ENLIGHT–DeepPT: Predicting Cancer Treatment Response from Histopathology

This study introduces ENLIGHT–DeepPT, a cutting-edge deep-learning framework that predicts cancer treatment responses using hematoxylin and eosin-stained tumor slides. The two-step approach involves DeepPT, which infers genome-wide mRNA expression from pathology images, and ENLIGHT, which predicts therapy response based on the imputed transcriptomics. Tested on 16 The Cancer Genome Atlas cohorts and validated across independent datasets, ENLIGHT–DeepPT demonstrated high accuracy in predicting treatment responses for targeted and immune therapies across six cancer types. Without requiring specific training on treatment cohorts, the model achieved an overall odds ratio of 2.28, with a 39.5% increased response rate among predicted responders. This framework highlights the potential of leveraging routine histopathology data for precision oncology, bypassing the need for direct genomic profiling or therapy-specific training.