Cancer is a non-Mendelian and heterogeneous disease. Multiple factors contribute to the initiation, development, and treatment of cancer, such as genomic, transcriptomic changes, and even histopathological differences. As a complex disease, the omics data serve as a roadmap for cancer investigation. The omics data in cancer studies, such as the genomics and pathological images, are highly complex with multiple variables. The deep learning algorithm, such as neural networks, can efficiently reduce the data dimension. [1] Therefore, deep learning approaches are employed to process these mega data for automated cancer diagnosis, novel biomarker discovery, prognostic morphological determinants, etc. [2] .In clinical practice, accurate annotation of tumor grades and cancer subtypes can overcome the variations caused by clinician expertise or potential fatigue and adds significantly to the efficacy of the initial diagnosis and the following treatment, such as surgical excision, radiation, and chemotherapy, with the aid of the deep learning model. [3] The deep learning models have been applied to endoscopic images [4] and ultrasound images [5] for early esophageal disease diagnosis, PET images for treatment outcome prediction, [6] and computerized tomography for esophageal fistula prediction. [7] However, the identified properties or interpreted hypotheses by deep learning are seldom verified by experimental approaches to facilitate further scientific research. Although histopathologic features are the gold standard for staging and grading human cancers, the regularly employed hematoxylin and eosin (H&E) staining is of little assistance with the novel-targeted therapy, which frequently targets epidermal growth factor receptor (EGFR), vascular endothelial growth factor (VEGF), and other surface antigens