Clinico-genomics analysis of shared cancer neoantigens

To establish a computational and experimental pipeline for neoepitope–HLA discovery, we first identified the most common recurrent point mutations across cancer types within a compendium of sequencing data from tumor and normal tissue samples9, filtering at a per-indication case prevalence of 2% (Fig. 1a). This led to a list of 36 shared cancer neoantigens (Supplementary Table 1). Next, we mined the Allele Frequency Net Database (AFND) and The Cancer Genome Atlas (TCGA) to catalog common haplotypes, narrowing to those with a carrier frequency of at least 10% in TCGA and an allele frequency of at least 5% in AFND. This analysis led to a list of 16 HLA alleles that combined with the 36 selected neoantigens to provide the foundation for development of our platform (Supplementary Table 1).

Fig. 1: A shared neoepitope discovery pipeline featuring characterization of neoepitope–HLA binding through a high-throughput TR-FRET assay and NetMHCpan-4.0 prediction.

a, Overview of the shared neoepitope discovery pipeline. b, Schematic diagram of the TR-FRET assay used to measure stable neoepitope–HLA binding. In brief, HLA monomers bound to UV-cleavable peptides are exposed to UV light in the presence of mutation-bearing candidate neoepitopes. Successful exchange of the candidate peptide will lead to complex stabilization and TR-FRET emission (top). Unsuccessful exchange will lead to aggregation and no TR-FRET emission (bottom). c, TR-FRET data for all controls measured within the screen. Z-score was calculated as compared to −Peptide/+HLA controls for each 384-well plate. The box represents the interquartile range; the line represents the median value; and the whiskers represent the minimum and maximum values (excluding outliers). Each dot represents an individual well measurement (−Pep,−HLA n = 1,477; −Pep,+HLA n = 368; pp65,+A*02:01 n = 623). d, Representative TR-FRET results for KRAS G12V/A*03:01 comparing NetMHC BA percentile rank (%Rank, blue) and RZ-score (red). e, Percent of neoepitope–HLA combinations that were determined to be stable binders by TR-FRET (red) or NetMHC (blue) across the HLA A, B and C alleles. The box represents the interquartile range; the line represents the median value; and the whiskers represent the minimum and maximum values (excluding outliers). Each dot represents the count of binders for a single allele (A allele,NetMHC n = 5; A allele,TR-FRET n = 5; B allele,NetMHC n = 4; B allele,TR-FRET n = 4; C allele,NetMHC n = 6; C allele,TR-FRET n = 6). f, Percent of neoepitope–HLA pairs found to be binders by NetMHC and/or TR-FRET across the A (green), B (purple) and C (orange) alleles, with each dot representing a single allele. The box represents the interquartile range; the line represents the median value; and the whiskers represent the minimum and maximum values (excluding outliers). Each dot represents the percent agreement for each allele (A allele n = 5, B allele n = 4, C allele n = 6). g, Scatter plot of TR-FRET RZ-score and NetMHC BA %Rank. The dashed red line represents the cutoff for stable binders as measured by TR-FRET, where values higher than the red line are considered a stable binder. The dashed blue line represents the cutoff for binders based on NetMHC analysis, where values lower than the blue line are considered binders. a and b were created with BioRender.com.

High-throughput TR-FRET analysis of neoepitope–HLA stability

To survey all potential neoepitopes between candidate cancer neoantigens and selected HLA alleles, a high-throughput time-resolved fluorescence energy transfer (TR-FRET) assay based on peptide-mediated stabilization of conditional HLA complexes was developed (Fig. 1b)5. Our neoantigen target set consisted of 36 shared cancer neoantigens identified above along with 11 additional tumor antigens. Separately, 15 of 16 prioritized HLA variants were viable in the conditional HLA complex format (Fig. 1b). Together, this permitted the characterization of 24,149 neoepitope–HLA complexes after eliminating overlapping cancer neoepitopes as well as one allele due to synthesis challenges (Supplementary Table 2).

