The T Cell Receptor β Chain Repertoire of Tumor Infiltrating Lymphocytes Improves Neoantigen Prediction and Prioritization

Reviewer #1 (Public Review):

Summary:

This paper reports a number of somewhat disparate findings on a set of colorectal tumour and infiltrating T-cells. The main finding is a combined machine-learning tool which combines two previous state-of-the-art tools, MHC prediction, and T-cell binding prediction to predict immunogenicity. This is then applied to a small set of neoantigens and there is a small-scale validation of the prediciton at the end.

Strengths:

The prediction of immunogenic neoepitopes is an important and unresolved question.

Weaknesses:

The paper contains a lot of extraneous material not relevant to the main claim. Conversely, it lacks important detail on the major claim.

(1) The analysis of T cell repertoire in Figure 2 seems irrelevant to the rest of the paper. As far as I could ascertain, this data is not used further.

(2) The key claim of the paper rests on the performance of the ML algorithm combining NETMHC and pmtNET. In turn, this depends on the selection of peptides for training. I am unclear about how the negative peptides were selected. Are they peptides from the same databases as immunogenic petpides but randomised for MHC ? It seems as though there will be a lot of overlap between the peptides used for testing the combined algorithm, and the peptides used for training MHCNet and pmtMHC. If this is so, and depending on the choice of negative peptides, it is surely expected that the tools perform better on immunogenic than on non-immunogenic peptides in Figure 3. I don't fully understand panel G, but there seems very little difference between the TCR ranking and the combined. Why does including the TCR ranking have such a deleterious effect on sensitivity?

(3) The key validation of the model is Figure 5. In 4 patients, the authors report that 6 out 21 neo-antigen peptides give interferon responses > 2 fold above background. Using NETMHC alone (I presume the tool was used to rank peptides according to bding to the respecitve HLAs in each individual, but this is not clear), identified 2; using the combined tool identified 4. I don't think this is significant by any measure. I don't understand the score shown in panel E but I don't think it alters the underlying statistic.

In conclusion, the paper demonstrates that combining MHCNET and pmtMHC results in a modest increase in the ability to discriminate 'immunogenic' from 'non-immunogenic' peptide; however, the strength of this claim is difficult to evaluate without more knowledge about the negative peptides. The experimental validation of this approach in the context of CRC is not convincing.

Reviewer #2 (Public Review):

Summary:

This paper introduces a novel approach for improving personalized cancer immunotherapy by integrating TCR profiling with traditional pHLA binding predictions, addressing the need for more precise neoantigen CRC patients. By analyzing TCR repertoires from tumor-infiltrating lymphocytes and applying machine learning algorithms, the authors developed a predictive model that outperforms conventional methods in specificity and sensitivity. The validation of the model through ELISpot assays confirmed its potential in identifying more effective neoantigens, highlighting the significance of combining TCR and pHLA data for advancing personalized immunotherapy strategies.

Strengths:

(1) Comprehensive Patient Data Collection: The study meticulously collected and analyzed clinical data from 27 CRC patients, ensuring a robust foundation for research findings. The detailed documentation of patient demographics, cancer stages, and pathology information enhances the study's credibility and potential applicability to broader patient populations.

(2) The use of machine learning classifiers (RF, LR, XGB) and the combination of pHLA and pHLA-TCR binding predictions significantly enhance the model's accuracy in identifying immunogenic neoantigens, as evidenced by the high AUC values and improved sensitivity, NPV, and PPV.

(3) The use of experimental validation through ELISpot assays adds a practical dimension to the study, confirming the computational predictions with actual immune responses. The calculation of ranking coverage scores and the comparative analysis between the combined model and the conventional NetMHCpan method demonstrate the superior performance of the combined approach in accurately ranking immunogenic neoantigens.

(4) The use of experimental validation through ELISpot assays adds a practical dimension to the study, confirming the computational predictions with actual immune responses.

Weaknesses:

(1) While multiple advanced tools and algorithms are used, the study could benefit from a more detailed explanation of the rationale behind algorithm choice and parameter settings, ensuring reproducibility and transparency.

(2) While pHLA-TCR binding displayed higher specificity, its lower sensitivity compared to pHLA binding suggests a trade-off between the two measures. Optimizing the balance between sensitivity and specificity could be crucial for the practical application of these predictions in clinical settings.

(3) The experimental validation was performed on a limited number of patients (four), which might affect the generalizability of the findings. Increasing the number of patients for validation could provide a more comprehensive assessment of the model's performance

Reviewer #3 (Public Review):

Summary:

This study presents a new approach of combining two measurements (pHLA binding and pHLA-TCR binding) in order to refine predictions of which patient mutations are likely presented to and recognized by the immune system. Improving such predictions would play an important role in making personalized anti-cancer vaccinations more effective.

Strengths:

The study combines data from pre-existing tools pVACseq and pMTNet and applies them to a CRC patient population, which the authors show may improve the chance of identifying immunogenic, cancer-derived neoepitopes. Making the datasets collected publicly available would expand beyond the current datasets that typically describe caucasian patients.

Weaknesses:

It is unclear whether the pNetMHCpan and pMTNet tools used by the authors are entirely independent, as they appear to have been trained on overlapping datasets, which may explain their similar scores. The pHLA-TCR score seems to be driving the effects, but this not discussed in detail.

Due to sample constraints, the authors were only able to do a limited amount of experimental validation to support their model; this raises questions as to how generalisable the presented results are. It would be desirable to use statistical thresholds to justify cutoffs in ELISPOT data.

Some of the TCR repertoire metrics presented in Figure 2 are incorrectly described as independent variables and do not meaningfully contribute to the paper. The TCR repertoires may have benefitted from deeper sequencing coverage, as many TCRs appear to be supported only by a single read.