Experimental Structures of Antibody/MHC-I Complexes Reveal Details of Epitopes Overlooked by Computational Prediction

This article is featured in Top Reads, p. 1255

The production and use of mAbs have proved revolutionary for addressing fundamental questions of a range of problems, including protein structure and function, cell differentiation, immunodiagnostics, and immunotherapy (1–3). mAbs directed against the cell surface histocompatibility Ags H2 in the mouse and HLA in the human have been particularly useful in a wide range of applications (4–7), including the exploration of structural aspects of MHC molecules (8, 9) and HLA typing for organ transplant (10). Anti–MHC class I (anti–MHC-I) mAbs have also been used for biosynthetic studies of the cell biology and assembly of MHC molecules as they proceed from synthesis in the endoplasmic reticulum through steps of folding, glycosylation, peptide acquisition, and trafficking to the cell surface for recognition by receptors on T cells, NK cells, and other immune and inflammatory cells (9, 11, 12). Recently, in efforts to target specific peptide/MHC (pMHC) complexes characteristic of tumor cells, mAbs that mimic TCR-mediated recognition have been isolated and are in development as part of the therapeutic armamentarium for a variety of malignancies (13). TCR mimic mAbs have potential value in treating autoimmune diseases as well (14).

Many anti–MHC-I mAbs were initially defined by their reactivity against genetically well-defined strains of inbred animals (15) or characterized by their reactivity against panels of well-known cell lines or purified HLA molecules (16) or domain-shuffled MHC-I transfectants to map the location of their epitopes (17–24). Although such approaches have contributed to an understanding of specific sites that define MHC polymorphism and control interaction with TCR, coreceptors, NK receptors, or other inhibitory or activating receptors (25–27), structural analysis offers to elucidate further details of the binding sites and to provide insight into which mAbs may compete for binding by ligands with known sites of interaction. These structures can provide a basis for engineering Abs with increased affinity or improved specificity. In addition, precise knowledge of the antigenic epitopic residues provides a structural basis for the transfer of specific recognition sites to other allelomorphs or even unique engineered proteins.

To understand better the structural details of the reactivity of anti–MHC-I mAb, we determined by X-ray crystallography the structures of complexes of four mAb Fab with pMHC-I: two that bind distinct regions of H2-D^d [34-5–8 (α2 domain) and 34-2–12 (α3 domain) (28)], one that binds the conserved α3 domain of both H2-L^d and -D^b [28-14-8 (29)], and one that discriminates a single–amino acid polymorphism of the β₂-microglobulin (β₂m) L chain subunit of MHC-I [S19.8 (30, 31)]. These experimental structures reveal details of the footprints of their respective Fab on MHC-I consistent with previous biochemical, genetic, and immunological studies. In addition, the structures pinpoint side chain interactions, explain allele specificity, and shed light on conformationally plastic regions of MHC, particularly with respect to changes observed in the α2 domain on peptide binding.

With the X-ray structures in hand, we evaluated the ability of one computational algorithm, AlphaFold-Multimer (AF-M) (32) to predict and visualize complexes of these selected Abs with their respective MHC-I protein Ags. Deep-learning methods, such as AlphaFold, have been remarkably successful for prediction of protein structures from amino acid sequence (33, 34), particularly with respect to individual domains of structured proteins. The release of three-dimensional models of the entire human proteome already promises rapid progress in rational approaches to drug discovery and understanding fundamental mechanisms of cellular biochemistry (35). Although computational determination of the organization of multidomain proteins and multimolecular complexes is clearly a more challenging problem than domain prediction alone (36), AF-M (32) offers an opportunity for predicting and evaluating higher-order interactions (37). The availability of AF-M implemented in ChimeraX (38, 39) and linked to Google Colab servers (40) permits rapid assessment of models of a variety of protein complex structures. We applied this modeling approach to these four mAb/MHC-I complexes. Although the resulting computation generated good models for the previously well-known MHC-I/β₂m complexes, the algorithms were less successful in the prediction of Fab and V region fragment (Fv) structures. Computational models failed to properly identify the sites where the Fv V_HV_L docked on the MHC-I. The discrepancies between experiment and prediction arise from difficulties in establishing proper domain relationships as shown by elbow angle, ambiguities in loop structures, particularly the Ab CDRs, and to complexities in the docking of Ab with protein Ags. Accumulation of experimental structural data on protein Ag/Fab complexes should provide a more extensive database for improvement of algorithms for structure prediction.

