The Banbury Center Workshop

Cold Spring Harbor Laboratory
September 19-22, 2004

Workshop Report

The CDC CFS Research Program sponsored a workshop, Integrating Disparate Data to Simulate Lymphocyte Function, at the Banbury Center, Cold Spring Harbor Laboratory, on September 19-22, 2004. The objective was to discuss current knowledge concerning lymphocyte function and to identify means by which computational modeling could be used to understand how this complex biologic system functions in persons with CFS. The workshop brought together experts in immunology, molecular biology, computer sciences, and molecular modeling. Specific aims were to:

1. Define the types of laboratory and clinical data involved in the current concept of lymphocyte function in normal and abnormal states.
2. Present approaches for integrating genomic, proteomic, clinical, and epidemiologic data in such models.
3. Define the level of abstraction and types of assumptions necessary to create the next generation of molecular models.

Assessment of Lymphocyte Function

Monday, September 20

The first day of the workshop was dedicated to discussing current knowledge derived from computational modeling concerning lymphocyte function. We chose to focus the workshop on lymphocyte function because interactions between the brain and immune system likely play a central role in CFS, there is a wealth of information concerning lymphocyte function in health and disease, and because lymphocytes travel between the brain and periphery in both normal and abnormal states. The day was divided into two sessions: the morning focused on normal lymphocyte function and the afternoon on lymphocyte function in an abnormal state (persistent infection).

Lymphocyte trafficking into the brain was the first presentation. Under normal conditions, the central nervous system (CNS) is largely shielded from the peripheral immune system by the blood-brain barrier. However, small numbers of activated lymphocytes continuously enter the brain as part of normal immune surveillance. Lymphocyte activation can be initiated by antigenic stimulation (e.g., infection) and non-antigenic stimulation (e.g., peripheral nerve damage). Activated peripheral lymphocytes interact with perivascular and microglial cells of the CNS. Microglia are similar to macrophages and secrete substances that affect the CNS (e.g., tumor necrosis factor-alpha, nitric oxide, interleukin-1beta). In short, the brain responds to signals from the peripheral immune system, and ongoing low-grade, lymphocyte activation affects brain function.

The next several talks involved how lymphocytes become activated and migrate to target tissues. Lymphocytes, such as T cells, become activated when an antigen-presenting cell engages the T cell receptor (TCR). TCR engagement is controlled by the immunologic synapse (IS), an organization of proteins where antigen-presenting cells engage the TCR. The IS amplifies signals, sustains signaling over time, and coordinates engagement of 'secondary' receptors. Dysregulation at the IS level affects coordination of the immune response.

Lymphocytes must move and adhere to targets as part of immune surveillance and the mounting an immune response. The integrin family of adhesion molecules participate at all steps of this process. Integrins serve as signaling molecules and communicate allostatic changes in both directions across cell membranes, depending on affinity (low affinity allows inside-out signaling and medium/high affinity allows outside-in signaling). Avidity changes work with affinity changes to regulate integrin function. In addition to these functions, integrins are crucial for the speedy delivery of messages, in both directions, across spatially segregated surface receptors (i.e., in the IS).

Immune cell migration and lymphocyte interactions can be studied visually by dynamic imaging using confocal 2-photon microscopy. Immune cell migration and lymphocyte interactions were compared in two systems: liquid culture and 3-D collagen matrix. In liquid culture, lymphocytes cluster around antigen-presenting cells and stay this way for hours. This implies that sustained stable contact is necessary for efficient T cell signaling. However, in 3-D collagen matrix models (similar to tissues), cells are highly motile and only brief interactions occur, yet T cell antigen-specific cell proliferation results from repetitive, serial contact. The 3-D matrix observations were validated with intra-vital imaging of exposed lymph nodes in anaesthetized animals. The relative importance of serial versus prolonged contact remains unresolved. The need for prolonged contact may be associated with the nature of the antigen-presenting cell (e.g., dendritic cells are very efficient compared with B cells) and activation status (e.g., activated B cells are far more efficient in activating T cells).

