Deconstructing the peptide-MHC specificity of T cell recognition

Deep sequencing of selections for TCR-binding peptides

Analysis of the pooled yeast library DNA after each successive round of selection with 2B4 via deep sequencing showed enrichment from an essentially random distribution of amino acids to a highly WT-like TCR recognition motif (Figures 2A, S2A). After the third round, there were non-homologous amino acids at P5 and P8 selected above background (Met and Ser for P5, Ile and Leu for P8) that were outcompeted by the WT-like motif by the final round of selection. Overall, the number of unique peptides observed via deep sequencing progressed from 132,000 unique in-frame peptides observed in the sequenced portion of pre-selection library to only 207 unique peptides after the 3^rd round of selection (Figures 2B, 2C, S2A, S2B). By the final round of selection, the library was dominated by a handful of sequences, matching the result obtained by manual curation (Figures 1D, 2B, 2C).

Open in a separate window

Figure 2

Deep sequencing of peptide selections on I-E^k converges on one dominant epitope for 2B4 TCR recognition

(A) Plots for amino acid prevalence at the three primary TCR contact positions (P3 (cyan), P5 (magenta), and P8 (orange)) show the peptide library enriches from even representation of all amino acids in the pre-selection library to a WT-like motif at each position. A secondary preference can be seen at P5 and P8 in round 3 but is outcompeted by round 4. (B) Sequence enrichment of 250 most abundant peptides show a convergence from a broad array of sequences to a few related clones. Area in grey represents all clones other than the most prevalent 250. (C) Comparison of total number of peptides and prevalence of 10 most abundant peptides for each round of selection. See also Figure S2.

We repeated the selections with two other TCRs reactive to MCC-I-E^k: 226 and 5cc7. We analyzed enrichment for each TCR after the third round of selection, where there is enrichment for a binding motif but before complete convergence to a small number of sequences (Figures 2A, ​,3A,3A, S2B, S3A). While all three TCRs retain a WT-like TCR recognition motif (indicated by the outlined boxes in the heatmaps), each TCR also shows some variation in positional preferences (Figure 3A). For example, where 2B4 can recognize P5 Met and Ser, 5cc7 can accommodate P5 Leu, Val, and Arg. The P3 TCR contact position showed the least variance across all three TCRs, with either Phe or Tyr being required for 2B4 and 5cc7, and Phe, Tyr, or Trp being required for 226 (Figure 3A). 226, as previously reported, showed a greater degree of cross-reactivity, able to recognize 897 unique peptide sequences. The larger number of peptides recognized was largely a function of a higher tolerance for substitutions on TCR-neutral and MHC-contacting residues, such as at positions P(-1) and P4 (Figure 3A, S3A) (Ehrich et al., 1993; Newell et al., 2011).

Open in a separate window

Figure 3

Three different MCC/I-E^k reactive TCRs require a WT-like recognition motif in the peptide antigens

(A) Heatmaps of amino acid preference by position for 2B4 (left, red) 5cc7 (center, green) and 226 (right, blue) TCRs after three rounds of selection. The sequence for MCC is represented via outlined boxes. TCR contact residues are labeled red on x axis. (B) Covariation analysis of TCR contact positions P5 (x axis) and P8 (y axis) show distinct coupling of amino acid preferences. (C) Minimum distance clustering of all TCR sequences selected above background show sequences for all TCRs form one large cluster with MCC (black circle, not represented in library but added for reference). Sequence cluster placed in a representation of whole-library sequence space (left: 1x magnification, center: 1000x magnification) for reference. See also Figure S3.

The large collection of peptides recovered via deep sequencing enabled us to apply covariation analysis to discover intra-peptide structure-activity relationships that were not previously accessible with traditional single residue substitution analysis (Figure 3B) (Ehrich et al., 1993; Newell et al., 2011; Reay et al., 1994; Wilson et al., 1999). By using co-variation analysis of the central P5 residue and the C-terminal P8 residue, a pattern emerged: the native, MCC-like ‘up-facing’ TCR-contact motifs for each TCR (P5 Lys, P8 Ser/Thr) were strongly correlated, while the altered residues (P5 Ser/P8Leu for 2B4, P5 Leu or Arg/P8 Phe for 5cc7) independently segregated (Figure 3B). These results highlight a degree of cooperativity in the composition of residues comprising a ‘TCR epitope’ that is clearly revealed with deep sequencing. Furthermore, such intra-peptide residue coupling reveals how cross-reactivity can occur through mutually compensatory substitutions to the parent peptide.

