Patient characteristics
In this study, we included a total of 637 patients with stage III-IV cutaneous malignant melanoma, consisting 192 cases in the training cohort (TCGA), 227 patients in the validation cohort (GEO1) and 218 cases in the anti-PD1 treatment group (GEO2). The clinicopathologic information of patients in the training and validation cohort was presented in Table 1. Based on these two groups of patients, we established an immune score for predicting patient survival. Detailed clinical characteristics of cases receiving anti-PD1 treatment were given in Table S1.
Table 1
Baseline patient characteristics
Variable
|
No. of patients or mean ± SD
|
Age, years
|
59.7 ± 15.6
|
Patients
|
|
Training
|
227 (54.2%)
|
Validation
|
192 (45.8%)
|
Sex
|
|
Female
|
163 (38.9%)
|
Male
|
255 (60.9%)
|
Race
|
|
White
|
179 (42.7%)
|
Asian
|
10 (2.4%)
|
Tumor location
|
|
Extremities
|
97 (23.2%)
|
Trunk
|
96 (22.9%)
|
Head or neck
|
24 (5.7%)
|
Other
|
22 (5.3%)
|
AJCC-TNM stage
|
|
III
|
288 (68.7%)
|
IV
|
131 (31.3%)
|
Clark classification
|
|
1
|
3 (0.7%)
|
2
|
4 (1.0%)
|
3
|
25 (6.0%)
|
4
|
91 (21.7%)
|
5
|
38 (9.1%)
|
Breslow depth, mm
|
6.6 ± 9.6
|
Immune score
|
-0.8 ± 0.4
|
Immune type
|
|
Immunoscore low
|
228 (54.4%)
|
Immunoscore high
|
191 (45.6%)
|
TCGA, The Cancer Genome Atlas; GEO, Gene Expression Omnibus; SD, standard deviation; AJCC, American Joint Committee on Cancer
|
Construction Of The Prognostic Immune Score
As shown in Fig. 2A, M0 macrophages, CD8 + T cells and M2 macrophages were the three most abundant immune cell types. The Pearson correlations between the row and column corresponding immune cell types were investigated in the training cohort (Fig. 2B). Significant correlations were observed between some immune subsets, such as M0 macrophages and CD8 + T cells. Given the obvious colinearity among immune cell types, the Lasso Cox regression was applied to identify the prognostic correlation immune cells, and build an immunoscore model in the training cohort (Fig. 2C-D). The immune score for predicting patient overall survival (OS) is as follows: -0.349 × B cells naive (≤ 0.01) + 0.026 × T cells CD8 (≤ 0.23) + 0.266 × T cells CD4 Memory activated (≤ 0) + 0.073 × NK cells activated (≤ 0.04) − 0.025 × monocytes (≤ 0.03) − 0.378 × macrophages M0 (≤ 0.45) + 0.245 × macrophages M1 (≤ 0.09) − 0.322 × dentritic cells resting (≤ 0.03) − 0.525 × neutrophils (≤ 0). In this formula, a value (coefficients shown in the formula) was only assigned when the fraction of one cell type was less than the corresponding cut-off value in the brackets. The cut-off values of all immune cell types were shown in Table S2.
The performance of immune score in the training and validation groups
Firstly, the predictive accuracy of the immunoscore was evaluated by the ROC curves at the time points 1, 3 and 5 years (assessed as a continuous variable). As presented in Fig. 3A (training cohort) and 3B (validation cohort), the immunoscore displayed acceptable predictive ability, especially in the validation cohort. In addition, with the cut-off value of -0.69, we divided patients into immunoscore-low (IS-low) and immunoscore-high (IS-high) groups (immunotype). In all of the training (Fig. 3C), validation (Fig. 3D) and total (Fig. 3E) cohorts, patients in the IS-low group showed significantly better OS than those in the IS-high group (all P values < 0.05). The restricted cubic spline of immune score in the training dataset showed that the immunoscore presented a linear profile (Fig. 3F).
