Yeast cell libraries displaying HLA-B*57:03-β2m linked to a randomized peptide library were generated to screen the breadth of AGA1 TCR binding (see Figure 1A (i) for overview of strategy). The peptide-β2m-MHC library construct was based on previous designs (Hansen et al., 2009; Birnbaum et al., 2014a; Gee et al., 2018) and comprised a single peptide test reagent (KF11) or randomized peptide libraries linked to human β2m and the heavy chain of HLA-B*57:03 encoding a Y84A mutation. A myc-tag incorporated between the MHC α3 domain and the Aga2 yeast protein allows the monitoring of peptide-β2m-HLA-B*57:03 expression on the yeast cell surface (Figure 1A (ii)). The functionality of yeast displayed HLA-B*57:03 was confirmed initially by staining a control KF11-β2m-HLA-B*57:03 yeast display with fluorescently conjugated AGA1 TCR tetramers (Figure 1B). Following validation, mutagenized peptide libraries were produced. For the preparation of these libraries, peptide residues at positions 2 (Ala/Ser/Thr) and 11 (Phe/Trp/Tyr) (CΩ) were restricted to retain the preferred HLA-B*57:03 anchor binding motifs (Barber et al., 1997), whilst the remaining nine amino acids were randomized using degenerate NNK nucleotide addition as described previously (Adams et al., 2011; Figure 1A (ii)). The theoretical diversity was 1.06E15 unique nucleotide sequences (4.61E12 unique protein sequences), of which the experimental library contained 2 × 10⁸ unique transformants.

To evaluate the peptide repertoire that allowed AGA1 TCR binding in the context of HLA-B*57:03, streptavidin-coated magnetic beads (MACS) conjugated to biotinylated AGA1 TCR were employed to perform iterated rounds of library selection. The enrichment of library peptides was tracked via deep sequencing analysis. By Round 3 of selection, clear evidence of peptide sequence enrichment emerged for variants that resembled the known KAFSPEVIPMF (KF11) peptide. Of the roughly 2.1 × 10⁶ reads detected from deep sequencing runs, the top 20 selected peptides represented 89.4% of recovered reads at Round 3, and three peptide sequences accounted for approximately 35% of these reads (Figure 1C).

Distinct patterns of individual amino acid enrichment/fixation within the peptide sequences had also emerged by Round 3 (Figure 2, Figure 2—figure supplement 1). The most striking distribution of amino acids included the near exclusive selection of peptides with a conserved central p5Pro-p6Glu motif, which contrasted with the greater diversity tolerated at auxiliary positions along the remaining peptide sequences (Figure 2A). Given that the majority of interactions are formed between CDR3 alpha chain amino acids and the KF11 peptide Glu6 side chain in the published HLA*57:03-KF11-AGA1 TCR structure (Stewart-Jones et al., 2012; Figure 2B), fixation of this central motif most likely reflects its critical importance for AGA1 TCR binding. Amino acid selection outside the p5Pro-p6Glu di-motif was variable, although patterns of conservation specific to the physiochemistry of residues were apparent; the strong enrichment of hydrophobic amino acids at p3 and p7, and to a lesser extent at p10, for example, reflected the chemistries of these residues in the KF11 peptide (p3Phe, p7Val, p10Met, respectively). Similarly, p4 position - a solvent-exposed Ser and an important AGA1 TCR contact residue in the KF11 peptide - displayed a specific enrichment of polar residues in library-derived peptides. Selection patterns were less obvious at positions 1, 8 and 9, where mixtures of both polar and non-polar residues were tolerated.