Cells producing 34-5-8, 34-2-12, and 28-14-8 (originally designated 34-5-8S, 34-2-12S, and 28-14-8S and occasionally referred to as 34.5.8, 34.2.12, and 28.14.8) were the gift of Drs. Keiko Ozato (Eunice Kennedy Shriver National Institute of Child Health and Human Development, National Institutes of Health [NIH]) and David Sachs (National Cancer Institute, NIH) (28, 29). 34-5–8 and 34-2–12 derive from C3H (H2^k) anti-BDF1 (H2^bxd) and 28-14-8 was derived from C3H/HeJ (H2^k) anti-C3H.SW (H2^b) responses. Cells producing S19.8 [SJL (H2^s/β₂m^a) anti-B10.S (H2^s/β₂m^b)] were obtained from Dr. Ulrich Hämmerling (Memorial Sloan Kettering Cancer Center) (30). Cells were maintained in culture in DMEM, high glucose, supplemented with 10% FBS, in a 7.5% CO₂ atmosphere at 37°C. Cell culture supernatants were collected, purified by passage over Protein A Sepharose (Cytiva), washed with 0.45 M NaCl, 10 mM Tris, pH 8.0, eluted with 0.1 M glycine HCl, pH 3.0, into 1 M Tris, pH 8.0, and dialyzed against 1× PBS. Fab fragments were then prepared and purified following papain digestion and protein A purification to remove undigested molecules and Fc fragments, followed by size exclusion chromatography on either Superdex 200 or Superdex 200 increase columns (Cytiva). In vitro expressed and refolded H2-D^d,-D^b were prepared with β₂m^b and peptides RGPGRAFVTI (HIVIIIB glycoprotein peptide, for H2-D^d) and ASNENMETM (influenza peptide, for H2-D^b) as described previously (41). The MHC-I molecules used for these structures were expressed in Escherichia coli and are not glycosylated. Glycosylation at N86 and N176 of H2-D^d and N86, N176, N256 of H2-D^b would not be expected to impinge directly on the binding of the mAbs studied here, because the recombinant E. coli–expressed MHC-I molecules all react well with the indicated mAbs. Binding of the H2-D^d–specific mAb, 34-2-12, however, because it makes contact with K256, would be expected to be sterically impeded from interacting with H2-D^b, H2-L^d, or H2-K^d due to their N256 carbohydrate.

Total RNA was extracted from 10⁷ hybridoma cells using the Monarch total RNA extraction kit (New England BioLabs, Ipswich, MA) following the manufacturer’s instructions. Two µg RNA served as a template for cDNA synthesis using oligo dT and murine leukemia reverse transcriptase as implemented in the OneTaq RT PCR kit (New England Biolabs). A panel of oligonucleotides designed to amplify mouse Ig V genes as described by Wang et al. (42) was used to PCR amplify and then sequence V_H and V_L from each hybridoma. The encoded protein sequences are shown in Supplemental Fig. 1.

Equimolar amounts of purified Fab and MHC proteins were incubated at 25°C for 2–3 h. Complexes were isolated on a Superdex 200 increase column in 1× PBS, concentrated and buffer exchanged into 25 mM Tris, pH 8.0, 50 mM NaCl in preparation for crystallization.

Crystallization conditions were identified by screening hanging drops at 18°C. Crystals of Fab28-14-8/H2-D^b were grown in 16% PEG 4000, 0.1 M HEPES, pH 7.5, 0.2 M MgCl. Crystals of Fab34-5-8/H2-D^d were grown in 18% PEG 4000, 0.1 M MES, pH 6.0, 0.12 M Ca acetate. Crystals of Fab34-2-12/H2-D^d were obtained in 12% PEG 8000, 0.1 M MES, pH 6.5. Crystals of FabS19.8/H2-D^d grew in 0.5 M ammonium sulfate, 0.1 M Na citrate, pH 5.6, 1.0 M LiSO₄, and were further improved by seeding. Crystals were cryoprotected in mother liquor containing 10% ethylene glycol and flash frozen in liquid nitrogen. Diffraction data were collected (at wavelength 1.033 Å, in N₂ stream at ∼100 K) at Southeast Regional Collaborative Access Team (SER-CAT) beamline 22ID at the Advanced Photon Source, Argonne National Laboratory, and processed with XDS (43) to 2.6 Å, 2.7 Å, 2.8 Å, and 2.9 Å resolution for Fab28-14-8/H2-D^b, Fab34-5-8/H2-D^d, Fab34-2-12/H2-D^d, and FabS19.8/H2-D^d/β₂m^b, respectively (see Table I). The structures were solved by molecular replacement with Phaser (44) using H2-D^b from Protein Data Bank (PDB) 1WBX or H2-D^d from PDB 5WEU as search models. For the Fab search, we used a model of DX17 Fab (not yet deposited) with the CDR loops trimmed as the initial search model. Molecular replacement solutions were subjected to several rounds of refinement with Phenix (45, 46) interspersed with manual building in Coot (47). R_work/R_free (%) values for final refined models of Fab28-14-8/H2-D^b, Fab34-5-8/H2-D^d, Fab34-2-12/H2-D^d, and FabS19.8/H2-D^d are 21.5/25.5, 22.7/25.7, 23.5/28.6, and 19.2/24.9, respectively. Data collection and refinement statistics are summarized in Table I. Graphics were generated with PyMOL (48) and ChimeraX (38, 39).