Last, the last presentation discussed the role of psychosocial effects on lymphocyte function. The nascent field of pyschoneuroendocrine immunology attempts to quantify the psyche by measuring neuroendocrine and immune correlates of acute and chronic sickness behavior. Several excellent examples were given, demonstrating that chronic life stress increases sympathetic arousal and that acute stress, such as a math examination, modulates immune response to vaccination.

Summary of Day 1

We began by considering lymphocyte trafficking and movement in the context of the blood and the brain. We then heard elegant analyses for dissecting lymphocyte function in order to understand it. We ended with the reminder that lymphocyte function and its properties are dependent and best understood on the organization of the system as a whole. A detailed understanding of molecular processes and mechanisms, such as lymphocyte migration and antigen recognition, serve as fundamental building blocks in piecing together the complexity of the immune system. The detailed molecular description of mechanisms such as the immune synapse helps to define events involved in the induction of disease. Linking mechanisms and processes begins to get at the mode of action (i.e., interactions of key events and sequential processes) and are most effectively interpreted when observed and described in a systems context.

Implications of Day 1 for CDC CFS Research Program

These presentations highlighted the enormous complexities of lymphocyte functioning. Many different proteins and molecules participate in multiple steps to enable numerous physiologic processes, and small changes cause significant perturbations of sub-systems and the systems as a whole. Intensive study of one aspect (i.e., one type of lymphocyte or one body function) may not necessarily enhance our understanding of CFS system biology.

The ultimate test of any simulation is to test its validity in the living organisms and to integrate models of particular sub-systems into the larger picture. For example, activated lymphocytes that routinely enter the brain and products of activated macrophages/glial cells are the most likely culprits for causing CNS disturbances. However, this model does not account for other means of communication between the CNS and the peripheral immune system (e.g., the autonomic nervous system provides direct neural signaling between the periphery and brain). A complex/complete model must account for the effects of integrated messages arriving by both lymphocytes and neural routes.

Top of Page

Managing, Integrating, and Modeling Data

Tuesday, September 21

The second day of the workshop focused on bioinformatics, databases, and tools for managing and integrating different types of data and the computational approaches used to describe and understand complex biologic networks. Computer modeling is pivotal to understanding complex systems by validating laboratory observations and to predicting biologic phenomena because millions of simulated experiments can be done in the time it takes to complete a single laboratory experiment. Continuing the theme of Day 1, presentations discussed computer modeling of the immune synapse and molecular aspects of lymphocyte function.

The presentations on computer modeling of lymphocyte function and the immune synapse involved two types of biologic phenomena known as switch-like and oscillatory behavior. We learned that T cell synapses can be switch-like and serve to enhance true signals, filter out noise, and then reverse the signal once its function is done. Computer simulations were also used to demonstrate the oscillating behavior observed in biologic networks.

One type of computer simulation uses unified modeling language state charts to model T cell motion in the thymus and provides a powerful visual means of capturing complex behavior. In the case of thymocyte motion, a few simple rules of cell-cell communication describe how T cells move through a collagen matrix, recapitulating thymocyte maturation in a lymph node.

The next presentation noted that while general behaviors can be modeled with classical methods like differential equation models, realistic quantification and prediction of biologic networks requires discrete, finite-element, stochastic methods based on rules rather than differential equations. As examples:

1. Simulation of T cell synapses in FEMLAB (modeling and simulation software) because of the paucity of antigen-specific MHC-1/peptide molecules on the antigen-presenting cell.
2. A cell model of 1000s of compact spheres because of the break-down of the free diffusion assumption in crowded cell environments.
3. Simulation of the clonal expansion of T and B cells from a single cell.

Other presentations discussed databases, database tools, and strategies for turning high throughput, data into useful visualizations about how complex biologic networks function. For example:

1. ArrayTrack is a system for visualizing genomics data and integrating it with databases to do a functional analysis on significant results so that simulation efforts can be directed to a certain area.
2. Object model formats were defined so that disparate data, high throughput or otherwise, can be put together and mined in one large database.
3. Approaches were explained for text mining large databases of scientific abstracts and for building simulations of complex networks directly from the text-mining results by using state chart/reactive-animation approaches and artificial intelligence.
4. Artificial intelligence approaches could be used to mine large amounts of disparate data to discover clusters, patterns, and presumably cause and effect in a large dataset, which would inform not only what to simulate but, to some extent, how to simulate it.