Conditional HLA complexes, pre-loaded with ultraviolet (UV)-cleavable peptides, were incubated with a neoepitope of interest at 100-fold molar excess and exposed to UV light for 25 min. This reaction leads to conditional ligand cleavage and conversion of the peptide from a stable high-affinity ‘binder’ to an unstable binder that dissociates from the HLA groove. In the presence of a binding neoepitope, peptide exchange would stabilize the HLA complex, whereas a lack of binding results in complex dissociation. Complex stability was monitored using fluorescence of a TR-FRET donor (europium) conjugated to an anti-β2M antibody and a TR-FRET acceptor conjugated to streptavidin, which bound to the biotin tag on the HLA alpha chain, where a TR-FRET signal would be observed only if the complex remained intact. TR-FRET signals were quantified based on the ratio of relative fluorescent units (RFUs), and signals were subjected to a double normalization to generate a robust Z-score (RZ-score). Any neoepitope–HLA combination with an RZ-score ≥5 was considered a ‘stable binder’, which was a conservative measure based on prior assessment with our positive control CMV-peptide/HLA-A*02:01 complex. A Z-score ≥5 captured 90% of positive control binding events without identifying false-positive binders (Fig. 1c).

NetMHCpan-4.0 (hereafter NetMHC) was employed to better understand how our TR-FRET results compared to computational prediction methods10,11. We considered NetMHC binding affinity (BA) percentile rank (%Rank) relative to TR-FRET results and eluted ligand (EL) %Rank to determine if a neoepitope was predicted to be presented (%Rank ≤2). Representative data for KRAS G12V peptides binding to A*03:01 showed two previously described neoepitopes12,13, VVGAVGVGK and VVVGAVGVGK, as binders with both approaches (Fig. 1d). Further examination of neoantigen–HLA combinations revealed variable concordances between the TR-FRET and NetMHC results (Supplementary Fig. 1a,b). Assessment of KRAS G12D peptides with C*08:02 found a known neoepitope (GADGVGKSAL)14 to be a binder by both methods (Supplementary Fig. 1c).

When measured as a percentage of all potential neoepitope–HLA complexes, TR-FRET generally identified more stable binders as compared to NetMHC (Fig. 1e and Supplementary Fig. 1d). We found that the percent agreement between NetMHC prediction and TR-FRET when classifying binders was generally less than 30% (Fig. 1f and Supplementary Fig 1e), whereas much stronger agreement was found for non-binders only (Supplementary Fig. 1f,g). For further comparison, TR-FRET RZ-scores and NetMHC %Ranks were plotted for all candidate neoepitope–HLA pairs, demonstrating that 0.63% of neoepitope–HLA pairs were probable binders by both methods (Fig. 1g and Supplementary Fig. 2a). The different methods identified a similar percentage of additional binding events for neoepitope–HLA pairs, demonstrating that each has the potential to uncover unique binding combinations (Fig. 1g and Supplementary Fig. 2b). These findings highlight the power of our high-throughput TR-FRET assay to identify an expanded and complementary set of neoepitope–HLA pairs relative to computational prediction and suggest that co-deployment of both approaches would be needed for comprehensive neoepitope discovery.

Generation of monoallelic cells co-expressing 47 neoantigens

Despite observed peptide–HLA stabilization in vitro or computational prediction of an interaction, mutant protein expression and processing may not result in neoepitope presentation in a cellular context15,16,17. For this reason, candidate neoepitope validation typically requires evidence of direct physical association with surface-bound HLA via HLA immunoprecipitation (HLA-IP) followed by MS. This process has been enhanced through the use of engineered ‘HLA monoallelic’ cell lines, although these have largely relied on endogenous mutant protein expression or expression of relatively few mutant transgenes, thus limiting throughput13,18,19.

