A machine learning-based method toward the prognostic analysis of CD38 with a pathomics model in cutaneous melanoma

Introduction CD38 possesses an essential role in cancer evolution, particularly in cutaneous melanoma. Herein, we aimed to evaluate the prognostic value of CD38 in patients with cutaneous melanoma (CM) and to propose a HE-based pathomics model for its expression prediction. Methods A total of 377 CM patients with transcriptomic data and 302 individuals with hematoxylin-eosin staining images were retrieved from the TCGA database for the prognostic analysis and pathomics feature extraction. Accordingly, a brand-new machine-learning method, namely the gradient boosting machine (GBM) algorithm, was used to develop a pathomics model to examine CD38 expression, accompanied by generalized linear model (GLM) and random forest (RF) analysis. Afterward, pathway enrichment analysis, immune-related genes and immune cell infiltration were examined to determine the underlying intrinsic mechanism of the pathomics score. Results CD38 expression was found to be significantly higher in tumor samples than in normal samples, which also showed statistical correlations with a variety of clinical variables, including tumor site, melanoma clark level, melanoma ulceration and sample type. Notably, elevated CD38 level was associated with favorable overall survival (OS). Furthermore, in the subgroup analysis, higher CD38 was identified as an important favorable prognostic for OS in melanoma clark level I/II/II, radiotherapy NO and metastatic melanoma patients. The newly established pathomics model successfully anticipated CD38 expression with excellent performance and exhibited a solid relationship between high pathomics scores and improved OS. Meanwhile, high pathomics scores corresponded with a decreased risk of mortality in the subgroups of age≥60, male and metastatic melanoma. In addition, pathomics scores were involved in particular hallmarks, immune-related characteristics and immune cell infiltration in cutaneous melanoma, such as NATURAL_KILL_CELL_MEDIATED_CYTOTOXICITY, KRAS SIGNALING UP, PDCD1, BTLA, CD86, CD80 and Macrophage M1 infiltration. Conclusion Our findings imply that CD38 and the generated pathomics scores hold promising potential as prognostic indicators for OS in CM patients.