Comparative performance analysis of neoepitope prediction algorithms in head and neck cancer

Abstract

Background: Mutations in cancer cells can result in the production of neoepitopes that can be recognized by T cells and trigger an immune response. A reliable pipeline to identify such immunogenic neoepitopes for a given tumor would be beneficial for the design of cancer immunotherapies. Current methods, such as the pipeline proposed by the Tumor Neoantigen Selection Alliance (TESLA), aim to select short peptides with the highest likelihood to be MHC-I restricted minimal epitopes. Typically, only a small percentage of these predicted epitopes are recognized by T cells when tested experimentally. This is particularly problematic as the limited amount of sample available from patients that are acutely sick restricts the number of peptides that can be tested in practice. This led our group to develop an in-house pipeline termed Identify-Prioritize-Validate (IPV) that identifies long peptides that cover both CD4 and CD8 epitopes.

Methods: Here, we systematically compared how IPV performs compared to the TESLA pipeline. Patient peripheral blood mononuclear cells were cultured in vitro with their corresponding candidate peptides, and immune recognition was measured using cytokine-secretion assays.

Results: The IPV pipeline consistently outperformed the TESLA pipeline in predicting neoepitopes that elicited an immune response in our assay. This was primarily due to the inclusion of longer peptides in IPV compared to TESLA.

Conclusions: Our work underscores the improved predictive ability of IPV in comparison to TESLA in this assay system and highlights the need to clearly define which experimental metrics are used to evaluate bioinformatic epitope predictions.

Keywords: bioinformatics; cancer; immunogenicity; neoepitope prediction; neoepitope screening.

Figure 1

Overview of project workflow. Schematic outlining the workflow of the project. Whole blood and tissue were isolated from each patient. Blood samples were used for whole exome sequencing and tumor samples were used for exome and RNA sequencing. IPV was used to identify highly expressed, tumor-specific mutations from this sequencing data. The prioritized mutations were then used to generate neoepitope candidates either using IPV or TESLA ranking metrics. Finally, epitope candidates were cultured in vitro with patient PBMC and an IFNg and IL-5 fluorospot was used to assess reactivity.

Figure 2

Systematic comparison of the immunogenicity of IPV and TESLA peptides. (A) The magnitude of the IFNg response from patient PBMC (one dot represents one patient) to neoepitope pools containing the top 10 TESLA and IPV peptides. The magnitude is represented as the number of spot-forming cells (SFC) per 10⁶ PBMC. A threshold of 100 SFC (dotted line) is used to distinguish between positive and negative responses. (B) IL-5 fluorospot data represented as SFC per 10⁶ PBMC. The same threshold as above is applied for the IL-5 data. (C) IFNg and IL-5 fluorospot-detected responses to IPV peptides from each patient within the cohort. IFNg responses are displayed as black bars and IL5 responses as grey bars.

Figure 3

Peptide length plays an important role in peptide in vitro immunogenicity. (A) The magnitude of the IFNg and IL5 responses from patient PBMC to neoepitope pools containing the top 10 IPV short and TESLA long peptides. The magnitude is represented as the number of spot-forming cells (SFC) per 10⁶ PBMC. A threshold of 100 SFC (dotted line) is used to distinguish between positive and negative responses. (B) The magnitude of IFNg and IL5 responses from patients 11057 and 11095 PBMC in response to stimulation with the TESLA long and IPV long pools. (C) IFNg and IL5 fluorospot data from short and long neoepitope candidate pool stimulation. IFNg responses are displayed as black bars and IL5 responses as grey bars. (D) 2x2 contingency table comparing the performance of short and long peptides for IFNg. (E) 2x2 contingency table comparing the performance of short and long peptides for IL5.

Figure 4

IPV long peptide pool deconvolution identifies immunogenic variants. (A) List of peptides generated from each somatic variant for patients 11043, 11095, and 11098. (B) Positive responses from the IPV peptide pool for patients 11043, 11095, and 11098 were deconvoluted to identify the variants contributing to IFNg and IL-5 signals. Magnitude of responses are displayed as SFC per 10⁶ PBMC for both IFNg and IL-5.

Figure 5

Application of IPV to an additional cohort of patients identifies additional immunogenic variants. (A) IFNg and IL-5 ELISpot was used to detect responses to IPV peptides in a second cohort of HNSCC patients (n=6) (one dot represents one patient). (B) List of peptides from each somatic variant used to deconvolute the second cohort responses. (C) Deconvolution data for patients 10193, 10197, 10198, and 10203.

Figure 6

Ranking metrics and epitope immunogenicity. The percentile ranks for TPM, DNA VAF, and RNA VAF values for all peptide candidates were calculated on a per donor basis. The percentile ranks for the ranking metrics employed by IPV to rank positive epitopes higher than negative epitopes was assessed. The TPM (A), DNA VAF (B), RNA VAF (C) and sum of all metrics (D) for all the variants tested in this study are plotted. One dot represents one variant. ‘*’ indicates a p-value less than 0.05. ‘ns’ indicates non-significant p-values.