Summary of Day 2

Well-designed databases are necessary to manage high throughput and disparate data and to construct valid complex computational models. The use of informatics, statistical physics, and mathematical models in integrating complicated, disparate data sets allow in-depth exploration of associations and connections among these variable sets. The more accurate and complete the data, the better the modeling.

Our knowledge of biologic systems can be expanded at a much quicker pace with computational models based on limited sets of rules that none-the-less display complex behaviors. Computational models are powerful tools that may suggest ways of how seemingly disparate data may fit together and provide guidance for new, more specific hypotheses. Solutions from simulated models must be examined for biologic plausibility, must be integrated with the appropriate biological and medical context for discovery of the most sensitive and specific markers and accurate predictions, and must be validated on patient samples.

Implications of Day 2 for CDC CFS Research Program

The CDC CFS Program has an incredible resource in the data that has been collected from population-based, model, and clinical studies of CFS. Key to using this data will be effective data management and integration and availability of analysis tools. Both qualitative and quantitative data integrated and interpreted in the appropriate context are necessary for effective public health intervention and disease mitigation.

Top of Page

Brainstorming Session

Wednesday, September 22

We began with a synopsis of presentations from the previous 2 days. The organizers then challenged participants to derive a solution for diagnostic marker discovery in CFS and formed 4 breakout groups. The groups were to describe how to use epidemiologic, clinical, and laboratory data from studies of CFS to understand that pathophysiology of CFS and identify markers. Organizers summarized the multi-system nature of CFS and the types of disturbances that have been observed and provided examples of data collected in the CDC CFS research program. They then posed the question, "Given a data set containing this information what solution would you derive for discovery of a diagnostic marker in CFS?" The groups presented the following suggestions:

Solution 1

1. Sub-grouping by gene expression - looking at patterns predictive of specific symptom patterns. For example, is the nature of fatigue and associated symptoms (as empirically measured by the multidimensional fatigue inventory [MFI] and Symptom Inventory [SI]) associated with a pattern of gene expression that would regulate muscle pathways. Using gene expression data, homogeneous groups could be identified for directed intervention studies.
2. Select individuals with the two major gene expression patterns, follow them daily by measuring MFI and SI and by measuring gene expression in blood. This would allow analyses to infer causality.
3. Flow cytometric sorting of populations of interest for gene expression studies (e.g., activated CD4 or CD8 subsets, the NK population). One might identify a subpopulation more "brain like" (e.g., NK cells) for the peripheral brain model. Assay other sample types, such as muscle. Starting from purified subpopulations would also improve the sensitivity of the analysis and subsequent pathway mapping and proteomics.
4. Mathematical modeling to simulate the network dynamics, suggesting a model that started with the brain sending out signals of chronic fatigue that would force an adaptive biologic response, which in turn would reinforce the perception of fatigue by the brain. In this type of modeling, the many two-directional arrows that we often observe between multiple immune variables, neuroendocrine variables, etc., would be taken into account and weighted, using modeling strategies that encompass a comprehensive literature review. Such an approach would allow the use of the entire Wichita data set to delineate the brain/periphery interactions.

Solution 2

1. Define structure in the data, using unsupervised classification on quantitative data only (i.e., gene expression values).
   1. Weed out irrelevant variance but do not get rid of it; down-weight the variance that is uninteresting (still may influence somewhere).
   2. Define relevant variance using supervised classification methods.
2. Partial least squares analysis using a filter, use a linear and simple model method.
3. Build a supervised classifier with filtered features.
4. Cross-validate using a leave-one-out analysis.
5. Model applied to test group.
6. Evaluate biologic plausibility and statistical significance.
7. This group also came up with the idea of issuing a challenge, coined C 3, for CFS Computational Challenge. Data from Wichita Clinical would be provided to willing participants, the objective would be to define genes and/or pathways that can serve as biomarkers and plausible mechanisms for disease pathogenesis; time constraints would be imposed and publications from each solution would be an outcome.

