An engineered switch in T cell receptor specificity leads to an unusual but functional binding geometry

Summary

Utilizing a diverse binding site, T cell receptors (TCRs) specifically recognize a composite ligand comprised of a foreign peptide and a major histocompatibility complex protein (MHC). To help understand the determinants of TCR specificity, we studied a parental and engineered receptor whose peptide specificity had been switched via molecular evolution. Altered specificity was associated with a significant change in TCR binding geometry, but this did not impact the ability of the TCR to signal in an antigen-specific manner. The determinants of binding and specificity were distributed among contact and non-contact residues in germline and hypervariable loops, and included disruption of key TCR-MHC interactions that bias αβ TCRs towards particular binding modes. Sequence/fitness landscapes identified additional mutations that further enhanced specificity. Our results demonstrate that TCR specificity arises from the distributed action of numerous sites throughout the interface, with significant implications for engineering therapeutic TCRs with novel and functional recognition properties.

Graphical Abstract

Introduction

Antigen specificity is a hallmark of T cell immunity. Specificity is dictated by the T cell receptor (TCR), which recognizes peptide antigens bound and presented by self MHC proteins using six complementarity determining region loops (CDRs) with varying degrees of diversity. Crystallographic structures have illustrated how TCRs utilize their CDRs to engage peptide/MHC complexes (pMHC). Often, although not exclusively, the most diverse hypervariable CDRs are aligned alongside the antigenic peptide, and the germline-encoded alongside the self MHC (Miles et al., 2015). Although early studies emphasized the role of the hypervariable CDR3α and CDR3β loops in determining antigen specificity, more recent work has emphasized the capacity for specificity to arise from both germline-encoded and hypervariable loops and the composite peptide-MHC surface (Cole et al., 2009; Piepenbrink et al., 2013). Further, although cross-reactivity is often discussed as a necessary feature of TCRs dichotomous to specificity, recent findings show that many cross-reactive ligands share key structural and/or chemical properties, such that even cross-reactivity reflects the specific nature of TCR binding (Adams et al., 2016).

Within this backdrop are ongoing efforts to more deeply understand and control TCR specificity. This is particularly important for emerging immunotherapeutic approaches that utilize exogenous TCRs, such as TCR-engineered T cells and soluble agents. The specificity of these TCRs and the impact of affinity-enhancing modifications remains a significant concern, as illustrated by adverse events in recent immunotherapy clinical trials that resulted from off-target TCR recognition (e.g., Linette et al., 2013).

To better understand TCR specificity and as a step towards productively manipulating it, we recently engineered a switch in the peptide specificity of a TCR by operating on a limited set of CDR residues. Using in vitro directed evolution, we engineered the A6 TCR, which recognizes the viral Tax nonamer (LLFGYPVYV) presented by HLA-A*0201 (HLA-A2), to specifically recognize the anchor-modified cancer MART1 decamer (ELAGIGILTV), also presented by HLA-A2 (Smith et al., 2014). We further engineered a high affinity variant of the RD1-MART1 TCR, termed RD1-MART1^HIGH that bound to MART1 complexes with nanomolar affinity and showed no recognition of Tax or unrelated peptides.

Here, we studied this specificity switch by solving the crystallographic structure of RD1-MART1^HIGH bound to MART1/HLA-A2. This was followed by deep mutational scanning to test the impact of substituting all 20 naturally occurring amino acids in every CDR position of A6 and RD1-MART1^HIGH on binding. The combined structural and mutational scanning results allowed us to concretely establish the determinants of binding and specificity in both interfaces. We found that a limited number of residue changes resulted in a striking change in the TCR binding mode, leading to altered interactions throughout the TCR-pMHC interface. These involved germline and hypervariable loops, and interactions with both peptide and MHC, the latter of which were of surprising importance for the switch in receptor specificity. Indeed, a key driver of the change was the disruption of interactions between the TCR and MHC that in humans biases αβ TCRs towards particular binding modes over HLA-A2 (Blevins et al., 2016). Specificity and affinity in both TCRs were also strongly influenced by ‘second-shell’ CDR residues that do not contact peptide or MHC, allowing the identification of even higher affinity variants of RD1-MART1 and A6.

Despite the change in binding orientation of the RD1-MART1^HIGH TCR, when transduced into T cells the TCR/CD3 complex was fully capable of initiating responses toward the MART1/HLA-A2 complex and not the Tax/HLA-A2 complex. Our findings provide an unprecedented level of insight into how TCR specificity is encoded, and demonstrate new strategies for generating TCRs with novel and fully functional recognition properties.

Results

The RD1-MART1^HIGH TCR binds with a docking mode distinct from other TCR-peptide/HLA-A2 complexes

We previously changed the specificity of the well-studied human TCR A6 from Tax/HLA-A2 to MART1/HLA-A2 by screening a library mutated at five amino acid positions: one in CDR1α, two in CDR3β, and two in CDR3β (Fig. S1). To examine the structural basis for this switch in specificity, we determined the crystallographic structure of the high-affinity variant of this TCR, called RD1-MART1^HIGH, bound to MART1/HLA-A2. The structure was solved at a resolution of 2.5 Å, with two TCR-pMHC complexes in the asymmetric unit (Table 1). The two complexes were nearly identical, with all atoms of the complex superimposing with a root mean square deviation of 0.7 Å (Fig. 1A). Clear electron density was observed throughout the structure, with density lacking for only the Vα/Vβ linker and a few side chains (Fig. S2).

Table 1.

