Structural and physical features that distinguish tumor-controlling from inactive cancer neoepitopes

Significance

Cytotoxic T cells recognize peptides presented by MHC-I (class I major histocompatibility complex) proteins, which leads to killing of the presenting cell. While recognition of pathogen-derived peptides is widely appreciated, T cells also recognize tumor-specific peptides incorporating amino acid mutations. Recognition of these “neoepitopes” underlies anticancer immunity and is of interest for immunotherapies such as cancer vaccines. However, neoepitopes that control tumor growth are rare and their features are poorly understood. As they are derived from self-proteins, such neoepitopes must overcome tolerance mechanisms that prevent autoimmunity. Here, we identified structural and physical features associated with tumor-controlling neoepitopes, including features that distinguish active from inactive neoepitopes as well as their self-counterparts. Our findings have implications for neoepitope prediction and cancer immunotherapy.

Abstract

Neoepitopes arising from amino acid substitutions due to single nucleotide polymorphisms are targets of T cell immune responses to cancer and are of significant interest in the development of cancer vaccines. However, understanding the characteristics of rare protective neoepitopes that truly control tumor growth has been a challenge, due to their scarcity as well as the challenge of verifying true, neoepitope-dependent tumor control in humans. Taking advantage of recent work in mouse models that circumvented these challenges, here, we compared the structural and physical properties of neoepitopes that range from fully protective to immunologically inactive. As neoepitopes are derived from self-peptides that can induce immune tolerance, we studied not only how the various neoepitopes differ from each other but also from their wild-type counterparts. We identified multiple features associated with protection, including features that describe how neoepitopes differ from self as well as features associated with recognition by diverse T cell receptor repertoires. We demonstrate both the promise and limitations of neoepitope structural analysis and predictive modeling and illustrate important aspects that can be incorporated into neoepitope prediction pipelines.

A central function of the adaptive immune system is T cell surveillance of peptides bound to class I major histocompatibility complex (MHC) proteins presented on the surface of nucleated cells. Peptide/MHC-I complexes are generated when intracellular proteins are degraded by the proteasome, yielding fragments that can be loaded onto MHC-I proteins and subsequently presented on the cell surface. When presented peptides are derived from foreign sources, such as viral proteins, cytotoxic CD8+ T cells can recognize and eliminate infected cells. In tumor cells, mutated proteins can also yield peptides that can be perceived as foreign. Such neoepitopes contribute to naturally occurring and pharmacologically induced anticancer immune responses and have been explored for use as peptide-based cancer vaccines. Although there has been some clinical promise (1, 2), most neoepitopes are poorly or nonimmunogenic, and bona fide, protective neoepitopes are exceedingly rare: only a small handful of neoepitopes that unambiguously protect from tumor growth and extend survival have been identified, all in mouse models (3–13). Moreover, neoepitopes which elicit immune responses in vitro or in vivo are often not protective against tumor growth, and vice versa (4, 7–9). These challenges have hindered the development of neoepitope-based therapies and prompted considerable effort to identify the determinants of neoepitope-driven antitumor immunity.

With viral epitopes, high affinity between the peptide and its presenting MHC-I protein is strongly associated with CD8+ T cell–mediated antiviral immunity (14–17). Whether peptide-MHC-I affinity is similarly dominant in neoepitope-driven antitumor immunity is less certain. While neoepitopes selected based on affinity for MHC-I have been associated with CD8+ T cell immunogenicity and sometimes tumor rejection (3, 5, 6, 18–20), low-affinity epitopes—including those categorized as “nonbinders”—can also mediate tumor rejection (4, 7–9). At the same time, neoepitopes selected for high MHC-I affinity have not always been active, or in other cases have led to CD4+ rather than CD8+ T cell responses (2, 21–23).

Factors in addition to MHC-I binding affinity clearly influence neoepitope activity. Unlike viral epitopes, which are not derived from proteins encoded by the host genome, neoepitopes are altered self-peptides and must overcome immune tolerance against their corresponding wild-type (WT) self-peptides. Because of this, we and others have suggested that differences between the neoepitope and its corresponding WT peptide, whether in peptide binding affinity, conformation in the groove, or other physical properties such as exposed surface area or molecular flexibility, will be important in determining activity (4, 7–9, 22, 24, 25). Such differences, acting alone or in concert, could promote productive TCR binding and allow a neoepitope to overcome T cell tolerance toward the WT peptide.

Recently, we performed a comprehensive assessment of the activity of mass spectrometry-defined neoepitopes in the Meth A murine cancer model (8). We identified and validated the expression and presentation of eight MHC-I presented neoepitopes, each of which differed from its corresponding WT peptide by a single amino acid and was predicted to bind its restricting MHC-I with moderate to strong affinity. When administered as a vaccine, two of these neoepitopes controlled tumor growth in vivo and extended survival. One remarkable neoepitope, Gtf2b^MUT, was fully protective, preventing tumor growth and prolonging the survival of every animal tested. Another neoepitope, Pdpr^MUT, showed more moderate albeit still significant protection. The activity of both the fully and moderately protective neoepitopes was confirmed to be dependent on CD8+ T cells. As the two neoepitopes are from the same tumor in genetically identical mice with distinct T cell repertoires, their existence provides an unprecedented opportunity to study the correlates between neoepitope features and tumor control.

Here, we used X-ray crystallography, binding measurements, computational modeling, and simulations to study the structural and physical properties of the Gtf2b^MUT and Pdpr^MUT neoepitopes. For comparison, we also selected an inactive neoepitope from the same study. To evaluate differences from self, we compared the properties of each neoepitope to its WT counterpart. We identified features that distinguish the fully and moderately protective tumor-controlling neoepitopes from each other as well as the inactive neoepitope. We also identified differences between the active neoepitopes and their WT counterparts. Altogether, the data permitted a structural and physical rationalization of the activity of each neoepitope. Structural modeling and other in silico tools were able to reproduce key aspects of this rationalization, suggesting routes for improved prediction of tumor-controlling cancer neoepitopes.

Results

The Gtf2b^MUT, Pdpr^MUT, and Prpf19-2^MUT Neoepitopes Bind MHC-I More Tightly than Their WT Self-Counterparts.