Analysis of X-ray structures and comparison of computational predictions with experimental structures was carried out with a variety of programs, including S_c (49), CNS1.3 (50), PDBePISA (51), PyMOL (48), and ChimeraX (39). Ab V_H and V_L sequences, determined from cDNA as described above, and the corresponding MHC-I and β₂m sequences were entered into the AlphaFold structure prediction module of ChimeraX 1.6 (Tools>Structure Prediction>AlphaFold>Predict), which queried Colab via Google servers. The resulting “best_model” was further analyzed and compared with our X-ray structure of the same complex. In cases where the X-ray structure contained more than one complex in the asymmetric unit, the first complex was used. Center of mass for the indicated chains was calculated in ChimeraX (39).

Surface plasmon resonance (SPR) experiments were performed as described previously (52) at 25°C on a Biacore T200 (Cytiva, Uppsala, Sweden) in 10 mM HEPES, pH 7.4, 150 mM NaCl, 3 mM EDTA, and 0.05% Tween-20 at a flow rate of 30 µl/min. mAb S19.8 (Santa Cruz Biotechnology, catalog no. SC-32241) was repurified on a Superdex 200 increase column (Cytiva) in PBS and was immobilized on a series S CM5 sensor chip (Cytiva) by amine [1-ethyl-3-(3-dimethylaminopropyl)carbodiimide and N-hydroxysuccinimide] coupling. A reference cell was mock coupled to allow for background subtraction. Various MHC-I proteins described in the figure legends were prepared with either human β₂m or mouse β₂m^b. The binding surface was regenerated with 5 mM phosphoric acid. Binding was performed with graded concentrations as indicated in the figure legends. Sensorgrams were globally fit in steady-state mode using Biacore T200 Evaluation Software version 3.1 and plotted with GraphPad Prism. Proteins used in SPR were HLA-A*02:01/hβ₂m/flu peptide, HLA-A*02:01/mβ₂m^b/flu peptide, H2-K^b/hβ₂m/ova peptide, H2-K^b/mβ₂m^b/ova peptide, mβ₂m^b or human β₂m. Flu peptide is GILGFVFTL, representing influenza matrix protein M1 (residues 58–66) and ova peptide is SIINFEKL, OVA (residues 257–264). The indicated complexes were refolded from bacterially expressed inclusion bodies and synthetic peptide and purified by standard methods (41).

To understand details of the interactions of the four anti–MHC-I mAb, we purified refolded recombinant H2-D^d/and H2-D^b/β₂m complexes containing high-affinity peptides and prepared Fab of the mAb as described in detail in Materials and Methods. Diffraction quality crystals of Fab/MHC complexes were obtained, and X-ray data sets were collected at resolutions from 2.60 to 2.90 Å (see Table I). Complexes of the four Fab/MHC-I complexes were readily solved by molecular replacement and refined as summarized in Materials and Methods and Table I. Previous serological and functional studies had mapped 34-5-8 to the α1α2 peptide-binding domain and 34-2-12 and 28-14-8 to the α3 domains of H2-D^d and H2-L^d/D^b, respectively (17). The 34-5-8 epitope [peptide dependent but not peptide specific (53)] was similarly mapped to particular residues of the α2 domain (54–56). S19.8 recognized MHC-I molecules containing the β₂m^b allelomorph (Ala85) and not β₂m^a (Asp85) (31).

