Adaptive individualized gene pair signatures distinguishing melanoma and predicting response to immune checkpoint blockade.

Distinguishing similar cancer subtypes and predicting responses to immune checkpoint blockade (ICB) are critical for improving clinical outcomes. However, existing gene expression signatures often suffer from batch effects and poor generalizability across cohorts. To address these limitations, we propose adaptive individualized gene pair signatures (AIGPS), a robust method that adaptively quantifies gene pair reversals and selects informative features using machine learning. AIGPS was validated on 850 samples from 24 cohorts for multiclass skin cancer classification and on 252 samples from 7 cohorts including both bulk and single-cell RNA sequencing (RNA-seq) data for ICB response prediction in melanoma. Compared to existing approaches, AIGPS improves classification accuracy by over 5% and enhances response prediction performance by 6%. By relying on relative rather than absolute expression levels, AIGPS demonstrates robustness to technical variability and enhanced transferability across datasets. This adaptive framework offers a flexible strategy for biomarker discovery and has broad potential in precision oncology.