A Deep Learning Framework for Predicting Patient Response to Immune Checkpoint Inhibitors Using Single-Cell RNA Sequencing Data in Metastatic Melanoma

Immune checkpoint inhibitors (ICIs) targeting PD-1/PD-L1 and CTLA-4 have revolutionized metastatic melanoma treatment. However, therapeutic response remains highly heterogeneous, with a significant portion of patients failing to achieve durable clinical benefits. Bulk transcriptomic profiling often obscures cellular heterogeneity, which is a critical determinant of therapeutic resistance. In this study, we present MelanoDL, a novel deep learning framework designed to predict patient response to ICIs using single-cell RNA sequencing (scRNA-seq) data. MelanoDL integrates a graph convolutional network (GCN) with an attention-based autoencoder to capture complex intercellular communication networks and high-dimensional transcriptomic signatures of tumor-infiltrating immune cells and malignant cells. Leveraging scRNA-seq profiles from 142 metastatic melanoma patients across three independent cohorts, MelanoDL achieved an area under the receiver operating characteristic curve (AUROC) of 0.89 in predicting non-responders versus responders, significantly outperforming conventional machine learning models and bulk-signature-based predictors. Our framework identified a subpopulation of CD8+ T cells expressing high levels of exhaustion markers (LAG3, HAVCR2) coupled with dysfunctional metabolic pathways, alongside CXCL13-expressing follicular helper T cells, as the primary cellular drivers of ICI responsiveness. By mapping these single-cell architectures to clinical outcomes, MelanoDL offers a robust, scalable computational tool for precision oncology, facilitating personalized patient stratification and the identification of novel therapeutic targets.