We anticipated that co-expression of all 47 candidate neoantigen sequences (concatenated ~25 amino acid segments centered on the mutated position) within a single HLA-null cell line would improve throughput of monoallelic cell line generation and subsequent validation of TR-FRET/NetMHC-identified neoepitope–HLA pairs by targeted MS (Fig. 2a). For this, we selected the HMy2.C1R (C1R) lymphoblast cell line, which lacks HLA-A and HLA-B20,21. To generate a full C1RHLAnull cell line, the HLA-C allele (HLA-C*04:01) was disrupted using CRISPR–Cas9, and the HLA-null population was enriched by fluorescence activated cell sorting (FACS) (Supplementary Fig. 3a–c).

Fig. 2: Generation of HLA class I monoallelic cell lines that stably express a polyantigen cassette containing 47 shared cancer neoantigens.

a, Process overview for the cell engineering and MS analysis of peptides presented by polyantigen-expressing HLA monoallelic cell lines. KO, knockout. b, Vector map of the piggyBac polyantigen expression constructs used in this study. A single transcript containing 47 tandem neoantigens followed by seven control peptides and an IRES-linked mTagBFP2 (BFP) reporter is driven by an EF1a promoter. Neoantigens were either directly concatenated (no-linker) or interspersed by short flexible linker sequences (linker). c, Flow cytometric detection of HLA expression (W6/32 antibody-APC) and polyantigen cassette reporter (BFP) in selected cell lines or HLA knockout parental line. Here, ‘−’ indicates the absence of linkers and ‘+’ indicates presence of linkers in the polyantigen construct. d, Targeted immunopeptidomic detection of a previously described A*02:01-presented TP53 neoepitope (HMTEVVRHC, position 39) as well as two neoantigen control peptides from pp65 (NLVPMVATV, control 1) and IE-1 (VLEETSVML, control 4) known to be presented by the A*02:01 allele. a and b were created with BioRender.com.

Local amino acid sequence context may affect antigen processing22. Accordingly, we engineered unique C1RHLAnull lines to stably express concatemers of all 47 prioritized neoantigens that were separated, or not, by short, flexible amino acid linkers (Fig. 2b). Subsequent introduction of the 15 HLA alleles as individual transgenes through stable lentiviral transduction of the linker and no-linker neoantigen-expressing C1RHLAnull cell lines resulted in 30 total cell populations (Fig. 2c and Supplementary Fig. 3d,e). To validate functionality of the polyantigen cassettes, the linker and no-linker neoantigen constructs contained an identical set of seven known HLA-A*02:01-presented epitopes. HLA-IP followed by targeted MS analysis confirmed presentation of two control peptides and a previously described TP53 R175H23 neoepitope in both the linker and no-linker HLA-A*02:01-engineered cells (Fig. 2d).

Detection of neoepitopes presented on engineered monoallelic cells

Both targeted and untargeted MS were applied for neoepitope discovery across the panel of monoallelic cell lines. Untargeted MS analysis enabled unbiased identification of peptides from the entire immunopeptidome, whereas targeted analysis facilitated detection of peptides presented at low copies per cell but was constrained to prioritized sequences from our TR-FRET/NetMHC analyses.

Untargeted MS analysis was performed with a semi-automated workflow resulting in 218–6,663 unique 8–11-mer peptides identified from each cell population (Fig. 3a,b). The number of 8–11-mer peptides and general sequence features for each allele overlapped regardless of the polyantigen linker status (Supplementary Fig. 4a) and confirmed that presented peptides fit known motifs (Supplementary Fig. 5). Expression of the polyantigen cassette was confirmed by detection of control viral epitopes from A*02:01 and A*11:01 monoallelic cells (Supplementary Fig. 4b) as well as epitopes from an integrated blue fluorescence protein (BFP) selection marker across eight different HLA alleles (Supplementary Fig. 4c).

