AlphaFold 3 prediction of TCR mimic Fv – pMHC complex structures and their validation and optimization

ABSTRACT

Monoclonal antibody-derived chimeric antigen receptors (CAR) that mimic T‑cell receptors (TCR) on binding to peptide-major histocompatibility complex (pMHC) and activating T‑cell functions hold great promise for the development of effective immunotherapy. This study applied AlphaFold 3 (AF3) to model the quaternary structures of TCR mimic antibody variable fragment (TCRm Fv) complexed with NY-ESO-1_157–165/HLA-A*02:01. Benchmark study suggested that reliable TCRm Fv–pMHC structures were achieved by AF3 with high confidence. Toward NY-ESO-1/A2-specific CAR clones, AF3 prediction revealed their intermolecular geometry resembling the canonical TCR engagement, and binding avidity tests confirmed the peptide-dependent recognition of pMHC. Molecular interactions between Fv and the peptide antigen and its binding groove on MHC were further pinpointed, and the predicted epitopes and paratopes were validated by mutagenesis studies. Aiming to generate TCRm Fv mutants of improved affinity, AF3 aided designs with larger Fv–pHLA contact surface areas and more hydrogen bonds toward the peptide antigen were successfully identified. However, these AF3‑designed mutants failed to deliver enhancement on binding avidity or functional potency in experimental tests. Overall, AF3 is a highly valuable tool for TCRm Fv–pMHC complex modeling, and its combination with force field-based methods will be desirable to aid optimization tasks.

Introduction

Recognizing antigens presented by peptide-major histocompatibility complex (pMHC) molecules, adaptive T‑cell immunity is essential for the development of effective therapeutics against cancer, infection, and autoimmune disorders.¹ Notably, monoclonal antibody scFv-derived chimeric antigen receptors (CAR) that mimic T‑cell receptors (TCR) on binding to pMHC and activating T‑cell functions have attracted increasing attention, given their advantages in avoiding mispairing with endogenous TCR and the facile use of T‑cell engagers (TCE).^2–4 Currently, TCR‑mimic (TCRm) CARs and TCEs targeting neoantigens, tumor-associated antigens (TAAs), viral peptide/oncogenes, and autoantigens have achieved significant in vitro and in vivo efficacies with promising translational potential.^5–16

Structural studies have revealed key features of desired interactions between TCRm variable fragments (Fvs) and pMHC. To prevent off-target reactivity and to stimulate efficacious T-cell responses, TCRm should have a broad and energy-balanced epitope on its cognate peptide with an intermolecular geometry replicating canonical TCR engagement with the peptide-binding groove of the MHC molecule.¹⁷ Our previous work has identified panels of highly functional TCRm CARs able to stimulate efficacious and persistent T-cell responses toward HLA-A*02:01-restricted TAAs including MAGE-A3_112–120 (KVAELVHFL), NY-ESO-1_157–165 (SLLMWITQC), and MART-1_{26–35(27 L)} (ELAGIGILTV).¹⁸ In this study, we conduct structural analysis and experimental validations aiming to understand the molecular mechanism of isolated TCRm CARs. Particularly, we focus on NY-ESO-1/A2-specific TCRm CARs because a rich collection of 34 functional NY-CAR clones with unique complimentary-determining region (CDR) sequences has been discovered,¹⁸ allowing us to survey a diverse sequence space of Fvs.

Powered by generative AI, recent breakthroughs in biomolecular modeling have revolutionized multiple research areas, particularly for protein sciences.^19,20 The latest advancement, AlphaFold 3 (AF3) can predict biomolecular interactions involving ligands, nucleic acids, peptides, and protein–protein complex.²¹ In this work, we use AF3 to model the quaternary structures of TCRm Fv–pMHC complexes and validate the predictions by mutagenesis studies on both epitope and paratope. In addition, we investigate AF3-aided design of TCRm CAR mutants, aiming to improve functional potency of CAR-T. Overall, applying AF3 to TCRm, this work presents a case study accentuating structure-function relationship of specific T‑cell immunity.

Results

AF3 achieved reliable TCRm Fv–pMHC complex modeling with high confidence scores

To evaluate the performance of AlphaFold 3 in modeling TCRm Fv–pMHC complex structures, simulations were conducted on 12 published TCRm antibodies for which complex structures with their cognate pMHC have been solved by crystallography or cryo-EM.^13,22–31 AF3 predicted complex structures and experimentally determined ones were superposed based on alignment of HLA α chain, and root mean square deviations (RMSD) for V_H and V_L domains were calculated. Four representative TCR clones that were HLA-A*02:01-restricted and NY-ESO-1_157–165- or MART-1_{26–35(27 L)}-specific were also analyzed for comparison.^32–34 Results indicated that accurate AF3 predictions were largely achieved for the four tested TCR–pMHC complexes with low to moderate RMSD values of 1.5 to 3.5 Å between the modeled and the solved V_β and V_α domains, and the AF3 interface predicted template modeling (ipTM) scores ranged from 0.85 to 0.93, indicating high confidence of prediction (Figure 1). In contrast, for TCRm antibodies, AF3 accurately modeled 6 of 12 tested clones with RMSD less than 5 Å for both V_H and V_L (Figure 1), but for the remaining 6 TCRm clones, incorrect predictions were obtained, especially for the ones with their binding geometry/epitope footprint distinct from that of canonical TCRs, such as TCRm antibodies ESK1, 2Q1, and V2 (Figure 1).

