Deep Learning Predicts Underlying Features on Pathology Images with Therapeutic Relevance for Breast and Gastric Cancer

Abstract

DNA repair deficiency (DRD) is an important driver of carcinogenesis and an efficient target for anti-tumor therapies to improve patient survival. Thus, detection of DRD in tumors is paramount. Currently, determination of DRD in tumors is dependent on wet-lab assays. Here we describe an efficient machine learning algorithm which can predict DRD from histopathological images. The utility of this algorithm is demonstrated with data obtained from 1445 cancer patients. Our method performs rather well when trained on breast cancer specimens with homologous recombination deficiency (HRD), AUC (area under curve) = 0.80. Results for an independent breast cancer cohort achieved an AUC = 0.70. The utility of our method was further shown by considering the detection of mismatch repair deficiency (MMRD) in gastric cancer, yielding an AUC = 0.81. Our results demonstrate the capacity of our learning-base system as a low-cost tool for DRD detection.

Keywords: DNA repair deficiency; biomarker; deep learning; digital pathology; mutational signature.

Figure 1

Overview of the end-to-end framework of training and testing for DNA repair assessment on whole slide images.

Figure 2

A whole slide image (32x) and corresponding heatmap for a single patient in the independent test set. The intensity of coloring in the heatmap represents the probability of being class-positive in the CNN (convolutional neural network) step.

Figure 3

Tissue classes learned from convolutional neural network (CNN) in stomach cancer. Examples from mismatch repair deficiency (MMRD) and mismatch repair proficiency (MMRP) groups. The bar represents the spectrum of histologic diversity in the test set and it contains the top-ranking tiles per slide learned from CNN. These tiles were manually annotated by two pathologists among the following ten tissue labels: ADI, adipose tissue; BACK, background; DEB, debris; DIS, dysplasia; INF, inflammation; LYM, lymphocytes; MUS, smooth muscle; NORM, normal gastric mucosa; STR, cancer-associated stroma; TUM, gastric adenocarcinoma epithelium.

Figure 4

ROC analysis of classifier predicting DRD (DNA repair deficiency) status on (A) TCGA-BRCA training cohort, (B) ACCC-BRCA independent validation cohort and (C) TCGA-STAD training cohort.

Figure 5

HR (homologous recombination)-associated molecular drivers. The distribution of lesions in known HR-pathway genes of the MMRD predicted patients among the training (A) and independent cohorts (B). Numbers above each bar count the patients for the corresponding lesion.

Figure 6

MMR (mismatch repair)-associated molecular drivers. This figure shows the molecular interplay among the MMRD and MMRP predicted groups, which are colored in red and blue, respectively. (A) Changes in methylation of MLH1 promoter, (B) MLH1 expression levels and (C) mutational activity of Signature 20 were detected by comparing MMRD and MMRP groups; Scatterplots showing (D) methylation and gene expression levels of MLH1, (E) methylation levels of MLH1 and mutational activity of Signature 20, (F) gene expression levels of MLH1 and mutational activity of Signature 20; (G) Changes of Tumor Mutation Burden and (H) genes related to pro-inflammatory markers across MMRD and MMRP groups. (I) Mutational activity of Signature 20 across MSI subtype and others molecular subtypes. Statistical significance was assessed by the two-tailed Wilcoxon Rank Sum and it is indicated by * p < 0.05, *** p < 0.001. The MSI (Microsatellite Instability) and other subtypes (GS: Genomically Stable; EBV: EBV-positive and CIN: Chromosomal INstability) were collected from The Cancer Genome Atlas (TCGA) study.