Can one use microarray measured gene expression (MMGE) data to predict pharmacological and toxicological phenotypes in animals and man? Can similar data be used to account for a significant portion of the individualized variability in patient drug response? In collaboration with the Computational Biology Group at Roche Palo Alto, we are evaluating and contrasting the computational approaches that have been proposed to address these questions. Two large sets of MMGE data are being used to compare and contrast the information that results from use of the various available statistical pattern extraction, clustering and evaluation tools and techniques. Our current focus is on a variable selection and testing strategy that is particularly driven by directly relevant archival phenotypic data, information and knowledge.

Key personnel: Yuanyuan Xiao and Donglei Hu.

Collaborators: Mark Segal, Steve Shiboski, Michael N. Liebman, and Roche Palo Alto.

This project is part of a larger undertaking that will assemble, test and validate a prototype decision support framework (DSF). We are using both informatics and simulation strategies to create a "mirror" population of hypothetical patients that enables linkage of the information from highly multiplexed molecular analyses through available systems knowledge to patient clinical data so that expected changes in treatment outcomes for new patients can be visualized (on a computer) interactively as various options and/or new data are considered. The DSF is intended to function as part of an institution-wide, patient centered system. With such a system, the patient, the physician, and researchers will be able to detect important, decision impacting relationships contained within the cancer research and clinical trial databases relationships that otherwise might remain undetected. The DSF will also allow the decision makers, the patient and physician, to ask a variety of probing "what if"? questions, and to visually compare and contrast the answers graphically represented, with expected treatment results in real time. Having such a capability will allow customized (and, ideally, optimized) individual treatments, thereby lowering costs while improving both outcomes and the patient's quality of life.

Project contact: C. Anthony Hunt

Collaborators: Oztech Systems, Management Sciences Associates, Dan H. Moore, Joe W. Gray

This project is a breast cancer specific realization of the project "Toward Genotype Phenotype Linkage to Optimize Cancer Treatments." How can emerging genetic and specific molecular information on breast cancer progression be linked to outcomes associated with different treatments, so that the individual needs of the patient physician interaction can be optimally coordinated? There are critical gaps between the collection of biomedical information (biological and medical informatics) and its most useful application for impacting selection of treatment options. Our goal is to provide an integrated informatics system focused on the key issues in breast cancer treatment that places decision making power in the hands of the clinician, while providing outcome probabilities for the options available to the patient. The DSF (decision support framework) design anticipates making optimum use of such emerging technologies as microarray measures of gene expression. The immediate goal is to demonstrate the feasibility of a DSF that will enable systematic individualization of breast tumor treatments. A computational database is being developed that represents a "mirror" population of breast cancer patients. The core experimental database includes over two dozen prognostic (i.e., breast cancer markers, such as ER status, Her2, etc.), risk, medical, and key patient factors for several hundred patients, plus results from recent pharmacogenetic clinical trials. Each new patient can be matched with a small set "surrogates" within the DSF database, which allows one to ask many "what ifɺ" questions about treatment options and possible outcomes. The next step is to apply and develop algorithms to predict uncertainty throughout the decision making process. In part, our effort is similar to the application of Bayesian Networks to breast cancer. The project anticipates a future clinical setting where breast tumor treatments can be individually optimized, errors will diminish, outcomes will improve, and there will be considerable cost savings for the individual and for society.

Key personnel: Christine Case

Collaborators: John W. Park, Laura J. Esserman, Laurence H. Baker, and Dan Moore.

Stochastic Activity Networks (SAN) may have advantages over traditional modeling approaches for understanding the complex processes. We are using this tool to better understand the interacting biochemical events involved in the coagulation and fibrinolysis processes that cause Disseminated Intravascular Coagulation! (DIC). We use the UltraSAN2 software environment to simulate these processes. Within the SAN we can inactivate or alter certain proteins to mimic genetic defects and disease, and then compare the results to clinical observations for validation. We are developing this simulation project through 7 steps: 1) trypsinogen activation and trypsin inactivation, 2) normal blood coagulation, 3) disorders of blood coagulation, 4) normal fibrinolysis, 5) disorders of fibrinolysis, 6) normal hemostasis, and 7) DIC. Results of the successful simulations of steps 1?3 by Mounts and Liebman [1]3 provides a promising foundation. A combination of the SAN that separately represents of steps 2 and 4 is expected to adequately represent normal hemostasis. We will systematically inactivate sets of proteins in the final, linked SAN to generate results that are consistent with the clinical observations of DIC. The final, linked SAN will be used to assist in the identification of new therapeutic targets, suggest strategies for development of new diagnostics, and provide a network hypothesis for connecting to emerging genetic information.