X-ray data collection and refinement statistics for the RD1-MART1^HIGH complex

Data Collection
Resolution range (Å)	20.0 - 2.5 (2.6 - 2.5)*
Space group	P21
Cell dimensions
a, b, c (Å)	50.4, 164.5, 95.2
α, β, γ (°)	90.0, 100.6, 90.0
Unique reflections	47848 (3596)*
R-merge	0.13 (0.55)*
< I/σ (I) >	1.9 (2.1)*
% Data completeness	92.4 (74.3)
Temperature (K)	100
Wavelength (Å)	1.0
Refinement
Resolution range (Å)	20.0 - 2.5
Rwork / Rfree	0.19 / 0.24
No. of protein atoms	9745
No. of water molecules	76
RMS deviation from ideality
Bond lengths (Å)	0.004
Bond angles (°)	0.97
Ramachandran statistics (%)
Most favored	97
Allowed	3
Disallowed	0
Average B-factor
TCR	53.0
peptide	46.0
heavy chain/β2m	69.7
PDB Code	5E9D

^*Values in parenthesis refer to highest resolution shell.

Figure 1. The switch in specificity of RD1-MART1^HIGH is associated with an altered docking mode.

A) Overview of the RD1-MART1^HIGH-MART1/HLA-A2 complex. The two complexes in the asymmetric unit (green and yellow) are essentially indistinguishable, as shown in the superimposition. B) The binding mode of RD1-MART1^HIGH is distinct from the parental A6 TCR (left) as well as the MART1-specific TCRs DMF5 and Mel5 (right), despite all three sharing TRAV12-2. The vertical dashed lines show the incident angle of the various TCRs over pMHC. Horizontal lines show the crossing angles. C) Comparison of the incident and crossing angles of RD1-MART1^HIGH with other HLA-A2 restricted TCRs reveals it is an outlier compared to A6, other MART1-specific TCRs, and HLA-A2-restricted TCRs in general. The graph on the left quantifies incident and crossing angles, while the image on the right compares the binding of RD1-MART1^HIGH to A6 and three other MART1-specific TCRs. The spheres represent the centers of mass of the Vα/Vβ domains of the TCRs, with horizontal and dashed lines reflecting crossing and incident angles, respectively. See also Figs. S1, S2, S4 and S8 and supplemental movie S1.

Surprisingly, we found that RD1-MART1^HIGH TCR adopted a different binding mode over pMHC compared to the wild-type A6 TCR or the A6-c134 high affinity variant, which incorporates some of the same mutations as RD1-MART1^HIGH in the CDR3β loop (Cole et al., 2013; Ding et al., 1999). Unlike the A6 and A6-c134 TCRs, RD1-MART1^HIGH is tilted back and away from the peptide N-terminus (Fig. 1B; left). This binding mode is also distinct from the MART1/HLA-A2-specific TCRs DMF5 and Mel5 (Borbulevych et al., 2011; Cole et al., 2009), a striking observation as the A6, RD1-MART1^HIGH, DMF5, and Mel5 TCRs all utilize the same TRAV12-2 gene segment (Fig. 1B; right) (Baker et al., 2012). A broader comparison showed that the altered binding mode is facilitated by an unusual incident angle, quantified as the tilt of the Vα/Vβ pseudo-symmetry axis with respect to the plane of the pMHC binding groove (Pierce and Weng, 2013). While the crossing angle is within the range observed for other TCRs, the incident angle of RD1-MART1^HIGH is distinct from all other TCRs bound to peptide/HLA-A2 complexes (Fig. 1C; an interpolated animation illustrating the change in binding mode is available as supplemental movie S1). We note that the unusual binding mode of RD1-MART1^HIGH is not associated with a lack of specificity, as we previously demonstrated that RD1-MART1^HIGH does not recognize Tax or other non-cognate peptides tested (Smith et al., 2014).

Architectural changes in the RD1-MART1^HIGH-MART1/HLA-A2 interface

Compared to the complex with the A6-c134 TCR, the amino acid changes and the different binding mode of RD1-MART1^HIGH substantially altered the footprint over pMHC and led to a reduced number of contacts across the interface (Fig. 2A). Using a 4 Å cutoff, RD1-MART1^HIGH forms only 75 interatomic contacts, compared to 171 with A6-c134. The reduction is greatest with the peptide, which is involved in 26 contacts with RD1-MART1^HIGH and 80 with A6-c134. In comparison, there are 124 contacts in the complex of DMF5 with MART1/HLA-A2, of which 48 are to the peptide. The smaller number of contacts in the RD1-MART1^HIGH interface translates into a smaller amount of buried solvent accessible surface area (1427 Ų for A6c134, 2124 Ų for A6-c134). As shown in Fig. 2B, differences in contacts are seen for every loop except for CDR2β, which does not contact pMHC in either complex.

Figure 2. The RD1-MART1^HIGH complex is architecturally distinct from the A6-c134 and DMF5 complexes.

A) Comparison of TCR and pMHC contacted residues in the RD1-MART1^HIGH, A6-c134, and DMF5 complexes. Complexes have been opened up “book” style and amino acids contacted on the pMHC (left) or TCR (right) colored. Red lines either outline peptide (left) or separate Vα and Vβ (right). The number of interatomic contacts and buried solvent accessible surface for each complex is indicated. The color scheme is used throughout the figure. B) Schematic showing contacts between CDR loops and peptide in the three complexes. Line widths are proportional to the number of contacts between each loop and peptide residue. Only those CDR loops making contacts are listed. C) The backbone of the MART1 peptide in the RD1-MART1^HIGH complex is essentially identical to that in the DMF5 complex and the free pMHC (left). However, this conformation differs from that of the Tax peptide (right). D) Detailed view of the altered conformation and side chains of CDR3β. The conformations of the remaining CDR loops are largely unperturbed between the RD1-MART1^HIGH and A6-c134 complex. E) Even though five of six CDR loops adopt the same conformation, the difference in binding mode between RD1-MART1^HIGH and A6-c134 places them in different molecular environments, as shown by superimposing the HLA-A2 peptide binding domains in the two structures and viewing the positions of the loops. See also Figs. S1-S2 and supplemental movie S1.