The top library peptide sequences from Round 3 enrichments were used as input to generate a position probability matrix, used previously to predict antigens to score peptide sequences using a sliding window along proteins from the ‘non-redundant’ (nr)-database containing both human and microbial proteins (Gee et al., 2018; Birnbaum et al., 2014b). From this search, six KF11-like hits were returned (Supplementary file 1 - Table 1), implying that the yeast display selection results could uncover the cognate specificity of this TCR. However, a substantial number of microbial peptide hits that do not resemble the KF11 peptide dominated the sequence homology searches (Supplementary file 1 -Table 2). From the top 20 peptides hits retrieved, six microbial peptide hits, (Mic 1 to 6), were selected for follow-up in T cell functional studies (Table 1); their selection was biased to include sequence from microbes potentially encountered by humans - either those associated with the environment (e.g. soil-derived), or those previously isolated from, or linked to infections in humans (Chávez de Paz et al., 2004; Chhour et al., 2005; Wolfgang et al., 2012). Peptides identified from the yeast display library (Lib 1 to 6) and individual ‘nr’-derived microbial peptides predicted to be recognized by the AGA1 TCR were synthesized and tested for their ability to induce IFNγ secretion from an AGA1-expressing T cell clone (Figure 3A (i-ii) and Table 1 for the sequence identity of peptides). All responses were compared to the index KF11 gag peptide epitope. In terms of the yeast display-derived peptides, five of the six tested resulted in a measurable response. These peptides elicited IFNγ production that varied significantly across the peptide titration range, and one library peptide (Lib 2) failed to induce IFNγ production despite its selection by the AGA1 TCR in yeast display enrichment screens. Of the ‘nr’-mined microbial peptides, three activated the AGA1+ T cell clone, of which two (Mic 1 and 3) peptides mapped to microbes previously linked with gingival (Chávez de Paz et al., 2004; Chhour et al., 2005) and gastrointestinal (GI) infections in humans (Wolfgang et al., 2012). In terms of IFNγ production, responses to these peptides approached the magnitude elicited by the index KF11 peptide. The remaining three ‘nr’-derived microbial peptides drove weak or no T cell responses.

We questioned if the differential ability of the various library and microbial peptides to elicit T cell responses primarily reflected binding differences to HLA-B*57:01. This is particularly relevant for the Lib-derived peptides, which were bound to β2m as part of the HLA-B*57:03 yeast display single chain trimer construct but were presented in cellular assays as non-linked peptide epitopes in complex with HLA-B*57:01 on antigen presenting B cells. To address this, we performed a sandwich ELISA-based UV exchange peptide binding assay to gauge the ability of the various peptides to bind and therefore ‘rescue’ HLA-B*57:01 pre-refolded with UV-labile peptide upon photo-illumination. In terms of the library-derived peptides, there was a strong concordance between the magnitude of T cell responses generated and the extent to which the peptides stabilised HLA-B*57:01 (Figure 3B). Lib 2 peptide, for example, demonstrated the weakest binding to HLA-B*57:01 and represented the only library-derived peptide that failed to elicit functional responses by AGA1-expressing T cell clones. A similar pattern existed for the microbial peptides, where the largest T cell responses were generated by the strongest HLA-B*57:01 binding peptides. Interestingly, four peptides exhibited enhanced binding relative to the index KF11 peptide, including the Mic 1 peptide, which corresponds to a Sporosarcina newyorkensis-derived halodehydrogenase (HdH) peptide, and its closest library sequence match, Lib 1.

We re-analyzed responses to the database-derived microbial peptides using a second T cell clone isolated from a different donor that expressed an AGA1-homologous TCR, namely clone 1.2 (Figure 3C and Supplementary file 1 - Table 3). The functional screens were extended to assess both IFNγ production and CD107a up-regulation in response to the microbial peptides. This T cell clone also demonstrated a clear preference for the ‘nr’-mined Olsenella uli, Candida orthopsilosis peptides and in particular, the S. newyorkensis peptide, both in terms of IFNγ production and up-regulation of the LAMP protein, CD107a (Figure 3C (i) and (ii), respectively).

