Nantomics LLC, of Culver City, Calif., reported that research based on the company’s deep learning system has been published in a peer-reviewed journal, highlighting the algorithm’s ability to discern which mutation drives a patient’s breast cancer. The company said their approach is a rapid and cost-effective way to establish the breast cancer subtype, thus giving clinician and patient alike a good understanding of which therapies would be ineffective for that cancer and maximizing the chances for a cure.
Nantomics and Nanthealth Inc., also of Culver City, won the FDA’s seal of approval for a whole exome sequencing test to determine overall tumor mutational burden for the Omics Core diagnostic. The Omics Core is designed to tally all the gene-coding mutations in a tumor, but the company has announced a novel artificial intelligence technique that will make a distinction between mutation subtypes for breast cancer. Scientists with Nantomics trained a deep neural network with whole slide images (WSI) from more than 440 breast cancer tumors that had already been run through PAM50 (Predictor Analysis of Microarray 50) subtyping in an effort to classify the tumors into one of four major breast cancer molecular subtypes.
Molecular subtypes predictive of response across tumor grade
The subtypes for this evaluation consisted of luminal A, luminal B, basal-like and HER2-enriched. The Nantomics team then validated the algorithm and used it to classify samples from another 222 tumors, an effort the company said required RNA expression profiling. The authors of the article, which appears in the Jan. 28, 2020, online issue of Breast Cancer Research, said that intrinsic molecular subtype is seen as a strong prognostic feature when determined by an expression-based PAM50 assay. This holds even when controlling for patient age, tumor grade and nodal status.
The problem with bulk assays, such as RNA sequencing, is that intratumoral heterogeneity might not be apparent, which would deprive the doctor and patient of the chance to maximize treatment opportunities. Assays for RNA-based signatures are more commonly used of late as supplementary prognostic indicators, including the Predictor Analysis of Microarray 50 (PAM50), although immunohistochemistry-based tests might still be more widely used.
The Nantomics team used only whole-slide images of hematoxylin and eosin (H&E)-stained biopsy tissues, and noted that one of the more useful aspects of WSI analysis is that it offers the ability to concurrently use high-zoom patches that provide cellular-level information along with lower-zoom patches that depict the interdependence of tissue structures. This approach has been used to design convolutional neural networks for distinguishing between invasive ductal carcinoma and benign ductal carcinoma, and for determining whether lymph nodes adjacent to breast tumors were positive for metastases.
Still, this approach demands much of the method used to analyze all possible multiscale patches when gigapixel slides are used, the authors noted. Thus, they said, their approach offers a method for minimizing the manual work needed to identify the cancer-rich patches in H&E-stained slides. However, another advantage to their method is that it allows the user to retain the ability to directly observe intra-tumoral heterogeneity without the need for numerical deconvolutional methods.
The authors said their classifier may offer some applications for detecting intra-tumoral heterogeneity thanks to the sensitivity to sub-clonal diversity. One example of the utility provided by this kind of processing power is depicted by the finding that patients with tumors that were classified as luminal A, but which had basal subclones, had poorer survival than those with homogenous luminal A tumors.
The authors concluded that the ability of their classifier to identify heterogeneity in cancer cell populations has significant prognostic implications. However, they also said that the widespread availability of WSI technology combined with the cost-effectiveness of this methodology suggest that the application of this approach “to prognostic procedures may be accelerated.”