In recent work, we characterized neoepitopes from the BALB/c mouse fibrosarcoma Meth A and tested their immunogenicity and capacity for control of tumor growth (8). A neoepitope from the Gtf2b gene (sequence TGAA[S→R]FDEF, mutation indicated by bold type, termed Gtf2b^MUT) was remarkably protective, preventing tumor growth and prolonging survival in all vaccinated mice. A neoepitope from the Pdpr gene (sequence IGPRA[V→L]DVL, termed Pdpr^MUT) was moderately protective, preventing tumor growth and prolonging survival in one-half of the vaccinated mice. In contrast to these, a neoepitope from the Prpf19 gene (sequence KY[R→L]QVASHV, termed Prpf19-2^MUT) was immunologically inactive and failed to impact tumor growth in any mice. The presentation of each neoepitope was validated via mass spectrometry. The three neoepitopes and their properties are listed in Table 1.

Table 1.

Characteristics of neoepitopes and their WT counterparts

Peptide	Sequence*	MHC-I	Tm (WT/Mut)†	Predicted affinity (WT/Mut)‡	Tumor protection§	rmsd¶	SASA/hSASA#	ΔhSASA||
Gtf2b**	TGAA[S→R]FDEF	H-2Dd	45/61 °C	3,265/499	Full (10/10)	1.1 Å	367/171 Å2	2 Å2
Pdpr	IGPRA[V→L]DVL	H-2Dd	58/59 °C	237/157	Partial (5/10)	1.1 Å	297/181 Å2	27 Å2
Prpf19-2	KY[R→L]QVASHV	H-2Kb	65/66 °C	22/9	None (0/10)	0.9 Å	286/93 Å2	−6 Å2

^*Peptide sequence, with the mutation in the neoepitope in brackets and indicated by the arrow and bold type.

^†Thermal stability (T_m) as measured by differential scanning fluorimetry; values are the averages of three independent experiments for each peptide/MHC-I complex as shown in Fig. 1 and SI Appendix, Fig. S1.

^‡Predicted affinity in nM for peptide binding via NetMHC 4.0.

^§Values in parenthesis give number of animals protected from tumor growth and death (3).

^¶Measured rmsd between the neoepitope and WT counterpart when all common atoms of the peptides are superimposed.

^#Total and hydrophobic solvent-exposed solvent-accessible surface area for the neoepitope.

^||Change in exposed hydrophobic solvent-accessible surface area resulting from the neoepitope mutation.

^**rmsd and surface area calculations for Gtf2b^MUT were performed with the repositioned p6F side chain as described in the text.

To experimentally assess peptide binding to the restricting MHC-I proteins, we generated recombinant peptide/MHC-I complexes for all three neoepitopes and their WT counterparts and measured the thermostability of each complex with differential scanning fluorimetry (DSF). DSF yields the melting temperature or T_m of the complex, a well-accepted proxy for peptide-MHC binding affinity (26–29). The DSF data indicated that each neoepitope binds tighter than its WT counterpart, although the only large improvement was seen for the fully protective Gtf2b^MUT neoepitope, where the mutation converted a moderate-affinity peptide into a high-affinity peptide (Fig. 1 and SI Appendix, Fig. S1; summarized in Table 1). The improvement in MHC-I binding affinity was much smaller for the moderately protective Pdpr^MUT neoepitope, with the Pdpr^WT peptide already exhibiting tight binding. A similarly small improvement was seen with the inactive Prpf19-2^MUT neoepitope, whose WT counterpart bound with the highest affinity of all three WT peptides.

Fig. 1.

Thermal stabilities of the fully protective Gtf2b^MUT, the moderately protective Pdpr^MUT, and the inactive Prpf19-2^MUT neoepitope/MHC-I complexes, along with the stabilities of the complexes with their counterpart WT peptides. As indicated by the T_m values, the mutation in Gtf2b^MUT substantially improves peptide binding. The Pdpr^MUT and Prpf19-2^MUT neoepitopes show much smaller improvements compared to WT. Notably, the highest affinity is seen with the inactive Prpf19-2^MUT neoepitope. Additionally, tumor protection correlates inversely with the affinity of the WT peptide but there is no correlation with the mutant. Each measurement was replicated three times as indicated by the data points; horizontal lines give the averages. Representative thermal scans are shown in SI Appendix, Fig. S1A. All pairwise comparisons between the averages are statistically significant as shown in SI Appendix, Fig. S1B.

Predictions of peptide binding using NetMHC (30) were fully consistent with the experimental results, with NetMHC reproducing not only the large (for Gtf2b^MUT) and small (for Pdpr^MUT and Prpf19-2^MUT) enhancements in binding affinity but also the values and rankings of the three mutant and WT peptides (SI Appendix, Table S1). We note that NetMHC classified Gtf2b^WT as a strong binder, but the ranking was at the threshold separating strong from weak binding, consistent with our experimental determination of the peptide as a more moderate binder. Altogether, these data reinforce that MHC-I affinity is not an exclusive determinant of neoepitope activity. At the same time, large enhancements in affinity are positively associated with activity. As discussed below, we found it particularly notable that the extent of neoepitope protection is inversely correlated with the MHC-I affinity of the WT but not mutant peptides.

Structural Features and Differences from WT for the Fully Protective Gtf2b^MUT Neoepitope.

To examine the structural features of these neoepitopes, including how they differ from their WT counterparts, we crystallized and solved the X-ray structures of each neoepitope/MHC-I complex, as well as each WT peptide/MHC-I complex (Table 1). The structures of the six complexes were of high resolution, with the peptides clearly discernable in minimally biased composite/iterative-build OMIT maps calculated with simulated annealing (SI Appendix, Fig. S2 A and B) (31). We note that the diffraction data for the best crystal of the Pdpr^WT complex suffered from moderate anisotropy, reflected in a moderately high R_merge statistic and lower completeness (SI Appendix, Table S1). However, the features of the peptide and the binding groove were clear in the maps for the two molecules in the asymmetric unit (SI Appendix, Fig. S2B). The existence of two fully resolved and nearly identical copies of the Pdpr^WT complex in the asymmetric unit further improved our confidence in the refined structure.

