Neoantigen quality predicts immunoediting in survivors of pancreatic cancer

To identify the edited neoantigens, we extended our previous neoantigen quality model^4,5 that quantifies the immunogenic features of a neoantigen to propose that two competing outcomes determine whether a neoantigen is high-quality—whether the immune system recognizes or tolerates a neoantigenic mutation (Fig. 3a). To estimate the likelihood the immune system recognizes a neoantigen, we measure the sequence similarity of the mutant neopeptide (p^MT) to known immunogenic antigens. This infers the ‘non-self’ recognition potential R of p^MT, a proxy for peptides within the recognition space of the T cell receptor (TCR) repertoire.

By contrast, we posit that the immune system can also fail to discriminate p^MT from its wild-type (WT) peptide (p^WT), and therefore tolerate it as ‘self’. The immune system must therefore exert greater self discrimination D (Fig. 3a) in tumours to overcome the principles of negative T cell selection, the adaptation that limits autoreactivity to host tissues. We approximate the D between p^WT and p^MT by two features—differential MHC presentation and differential T cell reactivity. Differential MHC presentation of p^WT and p^MT (\(K_\rmd^\textWT/K_\rmd^\textMT\)), previously introduced as the MHC amplitude A (refs. ^4,5), estimates the availability of T cells to recognize p^MT. If p^WT is not presented to T cells in the thymus or the periphery (as with a high \(K_\rmd^\textWT\), which implies poor p^WT–MHC binding), p^WT-specific T cells escape negative selection to expand the peripheral T cell precursor pool available to recognize a p^MT presented on MHC (low \(K_\rmd^\textMT\))²⁰. Here we extend this concept and introduce cross-reactivity distance C, a new model term that estimates the antigenic distance required for T cells to discriminate between p^MT and p^WT. Thus, self discrimination D = log(A) + log(C) is a proxy for peptides outside the toleration space of the TCR repertoire. In summary, we define neoantigen quality as Q = R × D (Fig. 3a), now with components that estimate whether a neoantigen can be recognized as non-self and discriminated from self.

To model C, we leveraged recent findings that conserved structural features underlie TCR–peptide recognition. Specifically, the binding domains of peptide-degenerate TCRs^21,22 and TCR-degenerate peptides²³ share common amino acid motifs, suggesting that T cell cross-reactivity between p^MT and p^WT could estimate the relative C of different neoantigenic substitutions (Fig. 3b). We selected an HLA-A*02:01-restricted strong epitope (NLVPMVATV (NLV)) from human cytomegalovirus²⁴ that was previously used to model TCR–peptide degeneracy^21,22 as a model p^WT, and three NLV-specific TCRs (Extended Data Fig. 4a–c). We then varied the NLV peptide by every amino acid at each position to model p^MT substitutions, and compared how TCRs cross-react between each p^MT and its p^WT across a 10,000-fold concentration range where p^WT changes maximally altered T cell activation (Fig. 3b). We observed that substitutions were either highly, moderately or poorly cross-reactive (Fig. 3c, d), and the cross-reactivity pattern depended on the substituted position and residue (Extended Data Fig. 5a). Interestingly, we found similar patterns of cross-reactivity between a model HLA-A*02:01-restricted weaker p^WT epitope in the melanoma self-antigen gp100^25,26 (Extended Data Figs. 4d and 5b), three p^WT-specific TCRs and single-amino-acid-substituted p^MTs, suggesting that conserved substitution patterns define C (Fig. 3e and Extended Data Fig. 5b). Thus, we quantified the cross-reactivity distance C between a p^WT and its corresponding p^MT as \(\,C\left(\bfp^\rmWT,\bfp^\rmMT\right)=\rmEC_50^\rmMT/\rmEC_50^\rmWT\). We chose the half maximal effective concentration (EC₅₀) to model C, as T cell activation to p^WT was consistently a sigmoidal function (Extended Data Figs. 4c, d and 6a, b) described by a Hill equation, where EC₅₀ determines how a ligand activates a receptor. We next estimated the EC₅₀ of all 1,026 TCR–p^MT pairs to infer a model for C that estimates whether a neoantigenic substitution is cross-reactive (and therefore tolerated) based on the substituted amino acid position and residue (Extended Data Figs. 6a, b and 7a, b). We then tested whether C predicted cross-reactive substitutions in an HLA-B*27:05-restricted neopeptide–TCR pair from an LTS (Extended Data Fig. 4e). Notably, C predicted cross-reactive p^WT, p^MT and p^MT, p^MT substitutions in this neopeptide–TCR pair (Fig. 3f and Extended Data Fig. 5c, 6c). Thus, we combined all 1,197 TCR–p^MT pairs to derive a composite C—the antigenic distance for a TCR to cross-react between amino-acid-substitution pairs (Fig. 3g and Extended Data Fig. 7c). Broadly, two factors promote cross-reactivity: substitutions at peptide termini²⁷ and within amino acid biochemical families (driven by amino acids of similar size and hydrophobicity; Fig. 3g). With this composite C, we now define self-discrimination D between a p^WT and its corresponding p^MT (Fig. 3a) as

where \(w\) sets the relative weight between the two terms. We chose the parameters of the neoantigen quality model to maximize the log-rank test score of survival analysis on an independent cohort of 58 patients with PDAC⁵ (Supplementary Methods and Extended Data Table 1a).