RETRACTED ARTICLE: Cross-HLA targeting of intracellular oncoproteins with peptide-centric CARs

The majority of oncogenic drivers are intracellular proteins, thus constraining their immunotherapeutic targeting to mutated peptides (neoantigens) presented by individual human leukocyte antigen (HLA) allotypes1. However, most cancers have a modest mutational burden that is insufficient to generate responses using neoantigen-based therapies2,3. Neuroblastoma is a paediatric cancer that harbours few mutations and is instead driven by epigenetically deregulated transcriptional networks4. Here we show that the neuroblastoma immunopeptidome is enriched with peptides derived from proteins that are essential for tumourigenesis and focus on targeting the unmutated peptide QYNPIRTTF, discovered on HLA-A*24:02, which is derived from the neuroblastoma dependency gene and master transcriptional regulator PHOX2B. To target QYNPIRTTF, we developed peptide-centric chimeric antigen receptors (CARs) using a counter-panning strategy with predicted potentially cross-reactive peptides. We further hypothesized that peptide-centric CARs could recognize peptides on additional HLA allotypes when presented in a similar manner. Informed by computational modelling, we showed that PHOX2B peptide-centric CARs also recognize QYNPIRTTF presented by HLA-A*23:01 and the highly divergent HLA-B*14:02. Finally, we demonstrated potent and specific killing of neuroblastoma cells expressing these HLAs in vitro and complete tumour regression in mice. These data suggest that peptide-centric CARs have the potential to vastly expand the pool of immunotherapeutic targets to include non-immunogenic intracellular oncoproteins and widen the population of patients who would benefit from such therapy by breaking conventional HLA restriction.

• Is the immunopeptidome LC-MS/MS data deposited in PRIDE or MassIVE?What is the total number of unique samples (is it 8 total samples or were there replicates?) Can the authors provide a .rawfile mapping table that includes both the sample information (name, amount, antibody, HLA typing) that maps to each unique .rawfile generated along with the location where these data can be downloaded from?• The spectra in Supp Fig. S2 is to low resolution to be able to read the fragment ion masses.Can this be replaced with a high resolution image?• Related to the LC-MS/MS methods: What collision energies, max injection times, and ion targets were used in these methods?I was unable to find these parameters described in the methods.
• Are there other HLA-I alleles that this peptide is predicted to bind to when using either NetMHC4.1pan(http://www.cbs.dtu.dk/services/NetMHCpan/) or HLAthena (http://hlathena.tools/)for predictions?• The observation that the target peptide QYNPIRTTF binds to B14:02 is really interesting given that its binding motif is divergent from the A24:02/A24:03---B15:01/B15:02 seems to have a more similar motif.I may have missed this in the manuscript, but did the authors confirm B14:02 binding with the canonical biochemical HLA-I binding assay (PMID: 23392640)?If not, that would be helpful to have to confirm this peptide binds using an orthogonal assay.• I have not seen a paper report total spectra as a claim before-my expectation is a report the unique number of peptide identifications instead of total MS2 spectra.A unique spectra does not equal a putative peptide, as other chemical noise can trigger MS2 during data collection, and the same peptides can be sequences more than once in a single analysis run.Can the authors replace the top of the 1b with the total number of unique peptide identifications would be more useful.o Adding the peptide level FDR used for peptide identification in Fig 1B would also be useful for readers, as this impact the total number of peptide identifications.
• The authors report that 83 peptides derived from uniquely expressed genes in neuroblastoma that have not been previously detected in any normal tissue.What is the number of unique source proteins represented in this 83 peptide set?Reporting both the unique number of HLA-I and the unique number of source proteins they represent would be helpful.
• The authors describe a new algorithm called sCRAP-is this algorithm/code available?
• Adding limitations to the discussion would be useful that both describe their target identification approach and the.• Have the authors considered using either NetMHC or HLAthena to see if this peptide is predicted to bind to additional alleles?I am not sure how many alleles are covered by the prediction software used here.
• The authors's database search methods appear to describe parameters for mass spectrometry data collected using high resolution MS1 and low resolution MS2.However, their methods describe that the MS2 data was collected in the orbitrap resulting in high resolution data.Therefore, the 0.2 Da mass tolerance for fragment ions is not appropriate for this data type.Instead, a 20 ppm tolerance is the standard (PMID: 30666599, PMID: 30666599, PMID: 27869121).If the authors did collect MS2 data in the orbitrap, and a 0.02 Da window was used, these data need to be reanalyzed with 20 ppm fragment ion mass tolerance for the fragment ions.
• It is my opinion that a 5% FDR is too high when using high resolution MS2 data, and a more stringent FDR of <5%, more commonly 1% FDR may be more appropriate (PMID: 27869121).
• If the authors continue to use the 5% FDR, this should be started both in Figure 1 and when they make claims around the number of unique peptides identified, as their total identifications will contain higher numbers of false positive.
• Most publications report HLA-II peptides as 12-25mers.In the Database search and spectral annotation section, the authors state they include peptides 8-25mer in length.Is this a typo?If not, can the author add a citation that shows 8mers are presented on HLA-II using an orthogonal technology such as x-ray structure?If there are peptides that are below 12mers in this dataset, it is possible these are HLA-I binding peptides that are contaminating the samples.Can the authors confirm that the HLA-II peptides they are reporting that are less than 12 amino acids in length do not match the HLA-I motifs of the alleles present in their patient samples to rule this out?Referee #2 (Remarks to the Author): The study submitted by Yarmarkovich et al focuses on identifying new potential immunotherapeutic targets in neuroblastoma, a pediatric tumor with few mutations.Interestingly, they identify an unmutated peptide from PHOX2B which is presented by HLA-A*24:02.It is notable that in performing this immunopeptidome screen, parent genes of all peptides not detected in normal tissues resulted in a single enrichment ontogeny of "sympathetic nervous system development".Furthermore, PHOX2B is one of two penetrant susceptibility genes and the third most significant dependency in neuroblastoma.The group then develop peptide-centric chimeric antigen receptors (PC-CAR) targeting this peptide and find that these PC-CARs can be recognized by presentation on other HLA allotypes.It should also be noted that the group developed an important algorithm which will be important for the entire field; this selective Cross-Reactive Antigen Presentation (sCRAP) algorithm was able to predict the HLA-A*01:01 presented MAGE-A3 peptide that cross reacted with TITIN (4th ranked prediction), resulting in fatality in the MAGE-A3 TCR trial.Thus, this study presents a real tour de force; using an innovative immunopeptidome strategy for identifying targetable peptides and showing that they can be therapeutically targeted by a novel PC-CAR molecule.