Figure 1.

Validation of AlphaFold3 by comparison with experimentally determined pHLA–TCR/Fv complex structures. TCR / TCRm Ab, MHC allele, peptide antigen, PBD file, AF3 ipTM scores, and RMSD values are shown. For predictions with low confidence (ipTM < 0.65), three independent runs were conducted, and the structures with the highest ipTM scores were applied for analysis. pMHC is shown as cartoon and TCR/TCRm is shown as ribbon. Peptide, orange; HLA α1-α2 domains, gray; PDB structure, blue; and AF3 prediction, red. By alignment on HLA α chain, experimentally determined (PDB) and predicted (AF3) structures were superposed and RMSD values were calculated for V_H and V_L. Representative TCRs targeting NY-ESO-1/A2 and MART-1/A2 were included for comparison. Images were prepared using PyMOL. AF3 training data cutoff is September 30, 2021.

To understand the correlation between prediction accuracy and confidence, RMSD values were plotted against ipTM scores for each modeled TCR/TCRm–pMHC complex. As shown in Fig. S1, when ipTM scores were greater than 0.85, AF3 generally produced accurate quaternary structures with RMSD values of 1 to 5 Å. In contrast, ipTM scores of 0.65 or less were usually associated with large RMSD, e.g., >5 Å and up to 30 Å in certain cases. Although exceptions exist, for example, modeling of the alpha‑fetoprotein peptide (AFP)/A2-specific TCRm Ab achieved medium accuracy (RMSD of 3.5 Å) with low confidence (ipTM of 0.50), most accurate predictions had ipTM scores of high confidence. Overall, reliable TCRm Fv–pMHC complex modeling can be achieved by AF3 at ipTM > 0.85.

Sequence analysis and structural prediction of NY-ESO-1/A2-specific TCRm clones

A panel of 34 NY-ESO-1/A2-specific TCRm functional clones with unique CDR sequences has been isolated from a human naïve antibody repertoire-derived monoclonal CAR-T cell library by activation-based positive and negative screenings.¹⁸ Germline analysis indicated that diverse immunoglobulin heavy/light variable (IGHV/IGLV) genes were used among the identified NY CAR-T clones, with IGHV-1 and IGLV-2 as the dominant alleles (Figure 2(a)). Weighted mismatch distance analysis of their CDRs further revealed that the isolated clones are evenly distributed in the sequence space without apparent clusters, indicating their independent clonality (Figure 2(b)).

Figure 2.

NY CAR-T clone sequence analysis and AF3 predictions of Fv–pHLA complex structures. (a) V_H and V_L germline usage analysis. (b) Weighted mismatch distance between CDR sequences calculated by using the BLOSUM62 substitution matrix. (c) Plot of AlphaFold pTM and ipTM scores (Table S1). Green indicates the Fv mimicking TCR on recognition of the peptide antigen and the peptide-binding groove of the HLA-A2 molecule in the predicated complex structure. Red indicates the Fv binding at HLA rather than its peptide-binding groove or the peptide antigen. AF3 results of TCR 1G4 and TCRm Fv 3M4E5 and 3M4F4, all NY-ESO-1/A2-specific, were also plotted for comparison. Complex structure predictions of high confidence with ipTM scores > 0.9 are indicated.

To understand the molecular mechanism of these isolated NY CAR-T clones, their Fv structures in complex with the NY-ESO-1 peptide, HLA-A*02:01, and β₂ microglobulin were modeled by AF3. As shown in the pTM–ipTM plot (Figure 2(c)), predictions achieved high (ipTM > 0.8), medium (0.6–0.8), and low (< 0.6) confidence for 16, 14, and 4 Fv clones, respectively (Table S1). Remarkably, structural analysis revealed that the majority of Fv clones with high ipTM scores (15 of 16) recognized the peptide antigen and the peptide-binding groove of HLA in a manner similar to TCR (green dots in Figure 2(c)), while many predictions with medium (8 of 14) or low (3 of 4) confidence ipTM scores showed non-canonical geometries, i.e., Fv binding to HLA at epitopes rather than the HLA groove (red dots in Figure 2(c)). In addition, nonfunctional or cross-reactive clones tended to associate with the non-canonical recognition and low confidence (Table S1).

To test the consistence of AF3 prediction, three independent runs were conducted on these NY CAR-T clones of high T‑cell activation efficiency without cross-reactivity. Results confirmed that AF3 performance was largely consistent across multiple runs (Table S2), and the models with the highest ipTM scores were further used. Collectively, we chose CAR-T clones NY-6, -14, and -42 for in-depth analysis due to their distanced CDRs (Figure 2(b)) and high ipTM scores (0.92–0.94), as well as promising biological functions including efficient T‑cell activation and no cross-reactivity.¹⁸ Notably, when NY-ESO-1/A2-specific TCR 1G4 and TCRm 3M4E5 and 3M4F4 were modeled by AF3, an ipTM score of 0.93 was obtained, which was on par with those of the chosen NY CAR clones (Figure 2(c)).

