Multiplex immunofluorescence (mIF) images reveal rich details of the tumour microenvironment in melanoma, offering new opportunities to predict immunotherapy response. Analysing these high-dimensional, multi-channel images is challenging due to their size and missing stains. This talk presents a deep learning framework that models inter-channel relationships to capture complex biological interactions and uses stain imputation to handle incomplete data. Together, these approaches enable more robust and informative predictions, advancing biomarker discovery and personalised treatment strategies in melanoma.