As shown in Fig. 1, 34-5-8 interacts only with residues of the H2-D^d α2 domain, exploiting contacts of its V_H and V_L. The overall structure of the pMHC-I complex is the same as that of some 19 previously determined H2-D^d structures in the PDB and reveals root-mean-square deviation (RMSD) values for H2-D^d ranging from 0.762 to 1.545. Similarly, the RMSD of β₂m varies from 0.434 to 0.824 Å. The basic folds of the Fab V_HV_L and C_H1C_L of 34-5-8 are clearly representative of a host of previously determined Fab structures. The footprint of the Fab V_HV_L on H2-D^d is determined by interactions of 19 residues of the α2 domain (but none of β₂m), including 17 that contact the Fab H chain and 7 that contact the L chain, of which 5 are in common. The buried surface area of the MHC-I chain is 920 Ų, of which 710 Ų is due to the Fab H chain and 210 Ų to the L chain (see Supplemental Table I), and the shape complementarity [S_c (49)] between V_HV_L and H2-D^d is 0.637, characteristic of Ag/Ab interfaces. The region of the footprint is illustrated in Fig. 1D with contacts of the Fab to residues of the β-sheet floor (residues 104–111 and 127–132) and to several of the α2-2 helix (154, 157, 161, 162, 165, and 169). Residues 104–108 form a tight turn (focused on E104 and R106) that is engaged by H chain residues of CDRH1 (A28, S31, and Y32), CDRH2 (residues 52–57), and CDRH3 (residues 99–104) (Fig. 1E, 1F). L chain residues of CDRL2 (Y52, R54, N57, D59, S60) and of CDRL3 (E97 and W101) also contact H2-D^d.

Examination of the Fab/H2-D^d interface explains the private specificity of the mAb. The unique H2-D^d residue E104 as well as G107 and R157 are contacts made by the H chain, whereas the L chain contacts residues conserved among a sampling of other murine MHC-I allelomorphs (see Fig. 1F). Previous mutational analysis revealed reduced 34-5-8 reactivity of E104G and G107W, and the suggestion of improved binding of W97R (57), a residue that sits in the peptide binding groove but does not make direct contact with mAb 34-5-8. Because the W97 side chain is directed into the peptide binding groove and is distant (W97 NE1 to H chain Y100 O is 16.8 Å away) from the site of interaction with the 34-5-8 Fab, this likely reflects a change in β-strand 5 that influences the conformation of the 103 to 109 loop, which serves as a direct site for interaction with the Fab H and L chains. Another region of H2-D^d that contacts Fab 34-5-8, residues 127–132 (see Fig. 1F), also plays a crucial role in recognition by this mAb, as evidenced by the observation that the H2-D^dm6 mutation W133R obliterates binding to 34-5-8 (58). Although the completely conserved W133 is adjacent to but not directly in contact with the mAb, our modeling of the R substitution at this position clearly indicates distortion of the β8 strand and the peptide binding groove. Thus, the X-ray structure of the Fab34-5-8/H2-D^d complex confirms the results of earlier exon-shuffling and mutagenic studies and explains in detail the specificity of the mAb. Furthermore, the structure indicates that the peptide-dependent, but not peptide-specific, recognition by the mAb reflects the sensitivity of the 104–111 loop as an indicator of peptide binding.

The Fab34-2-12/H2-D^d complex, illustrated in Fig. 2, shows that the Fab H and L chains recognize three loops at the membrane proximal surface of the H2-D^d α3 domain. The p/H2-D^d/β₂m structure closely resembles that of other independently solved H2-D^d structures (MHC-I heavy plus light, RMSD from 0.708 to 2.163; H chain RMSD from 0.692 to 1.775, and β₂m from 0.289 to 0.887 Å). The interface between the Fab and H2-D^d involves 21 residues of the H2-D^d H chain, 19 of which interact with the Fab H chain and two with the Fab L chain, burying some 757 Ų of the H2-D^d α3 domain with an S_c value (49) of 0.668. Of particular note is that 34-2-12 envelops the region of the α3 domain, residues 219 to 227 (Fig. 2F), a region that is bound by the costimulatory T cell molecule, CD8αβ, described crystallographically by PDB 3DMM (59). Additionally, Fab34-2-12 binds membrane proximal loops of H2-D^d (residues 194–197 and 247–257), suggesting that it might bind H2-D^d molecules lying in a supine position on the cell surface: The C-terminal strand from β17 would be expected to have access to a peptide strand connecting this to the transmembrane region of H2-D^d, which would run from residue 271 to about residue 288. [Evidence for such an orientation of H2-K^b on a lipid bilayer has been reported (60).] The observation that 34-2-12 binds only amino acid residues of the α3 domain is consistent with experiments showing that it blocks function of CD8⁺ cytolytic cells (21) and stains cells expressing recombinant truncated H2-D^d α1α2 deletion mutants (61) and MHC-II/MHC-I hybrid molecules (62). H2-D^d mutants that abrogate 34-2-12 binding also reveal diminished susceptibility to cytolysis by CD8⁺ cells (21, 63). mAb 34-2-12 makes contact with K256 of the H2-D^d H chain. In H2-K^d, -D^b and -L^d, N256 is a glycosylation site, and the oligosaccharide would be expected to sterically block the mAb interaction via the CDRH2 residues, as suggested by superposition of Fab34-2-12/H2-D^d with glycosylated H2-D^b of 6MP0 (64).