Canonical geometry of AF3-predicted TCRm Fv–pHLA complex

The topography of AF3-predicted NY CAR Fv structures complexed with NY-ESO-1_157–165/HLA-A2*02:01 revealed that NY-6, -14, and -42 bind diagonally across the peptide-binding groove of MHC class I molecule in a geometry mimicking canonical TCR engagement (Figure 3(a), front view). Specifically, their V_L domains recognize the α1-helix of HLA and the C-terminal portion of NY-ESO-1_157–165, while their V_H domains mainly bind to the α2-helix and the middle portion of the peptide antigen (Figure 3(a), top view). Contact analysis suggested that V_H and V_L have balanced footprints on pHLA with a full coverage of the peptide antigen and an overall Fv–pHLA contact area ranging from 1065 to 1204 Ų, comparable to that of TCR 1G4 at 1331 Ų and TCRm 3M4E5 at 1153 Ų (Figure 3(a), contact). Notably, TCRs use their V_α (resembling V_L) to recognize HLA α1-helix and V_β (resembling V_H) for HLA α2-helix, a well-documented difference from TCRm.¹⁷

Figure 3.

Topography of AlphaFold3 predicted NY CAR Fv complexed with NY-ESO-1/HLA-A2. (a) All complexes are positioned according to pHLA alignment. In front and top views, V_H (lime), V_L (blue), HLA α1-α2 (gray), and peptide (orange) are shown. The contacts of Fv on pHLA are shown as surface for V_H only (lime), V_L only (blue), and shared between V_H and V_L (cyan). The contact areas were calculated by using ChimeraX. Experimentally determined structures of TCRm Ab 3M4E5 (PDB, 3GIF)²² and TCR 1G4 (PBD, 2BNR)³² complexed with NY-ESO-1/HLA-A2 were analyzed for comparison. For TCR, V_β in blue and V_α in lime. (b) Binding avidity. NY CAR transduced Jurkat cells were stained with serially diluted NY-ESO-1_157–165/HLA-A*02:01 pMHC-APC tetramer and pMHC⁺ cells were assessed by flow cytometry. MHC-APC tetramer without loaded peptide was used as control. n = 2.

To validate AF3 predictions that NY CAR Fv clones recognize pHLA at the peptide and peptide-binding groove, Jurkat T cells were transduced to display NY CAR and their binding avidities toward NY-ESO-1/HLA-A2 tetramer were measured by flow cytometry. Results indicated that the tested TCRm CAR clones NY-6, -14, -42 exhibited their binding avidities to the target pMHC tetramer with EC₅₀ values of 16, 103, and 34 pM (Figure 3(b)). Importantly, their binding toward the non-peptide-pulsed HLA-A*02:01 tetramer was fully abolished (Figure 3(b)). This peptide-dependent recognition indicated that these TCRm Fvs indeed bind to the antigenic peptide and its surrounding groove on HLA.

Interaction between TCRm Fv and pHLA and validation of predicted epitope

Positioning CDRs over pHLA revealed that TCRm Fvs mainly use their H3 and L3 loops for the recognition of antigenic peptide, whereas their remaining CDRs interact with the α1- and α2-helices of HLA (Figure 4(a)). Particularly, CDR-L1 binds to α1-helix, and CDR-H1 and -H2 bind to α2-helix. A large portion of important interactions between Fv and pHLA are mediated by CDR3s and CDR-L1. CDR residues with greater than 100 Ų buried surface areas upon complex formation include TyrH32, IleH104, TyrH105, TyrL32, and TyrL34 of NY-14; and TyrH101, TyrH102, TyrL32, and SerL96 of NY-42 (Figure 4(b), blue and green). Notably, among these nine identified residues, five are within CDR3s, and three are from CDR-L1. Interestingly, for NY-6, strong interactions are also contributed by CDR-H2- three CDR-H2 residues (TyrH54, GlnH57, and ArgH59) and one CDR-L3 residue (SerL98) have their buried surface areas > 100 Ų.

Figure 4.

Interaction of TCRm Fv with pHLA and experimental validation. (a) Top view showing CDRs positioning over pHLA. CDR-L in blue, CDR-H in lime, HLA α1- and α2-helices in gray, and peptide in orange. (b) Buried surface areas (in Ų/residue) upon Fv–pHLA complex formation calculated by ChimeraX. Same color code as (a). The paratope residues identified for alanine mutation analysis are in red. (c) Functional specificity against alanine‑mutated peptides. n = 2. Upward-, downward- and sideways-facing amino acids are indicated by arrows (PDB, 1S9W).³⁵.

On HLA-A*02:01 α chain, the residues contacting TCRm Fv are mainly located in its α1- and α2-helices. Conserved Fv-binding positions include R89, K90, and Q96 in α1-helix, and Q179 and T187 in α2-helix (Figure 4(b), gray). On the antigenic peptide NY-ESO-1_157–165, 5–6 non-anchoring residues exhibit strong interactions with TCRm Fv. Across all three modeled complex structures, the protruded motif in the middle of the peptide, e.g., MelP4 and TrpP5, substantially accounts for TCRm Fv binding with a combined buried surface area > 300 Ų (Figure 4(b), orange). In addition, the C-terminal residues of ThrP7 and GlnP8 also contribute to TCRm Fv binding.

