Using machine learning algorithms to predict response and toxicity to immune checkpoint inhibitors (ICIs) in melanoma patients

• Published in: Journal of clinical oncology Vol. 37; no. 15_suppl; p. 2581
• Main Authors: Johannet, Paul, Coudray, Nicolas, Jour, George, Donnelly, Douglas MacArthur, Bajaj, Shirin, Moran, Una, Dawood, Zarmeena, Nomikou, Sofia, Kim, Randie H, Pavlick, Anna C., Weber, Jeffrey S., Tsirigos, Aristotelis, Osman, Iman
• Format: Journal Article
• Language: English
• Published: 20.05.2019
• ISSN: 0732-183X; 1527-7755
• DOI: 10.1200/JCO.2019.37.15_suppl.2581

Abstract only 2581 Background: There is growing interest in optimizing patient selection for treatment with immune checkpoint inhibitors (ICIs). We postulate that phenotypic features present in metastatic melanoma tissue reflect the biology of tumor cells, immune cells, and stromal tissue, and hence can provide predictive information about tumor behavior. Here, we test the hypothesis that machine learning algorithms can be trained to predict the likelihood of response and/or toxicity to ICIs. Methods: We examined 124 stage III/IV melanoma patients who received anti-CTLA-4 (n = 81), anti-PD-1 (n = 25), or combination (n = 18) therapy as first line. The tissue analyzed was resected before treatment with ICIs. In total, 340 H&E slides were digitized and annotated for three regions of interest: tumor, lymphocytes, and stroma. The slides were then partitioned into training (n = 285), validation (n = 26), and test (n = 29) sets. Slides were tiled (299x299 pixels) at 20X magnification. We trained a deep convolutional neural network (DCNN) to automatically segment the images into each of the three regions and then deconstruct images into their component features to detect non-obvious patterns with objectivity and reproducibility. We then trained the DCNN for two classifications: 1) complete/partial response versus progression of disease (POD), and 2) severe versus no immune-related adverse events (irAEs). Predictive accuracy was estimated by area under the curve (AUC) of receiver operating characteristics (ROC). Results: The DCNN identified tumor within LN with AUC 0.987 and within ST with AUC 0.943. Prediction of POD based on ST-only always performed better than prediction based on LN-only (AUC 0.84 compared to 0.61, respectively). The DCNN had an average AUC 0.69 when analyzing only tumor regions from both LN and ST data sets and AUC 0.68 when analyzing tumor and lymphocyte regions. Severe irAEs were predicted with limited accuracy (AUC 0.53). Conclusions: Our results support the potential application of machine learning on pre-treatment histologic slides to predict response to ICIs. It also revealed their limited value in predicting toxicity. We are currently investigating whether the predictive capability of the algorithm can be further improved by incorporating additional immunologic biomarkers.