The conformation of the MART1 peptide in the RD1-MART1^HIGH complex is essentially identical to the conformation in the DMF5 and Mel5 complexes and in the unligated MART1/HLA-A2 complex (Fig. 2C) (Sliz et al., 2001). This conformation, however, differs substantially from that of Tax peptide, reflecting the different surface recognized by the TCR. Although some side chain rotamers are altered (e.g., Trp95α), only the CDR3β loop of RD1-MART1^HIGH significantly alters its conformation relative to A6-c134 (Fig. 2D). This is a surprising observation given the amino acid substitutions and different molecular environments in the RD1-MART1^HIGH and A6-c134 complexes (Fig. 2E). The conformation of the CDR3β loop of the A6 TCR varies considerably in different structures, whereas the other loops remain static (Baker et al., 2012), perhaps reflecting a level of intrinsic flexibility or adaptability in the unbound TCR that is not altered by the various mutations.

Structural drivers of the change in RD1-MART1^HIGH specificity and binding mode

A detailed examination of the RD1-MART1^HIGH–MART1/HLA-A2 interface provided insight into how the amino acid substitutions in RD1-MART1^HIGH work together to alter TCR binding and specificity. An important observation is the alteration of electrostatic interactions between the CDR3α loop and Arg65 on the HLA-A2 α1 helix (Fig. 3A). Accommodation of Arg65 via electrostatic interactions is a characteristic feature of TCR structures with HLA-A2, resulting from the need to offset the desolvation penalty associated with charge burial. This need is strong enough to have influenced the co-evolution of TCR alongside MHC genes (Blevins et al., 2016). Thus, the mutations in RD1-MART1^HIGH broke an apparent “rule of engagement” between αβ TCRs and HLA-A2, with dramatic consequences.

Figure 3. Structural drivers of the change in binding mode and switch in specificity.

A) Cross-eyed stereo view comparing the A6-c134 and RD1-MART1^HIGH interfaces, viewed from the N-terminal ends of the peptides. The difference in the positions of Arg65 and various interacting side chains is apparent, as is the difference in position 30 of CDR1α, which when glutamine interacts strongly with the peptide in the A6-c134 TCR structure, but does not contact the peptide when threonine in RD1-MART1^HIGH. For this and panels B-C, green lines indicate hydrogen bonds, orange lines salt-bridges or cation-π interactions, cyan lines van der Waals contacts. B) As in panel A, but viewed from the C-terminal ends of the peptides. The placement of important tryptophans in RD1-MART1^HIGH is apparent, as are the different mechanisms for engaging the C-terminal portion of the peptide (CDR3β in RD1-MART1^HIGH, CDR1β in A6-c134). C) Cross-eyed stereo view showing how Y50 of CDR2α helps “cap” the p5 Ile side chain of the MART1 peptide in the RD1-MART1^HIGH interface. The p5 Ile side chain is still partially solvent exposed, as shown by the blue surface. D) Comparison of how CDR2α is aligned alongside HLA-A2 α2 helix in the complexes with RD1-MART1^HIGH, A6-c134, DMF5, and Mel5. Yellow lines show interatomic contacts (indiscriminate of contact type). See also Figs. S3 and S4 and supplemental movie S1.

In every known structure of the A6 TCR and its variants, Arg65 is accommodated by a hydrogen bond and salt-bridge from Thr92α and Asp93α. Our previous analyses of these interactions by double mutant cycles revealed they are both very strong, each contributing in excess of -2.5 kcal/mol in binding free energy (Piepenbrink et al., 2013). In RD1-MART1^HIGH, Thr92α is replaced with a lysine, and Asp93α with a tyrosine; thus the TCR is unable to form these same interactions with HLA-A2. In response, Arg65 on HLA-A2 swings around away from its position with A6, likely due to charge repulsion from the new lysine at position 93α (Fig. 3A and supplemental movie S1). In this altered position, Arg65 forms cation-π interactions with Trp95 of CDR3α, which adjusts its position to become sandwiched between Arg65 and the backbone of Gly4 of the MART1 peptide. With the Tax peptide, Trp95α would be unable to assume this position due to clashes with the peptide backbone, illustrating how the altered TCR interactions with the MHC protein contributed to the new peptide specificity. The positioning of the new Lys92 in CDR3α, which as noted above contributes to the different position of Arg65 via charge repulsion, is stabilized by the formation of a new salt-bridge with Asp26 of CDR1α. This interpretation is supported by the loss of binding seen in both the K92Eα and K92Aα mutants of RD1-MART1^HIGH when measured by yeast display titrations with monomeric MART1/HLA-A2 (Fig. S3A).

Another key change is the replacement of Gln30 of CDR1α with threonine (Fig. 3A). With the A6 TCR, the side chain of Gln30α forms a hydrogen bond with Lys66 of HLA-A2, another important interfacial charge that must be accommodated (Blevins et al., 2016). The shorter threonine in RD1-MART1^HIGH is unable to form this interaction. Instead, the new tyrosine at position 93α interacts with Lys66. This interaction can only occur with the altered binding mode; otherwise atomic overlap would occur. The energetic importance of this tyrosine was also substantiated by mutational analysis, as the Y93Fα mutation resulted in about 10-fold weaker binding and the Y93Aα mutation showed no detectable binding in a yeast display titration (Fig. S3B).

In the structures of the TRAV12-2 TCRs DMF5 and Mel5 with MART1/HLA-A2, Gln30α interacts with the glutamate side chain at p1 of the peptide. Thr30α in RD1-MART1^HIGH is unable to form this interaction. Instead, the altered binding mode tips the TCR back so pGlu1 is more exposed than in the DMF5 and Mel5 complexes (Fig. S4A). Rather than performing the functions of Gln30α, the side chain of the threonine at position 30α hydrogen bonds with the backbone of CDR3α, stabilizing the altered loop conformation.

