To quantitatively compare the properties of immunoglobulin:antigen interfaces, we computed the buried surface area (BSA) and patterns of formal interactions across our representative complexes (see Methods, Supplementary Methods).

It has been previously shown that TCRs converge around diagonal engagement of pMHCs with the centre of mass of the beta chain sitting over the C-terminus of the MHC α₁ helix and the centre of mass of the alpha chain sitting over the C-terminus of the MHC α₂ helix (15, 36). This is quantified by the ‘docking/crossing angle’, calculated as the intercept of the line connecting conserved centres of mass within the variable region and the axis of the peptide binding groove. Our 60 representative TCR:pMHC complexes almost entirely comply with this canonical binding definition (µ: 45.8°, sd: 17.0°, Supplementary Figure 2A ), with one outlier [Protein Data Bank (37) identifier (PDB ID) 4Y19 (38)] that engages pMHC with a diagonal but reverse polarity mode (136.7°; i.e. the VB chain sits above the MHC α₂ helix while the VA chain occludes the MHC α₁ helix, Supplementary Figure 2B ).

To visualise their binding orientations, we aligned our set of TCRm complexes along the canonical peptide binding groove axis ( Figure 2 ). While two of the TCRms bound in a non-diagonal fashion, the rest (80% of non-redundant TCRms) adopted a diagonal pMHC binding mode that fell within the range of absolute docking angles set by TCRs ( Figure 2 ). However, in contrast to the TCRs, TCRm diagonal pMHC binding is more frequently achieved using reverse polarity (where VH is in the position of VA and VL is in the position of VB), reinforcing the notion that antibody and TCR chains with similar gene recombination mechanisms are not necessarily analogous in terms of pMHC recognition.

Convergence upon diagonal pMHC binding across TCRs is thought to be driven by improved typical TCR specificity, achieved by positioning the most hypervariable loops within interaction distance of the peptide, the key locus of pMHC variability (35). We therefore surveyed the properties of diagonally versus non-diagonally engaging TCRms to investigate to what extent this property correlates with more TCR-like pMHC recognition ( Supplementary Figure 1 ).

The property distributions indicate that TCRm diagonal engagement is not systematically linked with total pMHC BSA ( Supplementary Figure 1E ) nor a higher fraction of peptide buried surface area ( Supplementary Figure 1F ). Some diagonal modes result in few formal interactions between the TCRm CDR3s and the peptide (e.g. Hyb3, with just two interactions to the peptide from CDRH3), while others result in numbers large even by TCR standards ( Figure 1J ; 3M4E5 has three formal interactions to the peptide from CDRH3 and seven from CDRL3). However, one property that appears to be systematically linked to TCRm diagonal engagement is the ability to bury a greater number of peptide residues; where N is the total number of peptide residues, non-diagonal pMHC class-I modes bury a maximum of N-4 residues, while diagonal modes frequently bury more, up to a currently-observed maximum of N-2 ( Supplementary Figure 1G ).

Overall these mixed results show that not all diagonal TCRm pMHC binding modes yield TCR-like pMHC engagement profiles, and that this property alone is likely insufficient for capturing the specificity of a TCRm.

We next performed molecular dynamics studies on three pMHC contexts for which we have a crystal structure in the PDB of at least one TCRm partner and at least one natural/affinity-enhanced TCR partner, allowing us to investigate their binding energetics. We selected three cases studies in which HLA-A*0201 presents a different peptide antigen:

(1) Wilms’ Tumor 1 (WT1) antigen (RMFPNAPYL), bound to two TCRms and one affinity-enhanced TCR; ESK1 (PDB ID: 4WUU), 11D06 (PDB ID: 7BBG), and a7b2 (PDB ID: 6RSY), respectively.

(2) New York esophageal squamous cell carcinoma 1 (NY-ESO-1) antigen (SLLMWITQV), bound to one TCRm, one natural TCR, and three affinity-enhanced TCRs; 3M4E5 (PDB ID: 3GJF), sp3.4 (PDB ID: 6Q3S), NYE_S1 (PDB ID: 6RPB), NYE_S2 (PDB ID: 6RPA), and NYE_S3 (PDB ID: 6RP9), respectively.

(3) p53_R175H neoantigen (HMTEVVRHC), bound to one TCRm and three natural TCRs; H2 (PDB ID: 6W51), 1a2 (PDB ID: 6VQO), 12-6 (PDB ID: 6VRM), and 38-10 (PDB ID: 6VRN), respectively.

