New directions for deep learning in cancer research through concept explainability and virtual experimentation.

Projektbeschreibung

Werkzeuge zum Deep Learning für die Krebsforschung in der Biomedizin

Deep Learning bedeutet für die Krebsforschung eine Revolution, denn molekulare Informationen können aus Bilddaten extrahiert werden. Die Wirksamkeit ist jedoch begrenzt, da es rein deskriptiv ist und nicht mit Wissen zu biologischen Mechanismen verknüpft ist. Finanziert über den Europäischen Forschungsrat werden im Projekt NADIR Deep-Learning-Modelle genutzt, um neue biologische Mechanismen vorherzusagen, zu verifizieren und sogar zu entdecken. Durch Integration von Programmierung, medizinischer Bildanalyse und biomedizinischem Engineering wird das Team Deep-Learning-Werkzeuge entwickeln, die biologische Konzepte extrahieren, biologische Mechanismen aufklären und mechanistische Hypothesen aufstellen und prüfen können. Der Hauptfokus bei NADIR liegt auf Wechselwirkungen zwischen Tumoren und dem Immunsystem bei Darm- und Magenkrebs. Die Werkzeuge sollen über Bildungs- und Öffentlichkeitsprogramme der Krebsforschung im Bereich Biomedizin zur Verfügung gestellt werden.

Ziel

Deep learning (DL) is rapidly transforming cancer research and oncology. DL can extract subtle visual features from preclinical and clinical image data. In my junior research group, I have developed end-to-end DL methods to predict molecular biomarkers and clinical outcomes directly from histopathology slides. Because histopathology slides are ubiquitously available for any patient with a solid tumor, DL is a broad tool for translational studies, enabling researchers to extract molecular information and make predictions about clinical outcome.
However, the potential of DL in cancer research is fundamentally limited because it is purely descriptive and, in many cases, a black-box system. Also, DL is currently disjoint from the vast amount of biological mechanistic knowledge in cancer research, and from the world of experimentation. In NADIR, I will close this gap. My hypothesis is that DL models can not only make predictions but can be used to verify
existing biological knowledge and to make new mechanistic discoveries. The main tools that allow me to address this are concept explainability and counterfactual virtual experimentation. For both, there exists a nonmedical proof of concept, but no systematic biomedical application yet. I approach this problem as a biomedical cancer researcher with training in programming, medical image analysis, and biomedical engineering. As such, I will develop DL systems that can extract biological concepts, can elucidate biological mechanisms, and can be used to create, and answer, mechanistic hypotheses. NADIR’s tools will be synergistic with and can be used together with other biological high-throughput experimentation pipelines such as transgenic animal experiments or tumor organoid cultures. The main use case of NADIR is focused on tumor-immune interaction in colorectal and gastric cancer, and through the educational and outreach program in NADIR, it will be made available as a general tool for cancer researchers in biomedicine.