High-accuracy prediction of 5-year progression-free melanoma survival using spatial proteomics and few-shot learning on primary tissue

Abstract

Emerging techniques in imaging-based spatial proteomics (ISP) enable in-depth insights into the architecture and protein abundance of tissue(s). Explainable machine learning (xML) models promise to yield substantial advances in ISP data-based diagnosis and prognosis. However, a clinical application of these new possibilities predicting the course of a tumor has not been suggested yet. Here, we use a few-shot learning workflow on histological multi-antigen images to predict 5-year progression-free survival (PFS) in melanoma. We address the problem of a relatively small cohort (n = 22), by utilizing a pre-trained convolutional neural network (CNN) model, which we further pre-train on a proxy task for which more samples were available (n = 39) before fine-tuning for PFS prediction. Our approach yielded a model achieving an accuracy of more than 90%, outperforming baseline models trained on clinical data by around 10%. Using an xML technique, we identified immune infiltration and proteins associated with tumor progression as crucial predictors. This indicated that our models' PFS predictions are not only highly accurate but also grounded in a relevant biological background.