Background: Cilia exert tumor suppressor effects in several tumor types including melanoma, and scientists found that blocking cilia growth in drug-resistant cancer cell lines may restore cancer cells' sensitivity to immunotherapy. Therefore, we hypothesized that cilia might be used as a predictive biomarker for melanoma immunotherapy. Methods: Through literature search, a total of six sets of RNA-seq and WES data of melanoma samples receiving PD-1/CTLA-4 immunotherapy were collected. GSVA was used to obtain an enrichment score between the samples and a set of 112 ciliary genes, based on which ciliary signatures were screened for prognosis. The ciliary characteristic gene set was further optimized by LASSO regression, then a risk model was constructed by combing stepwise regression and COX regression. Log-rank test was used to verify survival differences between high and low risk groups in other independent data sets, and Wilcox test was used to explore differences in other clinical characteristics between high and low risk groups. Imaging mass spectrometry (IMC) and spatial position imaging (SPI) data were used to analyze cell states in tumor tissues. All analyses were performed by R 4.0.3. Results: A total of 10 cilia pathways related to prognosis were screened out in the training set, and 13 genes were picked out from these 10 pathways as selected features. Combining LASSO and COX regression, a risk model consisting of 5 genes were finally obtained. Patients in the high-risk (HR)group classified by this model had a significantly worse prognosis than the low-risk (LR) group in both the training and test sets (p < 0.05, respectively). In addition, by analyzing immune infiltration and pathway enrichment, it was found that LR group was in the state of immune activation, and its immune score was significantly higher than that of HR group. Further analysis showed the model had no significant correlation with other clinical features, such as pathological stage, tumor mutation burden and NF1/KRAS/BRAF gene mutations, indicating that the model was relatively independent. By IMC data analysis, it was found that the number of tumor cells in LR and HR groups were similar (107082 vs 11928), but the number of lymphocytes in LR group was significantly higher than that in HR group (71578 vs 5313). Moreover, among the 35 protein markers related to immunotherapy, 16 (45.7%) proteins were up-regulated and 0 protein was down-regulated in LR group. By analyzing SPI of multiple regions of these samples, it was also observed that tumor tissues in LR group were mostly in a state of lymphocyte infiltration, while these in HR group were in a state of immune rejection or immune desert. Conclusions: The model constructed based on cilia-related features in this study can independently identify melanomas with high risk of recurrence or progression, and provide a theoretical basis for exploring whether targeted cilia structure therapy can be used as a treatment for melanoma.