While the selected ligands for all three TCRs possessed shared features, each TCR also selected for a subset of sequences that were not selected by the other two. We applied distance clustering to the peptides selected by all three TCRs to determine if all selected sequences were part of the larger MCC-like peptide family or were distinct families (Figure 3C). We found that while sequences recognized by individual TCRs clustered more closely to each other, essentially all of the selected sequences formed one large cluster of peptides no more than three amino acids different than at least one other peptide in the cluster (Figure 3C, S3B). Therefore, the selected peptides for all three TCRs are related via a common specificity domain, and importantly, to the parent MCC ligand. Even though we conducted unbiased selections of random libraries, the only ligands that were recovered were remarkably similar to the WT ligand at the TCR interface.

Functional characterization of I-E^k library hits

We synthesized 44 of the library peptides selected for binding to the TCRs and examined their ability to stimulate T cell blasts from 2B4 and 5cc7 transgenic mice as assayed by CD69 upregulation and IL-2 production. The majority of the peptides predicted to bind 2B4 (19/19) and 5cc7 (17/21) expressing T cells induced CD69 upregulation (Figures 4A, 4B, S4A-D). The peptides had a wide range of potencies, including ~50-fold more potent than the wild-type peptide MCC (colored red). When we compared the presence of the MCC-like TCR recognition epitope with TCR signaling, we found that in general, sequences that shared the MCC-like epitope at all three major TCR contacts (colored blue) were more potent in inducing signaling than those peptides that were more distantly related (colored black) (Figures 4A, 4B). We also tested the peptides selected for binding to one TCR for their ability to cross-react with the other MCC-reactive T cells. Surprisingly, a large proportion of these peptides potently activated TCR signaling (Figures 4A, 4B, S4A-D). In general, the sequences that showed the most robust activation were again the ones that most closely shared the MCC TCR binding epitope.

Open in a separate window

Figure 4

Relationships between affinity and activity of peptides selected for binding to IE^k-reactive TCRs

(A) EC50s of IL-2 release and CD69 upregulation for 2B4 T cells stimulated with peptides selected with 2B4 TCR, plus MCC (red) (left), or peptides selected with 226 or 5cc7 TCRs (right). Sequences with close homology to MCC at P3, P5, and P8 are represented in blue. Sequences that do not share 3/3 TCR contacts with MCC are in black. (B) EC50s as in A for 5cc7 T cells with peptides selected with 5cc7 (left) or 226/2B4 (right) TCRs. (C) Correlation between TCR-pMHC affinity and peptide signaling potency. Each data point represents one peptide. See also Figure S4.

We additionally chose nine peptides from our initial set of 44 and exchanged them into soluble I-E^k MHC for TCR affinity measurements via surface plasmon resonance (SPR). For 2B4 and 5cc7, TCR bound the pMHC of interest with affinities ranging from KD of ~1 μM (over 10-fold better than MCC) to those with binding only barely detectable at 100 μM TCR (Figure S4E-F). When we compared the activity and affinity of our selected peptides, there is a loose but positive correlation between strength of TCR-pMHC binding and potency of activation (Figure 4C).