To validate the predicted epitope, alanine scanning was applied for the peptide antigen. Notably, simulation suggested that, except LeuP2Ala, these alanine-mutated NY-ESO-1 peptides can effectively bind to HLA-A*02:01 (Table S3). Following co-culture with antigen presenting cells (APC) pulsed with mutated peptides, T‑cell activations were monitored. Results showed that alanine substitutions at MelP4, TrpP5, and IleP6 fully abolished T‑cell activation across all three tested TCRm CAR clones (Figure 4(c)), confirming this dominant central motif predicted for TCRm CAR-T recognition (Figure 4(b), orange). With ThrP7Ala-mutated peptide, NY-6 also lost its recognition, whereas the reduction on activation was less pronounced for NY-14 and NY-42 (Figure 4(c)). For these alanine mutations at the N-terminal positions of NY-ESO-1 peptide SerP1 to LeuP3, activation of NY-6 CAR-T was remarkably diminished (Figure 4(c)), consistent with the predicted interaction at these positions (Figure 4(b)). In addition, T‑cell responses were significantly reduced against GlnP8Ala mutated peptide antigen for NY-14 and NY-42 (Figure 4(c)). Collectively, these experimental validations supported the accuracy of most predictions of epitopes on the NY-ESO-1 antigenic peptide.

Structural basis for pHLA recognition and validation of predicted paratope

To pinpoint important paratope residues, hydrogen bonds (H-bonds) and van der Waals (vdW) interactions between Fv and the antigenic peptide were identified for TCRm clones NY-6, -14, and 42. As illustrated in Figure 5(a), heavy‑chain CDRs predominantly recognize the N-terminal part of NY-ESO-1 (SerP1–TrpP5), whereas light‑chain CDRs bind to its C-terminal portion (TrpP5–Gln8). Notably, the highly bulged MelP4-TrpP5 side chains are surrounded by multiple CDR residues from both V_H and V_L (Figure 5(a)), explaining the largest buried surface areas calculated for MelP4 and TrpP5 (Figure 4(b), orange). Most identified H-bonds involve the carboxyl and amide groups of NY-ESO-1 peptide backbone (Figure 5(a), top), presumably due to the overall hydrophobic nature of this peptide antigen. Particularly, GlnH57 and ArgH59 of NY-6, TyrH105 of NY-14, and TyrH102 and TyrL93 of NY-42 form H-bonds with the backbone of NY-ESO-1 peptide (Figure 5(a), top). Consistent to epitope alanine scanning results (Figure 4(c)), prediction indicates that the side chain of GlnP8 presents important moieties for the formation of H-bonds with light‑chain CDR residues, such as SerL95 of NY-6, TyrL32 of NY-14, and TyrL32 and SerL27 of NY-42 (Figure 5(a), top). In addition, an H-bond between SerP1 and AspH55 was also identified for NY-6.

Figure 5.

Detailed structural basis for peptide recognition by NY-CAR clones and paratope mutation studies. (a) H-bond and vdW contact. Peptide is depicted as a stick, with residues contacting the heavy chain and light chains colored lime and blue, respectively. (b) Tetramer binding toward alanine‑mutated NY-CAR clones. Transduced Jurkat cells were stained with 1 nM biotinylated NY-ESO-1/HLA-A2 tetramer, and signals were detected by FACS with streptavidin-PE. The wt NY-CAR clones were applied as controls. n = 3. (c) Activation of alanine mutated NY CAR-T cells. Transduced Jurkat NFAT-luc was co-cultured with 10 µM peptide pulsed 293F at E:T ratio of 1:1 for 16 hr, and luciferase activities were measured. n = 3. Statistical analysis by t-test: ns, not significant; *, p ≤ 0.05; **, p ≤ 0.01; ***, p ≤ 0.001; ****, p ≤ 0.0001. (d) AF3 prediction results of alanine‑mutated NY-CAR clones. Following alignment on HLA α chain, RMSD between wt NY-CAR and alanine mutants were calculated for V_H and V_L. The Ala mutants with RMSD > 5 Å on V_H or V_L are highlighted.

Analysis of vdW contacts indicated that the interactions are mainly centered at MelP4 and TrpP5 (Figure 5(a), bottom) – there are 5 and 2, 4 and 2, and 4 and 4 heavy‑ and light‑chain CDR residues of NY-6, -14, and -42, respectively, binding to the MelP4–TrpP5 motif via vdW forces. Several CDR residues giving multiple vdW interactions with this central motif were identified, including TrpH33, GlnH57, and ArgH59 of NY-6; AsnH59 and TrpH50 of NY-14; and AsnH59, TrpH50, AspH99, and TyrH102 of NY-42 (Figure 5(a), bottom). Notably, many of these paratope residues possess large areas of buried surface (Figure 4(b), blue and green). In addition to the hydrophobic central motif of NY-ESO-1, its C-terminal residues such as ThrP7 and GlnP8 also support vdW forces with CDR-L3 and -L1, e.g., SerL95 and SerL98 of NY-6; SerL96 and TyrL32 of NY-14; and SerL96, SerL27, and TyrL32 of NY-42 (Figure 5(a), bottom). These identified CDR residues (either involved in H-bonds or vdW contacts) and their predicted buried surface areas (Figure 4(b)) were collectively considered, and six paratope positions of each TCRm NY-CAR clone were selected for further studies (red in Figure 4(b)).