Consistent with previous mapping, the structure of the Fab28-14-8/H2-D^b complex, illustrated in Fig. 3, reveals direct contact of both the H and L chains primarily with residues of the α3 domain of H2-D^b. Additionally, two residues of the α2 domain (conserved R111 and E128) and also one residue of the β₂m L chain (conserved D59) are bound by the Fab. The Fab buries 850 Ų of the H2-D^b H chain and rather little (10 Ų) of the β₂m L chain (Supplemental Table I), and the S_c of the interface (49) is 0.704. The focus of the Fab is on an extended region in the center of the α3 domain, involving residues 212–226, a region that largely overlaps the main contacts of the CD8αβ heterodimer with H2-D^d (PDB 3DMM) (59). The minimal contact with β₂m is consistent with earlier findings that 28-14-8 binds β₂m-free surface MHC-I molecules as well as H2-D^b expressed in a β₂m-negative cell line and α1α2 deletion mutants of H2-D^b as well (65). Structural alignment of H2 allelomorph sequences with attention to residues that distinguish those molecules that bind 28-14-8 (H2-D^b, -L^d) from those that do not (H2-K^b,-D^d,-K^d,-K^k,-D^k) indicates that R260, bound by Fab H chain Y101 and L chain G91, is critical for the allelic specificity of the Ab.

Fab S19.8 represents a specificity focused on the β₂m L chain of H2 complexes rather than on polymorphic structures of the H2 heavy chains. The S19.8 hybridoma [SJL (H2^s/β₂m^a) anti-B10.S (H2^s/β₂m^b)] was originally designated anti-ly-m11.2 (30) and subsequently identified as anti-β₂m^b (31, 66, 67). Binding studies using purified S19.8 and a selection of recombinant mouse and human MHC-I molecules prepared with both allelic forms of β₂m indicate that S19.8 distinguishes β₂m^b from β₂m^a when in complex with H2-K^b (Fig. 4). Additionally, the mAb binds free β₂m^b but does not bind human β₂m. The K_D value for binding to H2-K^b/β₂m^b is 1.3 × 10⁻⁷ M, whereas that for β₂m^b alone is ∼5-fold weaker at 6.3 × 10⁻⁷ M (Fig. 4). The X-ray structure of the FabS19.8/H2-D^d/ β₂m^b complex (Fig. 5) reveals the detailed explanation for this binding behavior. S19.8 H and L chains engage 1090 Ų of the H2-D^d/β₂m^b complex, with the greater area contributed by β₂m, 900 Ų versus 190 Ų of the H2-D^d H chain (see Supplemental Table I). The overall disposition of the S19.8 Fab is toward β₂m [18 residues (Fig. 5G)] and six residues of H2-D^d (α1 residues S13, R14, P15, F17 and α2 G91, S92—Fig. 5F). β₂m residue A85 defines the polymorphism of β₂m^b, which is at the center of the S19.8 H and L chain contact area. Thus, the β₂m^a chain, with D at 85, would be expected to be sterically incompatible with the S19.8 interaction (see Fig. 5H, 5I). Reactivity of S19.8 with rat β₂m has been reported and is consistent with the identity of 13 contact residues (1, 2, 4, 32, 35, 36, 45, 81–84, 89, 90) and the similarity of 2 (85, 87). Of note are differences at positions 34, 38, and 88 that may be sufficient to affect the affinity of the interaction. The decreased S19.8 reactivity of MHC-I complexes containing human β₂m with S19.8 (Fig. 4) likely reflects major differences at positions 34 (human D for H), 38 (human D for Q), and 89 (human Q for E) (see Fig. 5G). Further binding, mutagenic, and structural studies will be needed to reveal the details of the contributions of each of the residues of the S19.8/β₂m interface in different species. The ability of S19.8 to distinguish a single–amino acid substitution at the center of its interface, despite a rather modest affinity, is a striking example of how low-affinity interactions may be the basis of clearcut molecular discrimination.

