Return to 2012 MSM CONSORTIUM MEETING
Panel Discussion 2: Can computational biology facilitate precision healthcare (MONDAY 12 noon – 1 pm)
1) Big Data/Correlative Data Mining and its role in Systems Biology (Discussion Leaders: James Schwaber, Xiaobo Zhou)
What are the challenges and opportunities in incorporating big data into a systems biology framework of disease?
What are potential solutions for the problem of heterogeneous data base structures related to patient-specific and disease-specific data bases?
2) Prediction of the dynamic patient state. (Discussion Leaders: Yorum Vodovotz, Scott Diamond)
How do we "anchor" models of the in silico patient state to the bed side patient?
How do we score the patient in a multiscale n-dimensional space?
3) What would FDA-approved Systems Biology models look like in the future? (Discussion Leaders: All)
Are standards possible for applying a model in a clinical context?
Do certain disease categories, like cancer, provide guidance to this issue?
Yoram Vodovotz noted the following:
In my opinion, the Precision Medicine document misses an important opportunity to weave multiscale modeling into the fabric of the proposed "new disease taxonomy." I would note that Gary An has stated much of the basic vision of using modeling for dynamic knowledge representation about (patho)physiological processes at multiple scales (and creating knowledge repositories as outlined in the Precision Medicine document) in a paper he published in Science Translational Medicine (G. An. Closing the scientific loop: Bridging correlation and causality in the petaflop age. Science Translational Med. 2:41ps34, 2010). For those that are interested, Gary and I, along with collaborators, have published two review/perspective articles that I think are relevant to this discussion. The DDR paper places MSM in the context of “parallelizing” the drug development process and touches on some of your points regarding in silico clinical trials as well as individualized patient models. The J. Crit. Care paper describes the emerging concept of the “dynamic patient state”, which I believe is the MSM answer to the Precision Medicine document’s attempt at going beyond disease taxonomy. We could also consider this figure from one of our recent papers, describing how one could obtain the appropriate types of data (reflecting disease dynamics and including patient to patient variability), and analyze the data with combined data-driven and mechanistic models File:From data to knowledge.pdf
Considerations for Panel 2 Discussion of computational biology:
- 1) computational biology already has successes with analysis of high-dimensional data and precision medicine, especially in diagnostics (pathology) in correlational results, well short of modeling. Most medicine is essentially correlation and clinicians are very comfortable with this.
- 2) bridging from correlational results to edges-nodes-topology is hard and lacks well established procedure. The disconnect of data-driven approaches and known network topology modeling is an issue that arguably should be an important focus in computational biology. As Denis Noble and Sydney Brenner note, making a model is more than only bottom up, top down constraints are essential, or as they say "middle out".
- 3) It is common to hear those working on high dimensional data wonder if the complexity in biology exceeds what the human brain can grasp, that our need to reduce and simplify is a mismatch to the actual systems observed.
- 4) models that interact with high dimensional data arguably need to be networks rather than "pathways" or "deterministically causal chains”: networks in which the function arises from interactions of elements in the network, the network can take alternative configurations with new or different members, and individual members can play different roles in different networks.
- 5) computational biologists focused on analysis of high dimensional data are very positive about modeling, have very high hopes for the understanding and target identification that it is expected to bring. At the same time they tend to be naive e.g. about time line, manageable dimensions, data requirements for different styles of modeling.
How to "anchor" models of the in silico patient with the bed side patient?
(SLD comments) 1) Large MSB Models become: Smart & Dynamic Diagnostic/Therapeutic Trees 2) Accurate prediction of homeostatic states (steady state) or Initial Condition. 3) Prioritize biomarkers to anchor Main Drivers of dynamics 4) Reducing n-dimensional state to LOW-space (patient bins) 5) Reducing ensembles of models to most useful models of disease 6) Ensemble of in silico patients: selecting the useful ones (Perhaps in advance). 7) How to annotate the therapeutic actions/data into a “live, executable data set”
from bill lytton, suny downstate: @disdain -- i've noticed over the years that sadly physiology, anatomy, biochem etc -- all basic sciences -- have been gradually downplayed in med ed in favor of ... hmm how shall i say -- family med type stuff (that's not PC) -- warm&fuzzy stuff .. ironically modeling is or should be the most holistic .. so what is my question?-- 'how do we reverse the neglect of basic sci and reassert it's relevance to clinical med?'
