Introduction
Immunotherapy has made a substantial impact on the treatment of nonsmall cell lung cancer (NSCLC) patients, but few treated patients show durable clinical benefit. There is a need for a biomarker capable of selecting potential responders with greater predictive power than the current clinical standard, the tumor proportion score. We investigated graph neural networks (GNNs) for predicting immunotherapy response based on the tumor micro-environment (TME) in H&E whole slide images of NSCLC patients.
Material and methods
We processed 93 retrospectively collected NSCLC H&E biopsies or resections by i) detecting tumor and immune cells within the tissue using HoVerNet, ii) finding the densest region of tumor cells, iii) taking a 500 micrometer crop around this hotspot, hypothesizing that it is representative of the TME, iv) building a cell-graph of the hotspot using cell coordinates/type as node features. We trained three GNN architectures (GCN, GraphSage and GIN) on the graphs using five-fold cross-validation on two binary endpoints: 1-year overall survival (OS) and progression-free survival (PFS), measured from treatment start to radiological progression, death or loss to follow-up.
Results and discussion
Median follow-up was 11 months. The mean C-index (using the model's softmax as risk score) across five folds for all endpoints was consistently highest for the GIN model (OS: 0.727 +- 0.067, PFS: 0.599 +- 0.113), as compared to using the TPS thresholded at 1% (OS: 0.511, PFS: 0.512).
Conclusion
GNN assessment of H&E slides holds promise for better identifying responders to immunotherapy in NSCLC. Further studies on using GNNs, which describe relationships between cell types in the TME, are necessary.