Editors' ChoiceCancer

From micrographs to microsatellites in one bold step

See allHide authors and affiliations

Science Translational Medicine  26 Jun 2019:
Vol. 11, Issue 498, eaax9570
DOI: 10.1126/scitranslmed.aax9570

Abstract

Machine learning predicts microsatellite instability from histology.

Cancers with microsatellite instability—a condition in which excessive mutations accumulate in the cancer genome due to deficient DNA mismatch repair—are relatively visible to the immune system. Mutated proteins give rise to neoantigens, which can subsequently be recognized by cytotoxic T cells and trigger an antitumor immune response. In 2017, the U.S. Food and Drug Administration approved an immunotherapy (the anti–programmed cell death protein 1 antibody pembrolizumab) for all advanced cancers with microsatellite instability, making evaluation of DNA mismatch repair deficiency an important and ubiquitous clinical task. Traditionally, microsatellite instability has been assessed by polymerase chain reaction or by immunohistochemistry. Now, a new paper shows that it can be diagnosed directly from histological images of a tumor with the help of a deep learning algorithm.

Kather and colleagues fed histological images of stomach and colon cancers which had previously been tested for microsatellite stability with orthogonal methods into a deep learning pipeline. They trained two neural networks: one to automatically detect areas containing tumor and one to predict the microsatellite stability status of those areas. Testing their approach on an independent data set, they found that their algorithm could efficiently detect microsatellite instability with an area under the curve (a measure of test performance, in which 1 signifies perfect accuracy) of 0.84. To understand the biological features underlying the method’s “black box” predictions, they correlated the algorithm’s microsatellite instability score with gene and protein expression data. In stomach cancer, a cytotoxic T cell signature most correlated with microsatellite instability, whereas programmed cell death ligand 1 expression and an interferon gamma signature were the main correlates in colon cancer.

The ability to predict microsatellite instability directly from histological images that are routinely generated during clinical care may help to reduce costs and increase the number of patients that can benefit from immunotherapy. It is conceivable that in the future, a clinician will only have the upload the image of a tumor to a cloud-based service to receive a treatment recommendation that would have cost thousands of dollars to obtain experimentally.

Highlighted Article

View Abstract

Navigate This Article