As the Mic 1 peptide, which mapped to the haloacid dehydrogenase (HdH) enzyme of S. newyorkensis, elicited the strongest T cell-mediated responses at concentrations comparable to the KF11 epitope, we chose this peptide for follow-up studies. We re-interrogated the ‘nr’ database using this peptide motif to hunt for related peptides. A number of highly homologous peptides were identified of which the majority mapped to HdH enzymes specific to distinct microbial genera. Ten additional HdH peptides (Table 2) were synthesized and tested in T cell functional assays. The majority of these peptides elicited IFNγ production at titrations comparable to the S. newyorkensis (HdH1/Mic1) peptide, and only marginally weaker than responses induced by the index KF11 epitope (Figure 4A). One peptide (HdH2), was recognized at the highest peptide concentration only, whereas two additional peptides (HdH5 and HdH6) failed to elicit T cell responses. To address if this related to HLA-B*57:01 binding or recognition by the AGA1 TCR, we tested all HdH peptides in the sandwich ELISA-based UV exchange peptide binding assay (Figure 4B). When the HdH and KF11 peptides were ranked according to the strength of peptide binding ELISA data, specific amino acids were prominent in the stronger binding peptides: position (p)2 Thr/Ser (preferred anchors), p3Iso, p7Iso, a Pro at p8 or 9 (but not both) and p10Try. Peptides in the medium- and low-binding categories, although incorporating some of these residues, had either single or combined differences at these positions (Figure 4B and Supplementary file 1 - Table 4). We also compared UV peptide-exchange peptide binding ELISA data to NetMHCPan4.1 (http://www.cbs.dtu.dk/services/NetMHCpan/) predicted binding affinities for KF11 and the HdH peptides (Figure 4, Figure 4—figure supplement 1). Although a negative correlation was observed, this association was very weak (R = −0.18), and most likely reflects a general lack of data-informed algorithm training for MHC class I with longer peptide using artificial neural network analysis-based systems such as NetMHC.

In terms of the individual peptides, HdH6 demonstrated weaker HLA-B*57:01 binding, potentially owing to the presence of a non-favored (Glu) anchor residues at p2. In contrast, HdH2 and HdH5 produced strong ELISA binding signals, and their impact was potentially mediated at the level of TCR recognition, either via interactions that directly abrogate the peptide’s central ‘p5Pro-p6Glu’ TCR recognition hotspot (via replacement of p5Pro with Glu in HdH5, for example) or mutations distal to the hotspot that potentially distorts the normal trajectory of the peptide required to facilitate AGA1 TCR binding (p9Pro to Ser mutation in HdH2).

To compare the overall sequence commonality between the canonical KF11 peptide, the top 20 library-derived peptides and the panel of HdH peptides recognized by the AGA1-expressing T cell clones, Seq2Logo motifs were generated. As illustrated in Figure 4C, the AGA1 TCR-selected yeast display peptide library logo reflected tolerance to a broad array of amino acids beyond the conserved, central dipeptide ‘p5Pro-p6Glu’ motif. The HdH-based logo allowed further refinement of the motif based on T cell functionality and demonstrated that enhanced recognition by AGA1 TCR-expressing T cell clones required further fixation of amino acids outside the central ‘dipeptide’ motif; most notably, with a preference for p9Pro in HdH peptides. This logo also illustrated that despite low KF11 sequence homology (40–50%), the majority of HdH peptides elicited functional T cell responses at magnitudes approaching that generated by the KF11 peptide.

Although antigen presenting cells pulsed with exogenous ‘nr’ mined microbial peptides elicited strong cross-reactive T cell responses, we next wanted to ascertain if T cell-mediated recognition of these peptides occurred under conditions that more accurately reflect the physiological processing of microbial antigens in vivo. To test this, we generated lysates of the three microbial species that elicited AGA1 TCR-specific responses in the peptide-based functional assays. The lysates were incubated with HLA-B*57:01-expressing HL60 cells (Zorn et al., 2002) pre-treated overnight with calcium ionophore and IFNγ to induce DC-myeloid-like cellular differentiation and enhance MHC class I expression (Wu et al., 2004a). Following 7 hr of incubation, the cells were washed extensively and their ability to stimulate AGA1-expressing clone 1.1 and the related clone, 1.2, was assessed by ELISpot to evaluate IFNγ secretion. As demonstrated in Figure 4D, T cell clones 1.1 and 1.2 specifically responded to lysates generated from S. newyorkensis whereas reactivities to the other lysates, including the negative control Ruminococcus gnavus lysate, were not detected. 