Fig. 3: Untargeted immunopeptidomic analysis of monoallelic cell lines expressing the polyantigen cassette.

a, Workflow for untargeted immunopeptidomic analysis of monoallelic cell lines containing the polyantigen cassette. b, Number of unique 8–11-mer peptides identified in untargeted immunopeptidomic analysis. The box represents the interquartile range; the line represents the median value; and the whiskers represent the minimum and maximum values (excluding outliers). Each dot represents a separate analysis beginning with a replicate cell pellet (A*01:01 n = 6, A*02:01 n = 13, A*03:01 n = 8, A*11:01 n = 6, A*24:02 n = 4, B*07:02 n = 4, B*08:01 n = 4, B*35:01 n = 2, B*51:01 n = 2, C*03:04 n = 2, C*04:01 n = 4, C*05:01 n = 4, C*06:02 n = 4, C*07:01 n = 4, C*07:02 n = 4, HLAKO n = 5). c, Identified shared cancer neoantigen epitopes. The color scale represents the log10 largest area across all analyses. d, Comparison of TR-FRET RZ-score and NetMHC EL percentile rank (%Rank) score for each epitope identified through untargeted immunopeptidome analysis. a was created with BioRender.com.

From our untargeted analyses, we observed 22 neoepitope–HLA pairs and several peptides from non-mutation-bearing regions of the polyantigens. Neoepitopes corresponded to 15 shared neoantigens across five HLAs, representing ~5.4% of neoepitope–HLA pairs predicted by NetMHC and ~3.7% of neoepitope–HLA pairs identified within the TR-FRET assay (Fig. 3c and Supplementary Table 3). Of the 22 neoepitope–HLA pairs, 10 were previously described in the literature, and the remaining 12 were thought to be novel based on a search of Tantigen24, CAatlas25 and NEPdb26 and an extended literature survey (Supplementary Table 3). TR-FRET and NetMHC showed excellent concordance for all 22 identified pairs; 17 were identified as binders by both approaches (Fig. 3d). One and three neoepitope–HLA pairs were uniquely identified as hits by TR-FRET and NetMHC, respectively, demonstrating that each approach can predict distinct neoepitope subsets (Fig. 3d). One neoepitope–HLA pair (TP53 R175H (HMTEVVRHC)/A*02:01) represented an exception. This pair had a TR-FRET RZ-score of 3.9 and a NetMHC EL %Rank of 3.98 and was not considered a hit by either approach, demonstrating that false negatives remain possible.

Although we surmised that targeted MS analysis would improve detection of presented neoepitopes, this relied on heavy isotope-labeled standard peptides. As such, a logistically challenging synthesis of 1,786 peptides (47 neoantigens × 38 possible mutation-bearing candidate neoepitopes) would be needed to screen all potential neoepitopes from our monoallelic cell lines. Therefore, we used the TR-FRET results as a preliminary screen and synthesized all 397 peptides with an RZ-score ≥5. Due to the complementarity of TR-FRET and NetMHC results, an additional 81 peptides were synthesized that had an RZ-score <5 and NetMHC %Rank ≤2. The 479 peptides were divided into 15 HLA allele-specific pools comprising 21–88 peptides (Fig. 4a,b).

Fig. 4: Targeted immunopeptidomic analysis of monoallelic cell lines expressing a polyantigen cassette.

a, Targeted immunopeptidomic workflow for the analysis of candidate neoepitopes within monoallelic cell lines expressing a polyantigen cassette. b, The number of targeted (blue) and detected (red) shared cancer neoantigen epitopes within each targeted assay. c, Comparison of TR-FRET RZ-score and NetMHC EL %Rank score for each epitope identified through targeted immunopeptidome analysis. d, NetMHC EL %Rank scores for neoepitopes detected in both untargeted and targeted (teal) analysis or targeted analysis alone (red). The box represents the interquartile range; the line represents the median value; and the whiskers represent the minimum and maximum values (excluding outliers). Each dot represents a neoepitope–HLA pair (untargeted & targeted n = 22, targeted only n = 64). The P value was calculated using a Wilcoxon test (two-sided). e, Same analysis as d but for TR-FRET RZ-scores. f, Summary of neoepitope–HLA pairs detected from shared cancer neoantigens. Color represents attomol of neoepitopes detected on column during analysis. Bolded squares with centered dots represent neoepitopes also detected in untargeted analysis. A square with an ‘X’ indicates A*02:01-specific neoepitopes that were detected in a cell line containing a polyantigen construct lacking control neoantigen sequences. a was created with BioRender.com.