For the fully protective Gtf2b^MUT neoepitope, the mutation of serine to arginine is at position 5 (p5) of the peptide. As p5 in nonameric peptides often serves as a secondary anchor (32), we previously hypothesized that the mutation would introduce significant conformational changes in the peptide, resulting in a large structural difference between the neoepitope and its WT counterpart, helping to explain the striking protective activity of Gtf2b^MUT (8). Comparing the crystal structures of the neoepitope and WT peptides does reveal structural differences; however, these are much smaller than we originally hypothesized. In both structures, the p5 amino acid points down into the base of the H-2D^d peptide binding groove (Fig. 2A). The larger p5 arginine extends deeper into the groove, forming electrostatic interactions the WT serine is unable to form and explaining the stronger affinity of the neoepitope (Fig. 2B). Rather than induce a new conformation though, these new interactions simply shift the peptide backbone by approximately 1 Å from p4 to p5.

Fig. 2.

Structural comparison between the Gtf2b^MUT neoepitope and its WT counterpart bound to H-2D^d. (A) Overlay of the two peptide/H-2D^d complexes with all common peptide atoms superimposed. The stronger anchoring of the neoepitope resulting from the Ser-to-Arg mutation at p5 leads to a shift in the backbone at p4 and p5, represented as an approximate 1 Å shift in the backbone of p4G. The alternate position of the p6F side chain for the neoepitope is shown in magenta, with the rotation around χ1 highlighted (for clarity, the rotation around χ2 is not highlighted). (B) As opposed to the WT serine, the p5R in the neoepitope forms electrostatic interactions with amino acids of the α1 helix at the base of the groove, including a salt-bridge with Asp77 and a cation-π interaction with Phe74. (C) In the structure of the neoepitope, the p6F side chain sits in a crevasse formed by symmetry-related molecules (Top). This does not occur in the structure with the WT peptide, with the p6F side chain instead lying adjacent to a neighboring molecule (Bottom). (D) Molecular dynamics simulations of the neoepitope and WT peptide/MHC-I complexes show the p6F side chain of the neoepitope predominantly populates the conformation seen in the WT structure. The p6F χ1 torsion angles seen in the structures are indicated (−64° for WT, −178° for the neoepitope). Accordingly, p6F of the neoepitope was considered to adopt the WT conformation as shown in panel A via rotation around χ1 and χ2 (SI Appendix, Fig S3). (E) Solvent-accessible surface areas of the neoepitope (with p6F adjusted) and WT peptide in the H-2D^d binding groove, colored by hydrophobicity (from blue for the most hydrophilic to brown for the most hydrophobic). The surface of p6F is indicated.

In addition to the small changes in backbone at p4 to p5, the side chain of the p6 phenylalanine also differs between the two structures, pointing up from the center of the groove in the Gtf2b^MUT but not Gtf2b^WT structure. While the small backbone changes could be rationalized from the arginine and its role as a secondary anchor, there was no apparent reason for the change in the position of the p6 phenylalanine side chain. This change was initially of particular interest to us, as centrally positioned aromatic residues are associated with strong TCR recognition (33), and our previous work has demonstrated that mutation-induced alterations in the position of a central aromatic side chain can enable the recognition of an immunologically active neoepitope by a TCR (25). Considering this, the observed rotation for Phe6 could mistakenly be considered a distinguishing conformational difference between the neoepitope and the WT peptide. However, a close analysis revealed additional complexity resulting from crystallographic contacts with symmetry-related molecules. Although such crystal contacts are often seen in protein crystal structures, they are usually tenuous (34, 35). In the Gtf2b^MUT structure, however, the crystal contacts with the p6 phenylalanine side chain are substantial, with the side chain occupying a crevasse formed by multiple residues of neighboring molecules (Fig. 2 C, Top). This was not the case in the Gtf2b^WT structure, which crystallized in a different form with only a single neighboring side chain in close proximity (Fig. 2 C, Bottom).

To assess whether crystal contacts were a determining factor in the differential position of the p6 side chain, we performed 1 μs molecular dynamics simulations of the Gtf2b^MUT and Gtf2b^WT peptide/MHC-I complexes, including water and excluding molecules of crystallographic symmetry as is traditional. We used our previously established protocols for comprehensive studies of mouse and human peptide/MHC-I dynamics (36, 37). We then compared the conformations present in the structures with the conformations sampled during the simulations. In the Gtf2b^MUT neoepitope simulation, the χ1 torsion angle of the phenylalanine quickly rotated to the position seen in the Gtf2b^WT structure and remained there for 85% of the simulation time. In contrast, the χ1 torsion angle seen in the neoepitope structure was far less populated (Fig. 2D). The distributions of the χ2 torsion angles were indistinguishable in the two simulations, with both peaking at the chemically identical ±90° rotamer typically seen for phenylalanine (38) (SI Appendix, Fig. S3). These results indicate that the crystallographic structures artifactually overestimate the conformational differences between Gtf2b^MUT and Gtf2b^WT. As indicated by the molecular dynamics simulation data, we placed the p6 phenylalanine side chain of the neoepitope in the same conformation as the WT (Fig. 2A, alternate conformation in magenta). With this change, the rmsd (root mean square deviation; all common peptide atoms superimposed) of the neoepitope from WT is only 1.1 Å.

Given that solvent-exposed surface area, and hydrophobic surface area in particular, has been linked to the immunogenicity of MHC-I presented epitopes (33, 39–41), we computed overall and hydrophobic solvent-exposed surface area for Gtf2b^MUT bound to H-2D^d, as well as the difference between the neoepitope and its WT counterpart. Accounting for the repositioning of the Phe6 side chain, Gtf2b^MUT exposes 367 Ų of surface area, 171 Ų of which is hydrophobic. The difference between the neoepitope and WT peptide leads to a negligible increase in hydrophobic surface area of only 2 Ų (Fig. 2E and Table 1). Thus, as demonstrated by Gtf2b^MUT, large peptide conformational changes relative to self and large increases in exposed hydrophobic surface area are not needed for a strongly protective neoepitope.

Structural Features and Differences from Self for the Moderately Protective Pdpr^MUT Neoepitope.

