Interpretable Biomarker-AI-based prediction of tumour radiosensitivity for future personalised cancer treatment
Project leaders:
Sona Michlikova (OncoRay, Dresden) sona.michlikova(at)uniklinikum-dresden.de
Dr. María José Besso (DKFZ, Heidelberg), mariajose.besso(at)dkfz-heidelberg.de
Funded since: July 2022 for 3 years
The development of patient-specific individualized treatment strategies is one of the main goals of modern clinical radiation oncology in the area of precision medicine. Consequently, there is an urgent need for the development of prognostic and predictive biomarkers for reliable patient stratification. Head and neck squamous carcinoma (HNSCC) is the sixth most common cancer type worldwide that comprise a very heterogeneous group of solid cancers with varying treatment success. Resistance against radiotherapy is tightly linked to the composition of the tumor microenvironment (TME).
The combination of different data types allows the extraction of meaningful biomarkers with increased robustness and translatability. In this project we aim to integrate signatures derived from histopathology images and molecular Omics data. Histopathology images are most frequently used in clinical cancer management. Artificial intelligence (AI) methods have been employed to detect tumor tissue, identify tumor subtypes, and predict the mutation status or patient survival. However, common bottlenecks for clinical translation are image biomarker reproducibility, model overfitting, biased data and lack of biological interpretability. Tumor samples are being sequenced to infer the course of the disease and the potential therapy outcome. The limitation is that any spatial information is lost, which means that low-expressed markers or markers present only at a specific site of the tumor are overlooked. By relating the molecular signature to histology-derived characterization of the TME, this challenge can be addressed.
This project aims to identify prognostic and predictive biomarkers with increased accuracy by integrating image-based (OncoRay) and molecular (DKFZ) features. Conventional radiomics, deep learning, and bioinformatics approaches will be applied to predict radiosensitivity and induced residual DNA damage in preclinical HNSCC models based on immunohistochemistry (IHC) and immunofluorescence (IF) images, combined with transcriptomics, epigenomics, and genomics data. Results will be translated to patient samples and assessed for biological interpretability.