General comments:
1) The authors use the CEF1 influenza peptide to evaluate the potential of a NB cell lines, with low MHC expression, to determine whether low level MHC expression is sufficient to elicit T cell activation.However, their data suggest that responsiveness is related to Myc ampification (low T cell response) rather than MHC expression per se (S7, d-e).While this is indicated in the figure legend, it is not addressed in the text.Two issues come up: -Flu peptide elicits strong T cell responses; would the authors expect that MHC expression would play a more critical role if a less immunogenic peptide was utilized?-Given the significantly lower T cell responses in the Myc-amplified tumors, did Myc amplification alter killing by PC-CAR-T cells (i.e.differences between MNA as compared to MNN tumors)?The authors do not appear to directly address this in the text and it is an important point; it would be of interest to comment on pMHC-TCR-based activation as compared to peptide-PC-CAR-based activation.
2) To specifically show the specificity of the PC-CAR for the PHOX2B peptide, the authors pulsed the HLA-A*23:01/PHOX2B-negative WM873 melanoma cell line with the PHOX2B peptide.Is there a reason why they didn't pulse and/or introduce PHOX2B into a negative NB line (SW620, HEPG2, and KATO III)?Additionally, the authors could evaluate their PC-CAR on the same NB line with absent PHOX2B expression (i.e.CRiSPR ko as presented in Figure 4).
3) The data presented in Figure 4 are critical for evaluating the efficacy of the PC-CAR T cells against PHOX2B+ neuroblastoma.It appears that some mice did not reach endpoint and as such, it would be important to show these data (? Showing tumor curves for each mice).Otherwise, it is not clear from evaluating the figure that not all mice regressed, but rather that they are not included in the analysis because they were sacrificed.Additionally, the upregulation of MHC is a fascinating piece of data but it is not clearly explained.Was MHC upregulated in all NB lines (the NB line evaluated in Fig 4h is not indicated)?Does this MHC upregulation change with time as this was evaluated at day 11?The authors hypothesize that MHC is upregulated due to IFNg release?Interestingly though, the authors show that regions of tumor that stain for CD3 segregate from regions with PHOX2B+ tumor (Fig S16).Is MHC upregulation detected on PHOX2B+ tumor and/or is there a difference between MHC upregulation on tumor with/without loss of PHOX2B?This last point may be beyond the scope of this study but can be discussed.