Having solved the structures of these four Fab/MHC-I complexes, we reasoned that they might provide a fair test of the ability of AF-M (32) to predict the structure of each of the Fab/MHC complexes determined experimentally. Thus, the amino acid sequences of the component chains of each of the four complexes were submitted for prediction as described in Materials and Methods. In general, the predicted individual models of the basic MHC-I and β₂m folds were good, and pMHC-I trimer or V_HV_L and C_H1C_L heterodimers were reasonable (as indicated by predicted aligned error plots in AlphaFold). The pMHC-I heterotrimers were compared with the AF-M predictions as summarized in Fig. 6, and Fab34-5-8/H2-D^d complexes showed good agreement of the determined versus predicted pMHC-I (overall all-atom RMSD of 1.814 Å and MHC-I H chain alone of 1.330 Å). For the β₂m, however, a larger RMSD (2.690 Å) was noted (Fig. 6A, 6B), indicative of some strand differences, and the peptide model also showed some distortion, evidenced by an RMSD of 2.574 Å. Differences among the α1α2 peptide binding domain (RMSD of 1.259 Å) and the α3 domain independently were smaller (1.456 Å). Evaluation of the pMHC-I structures and models of the other three complexes showed several individual variations. The H2-D^b complex with Fab28-14-8 revealed a distortion of the MHC-I H chain 103–110 loop (RMSD 5.428 Å), a region not contacted by the Fab (Fig. 6A, 6C). Analysis of the H2-D^d complex with S19.8 showed that although the MHC-I H chain and β₂m were rather similar between structure and AF-M model, loops of the peptide and α1 domain (residues 13 to 20) differed.

For the complete four domain Fabs, we examined the differences between the variation in elbow angles describing the relationship of V_LV_H to C_LC_H1, which varied widely. Differences in the elbow angles of the experimental structures as compared with those of AF-M models are summarized in Supplemental Table II. Thus, the X-ray–determined elbow angle for Fab34-5-8 is 150° as compared with the AF-M model of 136°. Similarly, Fab34-2-12, Fab28-14-8, and FabS19.8 differ by 1.0, 28.0, and 16.0°, respectively. Additional comparisons of the X-ray versus AF-M models were carried out by calculating all atom RMSD, which varied widely (Fig. 7). Clearly, wide variation exists among the residues of the V region of the Fab H chains, particularly in the CDR loops (Fig. 8), with 34-2-12 displaying the least variation and 28-14-8 showing the most. The final step, the prediction of the docking site of the Fab onto the MHC-I molecules, did not agree with the experimentally determined structures and was grossly incorrect for all four Fab/MHC-I complexes. This is evaluated graphically in Fig. 9 and quantitatively using DockQ (68) in Table II.

Although the crystallographic structure of each of the Fab/MHC complexes confirmed the general characteristics of the H/L chain associations of the Fv and C_LC_H regions and of the fold and association of the MHC H chain with β₂m, the final computational docking of the four Fabs to the pMHC/β₂m complexes contrasted starkly with the experimentally determined structures (Fig. 9). Although the Fab34-5-8 focuses on the α2 domain of H2-D^d, the AF-M predicted model docks this Fab to the α1 domain, focusing the V_HV_L onto the opposite side of the molecule (Fig. 9A–9C, gray).

The X-ray structure of Fab34-2-12 bound to the α3 domain of H2-D^d is not borne out in the AF-M prediction (Fig. 9D–F), which placed the Fab interaction on the α1α2 domains, poised approximately like a TCR or TCR-like Ab (Fig. 9F). Once again, the final docking of a reasonably well-predicted Fab onto the well-predicted H-2D^d is at a completely different and nonoverlapping site (Fig. 9E).

Fab28-14-8, an mAb that uniquely sees the α3 domain of H2-D^b and H2-L^d, verified crystallographically (Fig. 3), was computationally docked to a TCR-like binding site involving the α1α2 and peptide (Fig. 9G–9I). This docking placed Fab28-14-8 with its H chain poised over the α1 helix and its L chain over α2. Again, the final docking performed by AF-M tended to focus the Fab roughly in a TCR-like orientation (Fig. 9H).

