The latest generation of machine learning tools, and in particular graph-based methods, provide accurate predictions that take into account complex interactions at different granularity levels. This is key to harness the wealth of information contained in whole-slide digital pathology images, e.g. specific cell-cell or cell-tissue interactions, especially when the multiplexed immunohistochemistry (mIHC) technique is used, which allows locating cells of up to 7 different types. This essential capacity will allow the identification of key patterns of the biological processes underlying response to cancer immunotherapy, and eventually the development of interpretable biomarkers that will be better understood, trusted and adopted by clinicians.
Clinical outcome predictions, leading (i) to define actionable candidate biomarkers for melanoma aggressivity and immune-resistance; (ii) to recognize automatically specific cell types in the cheap Hematoxylin-Eosin (H&E) stained images that are routinely produced; and (iii) to develop next-generation H&E-based biomarkers exploiting predicted cell type information. Both arms of this proposal, namely the discovery of novel mIHC-based interpretable biomarkers in melanoma, and the potential to re-interpret existing H&E databases for biomarker discovery have strong translational potential for future patients. Indeed, with the advent of digital pathology in clinical practice, such H&E databases are bound to grow fast in the coming years.