Solution 3

1. This is a solution based on the theoretical overview of etiology of post-infective fatigue: the three Ps, predisposing (sex, age, etc), precipitating (infections), and perpetuating.
2. Potential study designs to study the three Ps included:
   1. Population-based with stratified samples,
   2. Cross-sectional and study pattern recognition, and
   3. Compare and contrast findings and patterns following therapy (i.e., following CBT).
3. Conduct laboratory studies aimed at biomarker discovery:
   1. serum proteomics
   2. toll receptors that turn on proinflammatory cytokines
   3. examine stress markers, such as the NKG2D family of receptors

Solution 4

1. This group hypothesized that CFS is defined by multiple steady states and that the change in steady states may be defined by stochastic noise.
2. This group also focused on using the gene-expression data and on first defining the structure of the data, removing the noise, and determining if there are properties of multiple steady states inherent in the data.
3. This group suggests the use of Boolean networks, reverse engineering and crude network structure to reveal specific genes and pathways, with the rationale being that multiple steady states must be characterized by regulatory pathways.
4. This group would focus on mononuclear cell subset data on samples used for gene chip analysis. It may not be possible to purify subpopulations due to the small samples, but a simple analysis of CD3/CD4/CD8/CD19/CD11b/CD56 cells to identify ratios of T, B, NK cells and monocytes could be useful since some changes in gene expression may correspond to changes in subsets in the blood rather than changes in expression within these subsets. The gene expression profiles could be worked out for T, B, NK cells and monocytes from controls to get a feel for subtype specific gene expression to help determine what gene expression patterns are associated with normal cell types.

Top of Page

Meeting Participants

• Eric Aslakson
  Centers for Disease Control and Prevention
• Roumiana Boneva
  Centers for Disease Control and Prevention
• Gordon Broderick
  Institute of Biomolecular Design, University of Alberta
• Joseph G. Cannon
  School of Allied Health Sciences, Medical College of Georgia
• Arup K. Chakraborty
  Chemistry & Chemical Engineering, University of California, Berkley
• Daniel J. Clauw
  Center for the Advancement of Clinical Research, University of Michigan
• Richard C. Craddock
  Centers for Disease Control and Prevention
• Michael Dustin
  Skirball Institute of Biomolecular Medicine, NYU School of Medicine
• Sol Efroni
  Lymphocyte Biology Section, National Institutes of Health
• Jennifer Fostel
  National Center for Toxicogenomics
• Ben Goertzel
  Biomind LLC
• Matthias Gunzer
  German Research Center for Biotechnology
• Brian Gurbaxani
  Centers for Disease Control and Prevention
• William F. Hickey
  Dartmouth-Hitchcock Medical Center
• James F. Jones
  Centers for Disease Control and Prevention
• Jonathan R. Kerr
  Imperial College of London
• Nancy G. Klimas
  Miami Veterans Affairs Medical Center
• Anthony L. Komaroff
  Harvard Medical School
• Sri Kumar
  Information Technology Office, DARPA
• Andrew Lloyd
  University of New South Wales
• Elizabeth M. Maloney
  Centers for Disease Control and Prevention
• K. Kimberly McCleary
  The CFIDS Association of America
• Mangalathu Rajeevan
  Centers for Disease Control and Prevention
• Ann Rundell
  Purdue University
• Andrey S. Shaw
  Washington University
• Renee Taylor
  University of Illinois at Chicago
• Weida Tong
  Center for Toxicogenomics, FDA
• Elizabeth R. Unger
  Centers for Disease Control and Prevention
• Suzanne D. Vernon
  Centers for Disease Control and Prevention
• Ute Vollmer-Conna
  University of New South Wales
• Toni Whistler
  Centers for Disease Control and Prevention
• Peter D. White
  Queen Mary School of Medicine, St. Bartholomew's Hospital
• Troy Wymore
  Pittsburgh Supercomputing Center
• Lingchong You
  California Institute of Technology

Top of Page