SCRIBE TRANSCRIPT OF SESSION 2 DISCUSSION
DIALOG DURING PANEL 2 (SCRIBE NOTES)
QUESTION 1 Big Data vs. Systems Biology.
Jim Schwaber: I want to propose a controversial viewpoint: How useful is big data (genomics and genotyping)? There exists a potential of big data to have an impact in areas of diagnostics or stem cell research. But a big challenge is how to bridge the “genotype” vs. “phenotype” in disease and treatment. And why is there disdain between big data/genomics scientists and multiscale systems biology modelers?
Fundamental Issue: Can we predict phenotype from genotype? How far are we from meaningful accurate and useful predictions? What are the guiding approaches to link genotype to proteome to phenotype?
As an example, consider Liver Consortium in Germany (and the European debate of “How to have multiple length scales vs. study at only the level of genes. Even Sydney Brenner appears to be coming around to the role of multiple length scales.”)
Also, recognize that real clinical trials that cost $500M to $1B are routinely simulated for estimation of benefit, risk, and statistical power based on models of pharmacokinetics(PK)/pharmacodynamics (PD)/patient population variability. Systems Biology extends classical PK/PD approaches, especially when predicting the dynamic patient state.
HOWARD: I have a Stat Mech point of view and synthetic biology. An experimental configuration can often has “entanglements” (i.e. limitations of context or interpretability). We ran a study in our lab of an intracellular process inside a cell to predict transcription based on sequence within a bacteria. This is a clear example of genotype predicting phenotype. So in a specific context of bacteria, it is possible to link genotype to phenotype.
RAJ (Jefferson) We like to put the network in the middle, between big data and the systems biology/physics based modeling. This often helps.
YORAM: Good data is needed. We may not really need every single piece of data. Big Data may be very expensive, less useful than desired, and obscure the real mechanisms and regulation critical to disease progression or treatment. Ideally, we want to get disease data that is dynamic. Embrace subject-to-subject variability. Good data allows for data-driven quasi mechanistic analysis. We want to infer regulatory mechanism, new hypotheses, and prioritize biomarkers.
JASON: Take note of the “Culture GAP.” What is the scale of the problem! Going from genotype to phenotype is a really big problem. There is an alternative: What are the general operating principles of genetic networks. What are the principles of robust networks?
Solving the inverse problem is really hard. One example is PluriNET for postulating topologies
Xiaobo NextGEN sequencing and imaging informatics are a part of systems biology. Quantifying images presents data that spans scales but is often hard to archive and execute algorithms upon. See WIKI for additional comments: Gene Markers to Clinical Phenotypes….Discovery of mutated genes…Incorporation of biomarkers remains hard in mechanistic models. Opportunity to use pathway model into drug targeting.
QUESTION 2 (How to anchor a MODEL to a dynamic patient state)
YORAM: The dynamic patient state requires moving beyond the static state (genotype). It also requires a recognition that binning patients may obscure important mechanisms and opportunities. Each patient has a trajectory that is high dimensional. Binning loses that resolution. For example, if is even hard to bin single cells or a cell population into single state or bin, how can we expect do this for a patient. Patient dynamics may not even play out at the level of the genes.
There is an opportunity to sue biomarkers as a way to bypass the need to do gene analysis. Biomarkers may be a shortcut to quantify phenotype and dynamic trajectories of the patient state.
COMMENTS: Disease is dynamic. Phenotype is dynamic. Difficulty of taxonomy (binning). Dynamic models still need to accurately predict initial condition or unperturbed state as a way of validation. Predictive and dynamic models may eventually drive the patient-tailured therapy, and drive event-driven adjustments to therapy.