The structural basis of TCR recognition of cross-reactive peptides

To determine the molecular basis of the TCRs’ ability to recognize the most diverse peptides selected from our I-E^k libraries, we determined the crystal structures of 2B4 in complex with a peptide termed 2A bound to I-E^k, as well as 5cc7 in complex with two peptides bound to IE^k, termed 5c1 and 5c2 (Table S1, Figures 5A and 5B). When these complexes were aligned with previously solved complex structures of TCRs (2B4 and 226) binding to MCC-I-E^k, very little deviation in overall TCR-pMHC complex geometry from the parent complexes was observed (Figures 5A and 5B) (Newell et al., 2011). Since the 5cc7-MCC-I-E^k complex is not solved, 5c1 and 5c2 were compared to 226-MCC-I-E^k, which shares the TCRβ chain with 5cc7 and therefore likely retains a close footprint. The contacts between TCR germline-derived CDR1/2 loops and MHC helices, which make up roughly 50% of the binding interface between TCR and pMHC, were essentially unchanged in the new peptide complexes versus MCC (Figure 5C).

Open in a separate window

Figure 5

Peptides distantly related to MCC show highly similar mechanism of recognition and linkages to the cognate antigen

Crystal structures of TCR-pMHC complexes for 2B4-2A-I-E^k and 2B4-MCC-I-E^k (PDB ID: 3QIB) (A) and 5cc7-5c1-I-E^k and 226-MCC-I-E^k (PDB ID: 3QIU) (B) compared. TCR contacts are shown in magenta (top, noted with triangles). There is very little change in overall binding geometry despite significant variation of peptide sequence. The TCRs accommodate differences in peptide sequence primarily through differences in CDR3β (bottom). (C) TCR CDR loop footprints for 2B4 recognizing MCC and 2A peptides, 226 recognizing MCC and MCC K99E peptides, and 5cc7 recognizing 5c1 and 5c2 peptides show very little deviation. (D) Relationship between MCC and 2A peptides revealed through intermediate selected peptide sequences. See also Table S1 and Figure S5.

When we examined the chemistry of MCC versus 2A, and MCC versus 5c1 peptide recognition by their respective TCRs, we saw the interaction between the TCRα CDR loops and the N-terminal half of the peptides are essentially invariant (Figures 5A and 5B, lower panels). Each peptide backbone makes a hydrogen bond at the P3 carbonyl with Arg29α in the TCR CDR1α loop (Figure S5A). The contacts of 2B4 CDR3α with P2 and P3 in MCC and 2A are essentially identical (Figure 5A, lower panels). While an exact analogy cannot be made between 5cc7 recognizing 5c1 and 226 recognizing MCC due to sequence differences in their CDR3 loops, 5cc7 and 226 CDR3α loop conformations and peptide contacts are extremely similar (Figure 5B, lower panels).

In contrast, 2B4 and 5cc7 β chain CDR loop interactions with the C terminal half of the peptides show marked changes to accommodate the non-MCC sequences. For 2B4, the CDR3β loop conformation completely rearranges to engage the alternate P5 and P8 residues on the 2A peptide (Figure 5A, lower panels). Gln100β, a residue that makes no contact with the peptide in the 2B4-MCC-I-E^k complex, flips its side chain by 180 degrees to form hydrogen bonds with the peptide backbone carbonyl oxygens at P5 and P6 (Figure 5A, lower panels). The side chains of Trp98β and Ser99β form hydrogen bonds with the P5 Ser hydroxyl moiety (Figure 5A). Asp101β, one of the main contacts with P5 Lys in MCC, forms a hydrogen bond with Ser95β on the other end of the CDR3β loop, significantly altering the overall topology of the loop (Figure S5B).

In the 5cc7-5c1-I-E^k complex, there are fewer hydrogen bonds formed between the peptide and TCR due to the replacement of P5 Lys with Leu in the 5c1 peptide (Figure 5B, lower panels). Asn98β changes its hydrogen bonding network from engaging only the carbonyl of P6 on the MCC peptide backbone to simultaneously interacting with the carbonyl oxygen of P6 and the amide nitrogen of P8 of the 5c1 peptide (Figure 5B). The second 5cc7-reactive peptide, 5c2, is recognized essentially identically by 5cc7 as 5c1 despite the substitution of P5 to Arg (Figure S5C). The substitution of a bulkier side chain at P8 (Phe instead of Thr) results in a rocking of 5cc7 such that the TCR Cβ FG loop is translated by 15 Å relative to the 226-MCC structure (Figure S5D-E). It is interesting to note that all tested peptides with P8 Phe signal less efficiently than MCC-like peptides, even when affinities are closely matched (Figure S4E-F). These structures raise the question if a minor tilt of the TCR relative to the MHC can have consequences for signaling.

