DeepTFtyper: an interpretable morphology-aware graph neural network for translating histopathology images into molecular subtypes in small cell lung cancer.

Abstract

Small cell lung cancer (SCLC) is a highly aggressive high-grade neuroendocrine carcinoma with a poor prognosis. Molecular subtyping of transcription factors (SCLC-A, -N, -P, and -Y) shows great potential for guiding treatment decisions. However, its clinical application are limited by insufficient samples and the complexity of molecular testing. In this study, we developed DeepTFtyper, a graph neural network-based deep learning model for automatically classifying SCLC molecular subtypes from hematoxylin and eosin-stained whole-slide images. DeepTFtyper was trained and tested on the Cancer Hospital, Chinese Academy of Medical Science cohort (n = 389) with 4-fold cross-validation, and achieved high performance with an area under the receiver operating characteristic curve above 0.70 for all four molecular subtypes identified by immunohistochemistry (IHC). Furthermore, the digital H-scores predicted by DeepTFtyper showed a significant correlation with IHC-based H-scores. Patch-level visualization and morphological analysis revealed that DeepTFtyper identifies interpretable and generalizable features corresponding to areas of relevant transcription factor expression as revealed by IHC staining and correlates well with morphological features. This study represents the first deep learning framework for predicting SCLC molecular subtypes from hematoxylin and eosin-stained histology slides, providing a scalable, accurate, and clinically relevant tool to improve patient management and guide personalized treatment decisions.