Preclinical cancer models, such as tumour-derived cell lines and animal models, are essential in cancer research. Consistently used as a platform to investigate mechanism of action, they can identify potential biomarkers prior to clinical trials where similar exploration is more complicated and expensive. However, whilst cell lines are the most used preclinical model, their applicability in certain settings is questioned because of the difficulty of aligning the appropriate cell lines with a clinically relevant disease segment. I developed a methodology for systematic cancer cell line scoring based on patient sample subtypes and analysis of the causative elements of the subtype differentiation in cancer.
Machine learning classifiers I tailored to multi-omics nature of cancer have been highly accurate in predicting the subtype of new patient samples. Applying those models to cancer cell lines reslted in a clinically based cancer cell line relevance score. The majority of cell line scores were in line with the literature, but there were several misclassified cells.
Exploring the causative elements of the underlying biology, I confirmed the oncogenic nature of the features driving the classification. Additionally, through differential expression analysis, the nature of some of the misclassified breast cancer cell lines was elucidated–they were poorly representative of their receptor-positive type despite having HER2 receptor expressed. One of those cell lines, JIMT-1, has been shown to be resistant to HER2-targeted treatment, thus making the misclassification of my model more clinically relevant than the receptor statuses of the cell line itself.
Through several distance metrics I have expanded on the binary nature of the classifying methods and identified more and less suitable cell lines not just by their score, but also by how close they are to the patient samples.
The core aspects of my methodology have been implemented as an online tool, a Shiny application, in order to allow others to leverage my methods and findings.