For the moderately protective Pdpr^MUT neoepitope, the mutation of valine to leucine is at p6 of the peptide (Table 1). Position 6 of nonameric peptides also sometimes serves as a secondary anchor (32), but the crystallographic structures revealed that for both Pdpr^MUT and Pdpr^WT, the p6 side chain points up and away from the H-2D^d binding groove (Fig. 3A). The larger leucine of the neoepitope is more exposed, oriented toward the center of the peptide/MHC-I platform where the hypervariable loops of TCRs typically focus. There is a curious ~2 Å extension and “flip” of the p1 isoleucine at the N terminus of the Pdpr^MUT peptide, contributing to an overall 1.1 Å rmsd when all common atoms of the Pdpr^MUT and Pdpr^WT peptides are superimposed. This repositioning of the p1 isoleucine, driven in part by a ~130° rotation of the p1 ψ bond, alters the typical hydrogen bond arrangement between the peptide N terminus and the H-2D^d heavy chain, with the nitrogen forming a salt-bridge with Glu163 of the H-2D^d α2 helix rather than hydrogen bonding with Tyr171 as usual (SI Appendix, Fig. S4A). There is no apparent reason for this conformational difference between mutant and WT; unlike the Gtf2b structures, there are no symmetry-related crystal contacts in these regions of the Pdpr peptides. As with Gtf2b, we performed 1 μs molecular dynamics simulations on both Pdpr complexes and observed during the Pdpr^MUT simulation that, although the p1 ψ bond remained rotated, the peptide N terminus quickly receded back into the groove due to rotations around other bonds but otherwise did not appear more dynamic on this timescale than its WT counterpart (SI Appendix, Fig. S4 B–D; see also Fig. 5 below). Conformational sampling of Glu163 in the α2 helix was identical in both simulations (SI Appendix, Fig. S4E).

Fig. 3.

Structural comparisons between the Pdpr^MUT neoepitope and its WT counterpart bound to H-2D^d. (A) Overlay of the neoepitope and WT complexes with all common peptide atoms superimposed. The difference at the N-terminal ends of the peptides, which reaches 2.1 Å for the p1 isoleucine Cα atoms, is indicated. (B) Solvent-accessible surface areas of the neoepitope and WT peptide in the H-2D^d binding groove, colored by hydrophobicity (from blue for the most hydrophilic to brown for the most hydrophobic). The additional surface from the leucine in the neoepitope compared to the valine in the WT peptide is indicated.

Fig. 5.

Impact of mutations on peptide fluctuations in the MHC-I binding groove. (A) Differences in rmsf values (WT—neoepitope) for Cα atoms of the three peptide pairs. Differences are in Å and were computed from 1 μs molecular dynamics simulations. The shaded area set off by the dashed lines shows the significance threshold of ±0.4 Å, determined from repeat simulations on identical peptide/MHC-I complexes as described in the text. The center of the fully protective Gtf2b^MUT neoepitope (black squares) is substantially more rigid than Gtf2b^WT, consistent with the stronger H-2D^d anchoring in the center of the neoepitope. For the moderately protective Pdpr^MUT neoepitope (red circles), p6 at the site of the mutation is slightly more rigid. There were no differences in flexibility for the inactive Prpf19-2 neoepitope (blue triangles). (B) Visualization of the greater rigidity of Gtf2b^MUT compared to Gtf2b^WT. For both the neoepitope and WT peptide, each atom is rendered as a sphere with radius scaled according to its RMSF. The atoms of the p5R (neoepitope) and p5S (WT peptide) are circled in magenta for reference. The larger radii for the atoms of the WT peptide indicate the peptide’s greater flexibility.

As discussed below, conformational differences of magnitude similar to that seen between Pdpr^MUT and Pdpr^WT can be seen even when different structures of the exact same peptide/MHC-I complex are compared. We believe this structural difference at the N terminus is therefore best attributed to influences that stem from crystallization. This includes the possibility that the Pdpr^MUT crystal captured a lowly populated state not sampled over the fast timescale of our simulations, stemming from a slow-moving N terminus that is attributable to H-2D^d’s reliance on a primary anchor pattern that includes a p2 glycine followed by a p3 proline (42).

From the structures, we computed overall and hydrophobic solvent-exposed surface areas for Pdpr^MUT bound to H-2D^d, as well as the difference between the neoepitope and its WT counterpart. The Pdpr^MUT neoepitope exposes 297 Ų of surface area, 181 Ų of which is hydrophobic. The neoepitope exposes 27 Ų more hydrophobic surface area than the WT peptide, attributable to the larger, more exposed leucine at p6 (Fig. 3B). Thus, similar to what was seen with Gtf2b^MUT, large conformational changes relative to WT are not required for the activity of the moderately protective Pdpr^MUT neoepitope, although there is an increase in exposed hydrophobic surface area in the center of the peptide.

Structural Features and Differences from Self for the Inactive Prpf19-2^MUT Neoepitope.

For the nonprotective Prpf19-2 epitope, we previously hypothesized that the p3 arginine-to-leucine mutation would have a negligible impact on peptide conformation (8). Comparing the Prpf19-2^MUT and Prpf19-2^WT structures shows this is indeed the case, with only very subtle, nonsystematic conformational differences seen for exposed side chains (Fig. 4). All common atoms of the two peptides superimpose with an rmsd of 0.9 Å. Total and hydrophobic solvent-exposed surface areas for the neoepitope are 285 Ų and 93 Ų, respectively, with a very small reduction of 6 Ų of exposed hydrophobic surface area resulting from the mutation, attributable to a slightly greater exposure of the long hydrophobic component of the original arginine side chain compared to the mutant leucine (Table 1).

Fig. 4.

Structural comparisons between the Prpf19-2^MUT neoepitope and its WT counterpart bound to H-2K^d. The figure shows an overlay of the Prpf19-2 neoepitope and WT complexes with all common peptide atoms superimposed. The structural similarity between the peptides is apparent.

The Fully Protective Gtf2b^MUT Neoepitope Is More Rigid in the MHC-I Binding Groove.

We next examined the flexibility of each neoepitope/WT pair in the MHC-I binding groove. Our goals were to assess how the mutations in the peptides altered movement around the conformations seen in the static crystal structures, as well as explore possible dynamic correlates with immunogenicity, as has been suggested by previous work (4, 43, 44). Building on the Gtf2b^MUT/Gtf2b^WT and Pdpr^MUT/Pdpr^WT simulations mentioned above, we performed molecular dynamics simulations for the remaining two Prpf19-2 complexes, with all six structures simulated for 1 μs in explicit solvent. To gauge changes in peptide flexibility, we computed root mean square fluctuation (rmsf) values for each α carbon (Cα) of each peptide and determined the differences in these values between the neoepitope and WT simulations. To determine a significance threshold for this analysis, we used the repeat set of data from our previously published library of peptide/MHC-I molecular dynamics simulations, which contained 38 repeats of 19 pairs of the same MHC-I protein presenting the same peptide (37). The average difference in peptide Cα rmsf values across these repeats was 0.4 Å. We therefore used ±0.4 Å as a threshold of significance for differences in Cα fluctuations.

