Predictive modelling of highly multiplexed tumour tissue images by graph neural networks

Abstract

The progression and treatment response of cancer largely depends on the complex tissue structure that surrounds cancer cells in a tumour, known as the tumour microenvironment (TME). Recent technical advances have led to the development of highly multiplexed imaging techniques such as Imaging Mass Cytometry (IMC), which capture the complexity of the TME by producing spatial tissue maps of dozens of proteins. Combining these multidimensional cell phenotypes with their spatial organization to predict clinically relevant information is a challenging computational task and so far no method has addressed it directly. Here, we propose and evaluate MULTIPLAI, a novel framework to predict clinical biomarkers from IMC data. The method relies on attention-based graph neural networks (GNNs) that integrate both the phenotypic and spatial dimensions of IMC images. In this proof-of- concept study we used MULTIPLAI to predict oestrogen receptor (ER) status, a key clinical variable for breast cancer patients. We trained different architectures of our framework on 240 samples and benchmarked against graph learning via graph kernels. Propagation Attribute graph kernels achieved a class-balanced accuracy of 66.18% in the development set (N=104) while GNNs achieved a class-balanced accuracy of 90.00% on the same set when using the best combination of graph convolution and pooling layers. We further validated this architecture in internal (N=112) and external test sets from different institutions (N=281 and N=350), demonstrating the generalizability of the method. Our results suggest that MULTIPLAI captures important TME features with clinical importance. This is the first application of GNNs to this type of data and opens up new opportunities for predictive modelling of highly multiplexed images.