In all structures of A6 variants with Tax/HLA-A2, Glu30 of CDR1β hydrogen bonds with the side chain of the tyrosine at peptide position 8. This interaction is also very strong (Piepenbrink et al., 2013), and indeed, replacement of pTyr8 with alanine substantially weakens peptide potency in functional assays (Ding et al., 1999). As this tyrosine is replaced with a threonine in the MART1 peptide, MART1 would be unable to interact with Glu30β. However, the binding mode of RD1-MART1^HIGH moves Glu30β away from the interface, allowing the backbone oxygen of Met98 of the remodeled CDR3β loop to hydrogen bond with the p8 threonine, and the new tryptophan at position 97β to sandwich between Met98β and Trp95α (Fig. 3B). These observations further illustrate the connection between the new TCR binding mode and the switch in specificity.

The central regions of the MART1 and Tax peptides differ significantly (Fig. 2C), and are accommodated differently in the A6 and RD1-MART1^HIGH interfaces. Tax has a central p5 tyrosine that points directly into a pocket between the CDR3α and CDR3β loops of the A6 TCR. The MART1 peptide has an isoleucine at p5. However, while modifications to p5 of Tax are accommodated by alterations in the size of the pocket via changes in CDR3β (Borbulevych et al., 2009; Gagnon et al., 2006), this does not occur with RD1-MART1^HIGH. Instead, the CDR3β loop is repositioned away from the peptide center (Fig. 3B). This leaves the isoleucine at position 5 of MART1 relatively exposed, pointed towards the HLA-A2 α2 helix. It is partially “capped” by Tyr50 of CDR2α (Fig. 3C), mimicking somewhat how p5 Ile is accommodated in the structure of the Mel5 TCR bound to MART1/HLA-A2 (Cole et al., 2009). The capping of the Ile side chain is important, as RD1-MART1^HIGH is unable to bind the MART1-I5A variant (Fig. S4B), and mutations that could improve capping lead to enhanced binding as discussed below.

Lastly, in all A6, DMF5, and Mel5 complexes, the TRAV12-2 CDR2α loop is aligned alongside residues 154–158 of the HLA-A2 α2 helix, forming a similar pattern of interactions (Baker et al., 2012). The shift in the binding mode of RD1-MART1^HIGH slightly alters this pattern, although Tyr50α is still aligned alongside Gln155, and Ser51α contacts Ala158 (Fig. 3D). As discussed below, these subtle differences in a conserved interaction pattern create opportunities for introducing new mutations that allow improvements in receptor affinity and specificity.

The unusual binding mode of the RD1-MART1^HIGH TCR permits antigen-specific T cell signaling

As noted above, the engineering process used to alter TCR specificity disrupted highly conserved interactions between the TCR and HLA-A2, ultimately leading to the altered orientation of RD1-MART1^HIGH. An important question is whether these changes prevent competent signaling through the TCR/CD3 complex (Adams et al., 2011). To determine whether the RD1-MART1^HIGH TCR could mediate activity, the full-length α and β genes (with the human V regions and mouse C regions) were transduced into the mouse T cell hybridoma 58^−/− which lacks its own αβ TCR but contains the genes for the CD3 subunits (Holler and Kranz, 2003). RD1-MART1^HIGH TCR-transduced cells, but not mock-transduced cells, were positive for staining with MART1/HLA-A2 tetramer (Fig. 4A). The RD1-MART1^HIGH TCR-positive cells were stimulated to release IL-2 by immobilized MART1/HLA-A2 tetramers but not Tax/HLA-A2 tetramers (Fig. 4B). When the RD1-MART1^HIGH TCR-positive cells were incubated with the HLA-A2⁺ T2 cells, the MART1 peptide was also stimulatory, whereas Tax was not (Fig. 4C). These results indicate that despite the altered orientation of RD1-MART1^HIGH over pMHC, its ability to assemble with CD3 subunits and signal efficiently in an antigen-specific manner was not impeded.

Figure 4. RD1-MART1^HIGH TCR mediates activity of T cells in the presence of MART1/HLA-A2 but not Tax/HLA-A2.

A) The coreceptor negative mouse 58^−/− T cell hybridoma cell line was transduced with the full-length RD1-MART1^HIGH TCR and stained with 50 nM MART1/HLA-A2 tetramer (black lined histogram). Mock-transduced cells were also stained with 50 nM MART1/HLA-A2 tetramer (gray filled histogram). B) IL-2 release from RD1-MART1^HIGH (filled bars) and mock (open bars) T cells incubated in wells coated with Tax/HLA-A2 tetramer, MART1/HLA-A2 tetramer, or anti-CD3 antibody. C) IL-2 release from RD1-MART1^HIGH T cells incubated with T2 antigen presenting cells and various concentrations of MART1 peptide (filled circles), Tax peptide (open circles) or no peptide (gray circles).

Deep mutational scanning of the A6-c134 and RD1-MART^HIGH TCRs

To obtain a comprehensive view of how the residues of the RD1-MART1^HIGH and A6-c134 CDR individually contribute to ligand binding, we performed deep mutational scanning with both TCRs, using yeast displayed TCR mutants (Fig. S5A) (Procko et al., 2013). Yeast-display TCR libraries were sorted with concentrations of monomeric peptide/HLA-A2 ligand that ranged from 5-10 fold below to 5-10 fold above the K_D of the parental TCR (Fig. S5B,C). After deep sequencing of naïve and sorted libraries, the change in frequency of each single amino acid variant was calculated. This calculated enrichment ratio is a proxy for sequence fitness: TCR variants that bind pMHC with high affinity are enriched, while deleterious mutations are depleted.