For the fully protective Gtf2b^MUT neoepitope, the p5 serine to arginine mutation reduced fluctuations in the center of the peptide (p4 to p6), consistent with the stronger anchoring formed through new electrostatic interactions to the peptide center (Fig. 5A, black). The flexibility of the moderately protective Pdpr^MUT neoepitope was mostly unchanged compared to WT, with only a slight reduction at the site of the mutation at p6 (Fig. 5A, red). As noted above, there was no difference in the N-terminal region where the conformational differences are seen. The differences in flexibility for the nonprotective Prpf19-2^MUT peptide were within the ±0.4 Å threshold for significance (Fig. 5A, blue). Mimicking the changes in affinity, the fully protective neoepitope Gtf2b^MUT is thus distinctive from the moderately protective Pdpr^MUT and inactive Prpf19-2^MUT neoepitopes in how the mutation enhances peptide rigidity in the MHC-I binding groove. The greater rigidity of Gtf2b^MUT compared to Gtf2b^WT is highlighted in Fig. 5B.

Benchmarking Peptide/MHC-I Conformational Differences.

Although the conformational differences between the three neoepitopes and their WT counterparts bound to MHC-I are small (rmsd ranging from 0.9 Å to 1.1 Å), they are nonetheless larger for the two protective neoepitopes than the inactive neoepitope (Table 1). Although in other work, very small changes in peptide conformation have been shown to influence T cell recognition (25, 45–47), we were concerned about possibly overinterpreting the small differences observed here. We thus sought to benchmark the significance of structural changes seen in peptide/MHC-I structures. Accordingly, we compared instances where the crystal structure of the same mouse or human peptide/MHC-I complex has been determined multiple times (excluding structures with TCRs or other proteins bound and similar but not identical structures designed to probe the impact of heavy chain mutations, variant β₂-microglobulins, posttranslational modifications, etc.). Such structures are often termed “redundant,” yet they offer insight into the impact of different crystal forms and packing, different resolutions and data quality, intrinsic flexibility, etc., and are particularly relevant in establishing the limit to which similar protein structures can be compared (48).

Across 43 structures with 20 different peptides, the average rmsd (all-atom peptide superimposition) between identical peptides was 0.8 Å, with a SD of 0.4 Å. Limiting this analysis to only nonamers as studied here yielded 30 structures with 14 peptides and an average rmsd of 0.7 ± 0.4 Å (SI Appendix, Fig. S5). This result is similar to an earlier analysis that examined replicate hIL1β structures and found an average rmsd of 0.8 Å (49). We thus defined 1.1 Å (the average rmsd of 0.7 Å from the nonameric structural replicates plus one SD) as a benchmark of significance for conformational differences between two structures of nonameric peptides presented by MHC-I. Accordingly, the experimental conformational differences from WT for the Gtf2b^MUT, Pdpr^MUT, and Prpf19-2^MUT neoepitopes as measured by all common atom rmsd are insignificant. Distinguishing structural features such as the greater exposed hydrophobic surface area for Pdpr^MUT remain significant, however, as these emerge from the additional atoms of the mutant side chains.

In Silico Structural Models Vary in Accuracy and Are Improved with AI-Assisted Modeling.

We next evaluated the accuracy of our prior structural models, which as noted above we previously used to explore the activity of the Gtf2b^MUT and Prpf19-2^MUT neoepitopes (8). For completeness, we generated models of the Pdpr mutant/WT pair using the same Rosetta-based procedure. We also modeled the six peptide/MHC-I complexes using two recently described peptide/MHC-I modeling procedures, PANDORA and TFold. PANDORA is built upon the widely used protein modeling package MODELLER (50, 51), whereas TFold is built on the recent and impactful AlphaFold structure prediction tool (52, 53).

The performance of the three approaches to structural modeling is shown in Table 2, with detailed comparisons in SI Appendix, Fig. S6. For the Gtf2b peptides, as described above, Rosetta modeling misplaced the secondary anchor for Gtf2b^WT at p5, significantly overestimating the predicted conformational difference from self. PANDORA did not predict major differences between the Gtf2b peptides, but struggled with side chains and, for Gtf2b^WT, the peptide backbone in the C-terminal half of the peptide. The AI-assisted TFold method, on the other hand, performed best, accurately capturing both the backbone and sidechain positions of the Gtf2b peptides. Generally, these conclusions were mirrored with the other two pairs of peptides, with TFold outperforming both Rosetta and PANDORA. For Rosetta modeling, this is not surprising, as we and others have recently demonstrated the limitations of Rosetta-based peptide/MHC-I modeling that is not assisted by training on known structures (54, 55).

Table 2.

Structural modeling performance

Peptide	Rosetta modeling*	PANDORA modeling†	AI modeling‡
Gtf2bMUT,§	1.4	2.2	1.1¶
Gtf2bWT	3.0	2.1	0.9¶
PdprMUT	2.2	3.4	2.0
PdprWT	1.8	2.7	1.9
Prpf19-2MUT	1.2	1.6	0.9¶
Prpf19-2WT	1.3	1.5	0.9¶

^*All-atom rmsd in Å between structure and model generated with Rosetta-based modeling.

^†All-atom rmsd in Å between structure and model generated via PANDORA.

^‡All-atom rmsd in Å between structure and model generated via TFold.

^§Comparisons to the structure of Gtf2b^MUT used the repositioned p6F side chain as described in the text.

^¶At or below the 1.1 Å accuracy threshold as described in the text.

