binding.…”

Designing Feature-Controlled Humanoid Antibody Discovery Libraries Using Generative Adversarial Networks

“…Dhanda et al, creators of the protein deimmunization engine on the IEDB website, also aggregate MHCII binding scores across alleles using the median. 53 We ignore negative scores in our sum across the sequence because peptides that certainly don't bind to MHCII (large negative scores) should not offset peptides that bind MHCII tightly (large positive score). Figure S5B shows the fraction of putative MHC binding peptides for all unique 15mers in each sequence set described in Figure 2B .…”

“…We demonstrate the GAN library biasing on such properties as a reduction of negative surface area patches, identified as a potential source of aggregation, thermal instability, and possible half-life reductions, 51 and away from MHC class II binding, which may reduce the immunogenicity of the generated antibodies. [52][53][54][55][56] We show, additionally, library biasing to a higher isoelectric point (pI) to reduce aggregation and prevent precipitation in therapeutic formulations, and towards longer CDR3 lengths which can increase diversity and has been known to create more effective therapeutics for a class of targets. 57 To demonstrate the viability of the Antibody-GAN to generate humanoid antibody sequences, the GAN was used to generate a proof-of-concept validation library of 100k sequences from 4 germline subgroups.…”