Targeted MS analysis identified 86 neoepitope–HLA pairs across 12 different alleles and 36 neoantigens, representing a ~4-fold improvement compared to untargeted MS analysis (Fig. 4b and Supplementary Table 3). After a search of the literature and relevant databases, we determined that 21 of the neoepitope–HLA pairs were described previously, and 65 were novel (Supplementary Table 3). Twenty of 86 neoepitope–HLA pairs identified across untargeted and targeted analyses were associated with A*11:01. This was likely due to the presence of eight distinct KRAS neoantigen sequences in the polyantigen cassette, as 14 of 20 A*11:01-specific and nine of 14 A*03:01-specific neoepitopes mapped to KRAS G12X or G13X neoantigens.

To assess the value of using TR-FRET and NetMHC results to select peptides for targeted MS, we plotted RZ-score versus NetMHC %Rank for each of the observed 86 neoepitope–HLA pairs (Fig. 4c). This revealed that 55 neoepitopes were stable binders by TR-FRET and predicted to be presented by NetMHC. Thirteen neoepitope–HLA pairs were found as hits in TR-FRET only, whereas 18 neoepitope–HLA pairs were hits in NetMHC alone (Fig. 4c). To understand the binding characteristics of neoepitope–HLA pairs identified by targeted analysis alone, we plotted RZ and NetMHC %Rank scores for peptides observed in both untargeted and targeted analysis compared to peptides found only in targeted analysis (Fig. 4d,e). Neoepitope–HLA pairs identified by targeted analysis alone had a broader range of NetMHC %Rank and TR-FRET RZ-scores relative to neoepitopes also detected in untargeted analysis (Fig. 4d,e). This suggests that targeted analysis can identify neoepitopes that are weaker binders compared to those observed by untargeted means.

Targeted MS permits absolute quantification of peptide presentation across neoepitopes. Overall, the measured amount of neoepitope presentation spanned from 60 amol to 2.5 pmol (Fig. 4f) and was consistent across independent replicates of cell line growth and sample preparation (Supplementary Fig. 6). Two peptides detected by untargeted MS, EGFR G719A (ASGAFGTVYK) and FGFR3 S249C (ERCPHRPIL), exhibited the highest absolute quantities (Fig. 4f). When the absolute amounts of neoepitopes detected were compared to RZ-score, NetMHC EL %Rank or NetMHC BA %Rank for each allele, no clear correlation could be found (Supplementary Fig. 7a–c). This suggests that each score has predictive value for neoantigen presentation but also that these cannot be used to estimate the absolute amount presented.

Polyantigen cassette design impacts neoepitope presentation

The polyantigen sequence included neoantigens with known A*02:01-binding epitopes to confirm translation, processing and presentation of the cassette. It was possible that the controls could compete with experimental neoepitopes, thus creating an avenue for false negatives. To evaluate this, a separate A*02:01 cell line was created that stably expressed a no-linker polyantigen cassette lacking control sequences. Upon analysis of the no-control line, two additional neoepitopes were detected: YVCNTTARA (SF3B1 R625C; RZ-score = 16; EL %Rank = 5.3) and QLMPFGSLL (EGFR C797S; RZ-score = 7; EL %Rank = 0.21) (Fig. 4f, squares with ‘X’). These results suggest that strong binding peptides could inhibit presentation of certain neoepitopes, and a revised workflow may omit control sequences from the polyantigen cassette.