To better compare modeling performance, we used an all-atom peptide rmsd of 1.1 Å between model and experimental structure as a benchmark of accuracy as described above. According to this metric, the AI-assisted modeling approach TFold was by far the strongest performer, with 4 out of 6 models indistinguishable from their corresponding structures (the two outliers are the Pdpr neoepitope/WT pair, with errors in the center of the peptide). By comparison, none of the Rosetta models or PANDORA models met the 1.1 Å accuracy threshold (Table 2). Although AlphaFold has shown difficulty with various immune complexes (56), as seen with other instances where tuning and optimization have been applied (57, 58), the TFold implementation shows considerable promise for accurate peptide/MHC-I modeling.

Evaluating a Fully In Silico Approach to Distinguish the Fully Protective from the Inactive Neoepitope.

With structural models of protective and inactive neoepitope/WT pairs that are indistinguishable from their experimental structures (Gtf2b and Prpf19-2), we asked how well a completely in silico approach reproduced our salient conclusions for these pairs. As already noted, NetMHC estimates of peptide binding were consistent with experiment for all peptides examined (Table 1), and the AI-assisted TFold models accurately captured the structural details of both the Gtf2b and Prpf19-2 pairs (Table 2 and SI Appendix, Fig. S6). Also matching experiment, the modeled conformational differences from self for these two pairs were insignificant (all common atom rmsds of 0.3 Å for Gtf2b and 0.5 Å for Prpf19-2). The surface areas from the models did not exactly duplicate experiment, but the trends matched: from the models, exposed total and exposed hydrophobic surface area for the Gtf2b^MUT neoepitope were the largest at 315 Ų and 142 Ų, respectively, compared to 367 Ų and 171 Ų experimental. These were smaller for Prpf19-2^MUT at 301 Ų and 116 Ų, compared to 286 Ų and 93 Ų experimental. Moreover, the change in exposed hydrophobic surface area from the Prpf19-2 model structures was negligible, with 4 Ų modeled and −6 Ų experimental.

To complete the in silico evaluation, we performed molecular dynamics simulations on the Gtf2b and Pdpr19-2 models using the same protocol we used on the experimental structures. The results reproduced the trends obtained from the experimental structures, showing reduced flexibility and increased geometric order in the center of the Gtf2b^MUT neoepitope and no significant changes in flexibility for the Prpf19-2^MUT neoepitope (SI Appendix, Fig. S7). The fully in silico evaluation of the Gtf2b and Prpf19-2 neoepitope/WT pairs thus matches the experimental conclusions in terms of peptide binding affinity, peptide conformation and structural features, and peptide flexibility within the MHC-I binding groove.

Discussion

Here, we have completed an experimental structural analysis of cancer neoepitopes that unambiguously mediate tumor control in vivo and prolong survival, comprising an analysis of two rare neoepitopes with different levels of protective capacity—including a remarkably potent peptide that prolonged survival in every animal tested—as well as an inactive neoepitope. Identified from the same tumor model in genetically identical mice with distinct T cell repertoires, our work allowed us to carefully examine features associated with antigen presentation and recognition. Our comparison of the neoepitopes with each other as well as with their WT counterparts revealed distinct differences between the tumor-controlling neoepitopes and the inactive neoepitope, as well as differences between the two tumor-controlling neoepitopes. Altogether, these differences allow us to rationalize the activity (or lack thereof) of the three neoepitopes studied.

Given its important role in viral immunity, MHC-I binding affinity is commonly discussed in neoepitope immunogenicity. Our data reinforce that high neoepitope affinity by itself is a poor predictor of activity, as the highest affinity belongs to the inactive Prpf19-2^MUT neoepitope. As we have previously suggested though (4), enhancements in affinity relative to WT can be important, as demonstrated by the fully protective Gtf2b^MUT neoepitope, where the mutation converts a moderately binding peptide into a strong binder. Biologically, affinity improvements will increase the amount of peptide presented on the cell surface (as well as duration, as dissociation kinetics correlate well with affinity). Mimicking viral escape mutants but in reverse (59–61), changes from weak or moderate affinity to strong affinity are expected to have a larger immunological effect, increasing the likelihood for a neoepitope to be “seen” compared to its weaker binding WT counterpart. In contrast, as population bound increases hyperbolically as a function of affinity, when a WT peptide already binds tightly, an increase in affinity is unlikely to have a significant influence. Indeed, we find it notable that neoepitope activity here is inversely correlated with WT affinity, consistent with a conclusion that the stronger the WT peptide binds, the more challenging it will be for a mutant peptide to distinguish itself from self and overcome tolerance. Indeed, high-affinity self-peptides are more likely to promote T cell negative selection or induce other tolerance mechanisms (62), increasing the need for large (and difficult to attain) structural differences from self for generating active neoepitopes.

A large improvement in binding compared to WT is not the only distinguishing feature of the fully protective Gtf2b^MUT neoepitope. As expected from its stronger anchoring to MHC-I, Gtf2b^MUT is more rigid than Gtf2b^WT in the MHC-I binding groove; this is not seen for the moderately protective Pdpr^MUT or inactive Prpf19-2^MUT neoepitopes. As found in other studies of neoepitopes as well as shared tumor antigens (4, 43, 63), this increased rigidity could favorably impact T cell recognition by reducing the entropic cost for TCR binding. This is a general “TCR agnostic” benefit that would be expected to impact most if not all of the neoepitope-specific T cells that comprise individual repertoires.

Notably, conformational differences from self are lacking for all three of the neoepitopes studied, a striking observation given our prior modeling efforts (8) and expectations about the importance of structural differences from self (22, 24, 25). Structural features are nonetheless important. For example, the structures show that the fully protective Gtf2b^MUT neoepitope is the most solvent exposed among the three. A large amount of the exposed Gtf2b^MUT surface is hydrophobic, including a central aromatic side chain. Hydrophobic and aromatic side chains are strongly associated with TCR recognition and peptide immunogenicity (33, 39–41). This is easily rationalized: hydrophobic surface drives molecular recognition via the hydrophobic effect and, unlike charged surface, does not require precise alignment of opposing charges to offset energetically expensive desolvation penalties, translating into fewer constraints on the composition of compatible TCRs (39). Aromatic side chains are multifunctional, capable not only of meshing alongside other hydrophobic surface but also acting as hydrogen bond acceptors, as seen in multiple TCR-peptide/MHC structures (33).

