Interpretable prototype learning for EGFR amplification prediction from whole-slide images in glioblastoma

The epidermal growth factor receptor (EGFR) is one of the key biomarkers for diagnosis, treatment, and prognosis in glioblastoma (GBM). Current EGFR diagnostic methods, including immunohistochemistry, fluorescence in situ hybridization, and next-generation sequencing, are costly, time-consuming, and not uniformly accessible. A major challenge in whole-slide images (WSIs)-based prediction is to capture the full morphological heterogeneity of the tissue in a way that supports not only precise classification but also interpretability. To address these challenges, we proposed an automatic, interpretable prototype-learning framework that integrates a Variational Autoencoder with the Dirichlet Bayesian Gaussian Mixture Model to determine optimal tissue prototypes from morphological features of hematoxylin and eosin-stained WSI, which guide the classification of EGFR amplification. The classification performance is evaluated using internal 5-fold cross-validation on the Cancer Genome Atlas dataset and external validation on the Clinical Proteomic Tumor Analysis Consortium (CPTAC) dataset and the private GB-UK dataset. The proposed model achieved an area under the curve of 0.8087 ± 0.0153 for internal validation and 0.7740 and 0.7870 for external validations of the CPTAC and GB-UK cohorts, respectively, surpassing multi-instance learning models and conventional predefined clustering settings. Specific prototypes distinguish EGFR-amplified from EGFR-non-amplified cases, and histological analysis of these prototypes, which are consistent with expert-recognized morphological patterns, suggests an EGFR-amplified infiltration pattern. These findings demonstrate that the automatic prototype learning framework provides a rapid, cost-effective, interpretable, and scalable AI model for EGFR prediction in GBM, linking the learned prototypes to recognizable tissue patterns that can facilitate clinical decision-making.