A Machine Learning-Based Histopathological Image Analysis Reveals Cancer Stemness in TNBCs with 17p Loss

Abstract

Artificial intelligence and machine learning based methods have incorporated scientific research into clinical decision, leading to great improvement in clinical diagnosis and therapeutics. Here we developed a Convolutional Neural Network based model to identify cancer stem-like cells (CSCs) on H&E-stained histopathological images. Combined with cancer genomics profiles, our analysis revealed that triple negative breast cancers (TNBCs) with heterozygous deletion of chromosome 17p (17p-loss) correlate with higher cancer stemness potential compared to TNBCs with neural copy numbers of 17p (17p-intact). 17p-loss TNBC cells also have an increased percentage of CSCs and are resistant to chemotherapies compared with the 17p-intact TNBC cells. Moreover, we built a bioinformatics pipeline to screen compounds that target the stemness of 17p-loss cancer cells, one of which is FK866. FK866 promoted the antitumor activity of doxorubicin in the treatment of 17p-loss TNBCs. Our study provides a powerful computational tool for cancer image analysis as well as a feasible approach for precision cancer medicine.