Upon closer inspection, we find that homologies between what appear to be unrelated peptide sequences emerge from sequence clustering and structural analysis. For example, close structural relationships between the interaction modes of the 2B4-reactive peptides MCC and 2A are apparent even though the peptides show little homology at 4/5 TCR contact positions (Figure 5A). We set out to determine if we could identify intermediate sequences that would ‘evolutionarily’ link these two peptide sequences, given that both reside in the same sequence cluster (Figure 3C). Using our dataset of peptide sequences selected for 2B4 binding, we were able to populate a family of peptides that would incrementally link MCC and 2A, with each peptide differing by only one TCR contact from the peptide before and after it (Figure 5D). Thus, connectivity can be established between MCC and 2A through stepwise single amino acid drifts from their parent sequences.

Collectively, despite differences in peptide sequences, all MCC and library-peptide derived complexes share many common features with regards to docking geometry and interaction chemistry. Changes in up-facing peptide residue sequence (e.g. P5, P8) are accommodated ‘locally’ in a structurally parsimonious fashion that preserves most of the parent MCC peptide complex features, as opposed to accommodation through large scale repositioning of the CDR loops on the pMHC surface.

Development and selection of a human MHC platform for yeast display

To exploit our technology to find ligands for TCRs relevant to human disease, we also engineered the human MHC HLA-DR15, an allele with genetic linkage to multiple sclerosis (Hafler et al., 2007; Patsopoulos et al., 2013). For yeast surface display, HLA-DR15 was constructed comparably to the murine I-E^k β1α1 ‘mini’ MHC (Figure 6A). We chose to examine two closely-related TCRs, Ob.1A12 and Ob.2F3, that were cloned from a patient with relapsingremitting multiple sclerosis and recognize HLA-DR15 bound to an immunodominant epitope of myelin basic protein (MBP, residues 85-99) (Wucherpfennig et al., 1994b). These two TCRs utilize the same Vα-Jα and Vβ-Jβ gene segments and differ at one position in the CDR3α loop and two positions in CDR3β. Ob.1A12 is sufficient to cause disease in a humanized TCR transgenic mouse model (Harkiolaki et al., 2009; Hausmann et al., 1999; Madsen et al., 1999). A structure of Ob.1A12 complexed with MBP-HLA-DR15 revealed an atypical docking mode, with the TCR shifted towards the N-terminus of the peptide and primarily interacting with a P2-His/P3-Phe TCR contact motif (Figure 6A) (Hahn et al., 2005; Wucherpfennig et al., 1994a).

Open in a separate window

Figure 6

Design and selection of HLA-DR15 based libraries for myelin basic protein (MBP)-reactive human TCRs

(A) HLA-DR15 library design based upon structure of Ob.1A12-MBP-HLA-DR15 complex. Residues P(-4)-P10 are fully randomized, except for the P1 and P4 anchors (in black). TCR contacts are colored magenta. (B) Heatmap of amino acid preference by position for Ob.1A12 TCR. The sequence for MBP is represented via outlined boxes. TCR contacts are labeled red on the x axis. (C) Design and selection results of library that suppresses central ‘HF’ TCR recognition motif at P2-P3 of peptide. Resulting register shift is shown in blue on x axis. (D) Sequence clustering shows distinct, related clusters of selected peptides. Sequence cluster placed in a representation of whole-library sequence space (left: 1x magnification, center: 1000x magnification) for reference. See also Figure S6.

Since the initial wild-type MBP-HLA-DR15 yeast display construct was not stained by Ob.1A12 TCR tetramers, we subjected the construct to error prone mutagenesis and selected for binding with Ob.1A12. Our final construct combined the most heavily selected mutation (Pro11Ser on HLA-DR15β) with two solubility-enhancing mutations on the bottom of the MHC platform (Figure S6A). The final construct stained robustly with Ob.1A12 and Ob.2F3 TCR tetramers (Figure S6B).