The sequence/fitness landscapes determined from library selections at different pMHC concentrations yielded similar profiles (presented as heat maps; Fig. 5 and Fig. S6). As a validation of this approach, correlation plots between the enrichment values for each codon library selected at different concentrations showed excellent agreement for the positively enriched variants, and significant, although quantitatively more variable, agreement for negatively selected variants (Fig. S6C). For further analysis, we focused on selections conducted at the lowest ligand concentrations for each TCR (Fig. 5A,B).

Figure 5. Deep Mutational Scanning of A6-c134 and RD1-MART1^HIGH.

All single amino acid substitutions were expressed in a yeast display library of A6-c134 (A) and RD1-MART1^HIGH (B). Mutants that bound with the top 0.5% fluorescence signal for Tax/HLA-A2 at 5 nM (A) and top 1% fluorescence signal for MART1/HLA-A2 at 10 nM pMHC (B) were collected by FACS. Sequences were then analyzed using next generation sequencing. The sequence enrichment compared to the naïve library was calculated as a log₂ ratio. Enrichment ratios are color-coded from ≤ 2⁻³ (orange) to ≤ 2⁵ (blue). Stop codons are denoted with an asterisk. See also Figs. S5-S8.

To examine the impact of substitutions on overall protein stability and the inherent variability in the deep sequencing approach, the analysis of the A6-c134 TCR included a collection of residues that were present in exposed framework (FR) residues, many at the opposite end of the TCR (Fig. S7). As expected, these residues showed weaker selection (either positive or negative) compared to most of the CDR positions (Fig. 5A). However, there were some residues whose substitutions were depleted, likely due to protein destabilization (Shusta et al., 1999). Substitutions of Gly40 in Vβ and Gly15 in Vα were nearly universally deleterious, as were substitutions of Asp77, Ser78 and Gln79 in Vα, suggesting stabilizing structural roles for these residues (Fig. S7). Two FR residues in RD1-MART1^HIGH were examined and substitutions in each had minimal effects compared to those in CDR loop residues (Fig. 5B).

Comparison of the impact of A6-c134 versus RD1-MART^HIGH CDR residues on binding

As expected, the most significant effects on pMHC binding were in the CDR loops, especially the two CDR3 loops. When examining either the alanine substitution results alone (Fig. 6A,B) or the collective results of all 20 amino acids, many of the key residues were focused at the TCR-pMHC interface. The identity of these residues was in excellent agreement with the conclusions drawn from the crystallographic structures. For example, the structures suggested that Asp26α in RD1-MART1^HIGH stabilizes the conformation of Lys92α, forcing the novel orientation of Arg65 in the α1 helix of HLA-A2. Supporting this interpretation, only the conservative D26E substitution at position 26α permitted equivalent binding to the MART1 complex, based on similar enrichment ratios. Lys92α was required, as was the neighboring Tyr93α, which as noted above offsets the charge of Lys66 of HLA-A2 via a cation-π interaction (Fig. 4A and Fig. S3).

Figure 6. CDR Residues critical in binding pMHC and those residues that can potentially enhance affinity.

A) CDR residues of A6-c134 shown to have the largest reduction in Tax/HLA-A2 binding upon mutagenesis, based on the alanine results of the deep mutational screen. B) CDR residues of RD1-MART1^HIGH shown to have the largest reduction in MART1/HLA-A2 binding upon mutagenesis, based on the alanine results of the deep mutational screen. C) CDR residues of A6-c134 where mutations were capable of increasing binding to Tax/HLA-A2. D) CDR residues of RD1-MART1^HIGH where mutations were capable of increasing binding to MART1/HLA-A2. For (A) and (B), alanine mutation enrichment values were averaged for all selection conditions (A6-c134 n=2, RD1-MART1^HIGH n=3); residues with the ten lowest values are highlighted in yellow. Glycine residues were excluded from analysis due to their likely disruption of backbone conformation rather than specific side-chain contribution. For (C) and (D) residues are highlighted in yellow where at least one mutation in the single codon analysis showed an enrichment of 8-fold or greater than that of the wild-type residue.

Position 30 in CDR1α is another key position underlying the specificity switch. Threonine is required for high affinity pMHC binding by RD1-MART1^HIGH, whereas the wild-type glutamine is required in A6-134 (Fig. 5). This is consistent with a role for Gln30α in “pinning down” the Vα domain of A6-c134, whereas the shorter threonine in RD1-MART1^HIGH permits the TCR to tilt away from the peptide N-terminus, exposing pGlu1 as shown in Fig. 4A and Fig. S3A.

In the case of CDR1β and CDR2β residues, a variety of substitutions permitted strong binding by RD1-MART1^HIGH, reflecting the lack of involvement of these loops in direct ligand recognition as shown by the structure. Similar results were found with A6-c134, with the exception of a requirement for Glu30 in CDR1β (Fig. 5A), consistent with the strong hydrogen bond it forms with the Tax pY8 side chain (Piepenbrink et al., 2013).

In addition to contact positions, for both TCRs there were also significant effects in positions that do not directly contact ligand. These amino acids appear to be important for binding site architecture. This was particularly obvious in RD1-MART1^HIGH (Fig. 6B), as mutations in several residues in CDR1β and CDR2β that are in proximity to the start of CDR3β showed significant effects on binding. These residues are likely critical in orienting key residues for their interactions with peptide and MHC. As noted with the scan of FR residues, some of these effects could also result from a reduction in TCR stability. We suspect that this is particularly true of the impacts seen with glycines or hydrophobic residues (e.g., I49α) that are packed in the center of a CDR loop.

Finally, it is worth noting the distributed nature of residues found to have the greatest impact on binding their pMHC ligands, for both RD1-MART1^HIGH and A6-134 (Fig. 6A,B). Some of these residues were shared between RD1-MART1^HIGH and A6-134, but many differences were observed even though they had not been mutated to yield the specificity switch. Thus, both binding and specificity are the product of multiple, distributed CDR residues and how they interact with both peptide and MHC.

