An attention-based mRNA transformer network for accurate prediction of melanoma response to immune checkpoint inhibitors.

Melanoma immunotherapy urgently requires approaches that can accurately predict drug responses to minimize unnecessary treatments. Deep learning models have emerged as powerful tools in this domain due to their robust predictive capabilities. Integrating functional characteristics with expression data from mRNA transcripts shows promise for enhancing prediction accuracy. We developed a deep learning model called AMU (Attention mechanism Model for melanoma immUnotherapy) that incorporates a self-attention mechanism to predict clinical responses to immune checkpoint inhibitors in melanoma patients based on mRNA expression profiles. We evaluated AMU's performance against established machine learning approaches including Support Vector Machine (SVM), Random Forest, AdaBoost, XGBoost, and classical Convolutional Neural Networks (CNN). In the validation set (pre-treatment tissue samples), AMU exhibited outstanding performance, with an AUC of 0.941 and an mAP of 0.960. In the test set (post-treatment tissue samples), its AUC was 0.672, and the mAP was 0.800. Model interpretation revealed that the TNF-TNFRSF1A pathway was a crucial pathway influencing the efficacy of melanoma immunotherapy. Additionally, the expression levels of CD80 and CCR3 were closely correlated with the survival rate (hazard ratios of 0.761 and 0.134, respectively) and the response to immune checkpoint inhibitors in melanoma patients. The deep learning model integrated with the self-attention mechanism has demonstrated strong efficacy in processing mRNA expression data for melanoma immunotherapy response prediction. After rigorous evaluation, including batch effect correction and cross-validation, AMU achieved superior performance compared to traditional machine learning approaches. Beyond prediction accuracy, our model interpretation work identified the TNF-TNFRSF1A pathway as potentially crucial in determining melanoma ICI response, a finding aligned with recent experimental evidence. The embedding architecture's ability to capture meaningful gene-gene relationships, partially consistent with established protein interaction networks, suggests broad potential for representation learning in transcriptomic analysis. While acknowledging the limitations of current sample sizes and the need for prospective validation, this work provides both methodological advances in applying transformer architectures to gene expression data and biological insights into immunotherapy response mechanisms. The integration of robust machine learning approaches with domain-specific biological knowledge represents a promising direction for developing clinically relevant biomarkers in precision oncology.