To characterize the energetic differences between the bound and unbound states of each complex, we ran sets of thirty 5-ns all-atom explicit solvent molecular dynamics simulations, using Molecular Mechanics Generalized-Born Surface Area (MMGBSA) (39–41) to compute overall free-energy changes upon binding as well as the per-residue contributions from both the pMHC and TCR/TCRm (see Methods, Supplementary Methods, Supplementary Figures 3–5 ). We used an ensemble sampling approach, i.e. a set of short statistically independent simulations initiated with different random velocities from the same starting structure, to ensure thorough sampling of the bound configurations.

Quantifying the relative contributions of the peptide, MHC α₁ helix, and MHC α₂ helix to the overall free energy change of the pMHC (see Supplementary Methods for equations), we find that both natural TCRs and affinity-enhanced TCRs tend to gain a larger proportion of their binding energy through the peptide than do TCRms ( Figure 3 , affinity-enhanced TCRs labelled TCR*). Within a given pMHC case study, no single TCRm achieved a higher proportion of binding energy from the peptide than any TCR. It is striking that some peptide antigens appear more tractable than others in terms of achieving a significant contribution to the overall binding energy. For example, even an affinity-enhanced TCR engaging the WT1 pMHC only reaches around 40% peptide contribution to binding free energy, while natural tumor-infiltrating lymphocytes complementary to p53_R175H achieve up to 65%.

We also considered the energies contributed by individual residues to identify interaction hotspots ( Table 1 ; Supplementary Figures 2–4 , Supplementary Table 4 ). Arbitrarily, we define interaction hotspots as individual residues that have a predicted free energy of -7 kcal/mol or stronger. All the natural TCRs assessed have at least one hotspot residue in the peptide, suggesting this may be a frequent feature of natural peptide recognition and consistent with current hypotheses on TCR recognition (43, 44). That the three p53_R175H specific TCRs found in tumor-infiltrating lymphocytes achieve more neoantigen recognition at R7 rather than H8 (the site of somatic mutation) shows that this binding hotspot does not necessarily lie at the residue that distinguishes self from non-self.

By contrast, only TCRms were found to have hotspots that lie within the MHC (3M4E5 and 11D06). Despite their origins in different light chain loci (IGLV2-11 and IGKV1-5, respectively; Supplementary Table 5 ), the hotspot occurs at the same residue in the MHC α₁ helix (R65), which in both cases forms a salt bridge to D56 of CDRL2. That such non-peptide hotspots were not readily observed in TCR formats might indicate that current TCRms are prone to higher inherent affinity for the MHC and thus poorer typical peptide specificity. We observed from the crystal structures that in TCRm binding events that led to an energetic hotspot (3GJF, 7BBG), R65 tends to adopt a more extended conformation that protrudes towards the immunoglobulin, whereas in the TCR binding events without hotspots (6RSY, 6Q3S, 6RP9, 6RPA, 6RPB, 6VRM, 6VRN, 6VQO), R65 tends to adopt conformations that orient the plane of its guanidino group in line with that of the MHC α₁ helix ( Supplementary Figure 6 ). This observation holds throughout our molecular dynamics simulations, as measured by the tilt of the plane of the guanidino group relative to the plane of the MHC α₁ helix over each of the 30 independent trajectories ( Supplementary Figure 7 , Supplementary Methods).

Finally, to characterize the differences in the energetics of CDR binding to the pMHCs, we computed the fractional contributions of each CDR to the total CDR free energy change ( Figure 3 ). All TCRms and most TCR contexts were CDR3-dominated in their binding energetics. The balance of CDR[H/B]3 to CDR[L/A]3 energy reflected the picture seen in the general analysis of interface properties: TCRms gain more binding energy through CDRH3 than CDRL3, while in TCRs either CDR3 can be energetically dominant.

To enable a direct comparison between antibodies and TCRs, and between heavy and light chains, the IMGT numbering scheme and CDR definitions were used throughout this work [CDR1: IMGT residues 27-38, CDR2: IMGT residues 56-65, CDR3: IMGT residues 105-117 (49)]. ANARCI was used to number all sequence inputs (50). Where necessary, MHC chains were renumbered to enable a direct comparison between TCRm and TCR complexes.

SAbDab (6, 7) and STCRDab (4) were downloaded on 30^th September, 2022. All complexes were stripped of explicit hydrogens, heteroatoms, and water molecules, and only immunoglobulins with protein antigens were considered. These databases were then mined for particular subcategories as follows:

1. TCR-mimetic antibody (TCRm) complexes were identified by filtering SAbDab with the search terms ‘HLA’ and ‘MHC’, followed by manual validation.