Distinct differences in the contribution of CDR residues in A6-c134 versus RD1-MART^HIGH

To examine the role of several key residues in more detail, we focused on four residues in A6-c134 and RD1-MART1^HIGH that were of particular significance: position 30 in CDR1α (Gln or Thr), position 50 in CDR2α 50 (Tyr), position 91 in CDR3α (Thr), and position 92 in CDR3α (Thr or Lys). The enrichment values for substitutions at these four positions were compared directly in A6-c134 and RD1-MART1^HIGH (Fig. 7A). As indicated above, for position 30 in CDR1α, glutamine in A6-c134 and threonine in RD1-MART1^HIGH each were important, and neither amino acid could substitute for the other in their respective TCR context. To confirm this, the individual mutations were introduced into A6-c134 and RD1-MART1^HIGH and titrations performed with monomeric pMHC (Fig. 7B). Consistent with the inferences from the mutational scanning and structural analysis, substitution of threonine into A6-c134 reduced binding to Tax/HLA-A2 and substitution of glutamine into RD1-MART1^HIGH eliminated binding to MART1/HLA-A2. These results confirm the differential contributions for the amino acid at this key CDR1α position to the specificity of both TCRs.

Figure 7. Enrichment and binding comparison of select mutants from A6-c134 and RD1-MART1^HIGH.

A) Enrichment ratios from single-codon library sorts of A6-c134 and RD1-MART1^HIGH were compared at residues Q30α (A6-c134) or T30α (RD1-MART1^HIGH), Y50α (A6-c134 and RD1-MART1^HIGH), T91α (A6-c134 and RD1-MART1^HIGH), and T92α (A6-c134) or K92α (RD1-MART1^HIGH). The data for several key mutations that were analyzed further using single-site mutations are indicated by the upward arrows. B) Mutant Q30Tα of A6-c134 or mutant T30Qα of RD1-MART1^HIGH. C) Mutants Y50Wα and Y50Aα of both A6-c134 and RD1-MART1^HIGH. D) Mutants T91Dα and T91Qα of both A6-c134 and RD1-MART1^HIGH . TCR mutants were expressed on the yeast surface, and cells were stained with Tax/HLA-A2 and MART1/HLA-A2 monomer respectively. Mean fluorescence intensity (MFI) is plotted against peptide/HLA-A2 monomer concentration. Titrations are representative of two experiments with similar results. See also Fig. S7

The enrichment data for position 50 in CDR2α also showed a significant negative impact with all substitutions, except that a tryptophan was well tolerated in RD1-MART1^HIGH (Fig. 7A). Separate alanine mutations in both A6-c134 and RD1-MART1^HIGH showed reduced binding of their respective ligands (Fig. 7C), but as predicted by deep mutational scanning, mutation of Y50α to tryptophan enhanced binding with RD1-MART1^HIGH but weakened binding with A6-c134 (Fig. 7C). The need for a tyrosine at position 50α is consistent with a previously determined role for Tyr50α in TRAV12-2 TCRs in interacting with the HLA-A2 α2 helix (Smith et al., 2013). As discussed above, these interactions are altered in the RD1-MART1^HIGH complex, with Tyr50α “capping” the isoleucine at peptide position 5 in addition to interacting with the α2 helix (Fig. 4C). The bulkier tryptophan likely buries additional isoleucine non-polar surface without disrupting the interactions with the helix, leading to improved binding. Thus Y50α makes key interactions with the MHC in both TCRs, but the altered binding mode of RD1-MART1^HIGH provides an opportunity for enhancing binding that is absent with A6-c134.

Although position 92 of CDR3α (threonine in A6-c134 and lysine in RD1-MART1^HIGH) contributed to peptide specificity as illustrated by the importance of these alternative residues in their respective TCRs (Fig. 7A), we noticed that the adjacent position threonine 91α could be substituted with amino acids that either improved or reduced binding. Two of the substitutions (glutamine and aspartic acid) had opposing effects with the two TCRs. To substantiate this and examine the magnitude of the effects, the two mutations were introduced into each TCR (Fig. 7D). The T91Q mutation nearly eliminated binding of Tax/HLA-A2 by A6c-134 but significantly enhanced binding of MART1/HLA-A2 by RD1-MART1^HIGH. Remarkably, the T91D mutation had exactly the opposite effect. In both structures, Thr91α is at the N-terminal start of the CDR3α loop. Beyond this though, the altered binding mode and loop structure leads to a local environment that is quite different between the TCRs, which can explain the opposing effects of the mutations in the two TCRs (Fig. S8A).

Second-shell positions that yielded improvements in binding to the peptide/HLA-A2 complexes

The sequence/fitness landscapes showed that a high fraction of substitutions (7.8% in A6-c134 and 8.3% in RD1-MART^HIGH) yielded improvements in binding or stability for RD1-MART1^HIGH and A6-c134. In both TCRs, these improvements were distributed at various positions among the CDR loops, and thus did not indicate the existence of a sub-optimal sequence for any single loop.

To gain a structural perspective in the ability of the substitutions to generate improvements, we highlighted the positions of the most improved residues (based on either enrichment of a single substitution or the ability of multiple substitutions at the same position to yield enrichment) (Fig. 6C,D). Two features were apparent from this analysis. First, in the cases of both A6-c134 and RD1-MART1^HIGH, the positions that yielded enhancements were often in ‘second-shell’ residues, i.e., those that were not in contact with ligand. This reflects what was seen with Thr91α (Fig. 7A,D) and has been observed in other affinity maturation efforts (Boder et al., 2000). Many of these changes likely allow subtle reconfiguration or stabilization of the CDR loops for optimal binding.