These points all come together to explain the strongly protective nature of the Gtf2b^MUT neoepitope. The mutation converts a moderately binding peptide into a strong binder, which emerges from better anchoring in the peptide center. The result is the enhanced display of a more rigid epitope that, via its exposed central phenylalanine side chain, is primed for recognition by TCRs in the repertoires present in multiple animals.

These same points also rationalize the more moderate protective capacity of the Pdpr^MUT neoepitope, which has fewer points of dissimilarity from its WT counterpart. It does not show a considerable enhancement in binding affinity relative to WT, nor does it show significant conformational differences relative to self, nor is it substantially more rigid in the binding groove. However, the valine-to-leucine mutation in Pdpr^MUT increases exposed hydrophobic surface area in the center of the peptide, differentiating it from its WT counterpart and providing a distinguishing hydrophobic feature for TCRs of repertoires tolerized to the WT peptide to identify. Unlike what is seen with Gtf2b^MUT, where the large enhancement in binding affinity and greater rigidity is likely to impact TCR recognition in general, this greater hydrophobic exposure in Pdpr^MUT compared to Pdpr^WT is more likely to be TCR-specific. These factors may help explain why Pdpr^MUT is moderately rather than fully protective. This leads us to hypothesize that a mutation that changed the WT valine into an aromatic phenylalanine, tyrosine, or tryptophan, further increasing exposed hydrophobic surface as well as providing hydrogen bonding or cation-π opportunities (33), would further distinguish Pdpr^MUT from its WT counterpart and create an even more protective neoepitope.

Turning to the nonprotective Prpf19-2^MUT neoepitope, it is virtually indistinguishable from its high affinity and likely strongly negatively selecting WT counterpart. The mutation does not substantially impact peptide binding affinity or rigidity. Conformational changes from self are nonexistent, and the neoepitope shows the lowest amount of exposed hydrophobic surface area. The neoepitope is thus not distinctive from its WT counterpart nor does it show features associated with recognition by a diverse set of TCRs, explaining its inactivity (although other explanations may be possible for the epitope’s lack of activity, we believe this to be the most parsimonious interpretation, relying on established features of TCR recognition and peptide presentation).

Lastly, our prior models exaggerated the role of conformational differences from WT in the activity of the fully protective Gtf2b^MUT, but the more recent AI-based approach corrected this error, and as judged by comparisons of structural replicates, led to more accurate models. These models and other in silico assessments reproduced the features that distinguish the most protective neoepitope from the inactive neoepitope, including binding affinity, structures and structural features, and changes in flexibility. This agreement suggests possible routes for improved computational prediction of tumor-controlling neoepitopes, incorporating not only predictions of binding affinity but also structural modeling and assessment of exposed features such as hydrophobic and aromatic side chains, with an emphasis on the differences between mutants and their WT counterparts.

Methods

Peptides and Proteins.

The Gtf2b, Pdpr, and Prpf19-2 mutant and WT peptides were commercially synthesized by GenScript at >90% purity, dissolved in DMSO to a concentration of 30 mM, and stored at −80 °C until use. Genes encoding the H-2D^d and H-2K^d heavy chains and murine β₂-microglobulin proteins were commercially synthesized by Genewiz. Soluble constructs of the heavy chains and β₂-microglobulin were individually expressed in BL21 E. coli cells as inclusion bodies, refolded, and purified by anion exchange followed by size exclusion chromatography as previously described (64).

Protein Crystallization.

Purified peptide/MHC-I complexes were concentrated to 5 to 6 mg/mL prior to crystallization. Crystallization was performed via hanging drop vapor diffusion using a Mosquito robot. Crystals of Gtf2b^MUT/H-2D^d were grown at room temperature in 20% polyethylene glycol 3350, 200 nM sodium tartrate dibasic dihydrate (pH 7.3); Gtf2b^WT/H-2D^d at room temperature in 15% w/v polyethylene glycol 1500; Pdpr^MUT/H-2D^d at room temperature in 0.2 M sodium thiocyanate and 20% polyethylene glycol 4000; Pdpr^WT/H-2D^d at room temperature in 0.2 M potassium thiocyanate and 20% polyethylene glycol 3350; Prpf19-2^MUT/H-2K^d at 4 °C in 20% w/v polyethylene glycol 3000, 100 mM HEPES (pH 7.5), 200 mM sodium acetate; and Prpf19-2^WT/H-2K^d at 4 °C in 10% w/v polyethylene glycol 8000, 200 mM magnesium acetate. Crystals were harvested in cryoprotectant consisting of mother liquor mixed with 5% glycerol prior to flash freezing in liquid nitrogen.

Data Collection, Structure Determination, and Analysis.

X-ray diffraction data were collected at the Advanced Photon Source at Argonne National Laboratory on the 24-ID-C (Gtf2b^MUT/H-2D^d, Gtf2b^WT/H-2D^d, Pdpr^MUT/H-2D^d, Pdpr^WT/H-2D^d, and Prpf19-2^MUT/H-2K^d) and 24-ID-E (Prpf19-2^WT/H-2K^d) beamlines. HKL2000 (65) was used for indexing and scaling. Initial phases were determined using Phaser (66) with PDB IDs 5KD7 and 1VGK with peptides removed as models for molecular replacement for H-2D^d and H-2K^d, respectively (67, 68). Initial coordinates were built using Buccaneer (69). Structures were refined over multiple rounds of automated refinement with REFMAC5 (70) followed by manual refinement in Coot (71). Structures were deposited in the PDB with accession codes shown in SI Appendix, Table S1. PyMOL 2.5 and Discovery Studio 2022 were used for visualization and coordinate analyses. Simulated annealing composite OMIT maps were calculated using CNX (72) as implemented in Discovery Studio 2022. Solvent-accessible surface areas were calculated in Discovery Studio 2022 with 960 grid points per atom and a 1.4 Å probe radius. In hydrophobic surface area calculations, all carbon atoms were designated as hydrophobic. Rmsds between peptides were calculated by superimposing all common peptide atoms, ensuring the symmetrical Phe and Tyr side chains were properly aligned (e.g., ensuring atom CD1 aligns with atom CD1 by flipping χ2 180° when necessary). As the Pdpr^WT structure had two molecules per asymmetric unit, rmsds, distances, solvent-accessible surface areas, etc., for the Pdpr^MUT/Pdpr^WT pair and Pdpr^WT structure/models were determined by calculating values and differences for both pairs and averaging. This had a negligible impact on the results, as the two copies of the peptide in the Pdpr^WT structure superimpose with an all-atom rmsd of 0.7 Å.

