Developing an AI-based approach to predict response to anti-EGFR treatment for metastatic colorectal cancer patients using super-resolution imaging of EREG

Every year, there are approximately two million new cases of colorectal cancer worldwide. Globally, there is a growing number of cases of early-onset (< 50 years old) colorectal cancer, which is often diagnosed at an advanced stage. The outcomes for metastatic patients are grim, with a 5-year net survival rate of only 1 in 10. For these patients, treatment that targets epidermal growth factor receptor (EGFR), a cell surface protein that is involved in cell signalling, division and growth, can help shrink metastases for resection or slow cancer progression for palliative care. However, approximately 40% of patients receiving this treatment do not respond. The objective of this study was to investigate whether the nanoscale spatial organisation of epiregulin (EREG), one of the ligands for EGFR, could help predict response to anti-EGFR treatment for metastatic colorectal cancer patients.

To achieve this objective, we imaged tissue samples from metastatic colorectal cancer patients using single-molecule localisation microscopy (SMLM), which could resolve the high-precision positions of EREG proteins. We then developed and tested artificial intelligence (AI) based pipelines, locpix and ClusterNet, to segment and classify large-scale structures in SMLM data, such as cells. These pipelines were then applied to the SMLM data from the patients to manually segment the cells and classify them by response to treatment.

This approach may improve over an existing method for predicting response, which uses the protein expression level of EREG, but was inconclusive due to the small sample size in this study. More broadly, this study showed that the organisation of EREG may help predict response to anti-EGFR treatment. Further, we anticipate that the two novel AI-based pipelines may be generally useful for the analysis of SMLM data. This includes the first example of a graph-neural network designed for whole-graph classification of SMLM data. These pipelines could help to realise the use of SMLM data to characterise phenotypes and predict response to treatment across a wide variety of disorders.