Second, the majority of positions that yielded improvements were not shared between the two TCRs, despite the high level of TCR sequence identity and their common origin (i.e., the A6 TCR). Key examples of the opposite effects of a single position in the two TCRs were the T91D and T91Q mutations in CDR3α and the Y50W mutation in CDR2α, as described above. This finding is consistent with our general conclusion that multiple sites working collectively are required for achieving specificity.

Discussion

The RD1-MART1 variant of A6 represents the first example of a TCR custom engineered to possess a novel specificity. The structure of the complex combined with deep mutational scanning allowed us to ascertain the underlying mechanism of the specificity switch, and in turn to gain unprecedented insight into the determinants of TCR specificity and affinity and how they may be productively manipulated.

A key observation is the unusual tilt or incident angle of RD1-MART1^HIGH over MART1/HLA-A2, which substantially alters interactions throughout the TCR-pMHC interface. Importantly its formation is not simply due to introduction of new “attractive” interactions, but also mutations that destabilize the binding mode of the parental A6 TCR. A prime example is the replacement of negative with a positive charge in a crucial region of the interface, catalyzing the rearrangement of Arg65 on the HLA-A2 α1 helix and the subsequent formation of several new stabilizing interactions. The consequences that follow from this substitution in CDR3β reinforce our recent finding that accommodation of Arg65 strongly influences how human TCRs bind class I MHC proteins, and HLA-A2 in particular (Blevins et al., 2016).

Another example of an altered interaction that involves the MHC protein is the “repurposing” of Trp95α to interact with Arg65 and pack between the peptide and the α1 helix. Thus, the new position of the TCR over the pMHC and the resulting change in specificity can be attributed to the interplay of mutated and non-mutated amino acids across multiple CDR loops and how they interact with the composite peptide/MHC surface. Our experiments thus reveal a more “distributed” view of TCR specificity than is usually considered. This point is reinforced by our finding through mutational scanning that multiple positions throughout the TCR binding site, and non-contact residues in particular, could be mutated to alter both the specificity and affinity of RD1-MART1^HIGH and A6-c134.

It is interesting that RD1-MART1^HIGH maintains an alignment between CDR2α and HLA-A2 seen in structures with all other TRAV12-2 TCRs (Baker et al., 2012). The residues comprising this motif, and Tyr50α in particular, are important in the binding of A6-c134 and RD1-MART1^HIGH as indicated from the mutational scanning of this region. However, tryptophan at position 50α disrupts binding with A6c-134 but improves binding with RD1-MART1^HIGH. This indicates that the structural differences in this motif between RD1-MART1^HIGH and the other TRAV12-2 TCRs are sufficient to alter the energetics of these conserved interactions. This observation highlights the difficulty of identifying energetically significant interactions between TCRs and MHC proteins from structure alone. However, it also indicates how mutational scanning can be used to alter or even improve the specificity of TCRs by introducing substitutions that simultaneously weaken canonical interactions but enhance binding to a specific peptide/MHC complex (e.g., the opposite effects of the Y50Wα substitution in A6-c134 and RD1-MART1^HIGH).

Another important finding is the ability of many mutations in residues that do not contact pMHC to strengthen TCR binding. The presence of such ‘second shell’ mutations has been noted in other applications of affinity maturation (Boder et al., 2000), and indeed, in many affinity-enhanced TCRs mutations are found in residues that do not contact ligand. In some high affinity TCRs, however, contact residues have been mutated (e.g., Holler et al., 2000), and efforts using structure-guided design to generate high affinity TCRs have almost exclusively emphasized contact residues (e.g., Pierce et al., 2014). As mutating contact residues could potentially introduce new specificities by improving chemical complementarity with new peptides, these ‘second-shell’ sites are potentially superior candidates for mutation to improve binding. While more work is needed to understand how these mutations impart their effects, the combination of structure-guided engineering and comprehensive mutational analysis provides a new and powerful strategy to improve both TCR specificity and affinity.

Lastly, the mechanisms of TCR triggering remain widely debated. One model of TCR triggering places fine constraints on TCR binding geometry, and indeed, unusual binding has been associated with a loss of signaling with the 42F3 TCR (Adams et al., 2011). We found here that despite the unusual binding mode of RD1-MART1^HIGH compared to other HLA-A2 complexes, it was fully capable of mediating antigen-specific signaling in T cells. This could indicate that T cell signaling mechanisms are more tolerant to incident rather than crossing angles, as the signaling-incompetent 42F3 TCR has an altered TCR crossing angle, which is not observed here. Intriguingly through, two TCRs from regulatory T cells that bind pMHC with “reversed polarity” still signal (Beringer et al., 2015). Although additional studies relating TCR binding geometries to signaling are obviously needed, our results nonetheless indicate that T cell signaling mechanisms are tolerant to engineering approaches that alter TCR binding modes.

Experimental Procedures

Reagents and flow cytometry

Reagents used to induce and detect yeast surface expression of single-chain TCRs have been described previously (Smith et al., 2014). Peptides used in complex with HLA-A2 were synthesized by the Macromolecular Core Facility at Penn State University College of Medicine. An ultraviolet-cleavable peptide, KILGFVFJV, was synthesized by GenScript. HLA-A2 heavy chain was expressed as inclusion bodies in Escherichia coli and refolded in vitro with an ultraviolet-cleavable HLA-A2 binding peptide and human β₂microglobulin as described previously (Rodenko et al., 2006). The HLA-A2 heavy chain contained a biotinylation substrate sequence for in vitro biotinylation (Avidity, BirA enzyme). Peptide exchange with complexes containing the UV-cleavable peptide was achieved by exposure to UV light in the presence of 100-fold excess peptide.