The univariate analysis of the immunoscore in the training, validation and total cohorts showed that both of the immunoscore and immunotype were significantly associated with overall survival in all three cohorts (Table S3). In addition, Table S3 also showed the results of univariate analyses of clinicopathologic variables and all 22 types of immune cells. Besides, to further confirm the relationship between immunoscore (immunotype) and patient survival, we performed subgroup analyses based on available clinicopathologic features in the total cohort. The results demonstrated that the immunoscore and immunotype were associated with patient survival in most of the subgroups (Table S4). Based on the best optimal cut-off value of the immunoscore (-0.69), we divided patients into IS-low and IS-high groups. In the training cohort, multivariable analysis revealed that the immunotype was an independent prognostic factor for overall survival in stage III-IV melanoma (HR = 3.25, 95% CI 2.11-5.00, P < 0.001) (Table S5).
Association of immune score with clinical parameters, molecules, and biological pathways
In the training cohort, the prognostic immunoscore for stage III-IV melanoma was found to be negatively correlated with gene expression of certain immune checkpoint regulators (CD274, CTLA4, HAVCR2, LAG3, PDCD1, TIGIT, TNFRSF4 and CD47) and effector molecules of cytotoxic activity (GZMB and IFNG) (Fig. 4A-J). In addition, we performed GSEA analysis to explore the biological functions of the immunoscore. The results revealed that genes highly expressed in the IS-low group showed significant enrichment in multiple immune biological processes such as antigen processing and presentation, toll like receptor signaling pathway and natural killer cell mediated cytotoxicity (Fig. 4K). Finally, in the training cohort, we analyzed the correlation between immunoscore and clinical characteristics. As shown in Fig. 4L-O, we found that the female patient, the White, melanoma without ulceration and non-primary metastatic tumor tended to have lower immunoscore.
Nomogram Based On Immune Score And Clinical Features
To provide a quantitative instrument to predict the probability of OS, a nomogram integrating both of the immunotype and clinicopathologic indicators was established based on cases from the training cohort. By using multivariate analysis, we identified four prognosis-relevant factors including age, race, AJCC-TNM stage and immunotype (Table S5). In the nomogram, each value or subtype of these factors was assigned a score on the point scale. After calculating the total score and locating this value on the total point scale, we could determine the estimated probability of the 2-, 3- and 5-year survival probability (Fig. 5A; Table S6). The 2-, 3- and 5-year ROC curves were plotted to reflect prognostic accuracy the nomogram (Fig. 5B). We found that, compared to TNM stage alone, our nomogram could predict patient survival more accurately (Fig. S1). In addition, the 2-, 3- and 5-year calibration curves depicted in Fig. 5C-E demonstrated that the nomogram performed well when compared to the performance of an ideal model in the training cohort. The decision curve analysis (Fig. 5F-H) also showed that using the nomogram could add benefit to the prediction of survival of melanoma patient.
We then validated the discrimination, calibration and usability of the nomogram in the GEO1 cohort. The 2-, 3- and 5-year AUC curves, calibration curves and DCA curves were shown in Fig. S2.
Establishment of an immune score associated with response of anti-PD1 treatment
Finally, in the GEO2 cohort, the correlation analysis showed that the prognosis-relevant immunoscore could not accurately predict the anti-PD1 treatment responses (data not shown). After k-means clustering of the 22 types of immune subsets, patients were divided into two groups with different immune profiles (FIG. S3A-B). However, the two clusters of patients also did not show significant distinctions in response rate of anti-PD1 treatment in stage III-IV melanoma (FIG. S3C). The above results illustrated that neither the prognosis-relevant immunoscore nor the whole immune landscape diversity could accurately predict the response status to anti-PD1 therapy.
Based on the above results, we next explored whether some specific immune subsets were associated with the response to anti-PD1 therapy. We used the Lasso logistic regression to identify immune subsets for predicting the response to anti-PD1 therapy (FIG. S3D-E). Finally, several types of immune cells were identified, and an immune score was constructed based on these immune cell subsets. The immune score predicting anti-PD1 therapy response is as follows: -2.665 × B cells naïve + 2.450 × plasma cells − 10.711 × T cells follicular helper − 5.721 × T cells regulatory + 1.022 × T cells gamma delta − 0.824 × monocytes + 1.435 × macrophages M0–10.820 × dentritic cells activated + 2.728 × mast cells resting. The AUC for the immunoscore is 0.78 (FIG. S3F).