Removing T-cell epitopes with computational protein design

Overview of Computational Method.

For a given target protein and set of host MHC alleles, potential T-cell epitopes are first identified using a support vector machine (SVM). These regions are then optimized to eliminate the T-cell epitopes while retaining structure and function using an extension of the Rosetta all-atom protein design methodology with modifications to both the energy function used in the design calculations and the optimization procedure.

The energy function used in the sequence optimization is supplemented with two terms that incorporate immunologically relevant data. The first term calculates predicted epitope content using SVMs trained with experimentally determined peptide–MHC binding data. Scores from SVMs for each MHC allele in each 15mer sequence frame are averaged and then summed over each frame. The second term uses known host genome 9mer data and known epitope data, rewarding each host 9mer in proportion to its frequency of occurrence in the host genome and penalizing known epitopes. Both deimmunization scores favor negatively charged residues on the surface of the protein; hence, we also introduce a net charge constraint into the total objective function, penalizing deviations from the input protein formal charge. By weighting this term appropriately against the deimmunization terms, negatively charged residues may be placed at critical epitope-disrupting surface positions while compensatory positively charged residues are introduced at other positions.

Sequence optimization is carried out using a protocol that focuses on solutions that reduce the energy with a relatively small number of mutations and allows use of computationally expensive objective functions in design calculations that otherwise might be unacceptably slow. The energies of all point mutants at each design position are first computed and then sorted. At each position, all point mutants within a certain threshold of the lowest energy mutant are saved for subsequent combination. These lowest energy point mutations are then combined using a greedy stepwise, steepest descent heuristic (Methods), allowing for structural relaxation at each step, until mutations have been attempted at all design positions. Multiple diverse near-optimal designs can be generated in parallel by stochastically accepting the placement of near-optimal mutations during the combination process.

T-Cell Epitope Prediction.

We initially trained our SVMs on experimentally measured MHC binding affinities for 26 allelic variants (11). For each MHC allele type represented in the training set, we collected all epitope sequences from the Immune Epitope Database (IEDB) known to elicit T-cell activation (5). We restricted additional analysis to alleles with at least 25 known T-cell epitopes in the IEDB. In the absence of sufficient data on nonbinding peptides for each allele, we assumed that most 15mers in the host genome would not constitute strong MHC binders and generated negative datasets by randomly choosing 1,000 15mers from the human genome (12). We compared our SVM-based epitope predictions with the predictions of previous methods using a subset of T-cell epitopes withheld from initial training. Sensitivity and specificity were evaluated over all alleles for our method, NetMHCII-v2.2 (13), and PROPRED (14), and predictive performance was evaluated by calculating the area under the receiving operator characteristic (ROC) curve (AUC) for each allele type both independently and over the entire set. Because PROPRED contains only matrices for human leukocyte antigen DR beta (HLA-DRB) alleles, we combined testing data for the eight DRB alleles covered by all three methods and generated a standard ROC plot (Fig. 1A). The highest AUC was achieved by NetMHCII (AUC = 0.759). Rosetta performed comparably but slightly worse (AUC = 0.752), whereas PROPRED achieved significantly lower prediction accuracy (AUC = 0.710). Because more T-cell epitopes have been characterized for some alleles in the test set, combining all testing data weights performance analysis with alleles with more data points. If we average AUCs with equal weight over the shared DRB allele set, again, NetMHCII performs best (AUC = 0.792), with Rosetta slightly lower (AUC = 0.785) and PROPRED lower still (AUC = 0.771) (Fig. 1B).

Large-Scale Benchmarking and Calibration of Design Method.

We first tested the ability of the method to eliminate putative human T-cell epitopes from eight proteins from pathogenic organisms that contain known MHC-binding epitopes. The computational design protocol was used to simultaneously eliminate all predicted epitopes for eight representative human DRB1 alleles (SI Methods), collectively covering almost 95% of the human population (15). To evaluate the tradeoff between epitope removal and protein stability, multiple design simulations were carried out, with an increasing weight on the SVM-based epitope scoring term. Increasing the weight on this term decreases the number of MHCII predicted epitopes and increases the Rosetta energy (Fig. 2A).