To validate the predicated paratopes, alanine substitutions were applied for these selected key residues. Both antigen binding and T‑cell functional response were tested using Jurkat cells transduced with mutated NY CAR-T. In the tetramer binding assay with NY-ESO-1/A2, results suggested that 5, 3, and 6 of the 6 tested alanine mutants showed significant reductions of pMHC⁺ cell populations for NY-6, -14, and -42 NY-CAR clones, respectively (Figure 5(b)). Particularly, L3:S95A and L3:S98A of NY-6; L1:Y32A, H1:Y32A, and H3:Y105A of NY-14; and L3:Y93A of NY-42 largely abolished their tetramer binding compared to the corresponding parental NY CAR clones (Figure 5(b)). In the functional assay, 6, 5, and 6 of the 6 tested alanine mutants showed significant reductions of T‑cell activation (Figure 5(c)). Notably, partially reduced tetramer binding signals (e.g., ArgH59Ala of NY-6) and less significant changes (e.g., TyrH101Ala of NY-42) became intensified in the T‑cell functional assays. This increased sensitivity can presumably be attributed to the avidity effect given by tetramer versus a more native pHLA recognition by TCRm CAR during co-cultures with APCs. Overall, the above epitope and paratope mutagenesis studies (Figures 4(c) and 5(b,c)) collectively suggested that AF3 predictions of TCRm Fv–pHLA complex structures were largely reliable to reveal structural insights, especially for the modeling, with high confidence, i.e., ipTM > 0.85.

To evaluate whether the reduced functionality can be predicted by AF3, simulations were conducted to the alanine-mutated TCRm clones experimentally validated above, and following alignment on HLA α chain, RMSD values between NY-CAR TCRm wildtype (wt) and its Ala mutants were calculated. We particularly paid attention to the 6 paratope alanine substitutions that fully lost their binding avidity to the cognate tetramer (e.g., L3:S98A of NY-6, and H3:Y105A of NY-14). Notably, AF3 correctly identified 4 of these 6 loss-of-function mutants (Figure 5(d)). For instance, modeling confidence was remarkably lower for NY-6 L3:S98A, with ipTM of 0.59, and the AF3 result indicated that this alanine substitution at Ser98 of CDR-L3 led to an Fv–pHLA complex geometry distinct from NY-6 wt, with RMSD of 34.3 Å for V_H and 15.3 Å for V_L. AF3-predicted complex structures of NY-14 L1:Y32A and H3:Y105A, and NY-42 L3:Y93A (ipTM = 0.60, 0.79, and 0.91, respectively), though sustaining the overall geometry of their parent clones over pHLA, exhibited substantial domain position shifts (e.g., RMSD = 9.2 Å for V_L of NY-14 L1:Y32A, 5.2 Å for V_H of NY-14 L3:Y105A, and 5.8 Å for V_H of NY-42 L3:Y93A), especially at their mutated CDRs, implying functional disruption. Increased RMSD values were also modeled for NY-6 H2:Q57A, an Ala mutant that showed significant drops on tetramer binding (p < 0.001) and T cell activation (p < 0.0001, Figure 5(b,c)). However, it should be noted that AF3 failed to identify two remaining loss-of-function clones: NY-6 L3:S95A and NY-14 H1:Y32A (Figure 5(b–d)).

AF3 aided mutation designs aiming for affinity and T cell function improvement

Encouraged by the overwhelming success of AF3 prediction of TCRm Fv–pHLA complex structures, we next investigated the feasibility that AF3‑aided design of CAR-T mutants can improve affinity and enhance T-cell function. Given the impressive performance associated with NY-6, including its high AF3 ipTM score at 0.92 (Table S2), high binding avidity of 16 pM (Figure 3(b)), and confirmed epitope and paratope by mutagenesis studies (Figures 4(b) and 5(b,c)), we focused on this CAR-T clone for our TCRm optimization efforts. Because H-bonds are much stronger than vdW forces, we reasoned that by introducing additional H-bonds between TCRm Fv and pHLA, the binding affinity will be improved, and in consequence, T-cell functional sensitivity will be enhanced.

Analysis of the H-bond maps between Fv CDRs and peptide NY-ESO-1 indicated that unlike NY-14 and NY-42, NY-6 only contains one H-bond forming residue in its CDR-L3 (Figure 5(a)). Local structural review suggested that residues SerL95, SerL96, and SerL98 have their side chains in proximity to the C-terminal residues of the peptide antigen, and thus, these positions were selected for mutagenesis. Although serine is frequently involved in H-bonds, acting as a donor or an acceptor, the strength of H-bonds mediated by serine is generally moderate. Therefore, we substituted these serines with residues able to render strong H-bonds, including tyrosine (a donor), histidine (an acceptor), and asparagine (a donor or an acceptor). Promoting interactions between Fv CDR and the groove of HLA molecule was also attempted. Structural analysis identified that Glu82 at N-terminal of α1-helix is in proximity to SerH30 of CDR-H1, and its substitution with a positively charged residues, e.g., Arg, Lys, or His, can potentially form a salt bridge to increase the recognition strength.

