In multicellular organisms, the same genetic material is shared by many distinct cell types. During embryonic development, tissue repair & growth, etc., cells can transition from one type to another. At least in "complex" animals (read "vertebrates") under normal conditions, cell type transitions across different species broadly share the following characteristics:
- Transitions are targeted (a cell can't just transit from any type to any other type) and directional (transitions are one way, i.e., bone marrow → blood, but not vice versa).
- Transitions are directed by a small set of molecular signals (morphogens). However, different cell types can respond to the same signal in distinct ways.
It also seems approximately true that organisms with smaller gene-regulatory networks (GRNs) have fewer cell types. There are ~250 DNA-binding transcription factors (TFs) and O(1) cell type in S. cerevisiae (yeast), ~1000 TFs and O(10) cell types in C. elegans (nematode), ~2500 TFs and O(100) cell types in humans. However, the scaling of cell type number with GRN size is actually kind of odd, since generically the number of states of a N-component network grows as eN. The problem is not that we humans have more cell types than yeast, but that we only have ~100 times more.
I think these facts are probably related on a deep, physical level, by which I mean they are probably independent of the detailed interaction between specific genes and specific proteins in specific organisms. I suspect any sufficiently complex GRN will exhibit these characteristics.
Let's consider a minimalist N-component GRN governed by dsidt=−siτi+ri(→s),i=1,…,N. Here si is the expression level of gene i. I make no distinction between mRNA and protein (more on this in a later post), and by "genes" I mean genes encoding TFs. The term −si/τi describes the degradation of gene products over time (for simplicity I will set τi=1 for all i), and ri(→s) is the rate of transcription/translation from gene i. The rate is →s-dependent due to possible gene-regulatory activity by j,k,…, as well as i itself.
Now I assume the net gene-regulatory activity on i from i,j,k,… is described by a mean field ϕi(→s), and that ri(→s)=r[ϕi(→s)] is an universal function of ϕi only. In particular, I take r(ϕi)=r(ϕi−ϕ0) to be a S-shaped function centered on a threshold ϕ0>0 and bounded by 0≤r(ϕi)≤1. The strength scale of TF-gene interactions is given by ϕ0: gene-regulatory interaction where |Jij|<ϕ0 can be regarded as weak, while interaction where |Jij|>ϕ0 can be regarded as strong. We will set ϕ0=1 without loss of generality (see update below).
I don't know what ϕi(→s) looks like, but I can expand it as a series, ϕi(→s)=∑jJijsj+12∑j,kKijksjsk+…. The terms of the series can be interpreted as binding of i's regulatory elements by one, two, or more TFs. I will truncate the expansion to retain only one-body TF-gene interactions parametrized by Jij. The sign (±) of Jij describes whether j promotes or inhibits the expression of i.
As defined, (1) describes a large, non-linear dynamical system, and in general we expect such a system will be multi-stable. I will assume experimentally observable cell types correspond to stable fixed points of the model. The state →s=0, i.e. death, where nothing is being expressed, is a stable fixed point of the model. GRN states that are not fixed points will evolve on a time scale set by the lifetime of gene products (mRNA and protein), which is ~1 hour. Cells corresponding to a non-stationary state of the GRN will not be around long enough for us to capture and study it.
Does this model do anything interesting? Of course it does; otherwise I wouldn't be writing about it. This is an on-going research project, and these posts are essentially a set of notes-to-self to collect my thoughts in a semi-organized fashion.
Updated September 6, 2015. If Jijs are i.i.d. with zero mean and variance σ2J, then ϕi=∑jJijsj will be i.i.d. with zero mean and variance σ2ϕ=Nσ2J/2. The typical size of |ϕi|≈σϕ will grow as √N; this is dynamically equivalent letting ϕ0→0 as 1/√N. Maintaining dynamical similarity between GRNs of different sizes requires either rescaling ϕ0→√N/2ϕ0, or rescaling σJ→σJ/√N/2.