Because amino acid substitutions predicted to disrupt peptide–MHC binding might destabilize the overall protein, a balance between the stability of the protein and disruption of possible MHC binding must be sought. A weight on the epitope scoring term of 2.0 eliminates 79% of the predicted epitopes and 84% of the known epitopes without increasing Rosetta energy above that of the native protein (Table S1). Because sequence changes are permitted only at critical predicted epitope positions, the number of mutations is minimized, thus allowing for substantial reduction in predicted immunogenicity while maintaining average sequence identity at 66%. Similar calculations were carried out varying the weight of the term favoring 9mer sequences found in the human genome (Fig. 2B). As the weight increases, the average number of human genome 9mers over the designed regions increases, and the Rosetta energy becomes more unfavorable. At a weight of 3.5, human 9mer sequences increase from 0% to 4.3% of redesigned epitopes, whereas the average Rosetta energy increases only 10% over baseline (Table S2). These weights were used for the remainder of this work unless otherwise noted in Methods.

Recapitulation of Previous Immunogenicity Reduction Data.

We next attempted to recapitulate the results of previous experimentally validated protein deimmunization efforts where immunogenicity was reduced without disrupting biological function. Cantor et al. (16) sought to remove T-cell epitopes from Escherichia coli l-asparaginase II, an enzyme approved for treatment of acute lymphoblastic leukemia, while maintaining native-like enzymatic activity and protein stability. Cantor et al. (16) first used a neutral drift selection scheme to identify allowable mutations in each of three predicted HLA-DRB1*04:01 epitopes and then combined mutations in each epitope to produce a fully functional enzyme with reduced immunogenicity in vivo. We applied our protocol to the biologically active tetrameric state of E. coli l-asparaginase II, constraining design to the residue positions chosen for randomization in the previous study, and compared our predictions against the mutations identified in the selection screen and the predicted immunogenicity of the resultant design sequence. After Rosetta redesign for all 24 epitope residues, two of five mutations assumed residue identities found in the sequences of experimentally isolated activity-preserving variants, with one mutation occurring at a position not tested in the above study. In addition, two of three redesigned epitopes have predicted HLA-DRB1*04:01 binding affinity lower than the peptides tested in the previous study (Table 1).

Tangri et al. (4) attempted the deimmunization of erythropoietin (Epo), a growth hormone used in treatment of anemia and myelodysplasia. Tangri et al. (4) screened a set of Epo-derived peptides for binding to a set of 15 different HLA alleles, targeted two potential epitopes for deimmunization by screening a small set of point mutants in each epitope for reduction of T-cell activation, combined point mutants, and screened for both immunogenicity and biological activity. We applied our deimmunization design protocol to this protein using the structure of Epo bound to its native receptor and targeting the same epitope positions experimentally explored by Tangri et al. (4). Because the HLA allele set of the human donors in the study is unknown, we designed simultaneously against all eight alleles in the HLA-DRB1 allele set. In the first epitope region, Rosetta introduces only one substitution (the same mutation found to be optimal in ref. 4). Rosetta similarly recovers the point mutation position in the second epitope (Table 1) but substitutes an alanine to preserve the protein’s net charge and makes three additional mutations to further disrupt MHC binding. In both cases, the Rosetta design minimizes the number of predicted MHC-binding peptides while minimizing the energy of the protein in the epitope regions.

Epitope Removal in Superfolder GFP.