These NY-6 mutagenesis designs and their combinations were subjected to AF3 prediction, and the numbers of H-bonds and buried areas upon Fv–pHLA complex formation were calculated. The in-silico screening started with positions L95 and L96 and then followed by additional mutations at L98 and H30. In total, 21 mutation designs were modeled, and AF3 results indicated that all tested NY-6 Fv mutants remained high prediction confidence, with ipTM ranging from 0.87 to 0.93 (Table S4). When SerL95 and SerL96 were independently replaced with residues of strong H-bonds, larger buried areas at the mutated positions were usually observed (e.g., L3:S95Y, Table S4). However, the overall contact surface areas between Fv and pHLA could be compromised as well as the total number of H-bonds (e.g., L3:S95Y, Table S4), indicating the global effect of single point mutations. Therefore, double and triple mutations at these CDR-L3 and CDR-H1 positions were further attempted (Table S4).

The following criteria were jointly examined to select lead candidates: (1) contact areas of Fv–pHLA binding; (2) numbers of H-bonds between Fv and pHLA; (3) footprint balance between V_H and V_L (and between CDR-H3 and L3), in terms of buried areas and H-bonds; and (4) number of CDR-L3 residues of significant interaction (defined as 100 Ų buried area). Accordingly, three NY-6 mutants were selected (Figure 6(a)): two double mutants L3:S96Y / L3:S98Y (denoted as YY) and H1:S30K / L3:S96Y (KY), and one triple mutant H1:S30K / L3:S96Y / L3:S98Y (KYY). Compared to their parent NY-6, all these mutants have larger contact areas over pHLA, with more balanced H-bonds between V_H and V_L (mutant KYY) or larger buried areas of key CDR-L3 residues (mutant YY) (Figure 6(a)). Importantly, all three mutants increased the numbers of light‑chain CDR-mediated H-bonds toward the peptide antigen, with multiple CDR-L residues involved (Figure 6(b)). Experimental validations were conducted on these three NY-6 mutants by testing their tetramer binding and T‑cell responses. Results showed that mutants YY and KYY lost their binding avidity toward NY-ESO-1/A2 (Figure 6(c)) and exhibited basal T‑cell activation upon co-culture with APCs (Figure 6d). Although mutant KY largely reserved its binding avidity and T‑cell stimulation functionality, its potency was significantly weaker compared to NY-6 wild type (Figure 6(c,d)). Note AF3 modeling suggested that these designed mutations had no significant position shifts for their V_H or V_L domains compared to NY-6 wt (Figure 6(e)).

Figure 6.

AF3 aided NY-6 mutation designs for improved recognition strength. (a) Prediction results of representative designs. CDR-L3 residues with ~100 Ų buried area over pHLA are shown in bold, and CDR-L3 mutations are shown in red. See Table S3 for full list. (b) H-bonds toward NY-ESO-1 peptide antigen. Peptide is depicted as a stick, with residues contacting the heavy chain and light chains colored lime and blue, respectively. Peptide residues with their side chains involved in H-bond formation are labeled. (c) Tetramer binding toward NY-CAR clones carrying designed mutations. Transduced Jurkat cells were stained with 1 nM biotinylated NY-ESO-1/HLA-A2 tetramer, and signals were detected by FACS with streptavidin-PE. The wt NY-CAR clones were applied as controls. n = 3. (d) Activation of mutated NY CAR-T cells. Transduced Jurkat NFAT-luc was co-cultured with 10 µM peptide pulsed 293F at E:T ratio of 1:1 for 16 hr, and luciferase activities were measured. n = 3. Statistical analysis by t-test: ns, not significant; *, p ≤ 0.05; ****, p ≤ 0.0001. (e) AF3 predicted V_H and V_L positions of designed mutations compared to NY-6 wt.

Discussion

AlphaFold 3 uses diffusion neural network algorithm modeling molecular interactions between proteins, nucleic acids, and small chemical compounds.³⁶ Since the release of AF3,²¹ its accuracy on prediction of biomolecular complexes such as metal–protein, GPCR–ligand, antibody/nanobody–antigen, and TCR–pMHC has been comprehensively assessed.^37–42 In general, these benchmark studies agreed that AF3 outperformed previous generative AI models. Importantly, AF3 confidence scores (e.g., ipTM, I-pLDDT) were correlated with modeling accuracies.^21,37–40 Consistent with these findings, our work on TCRm–pHLA complexes suggests that reliable predictions can be achieved by AF3 with high confidence scores (Figure 1, Fig. S1). One limitation of this study is that training data cutoff (only using unseen structures for validation) was not applied, due to availability of TCRm–pMHC structures determined after the release of AF3.