Specific comments:
1.In Figure 2d, it is not easy to identify the cross-reactive peptides (colors are not labeled).There also appears to be a mislabeling with 2g (legend is in 2h) and the crystal structure is referred to as Please check all the panels/legends here.2. As an additional control in Figure 3C, 10LH PC-CART cells should be stained with cells in the absence or presence of PHOX2B peptide loading (i.e using the non-pulsed and pulsed WM873 melanoma line).3.In figure 4g, it would be important to assure that differences are not due to the characteristics of a specific donor T cell.How many donors were evaluated in these experiments (neither donors nor mice numbers are indicated).This is the case for several of the panels.
Referee #3 (Remarks to the Author): In this manuscript, Yarmarkovich et al. combined genomics and immunopeptidomics to find novel targets for the development of CAR-T cell therapy against neuroblastoma.They identified an HLA-A24:02-binding unmutated peptide QYNPIRTTF derived from the transcription factor PHOX2B overexpressed in neuroblastoma.Next, they selected peptide centric scFv-based CARs by screening a scFv library for high binding to HLA-A24:02/PHOX2B and negligible binding to other HLA-A24:02/peptide complexes, including complexes with a high predicted cross-reactivity based on a SCRAP algorithm developed by the authors.The authors selected the clone 10LH with the highest specificity profile for further characterization.They confirmed that 10LH binding to HLA-A24:02/PHOX2B is highly dependent the PHOX2B peptide sequence by alanine screen.More interestingly they demonstrated that the binding of 10LH is independent of the HLA allele and 10LH also binds with QYNPIRTTF presented by HLA-A23:01 and HLA-B14:02 with high affinity.Finally, the authors showed that 10LH CAR T cells kill HLA24:02 and HLA-23:01 neuroblastoma lines expressing PHOXB2 in vitro and in a mouse xenograft model.Overall the approach is interesting, it not only enables the identification of high affinity scFv against self-antigens for which high affinity native TCRs are unlikely available, it also allows to broaden the eligible patient population by abrogating HLA restriction.A major limitation is that the manuscript is difficult to read in its current form.The figure legends are poor and lack information making it difficult to interpret the data.
The authors filter their tumor antigen candidates against an existing database of the normal tissue immunopeptidome, but absence of detection in this type of database doesn't necessarily mean that the peptide is not presented in normal tissues.The authors should also consider single cell RNAseq databases as a filter.
Fig2: Legends and graphs do not match up.Fig2b: Why is the scFv screen limited to the top seven pMHC predicted by SCRAP?What is the score of these pMHC by SCRAP?Fig2d: Explain what T and F are in the table.Graph: What is measured?is it scFv binding?What is the relationship between the graph and the table ?A more comprehensive off-target assessment would be good to provide.For example, is off-target reactivity at all altered across different HLA molecules?The authors may think that the specificity towards the off-targets is mostly driven by interactions of the PC-CAR with the peptide, but it would be nice to see data addressing that.
Fig3: the authors show staining of the PC-CAR with different pHLA tetramers/dextramers.A more comprehensive functional data across these different pHLAs would be interesting.For example, for the PHOX2B/HLA-A24:01, the authors show activity at low concentrations of peptide, so is there activity against HLA-C07:02-positive cells at higher concentrations (but still potentially physiological levels) of the cognate peptide?
Fig4 and FigS9: The killing data needs to be shown in a more consistent way.For example, for the same "killing" assay is shown as relative survival, relative viability, % viable, and relative confluence depending on the experiment.The % of caspase signal could be shown for all experiments.Controls are also missing, such as T cell death.The authors should also show data from multiple donors, and not just a single Incucyte image.
Fig4F: MFI of staining of pHLA is very low.Need to show these flow plots to be able to determine if the staining looks convincing.
Fig4g: The authors show that 302HL CAR T cells do not kill NBSD tumor cells in vivo.It is an interesting finding that should be further explored.Do 302HL CAR T cells recognize and kill PHOX2B/HLA-A23:02 expressing cells?The authors found a faster off rate for 302HL bound PHOX2B/HLA-A24:01but it does not appear associated with a different T cell functon(Fig2F).Can the authors comment on this finding.Error bars are cut.Tumor growth for each animal should also be shown.
FigS7: The authors are rightly concerned that the low MHCI expression in neuroblastoma cells may limit the activity of PC-CAR T cells.To evaluate the potential of their PC-CAR approach, they use an artificial system in which neuroblastoma cells overexpress a highly immunogenic viral antigen derived from Influenza are cultured with a T cell hybridoma.T cell killing is dependent on each peptide-MHCI/TCR or peptide-MHCI/PC-CAR pair.To me this experiment is low bar and does not warrant that T cells or PC-CAR T cells can kill neuroblastoma cells presenting target peptides in more physiological conditions.
FigS8B: What are the right and left panels?Did the authors do tetramer enrichment to increase the frequency of tumor antigen-specific T cells?Double tetramer staining to identify true T cells would also help as single tetramer stains are of often noisy.Fig8D:Why only a small fraction is tetramer stained?Are the high affinity TCRs determined by tetramer staining?FigS11A: How were the gates drawn and why does the lower one contains tumor peptide specific T cells?Reduced binding to both tetramers appears linear to me.S11c: The experimental setting needs clarification.How is affinity measured?What is the color code?What are the bars representing?Is there really a decrease in affinity?The background is also lower.

