Colon cancer is the third leading cause of cancer-related deaths in the United States in both men and women. A major clinical challenge is to obtain an effective treatment strategy for each patient or at least identify a subset of patients who could benefit from a particular treatment. Since each colon cancer has its own unique features, it is very important to obtain personalized cancer treatments and find a way to tailor treatment strategies for each patient based on each individual's characteristics, including race, gender, genetic factors, immune response variations. Recently, Quantitative and Systems Pharmacology (QSP) has been commonly used to discover, validate, and test drugs. QSP models are a system of differential equations that model the dynamic interactions between drug(s) and a biological system. These mathematical models provide an integrated “systems level” approach to determining mechanisms of action of drugs and finding new ways to alter complex cellular networks with mono or combination therapy to obtain effective treatments. Since QSP models are a complex system of nonlinear equations with many unknown parameters, estimating the values of the model's parameters is extremely difficult. Existing parameter estimation methods for QSP models often use assembled data from various sources rather than a single curated dataset. These datasets are usually obtained through various biological experiments, in vitro and in vivo animal studies, thus rendering QSP models hard to be practicable for personalized treatments. To the best of our knowledge, no QSP model has been developed for personalized colon cancer treatments. In this project, we propose a unique approach to develop a data-driven QSP software to suggest effective treatment for each patient based on gene expression data from the primary tumor samples. Since signatures of main characteristics of tumors, such as immune response variations, can be found in gene expression profiling of primary tumors, we use gene expression data as input. We develop an innovative framework to systematically employ a combination of data science, mathematical, and statistical methods to obtain personalized colon cancer treatment. We will use these techniques to propose an optimal treatment strategy for each patient and predict the efficacy of the proposed treatment. The model might also suggest alternative therapies in case of low efficacy for some patients.
Just following-up from your ITCR talk today, which I enjoyed a great deal. With respect to the issue of having many more parameters than you have equations, have you considered using a hierarchical modeling approach? For example, if you have a parameter which plays a similar role in each equation, let's call it beta (say), then you might be willing to assume that those beta values are all sampled from a Normal (e.g.) distribution with mean mu and std. dev. sigma. This allows the estimates of the individual level parameters (the betas here) to borrow strength from each other. Additionally, by placing a prior on the hyper-parameters, you can reduce the dimensionality of your problem. (e.g., by placing a prior on sigma that forces it to be small, you can force the betas to all take similar values).
Apologies if you are already well aware of this, (I didn't catch the introduction, so don't know what your expertise is). But if you are not aware of these approaches, and they sound like they might be useful, I'd be happy to talk further about it with you. I have an ITCR R21 myself, and can be reached at pmarjora@usc.edu
Cheers, Paul Marjoram (USC)
Just following-up from your ITCR talk today, which I enjoyed a great deal. With respect to the issue of having many more parameters than you have equations, have you considered using a hierarchical modeling approach? For example, if you have a parameter which plays a similar role in each equation, let's call it beta (say), then you might be willing to assume that those beta values are all sampled from a Normal (e.g.) distribution with mean mu and std. dev. sigma. This allows the estimates of the individual level parameters (the betas here) to borrow strength from each other. Additionally, by placing a prior on the hyper-parameters, you can reduce the dimensionality of your problem. (e.g., by placing a prior on sigma that forces it to be small, you can force the betas to all take similar values). Apologies if you are already well aware of this, (I didn't catch the introduction, so don't know what your expertise is). But if you are not aware of these approaches, and they sound like they might be useful, I'd be happy to talk further about it with you. I have an ITCR R21 myself, and can be reached at pmarjora@usc.edu Cheers, Paul Marjoram (USC)