Customization of Treatment for Cancer Patients: An Engineering Approach

Cancer is a disease associated with uncontrolled cell proliferation or reduced cell death, either of which can lead to tumorigenesis. A possible route through which cancer can develop is by breakdowns in the signaling cascade of proteins at the cellular level. Since there are many ways in which such breakdowns can occur, anti-cancer chemotherapeutic drugs show varying degrees of efficacy in different patients. Thus, there is an urgent need to personalize the drug treatment regimen for better response to treatment while trying to reduce the side effects of these drugs. One way to meet this need would be to try every possible drug combination on cell lines extracted from a patient and find the combination with the least number of drugs in the mix but providing the best possible output. Although this method may work it is tedious and time consuming as the number of combinations increase exponentially with every new drug that is introduced into the repertoire. First, we consider the problem where the tumor is homogeneous in nature but the mutations within the mutated cells are unknown. We use Boolean network models with monotonicity properties to reduce the number of test cases, while still getting the best possible combination with the least number of drugs in the mix. This approach is efficient both in terms of time required and the costs involved. This method has also been applied to both simulated and real-world data collected from fibroplasts using qPCR to demonstrate the usefulness of the method. Another important area of study in cancer research concerns the heterogeneous nature of tumors. The clonal evolution of tumors is the driving force leading to heterogeneity in cancer tissues. Thus, in order to customize the treatment of cancer we need to be able to better model the heterogeneous subpopulations in the tumor. This can be done by estimating the impact of the various sub-populations and by modeling the interplay of various sub-populations within the heterogeneous tumor. Prior works in the literature have already addressed the problems of estimating the proportion of the sub-populations within a tumor and of modeling the interaction between the various sub-populations. In this work we present a way to improve the accuracy of the Bayesian hierarchical model which helps in estimating the proportional breakup of the tumor population. Additionally, it looks at ways to use the knowledge of the proportional breakup of tumor subpopulations and the interplay between the various subpopulations to help customize the treatment for the patient by making use of evolutionary game theory. We demonstrate the improvement of the presented methods as compared to the existing Bayesian hierarchical model by applying these techniques to qPCR and fluorescent data. Finally, the problem becomes more challenging when the nature and the number of the subpopulations are variable and difficult to estimate. In this work, we present a feasible way to find the best possible drug combination for such a scenario by training two neural network models on synthetic and real-world cancer data. Then we test each model, to verify their effectiveness and to demonstrate their usefulness in choosing the appropriate combination therapy. The models were evaluated on synthetic qPCR data and fluorescent data obtained from experiments. The results obtained from these methods take us a step closer to the realization of customized treatment for cancer patients. This will not only make the treatment more effective but also help reduce the side effects of the drug treatment