Differential Scanning Fluorimetry and Peptide Binding Predictions.

Differential scanning fluorimetry experiments to measure thermal stabilities were performed using a NanoTemper Prometheus instrument operating in fluorescence mode as previously described (27). Approximately 10 μL of peptide/MHC-I complex at a concentration of 10 to 20 nM was used for each experiment. Temperatures ranged from 20 °C to 95 °C with an increment of 1 °C/min. T_m values were determined by fitting the temperature derivative of the fluorescence curve (truncated at 80 °C) to bi-Gaussian functions using OriginPro 2021b. Three experiments were performed for each complex. Unpaired, two-tailed Student’s t tests were performed using GraphPad QuickCalcs (https://www.graphpad.com/quickcalcs/ttest1/). NetMHC 4.0 was used with default options (30).

Molecular Dynamics Simulations.

Molecular dynamics simulations were performed as previously described (36, 37). Briefly, simulations were conducted using the Amber18 package with GPU acceleration (73). Initial coordinates for the Pdpr^WT structure were from the first molecule in the asymmetric unit. For the Pdpr^MUT structure, missing residues in the extended loops of the peptide binding groove were modeled from the coordinates of the WT structure. Missing residues for the α3 domain and β₂-microglobulin were modeled in from the coordinates of PDB deposition 1S7R for both Pdpr structures (74). For simulations on the TFold models, the missing α3 and β₂m residues were modeled in from the coordinates of their respective templates (5KD7 for Gtf2b and 4Z77 for Prpf19-2). Protonation states for ionizable side chains were determined via the APBS-PDB2PQR web service at pH 7.0 (75). Simulations were performed using the ff14SB forcefield (76). All structures were solvated in a cubic box of SPC/E water and charge neutralized with sodium. Following minimization, all systems were gradually heated in the NVT ensemble to 300K using a Langevin thermostat and with solute restraints. Restraints were gradually relaxed from 25 kcal/mol/Ų to 0 kcal/mol/Ų and subsequently equilibrated for 10 ns in the NPT ensemble. Following this, production simulations in the NVT ensemble were calculated with a 2 fs time step and an 8 Å cutoff for nonbonded interactions, employing the SHAKE algorithm to restrain bonds involving hydrogen (77). All simulations were calculated for a total time of 1 μs. Analysis of trajectories was performed with cpptraj (78). RMSFs were calculated via the atomicfluct command following global Cα superimposition, and torsion angles were calculated via the dihedral command. Order parameters were calculated using isotropic reorientational eigenmode dynamic analysis via the vector and matrix commands in cpptraj, with vectors defined from the Cα of each residue to the respective sidechain center of mass and the carbonyl carbon and its oxygen (79).

Identification of Structural Replicates.

To identify unique structures of MHC-I proteins presenting the same peptide, FASTA sequences were extracted from the PDB for all human and murine MHC-I structures as of July 2023. Sequence similarity across all peptide sequences was determined using the editDistance command in MATLAB R2022a, with unique structures with the same peptide indicated as those with a comparative distance of 0. The resulting list of structures with identical peptides was manually filtered to include only those where the same peptide was presented by the same MHC-I haplotype, as well as removing structures where one of the sets had a TCR or other protein bound, a peptide posttranslational modification, an unusual nonpeptidic ligand in the peptide binding groove, a variant β₂-microglobulin, or MHC-I heavy chain mutations. This yielded truly redundant structures. Rmsds between peptides were calculated by superimposing all common peptide atoms, ensuring the symmetrical Phe and Tyr side chains were properly aligned by flipping χ2 180° when necessary. When multiple molecules in the asymmetric were present, pairwise comparisons between all peptide chains were performed and all values were included in the analysis.

Structural Modeling of Peptide/MHC-I Complexes.

The Rosetta models of the Gtf2b and Prpf19-2 neoepitope and WT complexes were previously reported and used here (8). The Rosetta models of the Pdpr neoepitope and WT complexes were generated similarly. Briefly, modeling was performed using PyRosetta 4.0 in conjunction with the ref2015 score function and KIC loop modeling (80–82). Atomic coordinates from PDB entry 5KD7 were chosen as a starting template based on the similarity of peptide anchor residues, peptide length, and sequence (67). The template structure was brought to a local energy minimum using Rosetta FastRelax with harmonic restraints of 0.05 kcal/mol. Peptide side chains were then replaced with those of the target peptide sequence for both the neoepitope and WT peptide. Amino acid side chains were repacked to energetically favorable rotamers using Rosetta PackRotamersMover. Peptide side chain and backbone atoms were minimized using neighbor-sensitive dihedral angle sampling followed by simulated annealing KIC with a maximum segment length of 12 via LoopMover_Refine_KIC. For both mutant and WT peptide sequences, 200 decoys were generated and the model with the lowest total ref2015 energy score was chosen for analysis and comparison. AI-based TFold models were generated following the recommended procedure (52). The templates of 5KD7 (Gtf2b and Pdpr peptides) and 5TS1 (Prpf19-2 peptides) (67, 83) were automatically selected and verified to assure compatibility with the target sequence. Models generated by TFold were scored by averaging the predicted local-distance difference test over the peptide and subtracting from 100; the lowest-scoring models were output and used. PANDORA models were similarly generated with automatic template selection according to the reported documentation (50). The selected templates of 5KD7 (Gtf2b and Pdpr) and 4Z77 (Prpf19-2) were inspected to confirm target compatibility. A total of 20 models were produced and scored using the MODELLER molpdf score function. The best scoring model for each target was used. For all three modeling procedures, rmsds between models and structures were calculated by superimposing all common peptide atoms, ensuring that the symmetrical Phe and Tyr side chains were properly aligned by flipping χ2 180° when necessary.

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

Protein structural data are available from the RCSB Protein Data Bank (https://www.rcsb.org/) with accession codes 8D5F, 8D5E, 8FHL, 8FHU, 8D5J, and 8D5K. The DSF data shown in Figure 1 and Figure S1 have been deposited at the Zenodo repository at https://zenodo.org/ with record number 10056580 (also https://doi.org/10.5281/zenodo.10056580) (84).