Spatial lipidomics-based melanoma diagnostics by integration of mass spectrometry imaging and machine learning.

Melanoma is an aggressive skin cancer with a rising global incidence, yet its lipidomic alterations remain understudied. We applied desorption electrospray ionization mass spectrometry imaging to 137 frozen skin specimens (65 paired normal, 57 melanomas, 15 melanocytic nevi) and developed a lipidomics-based machine learning model for diagnostic evaluation. At the clinically relevant lesion level, the model achieved 90.24% accuracy (37/41; 95% confidence interval [CI] = 77.0%-96.3%) in distinguishing melanoma from nonmelanoma tissues, comprising normal skin and nevi. Pixel-level evaluation showed 94.71% accuracy (95% CI = 94.23%-95.15%), with 93.76% sensitivity (95% CI = 93.09-94.37%) and 96.14% specificity (95% CI = 95.47%-96.72%). In differentiating melanoma from melanocytic nevi, the model achieved a lesion-level accuracy of 80.95% (17/21; 95% CI = 60.00%-92.33%) and a pixel-level accuracy of 93.10% (95% CI = 92.42%-93.73%). Further analysis revealed distinct lipidomic signatures in melanoma, characterized by elevated phospholipids and decreased triglycerides compared with those in nonmelanoma tissues. Spatially resolved desorption electrospray ionization mass spectrometry imaging enabled generation of probability heatmaps that accurately delineated melanoma regions, including early-stage lesions. In summary, integration of desorption electrospray ionization mass spectrometry imaging with machine learning demonstrates potential as a diagnostic adjunct for melanoma lipidomic profiling.