Author Rebuttals to Initial Comments:
Yarmarkovich, et al.

Referee #1
Is the immunopeptidome LC-MS/MS data deposited in PRIDE or MassIVE?What is the total number of unique samples (is it 8 total samples or were there replicates?) Can the authors provide a .rawfile mapping table that includes both the sample information (name, amount, antibody, HLA typing) that maps to each unique .rawfile generated along with the location where these data can be downloaded from?

The project DOI information is as is: DOI: 10.6019/PXD027182
Reviewer Username: reviewer_pxd027182@ebi.ac.uk

Reviewer Password: JHiSKklT
The spectra in Supp Fig. S2 is to low resolution to be able to read the fragment ion masses.Can this be replaced with a high resolution image?
Response.We now provide a higher resolution Fig. S2.
Related to the LC-MS/MS methods: What collision energies, max injection times, and ion targets were used in these methods?I was unable to find these parameters described in the methods.

Response: An AGC target of 1.5x10 5 and a max injection time of 50ms was used for MS1. An AGC target of 7x10 4 and a max injection time of 150ms was used for MS2. The collision energy for CID fragmentation was 35%. We have added these details to the methods section on page 30, lines 789-791.
Are there other HLA-I alleles that this peptide is predicted to bind to when using either NetMHC4.1pan(http://www.cbs.dtu.dk/services/NetMHCpan/) or HLAthena (http://hlathena.tools/)for predictions?
Response: We thank the reviewers for this suggestion and have used these searches to identify four additional common HLA allotypes predicted to bind QYNPIRTTF.We have added these to the legend of Fig. S14.We will incorporate these alleles in followup modeling and screening and plan to publish these in a future publication we are planning focused on further characterizing the cross-HLA binding of PC-CARs.We have attempted to run a comprehensive search using HLAthena but have experienced server issues when running many HLA allotypes.However, we have verified binding on a limited set of HLAs and include this in the legend of Fig. S14.
The observation that the target peptide QYNPIRTTF binds to B14:02 is really interesting given that its binding motif is divergent from the A24:02/A24:03---B15:01/B15:02 seems to have a more similar motif.I may have missed this in the manuscript, but did the authors confirm B14:02 binding with the canonical biochemical HLA-I binding assay (PMID: 23392640)?If not, that would be helpful to have to confirm this peptide binds using an orthogonal assay.

Response: Please note that we showed size exclusion chromatography for the three non-HLA-A*24:02 alleles studied here in Fig. S14 in the prior version (and maintained here) to provide this orthogonal validation but failed to point to this in the main text.
We added a sentence on page 12, line 370 to help the reader find these important data.The authors's database search methods appear to describe parameters for mass spectrometry data collected using high resolution MS1 and low resolution MS2.However, their methods describe that the MS2 data was collected in the orbitrap resulting in high resolution data.Therefore, the 0.2 Da mass tolerance for fragment ions is not appropriate for this data type.Instead, a 20 ppm tolerance is the standard (PMID: 30666599, PMID: 30666599, PMID: 27869121).If the authors did collect MS2 data in the orbitrap, and a 0.02 Da window was used, these data need to be reanalyzed with 20 ppm fragment ion mass tolerance for the fragment ions.
Response: We indeed used 0.02 Da mass tolerance as reviewer #1 suggests, not 0.2Da.We would like to note that Proteome Discoverer requires the definition of mass tolerance in Da not ppm.The search parameters used in this study are the standard for several labs (PMID: 31147937, 33858848, 33141586, 33901176).We have repeated the search using a more stringent 0.02 Da mass tolerance, 5 ppm, and report peptides at 1% FDR, as discussed above.
It is my opinion that a 5% FDR is too high when using high resolution MS2 data, and a more stringent FDR of <5%, more commonly 1% FDR may be more appropriate (PMID: 27869121).
Response: As described above, we have now repeated the analysis on the mass spectrometry data using a more stringent 1% FDR.We have adjusted the downstream analyses in both PDX and primary tumors using 1% FDR.Importantly, we find all of our prioritized tumor targets in the 1% FDR analysis.
If the authors continue to use the 5% FDR, this should be started both in Figure 1 and when they make claims around the number of unique peptides identified, as their total identifications will contain higher numbers of false positive.