The final example of the test of the ability of AF-M to dock an Ab to its MHC/β₂m Ag is that of FabS19.8. The X-ray structure of the complex (Figs. 5, 9J–9L) shows that the Fab interacts predominantly with the β₂m^b (Ala85) MHC L chain and also with six residues of the H2-D^d H chain. The Fab L chain sees some nine residues of both β₂m and H2-D^d. AF-M placed the S19.8 Fab directly over the α1α2/peptide surface, again, much like a TCR (Fig. 9K–9L), although experimentally the Fab clearly focuses on β_{2 m}.

Thus, AF-M in all cases failed to properly identify the crystallographically defined epitopes identified by the Abs. In three of the four cases, the AF-M–dependent docking misconstrued the position of the Fab to be similar (but not identical) to that seen in dozens of TCR/MHC examples. Measurement of differences in the location of the center of mass of each of the chains in the superpositions is given in the legend to Fig. 9.

Anti-MHC mAb raised in precise genetic backgrounds, particularly in the mouse, have proved crucial in studies of Ab allotypes (69) and of genetic polymorphisms controlling histocompatibility and immune responsiveness (70). The exploitation of mAb directed against human (71) and mouse (28) MHC Ags has revolutionized tissue typing for transplantation and our understanding of the genetic basis of immune responsiveness. Structural studies of the MHC class I (72) and class II molecules (73, 74) have contributed to a clear illustration of how MHC molecules bind peptide as a necessary prerequisite for presentation of antigenic peptides to CD4 and CD8 T cells. Exactly what regions of MHC molecules are bound by mAb that characterize particular domains or conformationally labile regions of MHC molecules remains in many cases poorly defined. Understanding with precision the nature of the epitopic sites seen by anti-MHC mAb may provide a basis for further engineering to optimize Ab/MHC. Careful inspection of the sites of interaction provides new insights into aspects of the peptide-dependent plasticity of the MHC-I molecule, the conserved site of α3 that interacts with CD8, or the location of a single-residue polymorphism of β₂m.

The structures of the four Fab/MHC-I complexes reported here reflect the domain organization of MHC-I molecules, peptide-dependent conformations of MHC-I, and MHC-I and β₂m polymorphism. Early exon-shuffling experiments and subsequent mutagenesis and allele screening successfully identified domains of the molecules recognized—Fab34-5-8 binds the α2 domain, and Fabs34-2-12 and 28-14-8 bind α3, although they approach the MHC molecule from rather different perspectives. The peptide-dependent mAb 34-5-8 senses conformational changes in the 104–108 loop that result from peptide binding, because this Fab does not interact with bound peptide. 28-14-8 addresses the α3 domain via a footprint that closely mimics that of the CD8αβ coreceptor ligand, whereas 34-2-12 approaches α3 via the membrane proximal loops, with a somewhat peripheral interaction with the CD8 binding loop of the α3 domain. S19.8, an anti-β₂m mAb, is focused on a single–amino acid polymorphism and, as expected, centers its Ag binding site on the polymorphic residue Ala85 of β₂m. Although directed to Ala85 of β₂m, S19.8 binds β₂m both free and when in complex with an MHC-I H chain. Each of these four Fabs encounter residues that, on substitution, abolish tight interaction and thus account for their allelic specificity.

Of the Fab/MHC-I complexes studied here, S19.8 serves as an example of an mAb that distinguishes a single-residue substitution, Ala85 for Asp85 of β₂m. The relatively low affinity of S19.8 for MHC-I/β₂m (∼0.1 mM) suggests that the discrimination of the mAb is dependent on the lack of reactivity for the Asp85 variant rather than a strong association with Ala85. This result should also be considered in the context of a number of other Abs that discriminate single-residue differences, such as antiallotype Abs to Igs (75–77) and Abs that discriminate allelomorphs of cell surface receptors such as Thy-1 (78).

