Cancer is a mutation driven disease, that has to escape immune surveillance if it is to develop into a full fledged disease. Cancer is therefore naturally patient specific and there is a great deal of interest in developing methods to tune cancer treatment for the patient, rather than giving a generic treatment for each cancer type. In this talk I will discuss two projects in my group that try to develop ways to predict the properties of cancers by using tumor specific data. My research group, in a collaboration with Daniel Gustafson and Dawn Duval's groups in Veterinary Medicine has been studying whether publicly available transcriptomic data of cancer cell lines contain enough information to reliably predict the drug sensitivity of patient tumors. As a first step towards this goal we analyzed whether models built on a few cell lines do a good job in predicting the drug sensitivity of the rest of the cell lines. We built both linear and non-linear models, in particular utilizing PC regression as well as neural networks and support vector machines. We found, rather surprisingly, that PC regression outperformed both nonlinear methods in predicting cancer drug sensitivity. We also found significant heterogeneity in the results for different drugs, suggesting that precise details of the mechanism of action of the drugs may be significant. For some drugs the correlation was significant enough to merit further investigation for therapeutic use. The second project involves studying the characteristics of cell shape of cancer cells in two dimensional culture. Using Zernike moments and geometric parameters to represent cell shape, we compared the shape characteristics of four invasive cell lines with a corresponding less-invasive parental line on three substrates. We find that shape
differences are robust enough to enable a neural network to classify cells accurately as belonging to the highly invasive or the less invasive phenotype. The patterns of shape changes were also reproducible for repetitions of the experiment. Our paper strongly suggests that shape may provide a means to read out the phenotypic state of at least some cell types, and we also show that shape analysis can be usefully performed using a Zernike moment representation.