We also note that during our reanalysis of the data using 1% FDR, we were also able to use a few peptide matching feature, that allowed us to match m/z and retention times for peaks on which MS/MS was not performed to fragmented peaks. This allowed us to infer that the PHOX2B peptide was also in other samples (with other HLA allotypes), including NBSD, which we had already show is killed by PHOX2B-directed PC-CARs. We now describe this analysis on page 12, lines 373-378 and have added a supplemental figure, S14f. We thank Reviewer 1 for prompting us to perform the reanalysis which led to this finding.
Most publications report HLA-II peptides as 12-25mers.In the Database search and spectral annotation section, the authors state they include peptides 8-25mer in length.Is this a typo?If not, can the author add a citation that shows 8mers are presented on HLA-II using an orthogonal technology such as x-ray structure?If there are peptides that are below 12mers in this dataset, it is possible these are HLA-I binding peptides that are contaminating the samples.Can the authors confirm that the HLA-II peptides they are reporting that are less than 12 amino acids in length do not match the HLA-I motifs of the alleles present in their patient samples to rule this out?
1) The authors use the CEF1 influenza peptide to evaluate the potential of a NB cell lines, with low MHC expression, to determine whether low level MHC expression is sufficient to elicit T cell activation.However, their data suggest that responsiveness is related to Myc ampification (low T cell response) rather than MHC expression per se (S7, d-e).While this is indicated in the figure legend, it is not addressed in the text.Two issues come up: -Flu peptide elicits strong T cell responses; would the authors expect that MHC expression would play a more critical role if a less immunogenic peptide was utilized?

Response: We constrained the scope of the influenza antigen work reported in this manuscript to showing that low MHC expression does not preclude T cell activation in neuroblastoma cells. We agree that the CEF1 peptide is highly immunogenic and chose this as a low bar to isolate the contribution of MHC expression to immunogenicity, rather than the contribution of immunogenic antigen, and establish proof-of-concept before initiating efforts to develop CAR receptors. We include it as supplementary data that was important for us prior to launching a major pMHC discovery effort focused on non-mutated "self" oncoproteins. We think that ability of PC-CARs to recognize nonimmunogenic PHOX2B peptide in low MHC tumors addresses Reviewer #2's second question.
-Given the significantly lower T cell responses in the Myc-amplified tumors, did Myc amplification alter killing by PC-CAR-T cells (i.e.differences between MNA as compared to MNN tumors)?The authors do not appear to directly address this in the text and it is an important point; it would be of interest to comment on pMHC-TCR-based activation as compared to peptide-PC-CAR-based activation.

Response: We agree that MYCN amplification is an important mediator or immune exclusion and now highlight that COG-564x is a MYCN amplified PDX model that shows complete responses to our PC-CAR. We think that reviewer 2 raises an important question that we will be very interested in addressing once we are able to generate paired TCR/PC-CAR targeting the same antigen, and perhaps bispecific T cell engagers. However, since we were unable to identify TCRs to the non-immunogenic PHOX2B peptide, we think that comparison of pMHC-PC-CAR for PHOX2B against pMHC-TCR cannot be compared on equal ground across different antigens as responses will be biased by antigen density and immunogenicity. As reviewers point out, CEF1 is highly immunogenic and not natively expressed in neuroblastoma cells, thus limiting the utility of comparing TCR activation via CEF1 to PC-CAR activation through PHOX2B. We have added a sentence with a reference to the discussion on page 16, lines 497-500, to highlight the potent impact of MYCN in this disease.
2) To specifically show the specificity of the PC-CAR for the PHOX2B peptide, the authors pulsed the HLA-A*23:01/PHOX2B-negative WM873 melanoma cell line with the PHOX2B peptide.Is there a reason why they didn't pulse and/or introduce PHOX2B into a negative NB line (SW620, HEPG2, and KATO III)?Additionally, the authors could evaluate their PC-CAR on the same NB line with absent PHOX2B expression (i.e.CRiSPR ko as presented in Figure 4).

