Résumé
Cutaneous melanoma is a highly invasive tumour. Despite the development of modern therapies, most patients with advanced metastatic melanoma have poor clinical prognoses. The most frequent mutations in melanoma affect BRAF, a protein kinase of the MAPK signalling pathway. Therapies targeting both BRAF and MEK are effective in only 50% of patients and, almost systematically, generate drug resistance. In order to understand the mechanistic origin of the resistance, we build multiscale mathematical models describing intracellular dynamics of metabolic pathways and dynamics of melanoma cell populations in interaction with their microenvironment, taking into account both spatial and cellular heterogeneity. Our primary assumption is that under treatment melanoma cells undergo a series of non-genetic transitions, leading to drug tolerant and resistant cell states. This is consistent with single cell mRNAseq studies (Rambow et al., Cell 2018) and led us to a first mathematical model predicting the outcome of treatments (Hodgkinson et al., Front Oncol 2022). Furthermore, like in (Kumar et al., Cell Metab 2019), we expect that the spatial distribution of sensitive and resistant cells depends on the distance to these sources. In order to refine our models, we require 3D reconstructions of blood vessels and cell states in naïve and treated tumours. Starting with whole slide images of melanoma tumors from patient derived xenograft (PDX) mouse models, we build 3D vascular models and use them to predict zonation of hypoxia and metabolic states within the tumour. For this study, PDX samples underwent serial sectioning over 2mm depth. Every 12um depth, slides were stained with hematoxylin and eosin and the two next adjacent slides with cluster of differentiation 31 (CD31, a blood vessel marker) and CA9 (a hypoxia marker). The 3D reconstruction pipeline involves three steps: 1) Vessel segmentation in 2D sections performed by deep learning with U-Net architecture, 2) Image registration developed using the scale-invariant feature transform, a feature-based method, 3) Vessels 3D rendering performed using a marching cubes algorithm. An original feature of our pipeline is its ability to handle sparse data by generation of synthesized slides using a Generative Adversarial Networks algorithm. This addition is also useful within a clinical context where synthesized slides can be artificially created from a handful of existing, real clinical slides. The resulting 3D vascularization model is used to predict the distribution of hypoxia in the tumor using a partial differential equations (PDE) model pre-trained on adjacent CD31 and CA9 stained sections. Another original aspect of this work is that the PDE model is trained with a few 2D sections and validated using 3D reconstructions. Future work will include modeling of the cell population dynamics in 3D reconstructed tumors and validation of these results on a selection of slides using Imaging Mass Cytometry.