Multi-step screening of neoantigens' HLA- and TCR-interfaces improves prediction of survival

Abstract

Improvement of risk stratification through prognostic biomarkers may enhance the personalization of cancer patient monitoring and treatment. We used Ancer, an immunoinformatic CD8, CD4, and regulatory T cell neoepitope screening system, to perform an advanced neoantigen analysis of genomic data derived from the urothelial cancer cohort of The Cancer Genome Atlas. Ancer demonstrated improved prognostic stratification and five-year survival prediction compared to standard analyses using tumor mutational burden or neoepitope identification using NetMHCpan and NetMHCIIpan. The superiority of Ancer, shown in both univariate and multivariate survival analyses, is attributed to the removal of neoepitopes that do not contribute to tumor immunogenicity based on their homology with self-epitopes. This analysis suggests that the presence of a higher number of unique, non-self CD8- and CD4-neoepitopes contributes to cancer survival, and that prospectively defining these neoepitopes using Ancer is a novel prognostic or predictive biomarker.

Figure 1

BLCA mutanome analysis workflow. Mutations were retrieved for each patient sample and evaluated using three analysis workflows and then compared for overall survival and disease-free survival predictive accuracy. The three types of analyses were defined as follows: (A) “TMB analysis”: tumor mutational burden is evaluated from the count of mutations present in each tumor. (B) “NetMHCpan analysis”: mutation-bearing HLA class I and HLA class II ligands are identified with NetMHCpan 4.0 and NetMHCIIpan 3.1, respectively. This approach is similar to the one employed by the TCGA Research Network in their analysis of the BLCA cohort. (C) “Ancer analysis”: mutation-bearing HLA class I and HLA class II ligands are identified with EpiMatrix, compared to matched normal sequences to identify Ancer-defined neoepitopes, and filtered with JanusMatrix to remove neoepitopes homologous to self.

Figure 2

Association between mutational and neoepitope landscapes. BLCA mutanomes were analyzed with Ancer to determine counts of Ancer-defined HLA class I neoepitopes (a), Ancer-defined HLA class II neoepitopes (b), and Ancer-defined neoantigen candidates that could be used for precision immunotherapy purposes (c). Numbers of HLA class I neoepitopes, HLA class II neoepitopes, and neoantigen candidates observed in each patient is strongly correlated with observed tumor mutational burden (TMB). Each dot corresponds to one TGCA BLCA patient.

Figure 3

Stratification of cancer patients according to TMB analysis, NetMHCpan analysis, and the Ancer pipeline. TCGA bladder cancer patients were separated based on their median TMB (a,b), NetMHCpan neoepitope burden (c,d), or Ancer pipeline-defined neoepitope burden (e,f). Median disease-free survival (DFS) (a,c,e) and median overall survival (OS) (b,d,f) were evaluated with the Kaplan–Meier estimator. Double-ended arrows define the differences in median survival times between the groups for each of the analyses. Statistical significance was evaluated using the log-rank test. The largest differential in median overall survival, 70 months, was obtained with the Ancer pipeline and was more than double the difference in median overall survival observed when stratifying patients using NetMHCpan (34 months).

Figure 4

Univariate survival analysis forest plots. Univariate survival analyses were conducted separately for the TMB, NetMHCpan, and Ancer analyses while considering either disease free survival (DFS) (a) or overall survival (OS) (b). BLCA patients were separated based on their median TMB or neoepitope burdens. Association with DFS and OS is improved with Ancer compared to the other analysis. Hazard ratios (HR), confidence intervals (CI) and p-values (p) were calculated using univariate Cox proportional-hazard models. Overall log-rank p-value is provided for each model on the left.

Figure 5

Effect of Ancer's neoepitope and homology filters on survival analyses. Ancer neoepitope and homology filters improve association with disease free survival (DFS) (a) and overall survival (OS) (b). Hazard ratios (HR), confidence intervals (CI) and p-values (p) were calculated using univariate Cox proportional-hazard models. BCLA patients were separated based on their Ancer neoepitope burden (Ancer, all steps), similarly to Fig. 3, "non-matching" Ancer neoepitope burden without considering the JanusMatrix filter (Ancer, steps 1–2), or based on their "raw" Ancer neoepitope burden (Ancer, step 1 only).

Figure 6

Prediction of bladder cancer patient five-year survival rate. Ancer pipeline analysis improves prediction of five-year survival compared to TMB- or NetMHCpan-based predictors. BLCA patients were predicted to survive more or less than five years based on their mutational or neoepitope burdens. Predicted survival status was compared to observed overall survival. (a) Accuracy of the five-year survival predictions. *p-value < 0.05, McNemar's test. (b) Positive (PPV) and negative predictive values (NPV) obtained for the TMB, NetMHCpan, and Ancer predictors. (c) PPVs and NPVs obtained using truncated and full versions of the Ancer pipeline. The increased NPV as compared to NetMHCpan and TMB suggests that an analysis that identifies tolerated or tolerogenic epitopes (Ancer) may be better suited to identify patients at a greater risk of earlier mortality (~ 9 out of 10 correct predictions) than the other types of analyses.

Figure 7

Multivariate survival analysis forest plots. Ancer neoepitope burden remains a significant co-factor associated with overall survival when adjusting for TMB, age, and disease stage (a). NetMHCpan neoepitope burden's association with overall survival is lost when adjusting for TMB, age, and disease stage (b). Hazard ratios (HR), confidence intervals (CI) and p-values (p) were calculated using multivariate Cox proportional-hazard models. Ancer neoepitope burden remained a significant cofactor associated with overall survival once adjusted for TMB, age, and disease stage.