Melanoma Checkpoint Signatures Split Tumors Into Three Survival Classes

Melanoma tumors with high coordinated expression of 47 immune checkpoint genes carried a 66 percent lower risk of death than low-expression tumors in a re-analysis of 472 cutaneous melanomas from The Cancer Genome Atlas. Median overall survival separated by more than threefold across the resulting phenotypes: 4,930 days for the high-expression group against 1,487 days for the low group. The study, published in Frontiers in Genetics by a team led by Mao, Yang and Gong at hospitals in Sichuan, Shandong and Jiangsu, is a retrospective re-analysis — an association, not a predictive test.

How were the three phenotypes built?

The researchers standardized each of the 47 checkpoint-related genes across tumors and applied unsupervised consensus clustering (a method that groups samples by expression similarity without predefined labels) over cluster counts from two to six. Stability gains peaked at three clusters. The final groups, ordered by mean checkpoint z-score, split the cohort into 136 ICG-high tumors (28.8 percent), 203 ICG-mid (43.0 percent) and 133 ICG-low (28.2 percent). Mean consensus within phenotypes was 0.958, against 0.020 between them — the groups were internally coherent and cleanly separated.

The checkpoint score tracked immune infiltration closely. It correlated with a CD8 T-cell signature at a Spearman coefficient of 0.828 (P = 5.06 × 10⁻¹²⁰). Mean CD8 T-cell scores fell stepwise across phenotypes, from 0.211 in ICG-high to −0.134 in ICG-low, with the strongest separation of any measure tested (Kruskal–Wallis P = 3.95 × 10⁻⁷⁰). Tumors with high checkpoint expression were therefore not merely immunosuppressed; they also showed the strongest immune-cell infiltration, a pattern consistent with adaptive immune resistance, in which an active antitumor response induces its own regulatory brakes.

What does survival show?

After selecting one tumor per patient, 454 of 469 unique patients had usable follow-up. Multivariable Cox models adjusting for AJCC stage (the American Joint Committee on Cancer staging system), age and sex covered 404 complete cases with 190 deaths. ICG-high versus ICG-low tumors carried a hazard ratio of 0.34 (95 percent CI 0.22–0.51; P = 2.68 × 10⁻⁷); the mid group also outlived the low group (hazard ratio 0.57; 95 percent CI 0.41–0.80). A stage-stratified sensitivity model returned a nearly identical 0.35. The authors frame this strictly as prognostic association, not evidence that checkpoint expression predicts benefit from blockade.

Mutation data added texture and caution in equal measure. Annotations existed for only 66 tumors, four of them ICG-high. Within that subset, BRAF mutation rates ran 50.0, 55.6 and 42.3 percent across high, mid and low phenotypes; median total mutation counts trended higher in ICG-high (581.5 versus 283.0 and 235.0) without significance.

What happens during treatment?

The nivolumab cohort GSE91061 contributed 109 RNA-sequencing samples from 65 patients, including 43 with paired pretreatment and on-treatment biopsies. The checkpoint score itself rose during treatment, with a median increase of 0.115 (Wilcoxon signed-rank P = 0.012). CD8 T-cell scores climbed from 0.142 to 0.209, interferon-gamma scores from 0.304 to 0.358, and exhaustion scores from 0.047 to 0.115 — all significant. Effector and exhaustion signals rose in parallel, suggesting nivolumab triggers immune activation accompanied by regulatory feedback rather than a simple unleashing of antitumor activity.

The single-cell cohort GSE120575 supplied all 16,291 CD45-positive immune cells from 48 biopsies of 32 patients treated with anti-PD-1, anti-CTLA4 or combination therapy: 7,242 general T cells, 4,403 CD8 T cells, 2,438 B cells, 1,182 myeloid cells, 912 NK cells, 104 dendritic cells and 10 mast cells. After adjustment for time point, regimen and repeated biopsies, immune-cell interferon-alpha scores averaged 1.343 in 20 responder biopsies versus 2.084 in 28 nonresponder biopsies (adjusted difference −1.147; P = 1.29 × 10⁻⁴); interferon-gamma scores ran 1.104 versus 1.657 (adjusted difference −0.556; P = 0.004). Nonresponders carried the higher interferon signal — a paradox the authors read cautiously as persistent stimulation or resistance-associated feedback rather than effective tumor control.

A cell-line experiment tested the acute arm directly. Recombinant interferon-gamma at 10 or 100 ng/mL for 24 or 48 hours induced CD274 (the gene encoding PD-L1) 4.16-fold in A375 and 3.41-fold in SK-MEL-28 melanoma cells; CXCL10 rose 9.23-fold and 7.24-fold; HLA-A reached 2.42-fold and 1.73-fold. All six comparisons held after Holm correction at adjusted P values below 0.001.

What the study does not establish

The limitations are explicit. The single-cell cohort profiled immune cells, not tumor cells, so composition may partly explain biopsy-level differences; the baseline-only interferon-gamma estimate was imprecise (P = 0.055) while the on-treatment estimate was tighter (P = 0.012); and none of the four evidence streams — TCGA clustering, paired transcriptomics, single-cell comparison, cell-line stimulation — were linked at the patient level or establish causality. What the work delivers instead is a documented framework, with full gene-set definitions and sensitivity analyses, for reading checkpoint and interferon programs as a coordinated language of the tumor-immune interface — a vocabulary that may inform the search for a clinically usable melanoma biomarker.