Pan-cancer, morphology-based prediction of therapy response to immune checkpoint inhibitors using deep learning

点击翻译按钮获取中文摘要

Malignant transformation of somatic cells is accompanied by genetic changes that often lead to the expression of neoantigens and qualify tumor cells as a possible target of an adaptive immune response. Tumors can only grow and manifest clinically if they develop immunological escape mechanisms under the constant selection pressure by the patient’s immune system. Immune checkpoint inhibitors (ICI) aim to interrupt these escape mechanisms and restore functional anti-tumor immunity. Although ICIs are extremely effective in some patients, the average response rate is only about 13% and there are still no reliable biomarkers for predicting a therapy response to ICI.The aim of this project is to predict the therapy response to ICI using artificial intelligence (AI) and machine learning (ML). Various deep learning (DL) models will be trained on histopathological and immunohistochemical image data to predict response to treatment. The hypothesis is that various stages of anti-tumor immunity have distinct morphological features and can thus can be detected by the intelligent algorithms on a microscopic scale. Molecular changes relevant to the response to ICI, such as microsatellite instability and tumor mutation load, have already been successfully predicted using DL methods. In addition, there are a number of morphological categorizations of the tumor-associated immune reaction, each of which responds differently to ICI.We will first train a series of specially established network architectures on a retrospective cohort of 551 patients who were treated with a PD-1 / PD-L1 inhibitor in the period 2015-2020 (retrospective training and validation cohort). The six tumor entities most frequently treated with ICI at the Mainz University Medical Center are included in the study. The target variable will be the best objective response according to RECIST. We will then prospectively check the predictive accuracy of our DL models in all patients with the above-mentioned tumor entities who are treated with ICI during the two-year funding period (prospective test cohort). Furthermore, we will identify histopathological features that influenced the decision-making of the DL model with the aim of identifying new morphological characteristics that determine a therapy response to ICI.