Example: Epilepsy requires a model of seizure: dynamical vs. onset prediction. Real potential to make a difference.
RAJ: Tailoring to a patientm may not be what a clinician is thinking or wanting. What about opportunity to break paradigm? Identifying new measurements to make. Help refine what the exact data is needed for collection. Help refine clinical trial. Help refine measurement. Add to mechanistic causality-based knowledge. This is more powerful than correlational data.
Some of the resistance of the clinical community may be skepticism about predictive power of physics based models. There is a need to prove that a model will make a difference to the patient. Any examples ?? It is time to walk out of your doctor’s office when they “In my experience”: Their judgement is often weighted by outliers and most recent experience. In contrast a clinical trial is evidence-supported on a statistically relevant experience.
In Systems Biology, there are different types of model. Need to ask: what is our objectives?
There are 2 objectives: (1) optimizing treatments…help recommend dosages, risks, etc. (2) fundamental science: New knowledge that is mportant to understand. Curiosity. May not be immediately useful. Example, scientists studying flower coloration discovers mechanism of siRNA. Hard to know how and when basic science knowledge is extraordinarily important.
Parable: Russian vs. American Engineers building tunnels: Americans use GPS and connect tunnel in the middle. Versus tunneling from both sides and perhaps missiong: Well then you have two tunnels. Power of unexpected benefits of doing something in a particular way.
Comments on GWAS: GWAS: genotype to phenotype (small number of SNPs) This is best case. GWAS: 100s or 1000s of SNPs linked to disease. This is really not too useful. GWAS: A single SNP gives 2 % increased risk. This also is not too useful.
TONY: With large longitudinal studies with GWAS correlations of SNP to disease, they are actually hard to repeat. May get different result, even if you come back to the same cohort.
SANDRA: Genes are one thing. But genes are dynamically interacting with phenotype. We don’t know how. Gene expression changes phenotype, phenotypes change gene expression. Hard to model the dynamics of the genes at the cells, organ, organism level.
Question 3) FDA approved systems biology
From a regulatory point of view, a model looks like a Device. Just as a diagnostic assay is regulated like a Device. FDA evaluates internal algorithms in Devices (example: pump controllers) all the time.
If a model like Watson changes as the data on the internet changes, then it can’t become FDA approved because it will produce different results over time. This is a limitation of Watson. An approved piece of software has to unchanging; it produces the same result everytime. A look up table (LUT) is an excellent way to archive the results of a model, even if the LUT is a hundred different tables. A LUT is also computationally fast in that a query produces an immediate result.
FDA: A clinical study is actually underway to test a computational model as a medical device. The challenge is proving outcome. What is the bar to determine success of the model. In some cases, the trial requires the use of an invasive Gold Standard to check the model predictions. May need to define a “standard dosing” cohort vs. a “model-derived” cohort in order to compare the model to a preexisting standard of care. Ideally outcome can be measured quantitatively in order to score the accuracy of the model. The possibility may include the predicting biomarkers as an auxillary output of the model may be a surrogate metric of a clinical trial, used in conjunction with patient outcome.
What if we feel a model is ready for prime time? What is the first step? FDA guidance would be helpful and may already exist with respect to regulation of biomedical devices. Use biomarkers and a mathematical transformation.
A model may influence the design of a clinical trial (dosing regimen, patient numbers), but it doesn’t matter from FDA point of view because the trial is prescribed and outcomes are measured, independent of the systems biology model.
Consider MedCALC. This is an App on iphone and it goes to a diagnosic issues for clinicians standing in front of a patient. Is MedCALC FDA approved? Probably not.
How to apply to individualized medicine to a FDA trial? Precision medicine (bins of patients) vs. individualized medicine (each patient is a bin). Also, have to consider if the computer model is too complex to actually run in a patient-specific manner or in a timely manner (6 months of CPU per patient is probably not possible). Ideally, a model is not run for every patient. Need to run a Look up table.