Polyantigen cassette length is an important consideration when designing cancer vaccines, and a concern that translation of neoantigens at the C-terminal/3′ end of the cassette will be decreased may have factored into the use of shorter cassettes in clinical settings (for example, 10-mer or 34-mer)27. To characterize the translation efficiency of our 47-mer polyantigen transgene, we performed ribosome profiling (Ribo-Seq) on A*02:01 monoallelic cells containing either the linker or no-linker cassette with A*02:01 controls (Supplementary Fig. 8a,b). These analyses demonstrated consistent translation across the no-linker polyantigen cassette, whereas the cassette containing linkers had a substantial decrease in translation after ~20 neoantigen sequences (Supplementary Fig. 8a,b).

We next sought to understand if the difference in translation between cassette designs was reflected within our targeted immunopeptidomics results. For this, we plotted the highest attomole abundance of presented peptide for each neoantigen (irrespective of HLA) versus neoantigen position within the linker and no-linker polyantigen cassettes (Supplementary Fig. 8c). This revealed a potential bias toward presentation of peptides derived from the first ~20 neoantigen sequences regardless of format. Within the portion of the polyantigen cassette that exhibits lower translation, we detected six additional neoepitope–HLA pairs from cells expressing the no-linker cassette, suggesting that the no-linker format may be advantageous for assaying ≥20 target sequences (Supplementary Fig. 8c).

For neoepitopes detected in both the linker and no-linker cell lines, there was not a clear difference in the maximum presentation, suggesting that positional effects detected in the Ribo-Seq data could be buffered at the level of presentation (Supplementary Fig. 8c). This was further supported by roughly equivalent presentation of KRAS G12X and G13X neoepitopes (which are identical except for the mutated residue) across the polyantigen cassette (Supplementary Fig. 8d). To evaluate the impact of linkers more broadly, we plotted the highest absolute amount of neoepitope presented and found that presentation of some neoepitopes increased in the presence of linkers while presentation of other neoepitopes decreased (Supplementary Figs. 9a,b and 10). Together, these data demonstrate that the no-linker polyantigen cassette enabled detection of a greater number of neoepitope–HLA pairs. However, if a neoepitope was detected in linker and no-linker cells, the presence of linkers did not impact abundance of presentation in a consistent manner.

Validation of neoepitope presentation from full-length protein

Neoepitopes derived from a polyantigen construct may not reflect peptides processed from a full-length mutant protein. To address this, we developed four HLA-A*11:01 monoallelic C1R lines expressing an inducible, full-length wild-type, G12C, G12D or G12V mutant KRAS transgene and compared neoepitope presentation from these cell lines with a cell line expressing the same HLA and a no-linker polyantigen cassette. Expression of full-length variant proteins was confirmed using a whole-cell targeted proteomic assay comprising a peptide that can detect total KRAS as well as three unique peptides that measured individual KRAS mutants (Fig. 5a). Little to no mutant peptide signal was observed in total protein samples from the polyantigen cassette-expressing cell line (Fig. 5a).

Fig. 5: Presentation of KRAS neoepitopes derived from exogenous and endogenous expression of full-length mutant protein.

a,b, A*11:01 monoallelic cells were engineered to express doxycycline (dox)-inducible full-length (FL) KRAS mutant proteins (G12C, G12D and G12V). These were compared against an A*11:01 monoallelic cell line containing the no-linker polyantigen cassette. a, Absolute amount of KRAS wild-type (WT) and mutant proteins in the cell lysate by targeted MS. b, Copies per cell of presented KRAS 9-mer (VVGAXGVGK) and 10-mer (VVVGAXGVGK) neoepitopes as measured by A*11:01 monomers containing heavy synthetic neoepitope peptides spiked in before affinity purification and targeted MS. c, Targeted immunopeptidomic analysis of neoepitopes in cell lines that endogenously express both KRAS and A*11:01. Two neoantigens for KRAS G12C (VVGACGVGK and VVVGACGVGK) and KRAS G12D (VVGADGVGK and VVVGADGVGK) were analyzed in cell lines that harbor KRAS G12C (HOP62 and NCIH2030), G12D (HuCCT1 and SNU601) or G12V (SW527). These neoepitopes are either novel or were described previously in non-endogenous systems. ‘Treated’ samples were treated with interferon-gamma.

