CPHx: John Quackenbush, Dana-Farber Cancer Inst. & Harvard – Network & State Space Models

Network & State Space Models: Science and Science Fiction Approaches to Cell Fate Predictions

John Quackenbush, Dana-Farber Cancer Inst. & Harvard


Challenge the way  you think about biological systems.

“Science is built with facts as a house is with stones,  but a collection of facts is no more a science than a heap of stones is a house” – Jules Henri Poincare

What is a model?

“The purpose of the models is not to fit the data, but to sharpen the questions.”

The question in biology – Is the mean large, given the variance?

Example, determining gender by height.  There is a correlation, but the variance is huge.

We would like small variance compared to the difference in mean.

An alternative:  Is the difference in variance large independent of the mean?

Modeling cell fate transitions.  How does one cell morph into another cell type based on stimulus. Also want to identify pathways that underlie various cell types. All of this comes from building models.

Referee #3 always contests the use of the word model on all his papers.

Phenomenology tries to look at the past.  Ultimately we look to develop a theory that describes the interactions that dreive biological systems.  Build an approximate model that describes a body of knowledge that relates empirical observations of phenomena to each other, consistent with fundamental theory, but not derived from theory.

A journey through Variation.  Jess Mar’s PhD work.

Cells converge to attractive states.  Stuart Kauffman presented the idea of a gene expression landscape with attractors.  Great illustration of gene networks on a landscape.. distinct patterns of gene expression.  States are attractors, and pathways tend to self organize towards them.

There are only 250 stable cell types and each of them represent attractors.

Can we push cells from one state to another based on external stimulus.

An example of Promyelocytes (HL-60) transforming into another cell type.  Arrays done to profile the states of the gene expression between the two end points over several days.

Cells Display Divergent Trajectories That Eventually Converge as they Differentiate.  What accounts for the divergence?

There are multiple processes that are occurring during this observed change.  What you see is actually the sum of all of the different process.  You can, in fact, divide the genes into different groups: transients and core changing genes.  Transients tend to be related to external stimuli.

Waddington’s hypothesis.  A developmental biologist, with publications of attractor states, etc.

Waddington’s model calls for creation of “canalization” of the landscape, in which you move from start to end in paths.

The paths, however, don’t have to be straight.  You can get paths that wander up the walls of the canals.  Individual cells can follow random courses down that path…  thus when you look at the population, you see the canal, but if you don’t, you’d see a high amount of variation.

Had to come up with a method or pathways that characterize various cell types.  What are the signatures?  “Attract” soon to be published.  Finds core pathways that underlie cell fate transition.  Pull out pathways from KEGG – then built new method of gene set phenotyping.  Ranking pathways based on cell type informativeness.

Need to look at separate expression groups.  Some profiles are common across various states, so you need to deconstruct the pathway profiles to make sense.  This can then be used to define an “informativeness” metric, which in tern can be used for identification of core pathways that identify states.

A variational approach to expression analysis.

A stem cell model for neurological disease, based on olfactory cells.  Nasal biopsies, culture pluripotent stemcells, then allow the stemcells to differentiate.  9 healthy, 9 schizophrenia, 13 parkinsons.

What are the pathways that characterize the differentiation of the stem cells?

A bunch of pathways were identified that stood out with significant p-values. One can then ask if anything stood out between the control and the neurological disease patients.  There were no real difference in average pathways… but there were significant differences in their variance!

How important is the difference in variance in defining phenotype?

When overlaid, you can observe skews in the data for pathways.  If the change in variance is important, you should see an even greater skew in the pathways that are key in defining the phenotype.

Indeed, when looking at key pathways, the skew becomes more apparent.  Top 5 pathways show the same skew each time.  There is a robust difference in the profiles, then.

You can also observe the same type of phenomena when using 5% top/5% bottom cutoffs.

High variance genes are cell surface genes and nucleus, low variance tend to be kinases, signalling. etc.

Variance constraints alter network topology.  This suggests schizophrenia are opposite ends of spectra of neural disease.  (Referring to variance being high in one, and low in the other)

Now, trying to understand the mechanisms underlying this variance.

Path integral formulations of quantum mechanics… neutonian objects follow one path. subatomic molecules follow EVERY path.  You must consider cells in the same way, they follow many paths that converge to the average path.

[Ok, I really like this analogy.]

Where are we going?

  • Biology is really driving this
  • integrated data types must be considered intelligently
  • We may be in a position to start developing functional biology models. [My words.. it was expressed more clearly by the speaker.]

Genomics is here to stay.  Even bus drivers have DNA kits to help identify people who spit on them. (-:


4 thoughts on “CPHx: John Quackenbush, Dana-Farber Cancer Inst. & Harvard – Network & State Space Models

  1. Pingback: Copenhagenomics » Recap of Day 1 at CPHx

  2. Pingback: Copenhagenomics, in review | Fejes.ca

  3. I think this critical thinking is significant in that it changes the paradigm how to look at biology data, not only from the stand of point how data can be analyzed statistically but how biological variations should be modeled.

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.