Response: We did indeed show that introduction of PHOX2B (by both peptide pulsing and transduction of full-length PHOX2B mRNA) into HLA-matched, non-NB cells (initially SW620, but we have repeated the experiment using HEPG2 and KATO III as well) successfully induces killing through PC-CARs. Please see Figure 4D showing killing in SW620 following peptide pulsing with PHOX2B and Figure 4E showing killing of PHOX2B-transduced SW620 cells. We have also now repeated the killing assay using GFP depletion as a measure of killing, and have included the PHOX2B-negative lines in Figure 4b. While we agree that a CRISPR KO could potentially show abrogation of PC-CAR activity, neuroblastoma cells do not tolerate PHOX2B depletion (PHOX2B is the third-ranking dependency gene in NB), a primary reason that PHOX2B was selected as a target.
3) The data presented in Figure 4 are critical for evaluating the efficacy of the PC-CAR T cells against PHOX2B+ neuroblastoma.It appears that some mice did not reach endpoint and as such, it would be important to show these data (? Showing tumor curves for each mice).Otherwise, it is not clear from evaluating the figure that not all mice regressed, but rather that they are not included in the analysis because they were sacrificed.Additionally, the upregulation of MHC is a fascinating piece of data but it is not clearly explained.Was MHC upregulated in all NB lines (the NB line evaluated in Fig 4h is not indicated)?Does this MHC upregulation change with time as this was evaluated at day 11?The authors hypothesize that MHC is upregulated due to IFNg release?Interestingly though, the authors show that regions of tumor that stain for CD3 segregate from regions with PHOX2B+ tumor (Fig S16).Is MHC upregulation detected on PHOX2B+ tumor and/or is there a difference between MHC upregulation on tumor with/without loss of PHOX2B?This last point may be beyond the scope of this study but can be discussed.

Specific comments:
1.In Figure 2d, it is not easy to identify the cross-reactive peptides (colors are not labeled).There also appears to be a mislabeling with 2g (legend is in 2h) and the crystal structure is referred to as

Response: Thank you for pointing out a mismatch we had in versioning of Figure 2. We have clarified and added labeling and scores for cross-reactive peptides.
2. As an additional control in Figure 3C, 10LH PC-CART cells should be stained with cells in the absence or presence of PHOX2B peptide loading (i.e using the non-pulsed and pulsed WM873 melanoma line).S14e.

Response: We have performed peptide pulsing of WM873 with PHOX2B and mismatched peptide, followed by co-culture with 10LH PC-CAR T cells shown in Figure
3. In figure 4g, it would be important to assure that differences are not due to the characteristics of a specific donor T cell.How many donors were evaluated in these experiments (neither donors nor mice numbers are indicated).This is the case for several of the panels.
Response: We have performed in vitro studies using four different donors that show consistent activity (three donors were used in all cell lines as indicated in Figure 4 legend) and have now replaced the in vitro killing assay in Figure 4b to show replicates across three different donors.However, we selected a single donor to perform largescale T cell transductions to maintain consistency across in vivo studies in multiple lines.We note that the donor used for in vivo studies was included in the in vitro studies and performed nearly identically to other donors, thus we expect similar consistency in vivo.For practical reasons we used a single donor for these initial efficacy studies.Ongoing IND-enabling work will utilize a second donor, but these studies will take some time.

Referee #3
The authors filter their tumor antigen candidates against an existing database of the normal tissue immunopeptidome, but absence of detection in this type of database doesn't necessarily mean that the peptide is not presented in normal tissues.The authors should also consider single cell RNAseq databases as a filter.

Response: We will be incorporating single-cell RNA-seq into the next iteration of target discovery and have added a comment in the discussion on page 16, line 554. We also agree that lack of detection in normal tissue immunopeptidome does not necessitate the absence of presentation and have added a comment to make this clear on page 5, lines 121-122. Acknowledging the limitations of normal tissue immunopeptidomes, we find utility in employing these for 1) deprioritizing tumor antigens detection of peptides in normal tissue immunopeptidomes and 2) prioritizing the screening of antigens predicted to be cross-reactive by sCRAP for functional screening.
Fig2: Legends and graphs do not match up.

Response: Thank you for pointing out this mismatch. We have corrected the legend to match the figure.
Fig2b: Why is the scFv screen limited to the top seven pMHC predicted by SCRAP?What is the score of these pMHC by SCRAP?

Response: We have now added the scores generated by sCRAP for predicted crossreactive peptides (peptide score and overall score, with a brief description in the figure legend). The initial screen was done on 7 peptides across multiple CAR constructs (Fig 2b) and then expanded to a wider panel for prioritized CARs (Fig 2d). Additionally, screening of prioritized PC-CARs in multiple HLA-matched tissues shows a lack of cross-reactivity with the entire immunopeptidome presented across various cell types.
We are continuing to expand these screens in preclinical studies.
Fig2d: Explain what T and F are in the table.Graph: What is measured?is it scFv binding?What is the relationship between the graph and the table?