Key personnel: Keith Erickson

Collaborators: Michael Liebman and Adam P. Arkin

Mounts, W.M. and Liebman, M.N. Qualitative Modeling of Normal Blood Coagulation and its Pathological States Using Stochastic Activity Networks. Intl J of Biological Macromolecules 20:265-281 (1997).

We are systematically evaluating the applicability of graph theoretic techniques to facilitate utilization of large pharmacological, genetic and physiological system models. The conceptual or computational model (such the above Stochastic Activity Networks) in is first represented as a network graph having nodes and edges with respective weights. Computational graph analysis is then used to bring critical elements into focus by grouping closely related system elements prior to finding and ranking elements or paths between elements that have the greatest influence on the system. When applied to pharmacological and physiological system models, such "critical elements" represent potential drug targets. By modulating the function of these elements or specific interactions between elements, we expect to be able to identify the process or event where intervention is most likely to have the greatest effect on potential outcome functions, such as drug or treatment effectiveness or undesired side effects.

Key personnel: Abraham Anderson

Collaborators: Richard Ho, Director of Medical Informatics, R.W. Johnson Pharmaceutical Research Institute.

There is a critical need for new innovative approaches to understanding and modeling interindividual variability and how it impacts development of new therapeutic strategies, e.g., new drugs, as well as the utilization of existing therapeutic strategies. Individual patient variability, such as cardiovascular status, pharmacogenetics, etc. are known to contribute to pharmacodynamic variability. Such variability is difficult to capture using traditional modeling strategies. We are successfully using a novel approach that uses an agent-based programming platform (e.g., SWARM.) The approach can provide an in silico patient population that responds appropriately to drugs based on individual patient attributes. While current approaches in the field rely heavily on complex, often stiff statistical models that are drug specific, our in silico approach offers the ability to model random biological fluctuations inherent in human populations along with a program which may be easily modified to test an array of therapeutic products. The long-term and potential groundbreaking benefits of such an undertaking include targeting appropriate populations for drug safety and efficacy trials, and having a tool that can be used to optimize drug regimens for specific patients.

Key personnel: Amina A. Qutub

Collaborators: Alex Lancaster

We made significant progress in developing a method of mapping large datasets onto a two dimensional surface where every point corresponds to an Iterated Function Systems (IFS). These IFS can be used to generate an long sequence of scalable arbitrary numbers. There is a correspondence between the similarity of these sequences and their location on the surface of the map. The values contained in these generated sequences is compared against sequential values of target datasets. A point that corresponds to the "best fit" to each target dataset is then located. One can quantify the relationship between similarities of target datasets and the location and geometry of their corresponding points on such a map with a view to optimizing the technique for visualizing relative similarity (or difference) between large sets of gene expression data. Sandy Shaw leveraged the progress made into a new company, Fractal Genomics.

Key personnel: Sandy Shaw

Collaborators: Jenny Harrison (Math UCB).

Can one prolong graft survival in transplant patients by using novel therapeutics oligonucleotides to decrease the ability of the graft to trigger the host immune system. Such a strategy is in contrast to the more common approach involving host immunosuppression. We are investigating a series of oligonucleotides that were designed to downregulate the expression of MHC class I and class II genes. In particular, we are interested in the extent to which constitutive and induced expression of MHC genes can be reduced by therapeutic oligonucleotides, and the resultant effect on the host immune response. In Project One, the Hunt Group is implementing strategies to acquire, evaluate, and use microarray measured gene expression to facilitate the above approach.

Collaborators: Richard Shafer

Epithelial cells perform essential functions throughout the body, acting as both barrier and transporter and allowing an organism to survive and thrive in varied environments. Although the details of many processes that occur within individual cells are well understood, we still lack a thorough understanding of how cells coordinate their behaviors to create complex tissues. In order to achieve deeper insight, we created a list of targeted attributes and plausible rules for the growth of multicellular cysts formed by Madin-Darby canine kidney (MDCK) cells grown in vitro. We then designed in silico analogues of MDCK cystogenesis using object-oriented programming. In silico components (such as the cells and lumens) and their behaviors directly mapped to in vitro components and mechanisms. We conducted in vitro experiments to generate data that would validate or falsify the in silico analogues and then iteratively refined the analogues to mimic that data. Cells in vitro begin to stabilize at around the fifth day even as cysts continue to expand. The in silico system mirrored that behavior and others, achieving new insights. For example, luminal cell death is not strictly required for cystogenesis, and cell division orientation is very important for normal cyst growth.

Project Contact: Jesse A. Engelberg

Collaborators: Anirban Datta, Keith E. Mostov

In Silico MDCK Analogue Supplementary Material:

1. Source code (download)
2. User manual