Our study also indicates that the overall rate of correct TCRm–pHLA complex prediction (defined as RMSD < 5Å for both V_H and V_L) was at 50% (6 of 12 TCRm clones, Figure 1), on par with observations for AF3 modeling of antibody–antigen and nanobody–antigen complexes.^37,38 When the AF3 training data cutoff (on September 30, 2021) is considered, the rate of correct prediction modestly favors the pre-cutoff group at 57% (4/7), and the success rate among the post-cutoff TCRm–pHLA complex molecules is sustained at 40% (2/5). Besides the well-documented limitations of AF3 on modeling dynamic and disordered regions, one restriction particularly impacting TCR/TCRm–pMHC is that AF3 removes explicit water molecules from inputs during processing and does not predict their positions. Evidence supports that buried water molecules contribute substantially to the high shape complementarity and directly involve H-bonds as interconnecting molecules between TCR and pMHC.^43,44 Exclusion of water in complex modeling may partially explain the incorrect predictions and the ones with low confidence.

Applied to 34 NY-ESO-1/A2-specific TCRm Fv clones, AF3 obtained high confidence predictions (ipTM > 0.8) for 16 clones (Table S1). Among them, three TCRm Fvs with the highest ipTM scores (0.92 to 0.93) were chosen for in-depth analysis and experimental validation (Table S2). Results of peptide-dependent tetramer binding confirmed the predicted pHLA recognition via NY-ESO-1 peptide and its surrounding groove on HLA (Figure 3(b)). Our subsequent epitope and paratope mutagenesis studies further validated that the TCRm Fv–pHLA complex structural details revealed by AF3 predictions were largely accurate (Figures 4(c) and 5(b,c)). Notably, the discrepancy between the tetramer binding (Figure 5(b) and T‑cell activation assay results (Figure 5(c)) can be explained by the avidity effect versus native pHLA recognition – we argue that the tetramer binding test could be less sensitive to detect subtle change of interaction strength raised by a single alanine substitution, whereas the formation of immunological synapse in native pHMC recognition facilitates a serials of cellular responses and renders high sensitivity. In addition, T‑cell response also reflects kinetic characteristics of the interaction (e.g., formation of catch bond),⁴⁵ which is lacking in simple binding assays.

Previous studies suggested that the performance of deep learning-based approaches, e.g., protein language models, is suboptimal in predicting mutational effects on binding affinity.^45–48 Although AF3 highly synergizes with force field in prediction, no thermodynamic outputs (Gibbs free energy of binding) are explicitly calculated. In this study, TCRm CAR mutants with larger Fv–pHLA contact surface areas and more H-bonds toward the peptide antigen were successfully identified by AF3 (Figure 6(a,b)). However, the experimental tests indicated that the designed mutants failed to give enhanced binding affinity or functional potency (Figure 6(c,d)). In future studies, both H-bonds and vdW forces can be considered for CDR mutation designs, as well as polar amino acids Asp, His, and Glu, because their substitutions of CDR residues have shown success on promoting the catch bond formation between TCR and pHMC.⁴³ Furthermore, we note that the search path used in this study, starting with CDR-L3 then L1, may miss some beneficial mutations or their combinations. Recent advances in prediction of single domain Ab V_HH–antigen⁴⁹ and TCR–pMHC⁵⁰ complexes reported superior performance to AF3, and applying these new tools to TCRm–pHLA remains to be explored in future studies.

In conclusion, AF3 achieved reliable TCRm Fv–pMHC complex modeling with high prediction confidence. Toward NY-ESO-1/A2-specific CAR clones, AF3 modeled their intermolecular geometry resembling that of canonical TCRs, and the predicted peptide-dependent engagement was validated by experiments. Identified molecular interactions were further confirmed by epitope and paratope mutagenesis studies. TCRm Fv mutants of enhanced interactions were designed by AF3 but failed to deliver improvements in affinity or potency in experimental tests. Overall, AF3 is a highly valuable tool for TCRm Fv–pMHC complex modeling, and its combination with force field-based methods will be desirable to aid optimization tasks.^45,47,48

Materials and methods

Structure prediction

Simulations were performed using AlphaFold 3 (AF3). Five polypeptide sequences encoding V_H, V_L, HLA-A*02:01 α1-α3 domains, β₂ microglobulin, and NY-ESO-1_157–165 (SLLMWITQC) peptide were submitted to the server for three independent runs. Predicted complex structures were downloaded from the AF3 server and the associated interface-predicted template modeling (ipTM) scores were recorded. High confidence is generally indicated by ipTM > 0.8.²¹ Representative TCR–pMHC complexes with known structures were modeled by AF3 similarly.

Structure analysis

The simulations of TCRm Fv–pMHC complex with the highest ipTM scores among three runs were applied for analysis. The buried surface areas upon Fv–pHLA complex formation were calculated by ChimeraX, and H-bonds and vdW overlaps of the interaction were also identified. For AF3 validation, predicted complex structures were aligned with experimentally determined structures based on HLA α chain by PyMOL, and Fv domains were superposed for RMSD calculation between the modeled and the solved structures. Results were visualized using PyMOL or ChimeraX.

