Annotation-free prediction of immunotherapy response in melanoma using single-cell transcriptomic data.

Immune checkpoint inhibitors (ICIs) have transformed the advanced melanoma treatment landscape; however, a subset of patients achieve durable responses. Current biomarkers, such as PD-L1 expression and tumor mutational burden, offer limited predictive power due to the profound heterogeneity of melanoma. Accordingly, we developed an artificial intelligence (AI)-based model to predict ICI responsiveness using single-cell RNA sequencing (scRNA-seq) data without requiring cell type annotation. scRNA-seq data profiled using Smart-seq2 platform were downloaded from a public repository (GEO: GSE120575). From these data, we analyzed 16,290 tumor-infiltrating cells from melanoma scRNA-seq dataset. Various AI-based models, including Extreme Gradient Boosting, Random Forest, Logistic Regression, Support Vector Machine, Feedforward Neural Network, and Convolutional Neural Network were constructed, with the best-performing model achieving an area under the curve of 0.87. This AI-driven approach identified 29 key predictive biomarkers, including CCR7 and MTRNR2L2. Validation using three independent bulk RNA-seq datasets (cBioPortal: DFCI melanoma; ENA: PRJEB23709; GEO: GSE91061) suggested that CCR7 was associated with favorable ICI response and improved survival, whereas MTRNR2L2 showed a tendency toward enrichment in non-responders and poorer outcomes. Cell-type-specific expression analysis revealed that CCR7 was primarily expressed in B cells and memory T cells from responders, whereas MTRNR2L2 was elevated in exhausted and cytotoxic T cells in non-responders. CCR7-positive B cells exhibited activation of the NF-κB pathway and demonstrated prognostic significance independent of the melanoma primary site or histologic subtype. However, among the three molecular subtypes, including immune, keratin, and microphthalmia-associated transcription factor (MITF)-low, CCR7 expression was significantly associated with the immune subtype. Additionally, pathway-level deep learning models reinforced these findings, highlighting immune activation in responders and cell cycle-related signals in non-responders. Our study demonstrates that predictive modeling based on unannotated scRNA-seq data enables clinically relevant biomarker identification, offering a robust approach for patients with stratifying melanoma and guiding personalized immunotherapy.