Response: We now expand figure legend 2 to clarify that T and F denote true or false as to whether peptide detected in normal tissue peptidome, the graph reports measured binding of 10LH to tetramer loaded with each peptide, and that the table is a quantification of graph.
A more comprehensive off-target assessment would be good to provide.For example, is offtarget reactivity at all altered across different HLA molecules?The authors may think that the specificity towards the off-targets is mostly driven by interactions of the PC-CAR with the peptide, but it would be nice to see data addressing that.
Response: While we agree that expanding cross-reactivity screening to additional HLA alleles would be ideal, we note that we used empty HLA-A*24:02 tetramers to generate pMHC for cross-reactive peptides and that these molecules are not available for HLA-A*23:01 or HLA-B*14:02.We will be generating these reagents moving forward and expanding the screening across HLA allotypes, but this will be a significant effort and we think it is beyond the scope of the current manuscript.
Fig3: the authors show staining of the PC-CAR with different pHLA tetramers/dextramers.A more comprehensive functional data across these different pHLAs would be interesting.For example, for the PHOX2B/HLA-A24:01, the authors show activity at low concentrations of peptide, so is there activity against HLA-C07:02-positive cells at higher concentrations (but still potentially physiological levels) of the cognate peptide?

Response: We are working on generating monoallelic HLA lines that will be necessary to test these interactions. However, based on the lack of cross-HLA activity of the 302LH PC-CAR (as compared to 10LH) due to lower binding affinity of PHOX2B on HLA-
A*23:01, we speculate that high peptide concentrations will not be enough to overcome the low binding affinity of 10LH to PHOX2B on HLA-C*07:02.We are performing a wider screen to identify PC-CARs that can potentially recognize different regions of the PHOX2B peptide as presented on HLA-C*07:02, and plan to report these in future publications.
Fig4 and FigS9: The killing data needs to be shown in a more consistent way.For example, for the same "killing" assay is shown as relative survival, relative viability, % viable, and relative confluence depending on the experiment.The % of caspase signal could be shown for all experiments.Controls are also missing, such as T cell death.The authors should also show data from multiple donors, and not just a single Incucyte image.

Response: We have explored multiple methods for reproducibly quantifying killing assays and have found that caspase signal is useful for quantifying cell killing for experiments within a cell line, but does not accurately represent cell killing across cell lines. We have found that measuring loss of GFP in transduced target cells using the Incucyte is the most indicative metric of cell killing across different cell lines. Since the GFP assay is only quantifying target cells, T cell death is not relevant in the GFP assays (T cell death is also accounted for in caspase assays using HLA-matched/PHOX2B -cell lines)
. Killing assays performed using A7 were done in 2018 before assays were optimized.Since we are not moving forward with the A7 construct and focus on 10LH and 302LH, we think it would not be helpful to repeat the assays with the older constructs.We have repeated and expanded the killing assay in Figure 4b and

now show biological replicates across three different donors, showing killing through depletion of GFP-transduced target cells. This is an orthogonal metric of cell killing to the Incucyte images showing target cell death by cleaved caspase. We have also added an image of T cell only cell death in Figure 4d.
Fig4F: MFI of staining of pHLA is very low.Need to show these flow plots to be able to determine if the staining looks convincing.

Response: The values reported were relative values, not MFIs. We have now replaced Fig 4F with the flow plots and have added a plot showing tetramerized 10LH binding to SW620 cells +/-PHOX2B in Fig. S15a.
Fig4g: The authors show that 302HL CAR T cells do not kill NBSD tumor cells in vivo.It is an interesting finding that should be further explored.Do 302HL CAR T cells recognize and kill PHOX2B/HLA-A23:02 expressing cells?The authors found a faster off rate for 302HL bound PHOX2B/HLA-A24:01but it does not appear associated with a different T cell functon(Fig2F).Can the authors comment on this finding.
Error bars are cut.Tumor growth for each animal should also be shown.
Response: The 302LH CARs do not kill PHOX2B/HLA-A*23:01 cells, which we expect is attributed to the lower binding affinity of between the PHOX2B/A23 complex with 302LH as compared 10LH (Figure S14c).We will be performing additional molecular studies, including the co-crystallization of the PHOX2B/A23 complex with 10LH as compared to 302LH to discern the structural basis for the difference in cross-reactivity.We speculate that the difference in binding affinity is partly due to the differential interface by which 10LH and 302LH recognize PHOX2B pMHC (see alanine scans below).We think that these findings are too speculative for the current publication and would like to pair these with pMHC/PC-CAR structural data in a future publication.Thank you for pointing out the missing error bars.We have corrected this and added data for individual animals in Figure S17a.

