816 VOLUME 35 NUMBER 9 SEPTEMBER 2017 NATURE BIOTECHNOLOGY recombination of human TCR genes; however, the peripheral repertoire actually contains only~ 2.5× 107 clonotypes19 (Fig. 1b). To this, we must factor in the individual’s genetic and environmental experience, which further shapes the available T-cell repertoire. This influence is best exemplified by a largescale study of monozygotic twins discordant for cytomegalovirus status, in which differences in immunological parameters were largely determined by non-heritable factors20. The presence of a tumor could certainly limit the available tumor-specific TCR repertoire. We therefore argue that, at present, the composition and specificity of the available TCR repertoire cannot be determined by an in silico approach. Lastly, we note the relative dearth of in silico methods for predicting MHC-II-restricted putative neoepitopes. Activation and maintenance of a CD8+ T-cell response is dependent on concomitant activation of MHC-II-restricted helper T cells during priming21. Consistent with this, CD4+ T cells are known to play crucial roles in the anti-tumor response in vivo, and many reports have established that MHC-II-restricted neoepitopes can be immunogenic and elicit anti-tumor protection21. Indeed, peripheral CD4+ T cells are substantially expanded in some patients responding to of computational tools for a given MHC allele, the range is from 0.07% to 10%) 9, 10. Of the~ 1% binding MHC, only~ 50% will be recognized by a T cell, but only 30–40% are naturally processed, enabling target cell killing10 (Fig. 1). In a report that systematically interrogated patient responses to vaccination with predicted tumor neoepitopes, three melanoma patients were each immunized with seven peptides with in vitro-corroborated MHC-binding affinities< 500 nM11. Of the 21 peptides tested, only 9 induced a T-cell response. Three of the nine neoepitopes were ‘dominant’(the responding T cells were present in the patient before immunization), four were ‘subdominant’(the T cells were induced by neopeptide immunization), and two were ‘cryptic’(the responding T cells reacted to the neoepitopes but not to cells expressing the corresponding peptides). Thus, only 30% of the tested peptides (7 peptides) elicited a T-cell response in vivo, which is intriguingly similar to the findings with viral peptides. Although binding to MHC-I is currently the most effective computational filter for removing nonantigenic peptides12, methods that identify competing MHC alleles can also reduce the burden of experimental validation. Humans express 12 MHC alleles: 6 class I (HLA-A, B, and C) and 6 class II (HLA-DR, DP, and DQ). However, as different alleles compete for peptides13, 14 mono-allelic profiling of the immunopeptidome15 may not recapitulate the fate of a single peptide within the complexity of the cell haplotype in vivo and computational methods may need to be developed to allow for the complexity of the system.

We must also ask whether our knowledge of cancer biology could be exploited for in silico screening. The majority of tumors are highly heterogeneous, and distinct regions of a single lesion can have different mutational profiles. Should we limit selection to peptides resulting from clonal mutations (present in all cancer cells in the tumor) or include subclonal mutations (present in only a subset of cells)? Evidence from non-small-cell lung cancer and melanoma patients treated with immune checkpoint inhibitors suggested that only T-cell responses to clonal neoepitopes were associated with clinical benefit and prolonged survival16. Thus, cancer genomic computational methods could guide the work of …