We then performed HLA-IP and targeted MS to quantify presentation of previously identified 9-mer and 10-mer KRAS epitopes associated with HLA-A*11:01 (Fig. 5b)12,13. In cell lines expressing full-length mutant transgenes, clear induction of neoepitope presentation was observed for both G12V epitopes as well as the 10-mer epitope of G12D (Fig. 5b). From cells expressing the polyantigen cassette, all targeted mutant KRAS epitopes were detected and measured at higher absolute copies per cell compared to lines expressing full-length mutant proteins (Fig. 5b). Detection of KRAS peptides after HLA-IP but not from total cell protein suggested that the polyantigen concatemer was likely unstable and efficiently degraded, resulting in enhanced epitope presentation28,29. Therefore, monoallelic cells containing the polyantigen cassette provided a reliable, higher throughput and more sensitive system for discovery of neoepitopes from shared cancer neoantigens relative to cell lines expressing a full-length antigen.

Lastly, we sought to demonstrate that neoepitopes discovered by our pipeline can be identified within cells that endogenously co-express relevant proteins and HLA alleles. Targeted MS assays were used to quantify four neoepitopes—9-mer and 10-mer from KRAS G12C and G12D—within cell lines that express A*11:01 as well as KRAS G12C (HOP62 and NCIH2030), KRAS G12D (HuCCT1 and SNU601) or KRAS G12V (SW527) (Fig. 5c). One of these neoepitopes (KRAS G12C (VVGACGVGK)) has not previously been described, whereas the remaining three neoepitopes have been confirmed only within cellular systems that exogenously express the neoantigen12. We confirmed presentation of the four target neoepitopes within cell lines that harbor the target neoantigens (KRAS G12C and G12D), whereas there was no observed presentation in a control cell line that contained KRAS G12V (Fig. 5c). In both HOP62 and NCIH2030 cells, KRAS G12C 9-mer neoepitopes appeared to have higher absolute presentation as compared to the previously described 10-mer (Fig. 5c). Furthermore, whereas the presentation of the 10-mer KRAS G12D epitope was similar across HuCCT1 and SNU601 cells, presentation of the 9-mer KRAS G12D neoepitope was much higher within HuCCT1 (Fig. 5c). This suggests that presentation of slightly varying neoepitopes can differ substantially based on the cell line from which they are derived. In total, these data demonstrate that neoepitopes discovered through our pipeline can be found from both exogenously expressed full-length proteins and within systems that endogenously express both the HLA and neoantigen.

Functional validation of tumor-specific antigen–HLA pairs

To determine whether neoepitopes identified through our workflow could be recognized by human T cells, we employed a modified multiplexed TCR discovery method8. Using two of the identified neoepitope–HLA pairs (FLT3-p.D835Y/A*02:01, PIK3CA-p.E545K/A*11:01) as examples, neoepitopes were first allocated to peptide pools in unique combinations before healthy human donor CD8+ T cells were expanded using autologous monocyte-derived dendritic cells, restimulated with the neoepitope peptide pools, sorted for activation marker upregulation and subjected to TCRβ sequencing. This method was used for donors spanning a range of HLA genotypes, enabling the association of TCRs with a variety of peptide–HLA pairs. However, owing to the multiallelic nature of donor cells, the HLA restriction of identified neoepitopes was not initially disambiguated among the 3–6 donor HLA alleles.

