Computational design of the affinity and specificity of a therapeutic T cell receptor

Abstract

T cell receptors (TCRs) are key to antigen-specific immunity and are increasingly being explored as therapeutics, most visibly in cancer immunotherapy. As TCRs typically possess only low-to-moderate affinity for their peptide/MHC (pMHC) ligands, there is a recognized need to develop affinity-enhanced TCR variants. Previous in vitro engineering efforts have yielded remarkable improvements in TCR affinity, yet concerns exist about the maintenance of peptide specificity and the biological impacts of ultra-high affinity. As opposed to in vitro engineering, computational design can directly address these issues, in theory permitting the rational control of peptide specificity together with relatively controlled increments in affinity. Here we explored the efficacy of computational design with the clinically relevant TCR DMF5, which recognizes nonameric and decameric epitopes from the melanoma-associated Melan-A/MART-1 protein presented by the class I MHC HLA-A2. We tested multiple mutations selected by flexible and rigid modeling protocols, assessed impacts on affinity and specificity, and utilized the data to examine and improve algorithmic performance. We identified multiple mutations that improved binding affinity, and characterized the structure, affinity, and binding kinetics of a previously reported double mutant that exhibits an impressive 400-fold affinity improvement for the decameric pMHC ligand without detectable binding to non-cognate ligands. The structure of this high affinity mutant indicated very little conformational consequences and emphasized the high fidelity of our modeling procedure. Overall, our work showcases the capability of computational design to generate TCRs with improved pMHC affinities while explicitly accounting for peptide specificity, as well as its potential for generating TCRs with customized antigen targeting capabilities.

Figure 1. Representative binding affinity measurements.

(a) Steady-state binding equilibrium data for ELA/HLA-A2 binding wild type DMF5 and the βL98W and αG28P mutants. Solid lines represent a fit to a 1∶1 equilibrium binding model. (b) Kinetic titration data for ELA/HLA-A2 binding of the high affinity YW (αD26Y/βL98W) mutant of DMF5. Data are in black in the bottom panel; the red line is a fit to a 1∶1 kinetic titration model with drift. Residuals (difference between data and fitted curve) are shown in the smaller top panel.

Figure 2. ΔΔG (in kcal/mol) for DMF5 point mutants for nonameric (AAG) versus decameric (ELA) peptide bound to HLA-A2.

Solid line denotes equal ΔΔG values, while dashed lines denote a 4-fold affinity shift (0.82 kcal/mol) toward AAG (bottom dashed line) or ELA (top dashed line). AAG and ELA ΔΔG error bars are shown for each mutant, while solid points are the αG28 substitutions selected to shift preference toward the nonameric variant.

Figure 3. Predicted versus measured ΔΔGs for measured DMF5 point mutants binding to ELA/HLA-A2 (solid circles) and AAG/HLA-A2 (empty triangles), using the Rosetta (a, c) and ZAFFI (b, d) functions.

Mutations were modeled in Rosetta without minimization (a, b) or with minimization of interface backbone and side chains (c, d). For (a) and (b), four outlier points with poor measured ELA/HLA-A2 binding and highly unfavorable scores are not shown. For each plot, best fit lines and correlations (all calculated without the four outlier points for consistency) are given.

Figure 4. Structure of the DMF5 YW double mutant in complex with ELA/HLA-A2.

(a) Superposition of the YW/ELA/HLA-A2 and the DMF5/ELA/HLA-A2 complexes. DMF5 α chain is yellow, β chain is tan, peptide is magenta (shown as sticks), MHC is green, and β2m is cyan; residues that were mutated are shown as sticks. Close-ups of (b) wild-type αD26, (c) mutant αY26, (d) wild-type βL98, (e) mutant βW98 are shown. In (b–e), residues proximal to the mutation sites are shown as sticks, and in (c) hydrogen bonds involving the αY26 side chain and a bound water molecule are shown as dashed lines.

Figure 5. Predicted structures of mutant residues (a) αD26Y and (b) βL98W compared with the crystal structure.

Colors for α chain, β chain, peptide, MHC, and mutant side chains from the crystal structure are as in Figure 4. Mutant side chains are shown as sticks, with models in yellow (no minimization), cyan (with minimization), and green (no minimization, in the context of the mutant crystal structure). For simplicity, only the pMHC from the crystal structure is shown.