A transfer learning-based graph neural network model using mIF images to predict neoadjuvant immunochemotherapy response in patients with gastrointestinal cancer.

Abstract

e13592 Background: Neoadjuvant immunochemotherapy has great potential to revolutionize the treatment of gastrointestinal cancer. However, effective biomarkers to classify responders from non-responders are still lacking. We aimed to build a response prediction model using multiplex immunofluorescence (mIF) images from patients with gastrointestinal cancer. Methods: We proposed a graph neural network-based (GNN) approach based on the pre-treatment mIF samples collected from 77 patients, including 30 gastric cancer (GC) patients, 34 colorectal cancer (CRC) patients, and 13 esophageal cancer (ESCA) patients. All the tissue samples were stained by two panels of mIF agents, namely Panel-A (CD3, CD8, PD1, PD-L1) and Panel-B (CD56, CD68, CD163). Cell coordinates, agent densities, and cell morphology features were used to generate cell graphs for model training. Results: Initially, we trained a pan-cancer GNN model for each panel and assessed its performance on the patch- and patient-level using five-fold cross-validation. The Panel-A model achieved higher areas under the receiver operating characteristic curve (AUROC) (patch-level: 0.71 ± 0.036, patient-level: 0.65 ± 0.010) than the Panel-B model (patch-level: 0.59 ±0.017, patient-level: 0.62 ± 0.021) and conventional mIF-based markers (i.e., positive cell ratios/counts of fluorescence agents, AUROC: 0.34 ~ 0.54). Specifically, the performance of CRC was better in the Panel-A model (patch-level: 0.75± 0.053, patient-level: 0.76 ± 0.053) than that in Panel-B (patch-level: 0.52 ± 0.027, patient-level: 0.60 ± 0.028), while the performance of GC was better in Panel-B (patch-level: 0.64 ± 0.030, patient-level: 0.71 ± 0.034) than that in Panel-A (patch-level: 0.53 ± 0.019, patient-level: 0.51 ± 0.018). The AUROCs of both panels of ESCA were less than 0.5, reflecting feature differences associated with treatment response between ESCA and GC/CRC. Furthermore, discrepancies of GC and CRC between Panel-A and -B motivated us to further optimize the performance for individual cancer types using transfer learning (TL). The GC TL-model performance (patch-level: 0.691±0.025, patient-level: 0.751 ± 0.046) was improved compared with the pan-cancer model, but not for the CRC TL-model (patch-level: 0.636 ± 0.066, patient-level: 0.697 ± 0.039). Moreover, the best patient-level specificity and sensitivity under the optimal Youden-Index were 70% and 88% for CRC, and 93% and 62% for GC. Conclusions: The GNN-based approach using mIF images demonstrated promising potential in predicting the immunochemotherapy response of gastrointestinal cancer patients, despite the limited number of samples were curated. The performance variation among cancer types indicated the cancer heterogeneity and the necessity of building single-cancer models when more data becomes available.