For neoepitopes that elicited a T cell response, associated TCRβ and TCRα sequences were determined using a parallel multiplexed assay30 that enabled construction of paired TCR expression vectors and the selection of candidate neoepitope-specific TCRs. The specificity and potential efficacy of each TCR were then assessed through cellular assays. TCR encoding in vitro transcribed mRNA was introduced via electroporation into primary human T cells, which were then incubated with either an increasing concentration of the candidate neoepitope in the presence of A*02:01+ T2 cells or monoallelic K562 cells that co-expressed an HLA allele and neoantigen of interest.

We found dose-dependent upregulation of CD137 after 12-h co-culture of primary human CD8+ T cells transfected with predicted FLT3-p.D835Y/*02:01-specific TCRs in response to T2 cells incubated with exogenously delivered YIMSDSNYV peptide (Fig. 6a). Furthermore, these T cells were activated by and specifically killed monoallelic A*02:01-K562 cells expressing a mutant FLT3-p.D835Y transgene (minigene encoding 21 amino acids) but were not activated by and did not kill monoallelic A*02:01-K562 cells expressing a wild-type FLT3 transgene (Fig. 6b,c). These TCRs appear to be exquisitely specific for the mutant neoepitope, which is an important characteristic because a similar non-mutant epitope IMSDSNYVV was identified by untargeted analysis in A*02:01 monoallelic cells.

Fig. 6: Discovery of neoepitope-specific TCRs demonstrates immunogenic potential of discovered neoepitope–HLA pairs.

Human CD8+ T cells were transfected with either FLT3-p.D835Y-specific or PIK3CA-p.E545K-specific TCR RNA. a–c, CD137 expression was assessed after T cells transfected with FLT3-p.D835Y-specific TCRs were co-cultured overnight with YIMSDSNYV peptide-pulsed A*02:01+-T2 cells (a). CD137 expression (b) and specific lysis (c) were determined after co-culture with A*02:01+-K562 cells transfected with no RNA, transgene containing mutant FLT3-p.D835Y or transgene containing FLT3-D835 wild-type sequence. The box represents the interquartile range; the line represents the median value; and the whiskers represent the minimum and maximum values (excluding outliers) (n = 4). d–f, CD137 expression was assessed after T cells transfected with PIK3CA-p.E454K-specific TCRs were co-cultured overnight with STRDPLSEITK peptide-pulsed A*11:01+-K562 cells (d). CD137 expression (e) and specific lysis (f) were determined after co-culture with A*11:01+-K562 cells transfected with no RNA, transgene containing mutant PIK3CA-E545K or transgene containing PIK3CA-E545 wild-type sequence. The box represents the interquartile range; the line represents the median value; and the whiskers represent the minimum and maximum values (excluding outliers) (n = 3). mut, mutant; wt, wild-type.

As a second proof of concept, T cells were transfected with predicted PIK3CA-p.E545K/HLA-A*11:01 TCRs and mixed with monoallelic A*11:01-expressing K562 cells incubated with an increasing concentration of the predicted neoepitope STRDPLSEITK (Fig. 6d). Here, TCR-transfected T cells demonstrated dose-dependent activation as measured by CD137 expression. Furthermore, these T cells demonstrated higher levels of activation and cell killing when mixed with A*11:01 K562 cells expressing a PIK3CA-p.E545K transgene (minigene encoding 21 amino acids) as compared to cells that expressed a wild-type PIK3CA transgene (Fig. 6e,f). Mutations that introduce anchor residues are thought to have high immunogenic potential because the immune system has not built tolerance to a similar wild-type epitope. For PIK3CA-p.E545K/A*11:01, the E → K mutation introduces an anchor residue within the context of A*11:01, and the wild-type STRDPLSEITE epitope was not detected in untargeted MS analyses of A*11:01 monoallelic cells. Although false negatives are anticipated in our MS workflow, the wild-type epitope was also not predicted to bind A*11:01 by NetMHC (12.8). Taken together, these data provide a clear mechanism for the specificity of PIK3CA-p.E545K TCRs for recognition of mutant PIK3CA as compared to wild-type and lend support for these TCRs as potential therapeutic candidates.