A random peptide library displayed by the human MHC class I HLA-B*57:03 subtype, was used to generate the MHC class I expressing yeast display platform tested in this study. Although the AGA1 TCR is restricted by the KF11 peptide in complex with HLA-B*57:01, this TCR also binds with high affinity (Kd ~5 μM) to the near-identical HLA-B*57:03-KF11 complex (Stewart-Jones et al., 2012). The peptide-β2m-MHC heavy chain single chain trimer construct design was based on previously developed MHC class I yeast display scaffolds (Adams et al., 2011; Birnbaum et al., 2014a; Gee et al., 2018) and incorporated a single chain fusion display where HLA-B*57:03 heavy chain domains were included downstream of the KF11 epitope sequence and the β2m gene. Flexible GSSS linkers interconnected KF11, β2m ([GSSS]₃₎ and the B*57:03 heavy chain regions ([GSSS]₄₎. To minimize linker protrusions that could affect TCR binding, mutagenesis of position 84 (Tyr > Ala) in the HLA-B*57:03 alpha 1 (α1) region was performed to generate a larger linker-accommodating cavity (Hansen et al., 2009). A DNA construct library encoding an 11 amino acid (11mer) peptide-MHC library was generated where all peptide residues were fully randomized except for the preferred HLA-B*57:01/57:03 anchor binding residues, which were restricted to Ser, Thr and Ala at position two and to Trp, Phe and Tyr at position 11 to preserve binding to HLA-B*57:03 (Barber et al., 1997). All positions except for the anchor residues were generated using an NNK codon library. Position two utilized a DCK codon and position 11 used either a TGG codon for Trp or a TWY codon for Phe and Tyr. Two separate library inserts were generated to encode for the codon library NNKDCKNNKNNKNNKNNKNNKNNKNNKNNK(TGG/TWY) with theoretical nucleotide diversities of 2.1E14 and 8.4E14 for a total diversity of 1.06E15.

The peptide-HLA-B*57:03 construct library was used to generate a yeast display library detailed above as previously described (Gee et al., 2018; Birnbaum et al., 2014b). Briefly, freshly prepared EBY-100 electrocompetent yeast were incubated with a 5:1 ratio of pMHC insert to linearized pYAL vector by mass (~1 ug vector and ~5 ug insert per electroporation) and transformed to generate a randomized peptide library. The insert contained an equal mass mixture of the polymerase chain reaction product containing the two libraries detailed above. The final library, which was generated via 20 electroporations, was then quantified as containing ~2×10⁸ unique transformants via limiting dilution.

Each step described below was performed essentially as previously reported and was conducted with enough yeast to ensure >10 fold coverage of the previous step’s theoretical diversity. The yeast display library was seeded at an optical density of 1 (corresponding to 1 × 10⁷ yeast/mL) for growth in SDCAA media, pH 5.5, at 30°C and grown while shaking at 250 rpm until the culture reached saturation overnight. The library was then induced in SGCAA media, pH 5.5, and allowed to incubate while shaking at 250 rpm at 20°C for 48 hr.

After induction, 1 × 10⁶ yeast were assayed for induction by staining for a C-terminal Myc epitope tag placed between the MHC and Aga2. Next, yeast corresponding to >10 fold diversity of the initial library or the yeast from the previous round of selection were incubated with streptavidin microbeads (Miltenyi Biotec 130-048-102), washed in PBS + 0.5% BSA, and passed through a Miltenyi LS column (Miltenyi Biotec 130-042-401) to remove non-specific binders. Yeast were then incubated with streptavidin beads co-incubated with 400 nM biotinylated TCR, washed in PBS+BSA, and then run through an LS column. Retained yeast that were then eluted from the column, washed into SDCAA, and re-cultured. Selections were repeated for an additional two rounds. Finally, a more stringent final round of selection was conducted by staining yeast with 400 nM TCR streptavidin tetramers, incubating with anti-fluorophore microbeads, and sorting via magnetic selection.

