Research Highlights

LUNG SKIN CANCER BCC identity switch breaks AI to assess restraints of Hedgehog images pathway inhibition The distinction between adenocarcinoma and squamous Basal cell carcinoma (BCC), a form cell carcinoma (SCC), the most prevalent subtypes of of skin cancer, is the most common non-small-cell lung carcinoma, requires visual malignancy in humans and typically examination by a pathologist, and is key to determine arises owing to genetic aberrations the best treatment course for each patient. Aristotelis resulting in constitutive Hedgehog Tsirigos and colleagues have developed an artificial (Hh) signalling. is a intelligence (AI)-based strategy to assist evaluation, drug that inhibits Smoothened, a with results now published. key component of the Hh pathway, Publicly available whole-slide images from The Cancer and is approved for the treatment of Credit: Simon Bradbrook/ Springer Nature Limited Genome Atlas (TCGA) were used to develop the model: advanced-stage BCC. Two preclinical 1,176 from tumour-derived tissue and 459 from non-tumour studies provide new insights into the tissue. The images were split into three groups: training, responses of BCC to this drug and the effect seemed to differ, probably validation and testing. The researchers developed a model uncover a clinically relevant strategy owing to variation in the level of that enables the distinction between non-tumour and tumour to overcome disease relapse. Wnt inhibition achieved with an tissue with an area under the curve (AUC) of 0.99–0.993, The majority of patients with anti-LRP6 antibody in one study and between adenocarcinoma and SCC with an AUC of BCC benefit from vismodegib, versus a Porcupine inhibitor in the 0.949–0.952. although cures are difficult to achieve other. “The next step would be to In addition, three pathologists and disease relapse can occur after conduct clinical trials using such assessed the images: 50% of the treatment withdrawal. “We showed drug combinations in patients with images incorrectly classified with the in mouse models that, indeed, some relapsing BCC, and possibly other AI-based model were also there is tumour cells manage to escape the characterized by activation misclassified by at least one of the potentially activity of vismodegib and ‘hide’ of the Hh and/or Wnt pathways, such pathologists, but 83% of images under a different skin stem cell as subtypes of ,” much more identity,” states Frederic de Sauvage, Blanpain opines. “We believe that incorrectly classified by at least one information of the pathologists were correctly who led one of the studies. “By doing this mechanism is likely to be at play classified using AI. The classification … to predict so, the residual tumour cells become in other tumour types, such that of the samples was compared with the mutational drug tolerant but are able to switch combinations of existing targeted back to their original identity when agents might lead to more complete that in TCGA, with AUCs for the status algorithm and the pathologists’ treatment is discontinued and the responses,” adds de Sauvage. consensus of 0.82 and 0.78, driver pathway is reactivated; tumour “The two studies have many respectively. growth then resumes.” similarities and, at the same time, “We also show that there is potentially much more Vismodegib-induced tumour nicely complement each other,” information in these slides to enable the same algorithm to regression was mediated by says Blanpain. “Even though predict the mutational status of frequently-mutated cancer differentiation of BCC cells from a we use slightly different genetic driver ,” explains Tsirigos. On the basis of previously hair follicle stem cell-like phenotype to mouse models, we describe the described associations between mutations and specific an interfollicular epidermis or isthmus same mechanism of resistance to patterns of lung adenocarcinoma, the system was trained to cell fate. The cells that escaped vismodegib mediated by a quiescent predict the probability of mutations in the ten most commonly vismodegib-induced differentiation persistent cell population, which can mutated genes in lung adenocarcinoma using matching had activation of Wnt signalling. be overcome through combined Hh genomic data available in TCGA. AUC results ranged from This switch reflects cellular plasticity and Wnt pathway inhibition. In other 0.640 (0.419–0.845) for NF1 to 0.856 (0.709–0.964) for STK11. via chromatin remodelling, probably words, the same conclusions were “We anticipate that AI will be useful in clinical in a slow-cycling subpopulation of reached in both papers, although practice because it will facilitate the diagnosis process, BCC cells, rather than selection of through different experimental it will potentially help predict clinical outcomes and it will pre-existing Wnt-expressing cells. approaches and using different drug improve patient selection for clinical trials, particularly for “We demonstrated that combinations, thus reinforcing targeted therapies,” summarizes Narges Razavian, adding combining vismodegib with Wnt the validity and relevance of our “However, it is important to emphasize that a doctor should pathway inhibitors, which are already discoveries,” he concludes. always review any diagnoses and predictions coming from an being tested in clinical trials, leads to David Killock AI-based system.” the eradication of resistant tumour Original articles Biehs, B. et al. A cell identity switch allows residual BCC to survive Diana Romero cells and avoids tumour relapse in mice,” explains Cédric Blanpain, who Hedgehog pathway inhibition. Nature 562, 429–433 (2018) | Sánchez-Danés, A. et al. A Original article Coudray, N. et al. Classification and mutation prediction from led the second study. This finding slow-cycling LGR5 tumour population mediates non-small cell lung cancer histopathology images using deep learning. Nat. Med. was broadly consistent across both basal cell carcinoma relapse after therapy. Nature https://doi.org/10.1038/s41591-018-0177-5 (2018) , 434–438 (2018) studies, although the magnitude of 562

724 | december 2018 | volume 15 www.nature.com/nrclinonc