Single-codon library design and selection

Single-codon libraries of the A6-c134 and RD1-MART^HIGH TCRs were cloned as single chains (Vβ-linker-Vα) for yeast display (Aggen et al., 2011). Primers were synthesized by Integrated DNA technologies and PCR products containing degenerate NNK codons were generated via overlap extension PCR. Pooled single-codon PCR products for each TCR were introduced along with digested pCT302 vector into yeast and libraries of independent transformants generated by homologous recombination. The A6-c134 and RD1-MART^HIGH libraries contained 3 million and 1 million independent clones, respectively, which exceeded by several orders of magnitude the potential diversity of the libraries (e.g., 40 positions × 32 codons, or 1280 for the RD1-MART^HIGH library). Libraries were stained using pMHC monomer and sorted by FACS.

Transduction and activation of T cells

The RD1-MART1^HIGH full-length TCR construct (VαCα/VβCβ with mouse constant region genes) was cloned into the pMP71 retroviral vector and transfected in the platinum-E retroviral packaging cell line. Retroviral supernatants were harvested after 48 h and 58^−/− cells were transduced as described previously (Stone et al., 2014). After cells were transduced they were sorted via FACS for binding to MART1/HLA-A2 tetramer to obtain a positive staining population. To measure IL-2 release, 7.5×10⁵ cells were incubated in wells coated with MART1 or Tax tetramer (50 nM), anti-CD3 antibody (10 μg/mL) or with 7.5 × 10⁵ T2 cells pulsed with various concentrations of peptide. After 24 h at 37°C, supernatants were collected and analyzed for IL-2 via ELISA (Stone et al., 2014).

Deep sequencing analysis

Unsorted and sorted yeast libraries (4×10⁷ cells) were lysed and DNA was isolated according to manufacturer's protocol (Zymo Research). Genomic DNA was subsequently removed by incubation with 2000 U/mL Exonuclease I and 250 U/mL Lambda Exonuclease for 90 minutes, and plasmid DNA was purified using a PCR cleanup kit. Deep sequencing analysis was performed as described (Procko et al., 2013). Single-chain TCRs were amplified in two fragments (Vβ and Vα) through two sequential rounds; the first round added overhangs for annealing to Illumina sequencing primers, and the second round added adaptamers for annealing to the MiSeq chip. PCR products were gel purified and sequenced (MiSeq sequencer) using 250 nt paired-end sequencing runs. The naïve libraries had 1.2×10⁶ (RD1-MART1^HIGH) and 2.4×10⁶ (A6-c134) reads, and targeted single amino acid mutations ranged from 15–3.4×10⁴ (RD1-MART1^HIGH; median 798) and 38–4.7×10⁵ (A6-c134; median 1016) reads. All mutations were therefore well sampled in the libraries. Sorted libraries had 4.2×10⁵ – 1.9×10⁶ reads. Enrichment ratios, E^x_i = log₂ (f^x_i[sorted] / f^x_i[naïve]) where x is the amino acid identity, i is the position and f is the sequence frequency, were calculated using modified scripts from Enrich software (Fowler et al., 2014).

Crystallization, data collection, structure refinement, and analysis

Single chain, soluble RD1-MART1^HIGH, HLA-A2 heavy chain, and β₂m for crystallization were generated from bacterially produced inclusion bodies as previously described (Smith et al., 2014). For producing the TCR-pMHC complex for crystallization, pMHC was first generated in a standard 1L refolding reaction, then inclusion bodies for the TCR were added (TCR:pMHC molar ratio of 1:3). The refolding buffer was 50 mM Tris (pH 8), 2 mM EDTA, 2.5 M urea, 400 mM L-arginine, 9.6 mM cysteamine, 5.5 mM cystamine, 0.2 mM PMSF. The TCR was allowed to refold in the presence of pMHC overnight. The reaction was then desalted by dialysis at 4°C and the complex purified by anion exchange followed by size-exclusion chromatography.

Crystals of the RD1-MART1^HIGH-MART1/ HLA-A2 complex were grown in 15% PEG4000, 0.1 M sodium citrate (pH 5.6), 5% 2-propanol, and 2mM zinc acetate at 25°C. Crystallization was performed using hanging drop/vapor diffusion. For cryoprotection crystals were transferred into 20% glycerol/80% for 30 seconds and immediately frozen in liquid nitrogen. Diffraction data were collected at SER-CAT (22-ID) at APS, Argonne National Labs. Data reduction was performed with HKL2000. The structure was solved by molecular replacement using MOLREP, using PDB entry 1QRN and the HLA-A2 heavy chain and TCR variable domains as a search model. Rigid body refinement followed by NCS torsion angle restraints, translation/libration/screw refinement, and multiple steps of restrained refinement were performed using Phenix. Anisotropic and bulk solvent corrections were taken into account throughout. Post refinement, it was possible to unambiguously trace the peptides and TCR CDR loops against σ_A-weighted 2F_o-F_c maps. Evaluation of models and fitting to maps were performed using COOT. MolProbity was used to evaluate the structure during and after refinement. Iterative build OMIT maps were calculated in Phenix. Analysis of hydrogen bonds used a donor-acceptor maximum distance of 3.6 Å. Contacts were calculated with a 4 Å cutoff. Surface areas were calculated using a 1.4 Å radius probe. TCR crossing angles were calculated as described by Rudolph et al., 2006, using a line between the centroids of the Vα/Vβ disulphide bonds relative to a line through the α carbons of the MHC α helices. Incident angles were calculated as described by Pierce and Weng, 2013, using the MHC helix α carbons to define the plane of the binding groove and the TCR Vα/Vβ symmetry axis determined using the FAST structure alignment algorithm. The supplemental movie was generated with Chimera using corkscrew interpolation between the A6-c134 and RD1-MART1^HIGH structures with a sinusoidal interpolation rate.

Highlights.

• Switching TCR specificity via in vitro evolution leads to an unusual binding mode

• The engineered TCR still signals in an antigen-specific manner

• Disruption of canonical TCR-MHC interactions catalyzes the specificity switch

• Further contributions to the switch are distributed throughout the interface