In addition to precise identification of the epitopes recognized by these four Fabs, the structures of the Ab/MHC complexes provide excellent test cases for the advanced prediction program AlphaFold2 and its computational progeny, AF-M. AlphaFold2 was highly successful in predicting the individual domains of the two murine peptide/MHC-I/β₂m (H2-D^d and H2-D^b) heterodimers studied and also of the four Fv segments of the Fab heterodimers examined. The interactions of the MHC-I with β₂m predictions were considered acceptable to high (exhibiting DockQ scores greater than 0.8; see Table II), likely due to the large number of identical or similar experimental structural models in the protein database. The predictions of the interactions of the H and L chains of the Fab were also acceptable, but the docking scores of the four Fab binding to their pMHC Ags were incorrect (scores <0.2). Thus, when presented with sequences of heterodimers from a well-represented class of molecules, such as MHC-I, AF-M performs well. Also, AF-M performed well with respect to the prediction of the basic fold of the Fv region of the four Fab, consistent with prior assessment of the veracity of such computational predictions (79). As expected, however, the backbone and side chains of the CDRH3 regions were less accurately ascertained (Fig. 8). Most important, the ability of AF-M to predict the docking/binding of the Fab on the surface of the pMHC/β₂m (i.e., the identification of the epitopic sites) was consistently incorrect.

Thus, our results examine the ability of the structure prediction algorithms (1) to recognize and predict structure of individual domains, (2) to assemble and orient individual domains of a protein chain, (3) to generate the heterodimers of complexes frequently represented in the structural database, and (4) to produce models of the docking of the Fab V_H/V_L heterodimer (paratope) onto its MHC-I heavy, heavy/β₂m, or β₂m epitope. Current views of Ab structure prediction recognize both the power of AlphaFold2 and its limitations, particularly in assessment of the CDR loops of Abs (79–82). Protein Ag/Ab complexes continue to pose a major challenge for artificial intelligence (AI)-powered prediction programs in several respects. Although the fold of the Ab itself is highly conserved and the framework of Fab structure is almost identical for all Abs, the six Ag-binding sites established by the CDR loops that engage the epitope vary widely in length and composition. These six hypervariable loops are flexible and present a repertoire of dynamic states in solution that impose difficulties in accurately predicting a stabilized structure (83). In the cases we examined, although the folds of the Fv portion of the Fab were predicted correctly, the hinge angles between the variable and constant domains varied widely, and the CDR loops revealed a larger RMSD than the whole domains. The failure of AF-M to predict the epitopic sites (i.e., the detailed footprints of the Fab/MHC-I interactions) suggests that critical parameters such as binding energy or environmental conditions have been overlooked (84). Additional training with parameters such as elbow angles (85) or secondary structural features of epitopes (86) may help to improve the AI-based programs. Recent efforts at modeling TCR/pMHC complexes suggest that this may be a more tractable problem, in part because of the focused pMHC footprint (87).

We must point out that predicting binding, epitopic, or docking sites in a complex structure is distinct from predicting the structure of a single protein domain. The success of AlphaFold structural prediction depends on the advanced AI algorithm and high-performance computing. It also depends on the experimentally determined structures currently deposited in the PDB (∼200,000). AI training or deep learning from the PDB database in Alphafold2 creates ∼2 million parameters (84) in addition to the known structural parameters that greatly assist multiple sequence alignment for recognizing/constructing a similar structure from the PDB. AI training programs, such as Alphafold2, may evolve by learning additional parameters from the PDB as the number of experimental structures grows. Such training incorporates information from the experimental structures but also may incur biases from related but distinct complex structures. For example, AF-M positioned three of the Fab in the Fab/MHC-I complexes (34-2-12, 28-14-8, and S19.8) to sit atop the α1α2 helices of the MHC-I, atop the peptide binding groove (Fig. 9F, 9I, 9L). There are many (more than 1000) TCR/MHC structures in the PDB (88, 89) in addition to several structures of complexes of TCR mimic/MHC [PDB 1W72 (90), 3CVH (91), 4WUU (92), and 7TR4 (93)]. By and large, these structures indicate that the TCR or Fab binds over the α1α2 helices, whereas none of these complexes are similar to any of our mAb/MHC-I. Thus, AF-M predictions incurred a bias because of the available structures in the PDB. Once the Fab/MHC-I structures reported here are available in the PDB, they may be expected to contribute to improvement in the prediction of Fab/protein Ag complexes.

In summary, X-ray structures of mAb/protein Ags, as demonstrated here for a limited set of anti-MHC-I/MHC-I complexes, continue to provide detailed information describing the docking of Ab to their Ags and explain the reactivity and specificity profiles of such Ab. As more mAb/Ag structures are experimentally determined, this should contribute to the elucidation of the parameters that determine mAb structure and Ag recognition as we approach the goal of predicting antigenic specificity from Ab sequence.

The authors declare no conflicts of interest.

We thank John Altman, Christopher Boughter, Ted Hansen, Eduardo Padlan, T.V. Rajan, and Sebastian Springer for comments.