As an experimental proof of concept, we chose the fluorescent reporter protein superfolder GFP (sfGFP) (17). GFP is used to identify and track genetically modified stem cells in vivo for numerous applications (18) but concern about the immunogenicity of cells expressing GFP remains (19). We sought to redesign sfGFP to eliminate T-cell epitopes without disrupting fluorescence. To do so, we targeted the top four predicted H-2-IAb epitopes in the sfGFP sequence (Fig. 3A). None of these epitopes were present in our known epitope database, although epitope 84 had been previously identified as immunodominant in WT GFP. The design algorithm is nearly deterministic; to generate multiple candidates for testing, we stochastically sampled alternative sequences by random inclusion of locally near-optimal mutations at each design position. Eight designs were chosen for testing based on sequence diversity and predicted stability (Table 2 and SI Methods, sfGFP Deimmunization Design Sequences Alignment). sfGFP and all eight designs were expressed in E. coli and purified as soluble protein. To determine whether fluorescence was affected by the design mutations, emission and absorbance spectra were obtained for sfGFP and all eight deimmunized proteins. All eight designs showed fluorescence absorbance and emission spectra comparable with sfGFP, with fluorescence excitation peaks at 485 nm (Fig. 4A and Fig. S1).

We then investigated whether existing epitopes were correctly identified and removed and whether new epitopes were not introduced by the design mutations. sfGFP and the deimmunized variant (sfGFP.di.v3.2) were chosen for immunological testing. This variant was selected as the most aggressive design, because it had the highest Rosetta energy but still maintained function. GFP:I-Ab tetramer reagents were generated using both native and design peptide sequences for three of the predicted epitope regions. For both constructs, five mice were injected with the protein in complete Freund’s adjuvant (CFA). After 6 d, spleens from all 10 mice were stained with a multicombinatorial panel of GFP:I-Ab tetramers corresponding to both the native and design sequences of all three predicted epitope regions (Table 2), and tetramer-positive cells were magnetically bead-enriched. For mice challenged with WT sfGFP, flow cytometry confirmed epitope 84 as the immunodominant epitope, with epitope 223/224 recognized by a smaller number of T cells (Fig. 4B). For mice challenged with the deimmunized protein, all three tetramers corresponding to the three redesigned epitope regions failed to isolate T cells above background levels, except for one mouse that responded weakly to the mutant epitope 223/224 (Fig. 4C). These two experiments confirmed that the design mutations effectively eliminated or greatly reduced the antigenicity of WT T-cell epitopes without creating new immunodominant epitopes or disrupting fluorescence.

Epitope Removal in Domain III of PE38.

We next sought to remove T-cell epitopes from the toxin domain of the cancer therapeutic HA22, a recombinant immunotoxin containing a 38-kDa fragment of PE38 (20), while maintaining cytotoxic activity. HA22 has been used successfully to treat refractory hairy cell leukemia in a recent phase 1 clinical trial (21), and it has produced complete remission in several patients with acute lymphoblastic leukemia (22). However, HA22 has shown limited effectiveness in treating patients who are not immune-compromised, because the presence of the bacterial PE38 moiety leads to host immune response and the production of neutralizing antidrug antibodies (23). To address this issue, we targeted three epitope regions in PE38 previously identified in humans (24) and designed five mutations predicted to eliminate binding to a diverse set of 14 human HLA alleles while maintaining favorable interactions with the toxin substrate, eukaryotic elongation factor 2 (Fig. S2). The five mutants were expressed and purified for subsequent testing of cytotoxicity in two Burkitt lymphoma cell lines. Two mutants in the region 466–480, A476D and A476D+D474Y, had cytotoxic activity reduced by ∼80%, but three mutants in the region 547–564 displayed equal or greater cytotoxicity than the WT toxin (Fig. 5A). The two most active mutants (L552N and L552E) were chosen for additional characterization of immunogenicity. Peripheral blood mononuclear cells (PBMCs) derived from two patients and one naïve donor were stimulated with mutant antigen, and IL-2 response was measured after restimulation with the WT and mutant epitope peptides. Both mutants caused a significant reduction in T-cell response for both epitope peptides in all three samples (P > 0.01 in Student t test) (Fig. 5B).

Penalizing Predicted Epitopes.

Rewarding 9mer Sequences That Occur in the Host Genome and Penalizing Known Epitopes.

Rosetta Design Calculations.

sfGFP Activity and Immunogenicity Assays.

Exotoxin A activity and immunogenicity assays.