Tumor neoantigenicity assessment with CSiN score incorporates clonality and immunogenicity to predict immunotherapy outcomes

Abstract

Lack of responsiveness to checkpoint inhibitors is a central problem in the modern era of cancer immunotherapy. Tumor neoantigens are critical targets of the host antitumor immune response, and their presence correlates with the efficacy of immunotherapy treatment. Many studies involving assessment of tumor neoantigens principally focus on total neoantigen load, which simplistically treats all neoantigens equally. Neoantigen load has been linked with treatment response and prognosis in some studies but not others. We developed a Cauchy-Schwarz index of Neoantigens (CSiN) score to better account for the degree of concentration of immunogenic neoantigens in truncal mutations. Unlike total neoantigen load determinations, CSiN incorporates the effect of both clonality and MHC binding affinity of neoantigens when characterizing tumor neoantigen profiles. By analyzing the clinical responses in 501 treated patients with cancer (with most receiving checkpoint inhibitors) and the overall survival of 1978 patients with cancer at baseline, we showed that CSiN scores predict treatment response to checkpoint inhibitors and prognosis in patients with melanoma, lung cancer, and kidney cancer. CSiN score substantially outperformed prior genetics-based prediction methods of responsiveness and fills an important gap in research involving assessment of tumor neoantigen burden.

Fig. 1

Motivation for CSiN score. (A) Illustration showing the motivation of examining pairings of neoantigens and the tumor mutations with which they are associated. We demonstrated two hypothetical patients, one with an unfavorable distribution of neoantigens and the other with a favorable distribution. The actual mutations and neoantigens shown are based on real data. The outermost circle indicates the whole tumor. Each circle indicates a population of tumor cells with certain mutations. Each different color indicates a distinct mutation, and the area of each circle indicates the proportion of cells having the mutation. For the formula, on the left of each multiplication sign “x”, is the normalized VAF, and on the right of each “x” is the normalized per-mutation neoantigen load. The colorings in the formula correspond to the tumor mutations shown above with the same colorings. The two bigger tables on the right show the neoantigen sequences, registers (“Dist”), and the HLA alleles for each neoantigen. For neoantigens of missense mutations, “Dist” refers to the distance between the altered amino acid and the left end of neoantigen; for neoantigens of insertions/deletions and stoploss mutations, “Dist” refers to the distance between the left end of the mutation and the left end of neoantigen. The “+” sign indicates the left end of neoantigen is on the right of the altered position and vice versa. (B) The distribution of the CSiN scores in the RCC, LUAD, LUSC, and SKCM cohorts. T-tests were used for comparison of CSiN scores between different subtypes of the same tumor cohort. (C) A scatterplot showing the relationship between CSiN and the expression level of the IFN-γ signature in the RCC cohort. Spearman correlation between them is shown. (D) Heatmaps of the pairwise Spearman correlations of the CSiN, mutation load, neoantigen load, and the transcriptomics-based features are shown for the RCC, LUAD, LUSC and SKCM cohorts, which are calculated as in (C).

Fig. 2

Association of CSiN score with checkpoint inhibitor treatment response. (A) The VanAllen cohort. 11 patients with clinical benefit (response group), 6 patients with long-term survival with no clinical benefit (long-survival) group, and 20 patients with minimal or no clinical benefit (nonresponse) group. (B) The Snyder cohort. 27 patients with DCB, and 34 patients with NDB. (C) The Riaz cohort. 3 patients with complete response (CR), 12 patients with partial response (PR), 23 patients with stable disease (SD), and 27 patients with progressive disease (PD). (D) The Hugo cohort. 3 patients with complete response, 10 patients with partial response, and 13 patients with progressive disease. (E) The Miao cohort. 12 patients with clinical benefit, 8 patients with intermediate benefit, and 13 without clinical benefit. (F) The IMmotion150 cohort. There were 8 patients with CR, 15 patients with PR, 16 patients with SD, and 16 patients with PD. These patients were treated with atezolizumab and possess high T_eff signature expression. (G) The Hellmann cohort. There were 23 PD-L1⁺ (IHC>=3) patients with DCB, and 16 PD-L1⁺ patients with NDB. (H) The Acquired cohort. There were 8 patients with short term progression (progression<12 month) and 6 patients with sustained response (progression>12 month). (I) The Rizvi cohort. 11 patients with DCB and 15 patients with NCB. Biopsy and genomics data were obtained close to time of progression for all patients, while baseline biopsies were lacking for many patients. For (A)-(I), we tested the association of the dichotomized CSiN scores with the ordered response categories using an ordinal Chi-Square test. (J) Boxplots of bootstrap P values evaluating the robustness of the predictive performance of CSiN, neoantigen load and the neoantigen fitness score, with each P value generated from a bootstrap resample of each cohort. Two-sided Wilcoxon signed-rank test was used to compare the bootstrap P values. NS: P>0.01, *: P=0.01–0.05, **: P=0.001–0.01, ***: P=0.0001–0.001, ****:P<0.0001.

Fig. 3

Association of CSiN score with overall survival of patients. (A-E) Kaplan-Meier estimator was used to visualize patient overall survival. P values for logrank tests are shown. (A) The RCC cohort. (B) The LUAD cohort. (C) The LUSC cohort. (D) The SKCM cohort. (E) The patients identified as having “High T cells” are extracted from each cohort, combined, and tested together. The high and low CSiN score designations follow those in (A-D). The top 140 RCC patients, 100 LUAD patients, 100 SKCM patients, and 40 LUSC patients with the highest T cell infiltration were designated as having “High T cells”, so the more immunogenic tumor types have more patients selected (83). (F) Forest plot for the coefficients of the multivariate CoxPH analysis of the combined cohort in (D). Disease type, pathological stage, gender, age and the binarized CSiN were included as covariates. The dotted line shows the no effect point. 95% Confidence intervals were shown as bars. (G) Boxplots of bootstrap P values evaluating the robustness of the prognostic performance of CSiN, neoantigen load and the neoantigen fitness score, with each P value generated from a bootstrap resample of each cohort. Two-sided Wilcoxon signed-rank test was used to compare the bootstrap P values. *: P=0.01–0.05, **: P=0.001–0.01, ***: P=0.0001–0.001, ****:P<0.0001.