Peptide enrichment and overall motifs were analyzed via next-generation sequencing as previously described. Briefly, plasmids from 5 × 10⁷ yeast per round of selection were recovered from selected yeast (Zymoprep II kit, Zymo Research). Peptide sequences were then amplified via PCR. During the course of PCR, individual barcode sequences (to deconvolve rounds of selection) and random 8mer sequences (to increase sequence diversity) were added to the amplicons. The product of these PCRs were used in a second PCR to affix Illumina sequencing adapters. Final PCR products were gel purified, pooled, quantitated, and sequenced on an Illumina MiSeq using a v2 2 × 150 nt kit with PhiX DNA spiked in to ensure sufficient sequence diversity for high-quality reads. As detailed previously data processing, analysis, and predictions were as previously described (Birnbaum et al., 2014b). The forward and reverse sequencing data were used to create a single sequence of the paired-end reads using PandaSeq (Masella et al., 2012), deconvoluted into each round of selection by barcode, and trimmed for analysis using Geneious version 6 (Biomatters Inc). In order to correct for potential PCR or sequencing errors, sequences differing by one nucleotide from the most abundant were coalesced. Amino acid frequencies for each round were determined after translation and removal of sequences containing frame shifts or stop codons. Predictions from the profile-based search of the NCBI non-redundant protein data database (downloaded June 21, 2013) was preformed using the statistics from the third round using an amino acid frequency cut-off of 0.01. The custom scripts for data processing and analysis are available on GitHub (https://github.com/jlmendozabio/NGSpeptideprepandpred; copy archived at https://github.com/elifesciences-publications/NGSpeptideprepandpred; Mendoza, 2020).

Soluble AGA1 TCR was expressed and refolded as described previously (Boulter et al., 2003). To facilitate stable chain pairing, an extracellular C region disulfide bridge was introduced into both the alpha and beta TCR chains by site-directed mutagenesis. The modified TCR alpha chain was cloned upstream of a hexahistidine tag and a BirA biotinylation substrate motif was inserted downstream of the modified TCR beta chain. Both chains were cloned into the PET22b+ vector (Novagen), and were expressed in the E. coli strain Rosetta-2 (Novagen), isolated as inclusion bodies, purified, re-solubilized, and refolded as described (Stewart-Jones et al., 2012; Boulter et al., 2003). Refolded TCR complexes were purified by anion exchange, Ni⁺² and size exclusion chromatography. The functionality of the AGA1 TCR was confirmed by cellular and SPR-based studies prior to inclusion in HLA-B*57:03 yeast display screening runs.

Conventional peptide epitopes that included a panel of library-derived peptides (Lib 1–6) and their closest non-redundant (nr) database-mined microbial sequence hits (Mic 1–6) were synthesized commercially using Fmoc chemistry (Genscript, USA). A photo-labile version of the ISPRTLNAW epitope (9MT4), where a UV sensitive 3-amino-3-(2-nitrophenyl)-propionic acid residue replaced the C-terminal penultimate Ala residue, and validated previously (Brackenridge et al., 2011), was sourced from LUMC, Leiden, The Netherlands.

Refolding of HLA-B*57:01, β2m and 9MT4, and subsequent UV-photocleavage/peptide exchange were performed as outlined previously (Brackenridge et al., 2011; Toebes et al., 2009). 0.5 μM (~25 μg/mL) of UV-sensitive HLA-B*57:01-β2m-9MT4 refolded monomers were incubated with 50 μM of the various test peptides in polypropylene 96 (V-shaped) well plates (Greiner Bio-one), and the final volume was adjusted to 125 μL by adding UV exchange buffer (20mMTris pH7, 150 mM NaCl). Samples were incubated on ice in a CAMAG UV cabinet and photo-illumination at 366 nm was performed for 60 min. Following centrifugation at 4000 g for 10 min to remove aggregated material, 90% of the sample volumes were transferred to a fresh plate, of which 3 μL was subsequently diluted 1:50 in 2% bovine serum albumin (BSA)-PBS. 50 μL of this mix was transferred (in duplicate) into an ELISA plate pre-coated overnight at 4°C with 10 μg/mL of the anti-human MHC monoclonal antibody W6/32 (Biolegend) and blocked in 2% BSA-PBS prior to sample addition. Following immobilization for 2 hr at 4°C, and extensive washing in 0.01%Tween-PBS (referred to herein as ELISA wash buffer), the plates were incubated with 50 μL of biotinylated anti-human β2-microglobulin (2M2) antibody (BioLegend, RRID:AB_493689) diluted 1:100 in 2% BSA-PBS for 1.5 hr at 4°C. Plates were washed for a further six times in ELISA wash buffer, prior to the addition of Extravidin peroxidase (Sigma) diluted 1:1000 in 2%-BSA PBS. Following 30 min at RT and a subsequent round of plate washings, 100 μL of tetramethyl benzidinesubstrate (TMB) (BioSource) was added, and plates were developed for 10 min. 100 μL of H₂SO₄ STOP solution (BioLegend) was added to terminate the reactions and the plates were read immediately at 450 nm absorbance using a FLUOstar OMEGA plate reader.