302LH alanine scan:
10LH alanine scan: FigS7: The authors are rightly concerned that the low MHCI expression in neuroblastoma cells may limit the activity of PC-CAR T cells.To evaluate the potential of their PC-CAR approach, they use an artificial system in which neuroblastoma cells overexpress a highly immunogenic viral antigen derived from Influenza are cultured with a T cell hybridoma.T cell killing is dependent on each peptide-MHCI/TCR or peptide-MHCI/PC-CAR pair.To me this experiment is low bar and does not warrant that T cells or PC-CAR T cells can kill neuroblastoma cells presenting target peptides in more physiological conditions.
Response: As discussed above, the flu experiment is a low bar to see if any T cell response can be elicited.PC-CAR recognition of non-immunogenic self-antigens in vitro and in vivo is a much higher bar as these are natively expressed peptides, expressed under physiological conditions.
FigS8B: What are the right and left panels?Did the authors do tetramer enrichment to increase the frequency of tumor antigen-specific T cells?Double tetramer staining to identify true T cells would also help as single tetramer stains are of often noisy.
Fig8D:Why only a small fraction is tetramer stained?Are the high affinity TCRs determined by tetramer staining?

Response: Shown in each panel are individual donors screened for quantity of antigenspecific T cells with unstained on the left for each donor and tetramer stained on the right. We have clarified this in the FigS8b legend and have annotated the samples in the figure. We have indeed performed magnetic bead enrichment of tetramer + cells and co-cultured these with antigen-pulsed DCs to select for functional cells; we plan to report this work in a separate publication describing TCR development for other pMHC antigens. We used double tetramer staining with two antigens on each HLA to remove noise and non-specific MHC binding (shown on diagonal). Fig S8d shows a small fraction of tetramer positive cells as these TCRs are poor binders to PHOX2B, hence the necessity to develop synthetic scFv-based receptors. All TCR binding affinity was determined by tetramer or dextramer staining. We have clarified the figures and legends.
FigS11A: How were the gates drawn and why does the lower one contains tumor peptide specific T cells?Reduced binding to both tetramers appears linear to me.

Reviewer Reports on the First Revision:
Referee #1 (Remarks to the Author): The authors have thoughtfully addressed almost all of my concerns.There still remains a concern around sharing the data on all data levels and specific pieces of metadata.Please see specific below.After the authors have provided the peptide identification lists in the form of a table and a .rawfile mapping annotation table that contains sample IDs and HLA typing specific to each cell line that was analyzed, then they will be in compliance with Nature's data sharing standards.
• Two tables, Supplemental Table 1 and Table S1 are described in this manuscript.Are these two tables the same?• I was not able to find a table that has all HLA-I identified peptides in Figure 1b.This type of table containing the 9,117 peptides, and each list from the filtering shown in this panel should be available with this manuscript.• I was not able to find a file mapping table that has a column that lists each .rawfile in the data repository deposit, the assigned sample, and that samples HLA type.This data metadata is critical to ensure reproducibility of results.
• In figure 1d, is the affinity reported measured or predicted?Can this be added to the figure legend?
Referee #2 (Remarks to the Author): The revision by Yarmakovich should be commended for their close attention to the reviewers' comments.
Referee #3 (Remarks to the Author): The authors' responses and revisions have overall satisfactorily addressed my comments.Although I could not find their added comments on page 5 and 16.I suspect I have one more request.Could the authors clarify their prioritization method: "we prioritized peptides based on pMHC binding affinity, HLA allele frequency, degree of differential expression of parent genes, relative abundance on MHC as compared to other peptides, recurrence across multiple neuroblastoma tumors, and relevance to neuroblastoma biology based on the published literature".Does the order of the criteria reflect the order of priority?How much weight each criteria has?
Fig 2b, but it is actually Fig 2h in the figure (but not the legend).
Fig 2b, but it is actually Fig 2h in the figure (but not the legend).Please check all the panels/legends here.

S11c:
The experimental setting needs clarification.How is affinity measured?What is the color code?What are the bars representing?Is there really a decrease in affinity?The background is also lower.Response: In Fig S11A we show where main populations should fall if a receptor are antigen-specific or cross-reactive with MHC.Since the A7 receptor preferentially binds to the PHOX2B on HLA-A24 over mismatched peptides on the same HLA, some of the cells fall into the tumor peptide specific gate, but the main population is cross-reactive.The shifts towards tumor specificity were calculated as follows: fold-change reported incorporates the shift towards single specificity relative to the WT receptor.The coloring represents the contribution to non-specific MHC binding (red being cross-reactive).We have added a description to the figure legend and annotation to the figure.The double mutant has significantly lower overall binding compared to WT.