Germline analysis and CAR distance measure and visualization

The germline usage of V_H and V_L of isolated TCRm clones was analyzed by IMGT V-QUEST. The distance between any two CAR clones is defined as accumulated similarity-weighted mismatch distances between their respective CDRs. For each CDR pair, alignment was first performed by using pairwise2 module (Biopython) with the gap opening penalty at -10 and the gap extension penalty at -0.5. From the resulting optimal alignment, mismatch distances for each CDR pair were then calculated using the BLOSUM62 substitution matrix with the following rules of TCRdist⁵¹: distance (a,a) = 0; distance (a,-) = 4; distance (a,b) = min (4, 4-BLOSUM62 (a,b)), where 4 is 1 unit greater than the most favorable BLOSUM62 score for a mismatch, a and b are amino acids, and “-” is a gap. For each CDR pair, the total mismatch distance was computed as the sum of individual distance across the alignment. These distances were compiled into an array for each CAR pair. To account for the greater role of the CDR3 regions in antigen recognition, a weight factor of 3 is applied to CDR3s. A distance matrix between all CARs was computed accordingly and used for multidimension scaling to visualize the distances among CARs in a two-dimensional space.

Molecular cloning

The signaling peptide and 28z-mCherry construct were cloned into SFG retroviral plasmid at XhoI/SphI sites,^52,53 resulting in SFG-28z-mCherry. The scFv genes of NY-CAR were cloned into SFG-28z-mCherry with NgoMIV/NotI to make SFG-CAR-2A-mCherry. Genes of mutated scFv were constructed by overlapping PCR and cloned into SFG-CAR-2A-mCherry with XhoI / SalI, and resulted mutations were confirmed by DNA sequencing.

Cell culture

Jurkat E6-1 T cell line (ATCC, TIB 152) cells were cultured in R10 media composed of RPMI1640 medium (Corning, 10-040-CV), supplemented with 10% fetal bovine serum (FBS, Gibco, 10437028), 50 U/ml penicillin, and 50 µg/ml streptomycin (Gibco, 15140122). Jurkat-Lucia NFAT cells (Invivogen, jktl-nfat) were cultured in IMDM media (Cytiva, SH30228.02), supplemented with 10% FBS, 50 U/ml Pen-Strep, and 100 μg/ml zeocin was used every alternative passage. HEK293T cells were maintained in DMEM. All of the above cells were cultured at 37°C, in 5% CO₂ humidified atmosphere. Expi293F (Gibco, A14527) was cultured in Expi293 medium (Gibco, A1435101) supplemented with 40 U/ml penicillin and 40 µg/ml streptomycin at 37°C, in 8% CO₂ with shaking at 130 rpm. Cell numbers and viabilities were measured by Countess 3 (Invitrogen) with 0.4% trypan blue staining.

Retroviral package and transduction

The CAR-T genes encoding wild type and mutated TCRm NY-CAR clones were packaged to retroviruses in a 24-well plate by Gene Juice transfection reagent (Milipore Sigma, 70967). For transduction, a 24-well non-treated tissue culture plate (StemCell, 100-0097) was coated with retronectin (Takara, T100A) overnight at 4°C. After removal of retronectin, prepared retroviruses were added and spun at 4600 × g for 1 hr to facilitate virus binding. 5 × 10⁵ Jurkat-Lucia NFAT cells were then added to the plate and centrifuged at 800 × g for 10 min at 30°C. Two days after transduction, media was changed, and transduction efficiency was examined by FACS on mCherry expression. If needed, the procedure was repeated to reach transduction efficiency of 80–95% for each clone.

Tetramer binding assay

NY-ESO-1_157–165 (SLLMWITQC) and its alanine-substituted peptides were synthesized (GenScript) and dissolved in dimethylsulfoxide at 10 mg/ml. To obtain biotinylated pHLA monomers, 20 µl of 1 mg/ml peptides in PBS was slowly added to 100 µl of 500 µg/ml biotinylated human HLA-A*02:01 (A2) monomer (AcroBiosystems, HLM-H82W3), and the peptide loaded MHC monomer was purified by ultrafiltration. Tetramers were prepared by adding streptavidin-APC (Agilent, PJ27S) to the peptide/A2 monomer in dark. 10⁵ transduced Jurkat CAR-T cells were stained with prepared pHLA-APC tetramer at 0.1 pM − 18 nM for 30 min at 4°C in dark and then washed with PBSA three times before flow cytometry analysis in a Beckman CytoFlex SRT. Jurkat CAR-T cells stained with HLA-APC tetramer without peptide pulse were used as negative control.

Jurkat NFAT-luciferase activation assay

5 days after transduction, 2.5 × 10⁴ transduced Jurkat-Lucia NFAT cells were cocultured overnight with equal number of 10 µg/ml peptide pulsed Expi293F cells in a 96-well plate. Next day, the plate was centrifuged at 250 × g for 5 min. 20 µl supernatant from each well was transferred to a 96-well black plate mixed with 50 µl of prepared QUANTI-Luc™ 4 Lucia/Gaussia reagent to detect secreted luciferases per manufacturer’s instruction. For alanine substituted peptides activation assay, similar method was used. All measurements were performed on a BioTek Cytation 5 plate reader.

Funding Statement

The work was supported by the National Institute of General Medical Sciences, grant R35GM141089 to X.G.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The data that support the findings of this study are available from the corresponding author, XG, upon reasonable request.