The following microbes were grown aerobically in the specified culture conditions: Adherent-Invasive Escherichia coli LF82 in Luria Broth at 37°C; Candida orthopsilosis DSM 24508 in potato dextrose medium at 30°C; Sporosarcina newyorkensis DSM-23544 in Caso medium at 37°C. Strict anaerobes were cultured in a Whitley DG250 workstation at 37°C with 10% H₂, 10% CO₂, and 80% N₂: Olsenella uli DSM 7084 in DSMZ Medium 104 and Ruminococcus gnavus CC55_001C in filtered Reinforced Clostridial Medium (RCM). Filtered RCM was generated by centrifugation of the media at 10,000 x g for 10 min to remove agar components, followed by filtration through 0.2 μm filters. Optical density (OD600) to colony-forming unit (CFU) conversions were calculated from overnight cultures. Individual microbes were cultured for 24–48 hr in liquid media. Bacterial cells were harvested by centrifugation at 10,000 x g for 10 min, followed by three washes in sterile PBS. Cells were re-suspended in sterile PBS and heat-inactivated at 70°C for 1 hr, followed by two freeze-thaw cycles. Protein concentration was determined by BCA assay (Pierce BCA Protein Assay Kit, Thermo Fisher Scientific).

The standard ELISpot assay was used to detect IFNγ release by AGA1 CD8+ T cells clones, but with modifications (Stewart-Jones et al., 2012). 5000 peptide-pulsed HLA-B*57:01 homozygous immortalized (mycoplasma negative) B cell lines (BCLs) were incubated with 500 CD8+ T cell from two clones expressing the AGA1 TCR in the presence of 10-fold serial dilutions of KF11 and test peptides ranging from 10⁻⁵ to 10⁻⁹M final concentration. Assays were harvested following 15 hr incubation at 37°C. Results are reported as spot forming cells (SFCs) per 500 T cell input, and SFCs double that observed with media alone were considered positive.

CD107a expression following T cell stimulation was evaluated as previously described, but with slight protocol alterations (Stewart-Jones et al., 2012). 40,000 T cell clones were incubated with allogeneic HLA-B*57:01+ BCLs (mycoplasma negative) at a ratio of 4:1, pulsed with KF11 and the ‘nr’-derived Mic peptides at final concentrations of 1 and 10 μM. Cells incubated with media alone, or in the presence of 10 ng/mL PMA were included as negative and positive controls, respectively. 5 μL of CD107a PE labelled monoclonal antibody (BD Biosciences, clone H4A3, RRID:AB_396135) was added to each well and the samples were incubated at 37°C. Golgi-Stop (BD Biosciences) was added to individual samples 1 hr post incubation at a concentration recommended by the manufacturer. Following a total of 5 hr, cells were stained with anti-human CD8 APC (BD Biosciences, clone RPA-T8, RRID:AB_398595) and fixed in BD Cell Fix prior to flow cytometry analysis using a CyAn. Data analysis was performed using FlowJo software.

2 × 10⁶ HL60 cells (ATCC CLL-240/RRID:CVCL_0002, STR profiled and mycoplasma negative) were treated overnight with calcium ionophore and IFNγ at 37°C to induce DC/myeloid-like differentiation and to enhance MHC class I expression (Wu et al., 2004a; Wu et al., 2004b). The cells were subsequently divided into 48 well NUNC plates and 20 μg/mL of the test bacterial cell lysates were added to individual wells. Following a further 7 hr of incubation, the cells were washed extensively and their ability to stimulate AGA1-expressing T cell clones was assessed by ELISpot assay. 15,000 lysate-pulsed HL60 cells were incubated with 500 CD8+ T cell from two clones expressing the AGA1 TCR for 15 hr prior to development. Positive control included non-lysate-pulsed HL60 and T cells incubated with 10ng/mL PMA. Results are reported as spot forming cells (SFCs) per 500 T cell input, and SFCs double that observed with media alone were considered positive.