2. Representative sets of non-redundant high quality antibody:antigen and TCR:pMHC (αβ only) complexes were derived by first filtering SAbDab and STCRDab for structures of complexes solved by X-ray crystallography to 2.5 Å resolution. For each class of immunoglobulin separately, the IMGT-defined CDR sequences (49) were concatenated in a consistent order and used as inputs to greedy clustering by cd-hit (51) (80% sequence identity threshold), to create ‘paratope clusters’. To identify cases where chemically-similar paratopes bind to significantly different antigens, we performed a second round of clustering over the paratope clusters using the concatenated sequence of the antigen(s) associated with each antibody (concatenated in descending length order), or the sequence of the presented peptide for the TCRs. cd-hit was run at an 80% sequence identity threshold, allowing a minimum alignment length of as little as 20%. This ensured truncated antigens were not considered as different targets (e.g. an anti-coronavirus antibody solved in complex with the receptor-binding domain would be considered the same context as a complex of the same antibody binding to the full-length spike protein). This resulted in sets of 824, 52, and 8 representative antibody:antigen, TCR:pMHC (class 1) and TCR:pMHC (class 2) complexes respectively.

3. Complexes of TCRs and TCRms engaging the same peptide were obtained by searching SAbDab and STCRDab by antigen sequence for peptide fragments with a 100% sequence identity match.

Complete immunoglobulins were separated from their antigens (i.e the co-ordinates of the partner were deleted in a copy of the original file), and the difference in solvent-accessible surface area between each residue in the original complex and artificially-generated ‘apo’ state was recorded. Per-residue solvent-accessible surface area (SASA) was calculated using an in-house implementation of the Shrake and Rupley algorithm (52), applying a probe radius of 1.4Å. Residues were defined as buried upon binding if their SASA decreased when the co-ordinates of the binder partner were re-introduced into the system.

The Arpeggio (53) software package was used to assign interactions across each set of complexes. Arpeggio featurises input protein complexes with predicted interaction types based on the relative distances and orientations of PDB atom types; see (53) for definitions. In our analysis, we only considered the four most attractive interaction types: hydrophobic, aromatic, hydrogen bond and salt bridge interactions. We disregarded weak hydrogen bond assignments, as they ignore the relative angular dependence of hydrogen bond acceptor and donor. To interpret arpeggio interaction outputs on a per-residue level requires a consolidation of atom-level interaction profiles, which we performed according to the following post-processing steps:

1. Where two residues had multiple hydrophobic atom-atom contacts this was recorded as one residue-residue “hydrophobic” interaction.

2. Where two residues had multiple aromatic atom-atom contacts this was recorded as one residue-residue “aromatic” interaction.

3. A residue-residue pair can be labelled as contributing both a hydrophobic and an aromatic interaction due to their different electrostatic origins.

4. Polar interactions between oppositely charged residues were initially labelled as ‘salt bridge/hydrogen bond’. Once all interactions were recorded the salt bridge was assigned to the pair of residues with the closest atom-atom pair and all other interactions were considered hydrogen bonds, ensuring only one salt bridge was recorded per positive/negative charge pair.

pMHC-TCR and pMHC-TCRm complexes were prepared in AmberTools21 (54) (see Supplementary Table 6 for rebuilt sections) and all simulations run in OpenMM v7.5 (55). Cis peptide bonds and chiral centers with incorrect stereochemistry within rebuilt sections of protein structures were identified and corrected using the Cispeptide and Chirality (56) plugins of Visual Molecular Dynamics (57) v1.8.3. Corrected peptide bonds were maintained in the trans configuration with periodic torsion restraints with force constant 500 kJ/mol. These restraints were removed during the final unrestrained equilibration along with positional restraints on C_α atoms. Complexes were then solvated in orthorhombic boxes with a minimum of 1.4 nm between protein atoms and the box edge. Sodium and chloride ions were added to neutralize each system and bring the salt concentration to 0.15 M.

All simulations were carried out using FF14SB (58) and TIP3P (59) forcefield parameters with a Langevin thermostat (friction coefficient 1 ps^-1) and, for constant-pressure simulations, a Monte Carlo barostat (pressure 1 bar). Complexes were minimized and then heated to 298 K at constant volume with protein atoms restrained. These restraints were relaxed over a set of five 100-ps simulations at constant pressure, culminating with 100-ps of unrestrained equilibration. Finally, 5-ns production simulations at constant pressure were run. Thirty independent replicas of this protocol were executed for each complex (60). MMGBSA calculations were performed with MMPBSA.py (42) using 100 frames collected every 40 ps from the final 4 ns of each 5-ns trajectory (i.e., 3000 frames per complex). Per-residue energy contributions to the binding energy were computed using the MMPBSA.py DECOMP functionality (42). Further detail on all simulation methods can be found in the Supplementary Information.

All manuscript visualizations were created using open-source PyMOL or matplotlib version 3.5.2.