We designed a peptide library within the HLA-DR15 mini MHC scaffold to find novel Ob.1A12- and Ob.2F3-reactive peptides (Figure 6A). Since Ob.1A12 binds its cognate pMHC shifted towards the N terminus of the peptide, we extended the library, randomizing from P(-4) to P10 (Hahn et al., 2005). The P1 and P4 peptide anchors for HLA-DR15 were afforded limited diversity. When we selected with Ob.1A12 and Ob.2F3 TCRs, we observed a strong convergence to a wild-type MBP-like TCR recognition motif for the primary TCR contacts (P2 His, P3 Phe, and P5 Lys) (Figure 6B, S6C, S6D).

Given the dominance of the ‘HF’ motif in the selection results, we sought to determine if alternative cross-reactive TCR epitopes would emerge if the motif were suppressed. We made a library that allowed every amino acid except for His at P2, Phe at P3, and Lys at P5 (Figure 6C). After selection, the TCR-binding clones still converged to a central HF motif by register shifting towards the C-terminus of the peptide by one amino acid, allowing the previous P4 Phe anchor to be repurposed as the P3 TCR contact, and the P3 position of the library to become the new P2 His TCR contact (Figure 6C). Furthermore, when we subsequently prevented both His and Phe at P2 and P3 in a new library to suppress potential register shifting, we did not isolate any Ob.1A12- or Ob.2F3-binding peptides (data not shown). These results show that the ‘HF’ motif is required for TCR recognition and its enrichment is a function of TCR preference, not any inherent biases caused by the library or MHC anchor positions of the peptide.

Clustering analysis of the selected peptides for both Ob.1A12 and Ob.2F3 showed distinct clusters consisting of peptides no more than 4 amino acids different from each other (Figure 6D). When the stringency of clustering is increased to allow no more than 3 amino acid differences, matching the analysis done for I-E^k, there were several more sparse clusters (Figure S6E). Since Ob1.A12 and Ob.2F3 are so focused on the HF motif, there are fewer total hotspot residues distributed on the peptide compared to the MCC-reactive TCRs we studied.

Deep sequencing of pMHC libraries

Pooled plasmids from 5×10⁷ yeast from each round of selection were isolated via yeast miniprep (Zymoprep II kit, Zymo Research) and used as PCR template to prepare sequencing samples. The adapter and barcode sequences were appended via nested 25-round cycles of PCR of the purified plasmids using Phusion polymerase (NEB). Deep sequencing was conducted on an Illumina MiSeq sequencer at the Stanford Stem Cell Institute Genome Center.

Profile-based searches for naturally occurring peptide ligands

The positional frequencies from round 3 of the fixed HF library were used to generate a 14×20 substitution matrix. Each protein in the NR (NCBI) or human protein (Uniprot) databases was scanned using a 14 position sliding window and scored as a product of the positional substitution matrix (Cockcroft and Osguthorpe, 1991). In this way, a candidate peptide containing even a single disallowed substitution would be excluded as a possible hit.

Structural determination of pMHC-TCR complexes

All crystallographic data was collected at the Stanford Synchroton Radiation Lightsource (Stanford, CA) beamlines 11-1 and 12-2. Data were indexed, integrated, and scaled using either the XDS or the HKL-2000 program suites (Kabsch, 2010; Otwinowski et al., 1997). All structures were solved via molecular replacement using the program Phaser (McCoy, 2007) and refined with Phenix (Adams et al., 2010).

We thank Suzanne Fischer, Gary Mantalas, and Nelida Prado for technical assistance, and Jarrett Adams, Lauren Ely, and Evan Newell for helpful discussions. M.E.B was supported by a Regina Casper Stanford Graduate Fellowship, a Gerald J. Lieberman Fellowship, and a National Science Foundation Graduate Fellowship. J.L.M was supported by NCIK01CA175127 and a Helen Hay Whitney Foundation postdoctoral fellowship. This work was supported by the NIH (PO1 AI045757 to K.W.W., and R01 AI03867 and U19 4100041120 to K.C.G.), and HHMI (to M.M.D and K.C.G.).