Directed evolution using engineered cells

Directed evolution of cells is generally done by setting up a screening process. An example is a binding assay for evolving cell surface receptors.

It is very difficult to build screening procedures that select for functions. However, it might be possible to engineer "killer" cells that attack cells with particular types of behaviors. Thus the killer cells provide the screening process. And why limit to a single type of killer cell.. of course, the killer cells should not evolve (might be an issue)

Imagine this scenario: a population of cells evolving in an environment with populations of 3 or 4 different types of killer cells. Each killer cell targets a particular type of behavior. Further, another population of "helper" cells excrete specific nutrients in response to particular behaviors. The target population of cells should evolve to avoid specific functions that are targeted by the killer cells and acquire specific functions targeted by the helper cells. Due to the existence of multiple criteria, the evolution might be more gradual as well.

Signaling - specifity and decoding

Lets consider wireless signals. The signals themselves travel in all directions, so there is no specificity. The frequency provides the sender-receiver specificity. The signal pattern contains information that the receiver can decode, i.e. the receiver must expect a specific type of pattern.

Comparing the general idea to biological signaling... the specificity usually comes from binding affinity, so that aspect of signaling is clear. Now for decoding the information. A pathway probably has multiple molecules serving as signal carriers. The pattern of concentrations of those input molecules *might* serve as the encoded information that the receiver, i.e. the pathway, is able to decode. The pathway then sends a new set of molecules as output signals . Note that this results in a conversion of signal carrier, which is analogous to the wireless signaling analogy where the wireless signal is decoded into some other form such as digital signals.

Modules should span multiple layers

In biology, something like a protein domain is an ideal candidate for a "module" because it is a tool that can be reused at different places to serve different purposes. Nature does not want to reinvent tools; it would be more efficient to reuse existing ones -- protein domains are hard to invent, so it is a perfect item to reuse.

Moving on to network structure. Patters, such as feedback or feed-forward motifs, are not too difficult to reinvent (depending on how hard re-wiring is). For example, re-wiring genetic networks is easy. So it does not make sense to call any genetic network a "module". However, a combination of protein interactions combined with gene regulation might be a module.

For example, lets consider a network composed of a protein that responds to a small molecule and activates a protein that then upregulates a gene. This network is difficult to reinvent because it has multiple interactions that are very specific. It would be a module that is worth reusing. For example, the final gene product can be replaced with some other gene -- a simple way to reuse the module.

As an additional observation, I think it makes sense to say that modules span multiple "layers". For example, in electronics, logic gates convert analog circuits to digital. The example in the previous paragraph is a module that converts small molecule concentrations to gene regulation.

Microbial ecosystem programming

There is a lot of hype about programming single cells by editing their genetic code, which in turn alters the dynamics of their regulatory or metabolic networks. However, using these engineered microbes in the real world is a very skeptical step, simply because we cannot predict exactly what can happen.

An alternative is to not edit the microbes themselves at all. Instead of building networks using enzymes inside the microbe, why not see a cell itself as a complex catalyst. A living cell converts some chemicals into others (environmental conditions apply).

Using microfluidics, it might be possible to completely characterize the "catalytic" profile of hundreds of microbial species, including bacteria, fungi, amoeba, algae, archaea. Then, build a "network" of different species such that the whole system is stable and performs some metabolic process that is of use to us, such as bio-remediation. This "engineered" network should be safer in the real world.

Reducing stochasticity in biology

Suppose we are diagnosing a set of symptoms of a disease. If there is only one symptom, we will give our conclusion very little weight because that one symptom could be due to random chance. However, if we see multiple symptoms, then it is probably not due to random chance but due to some cause.

This simple rule is sufficient to eliminate stochastic effects in the final decision: make decisions based on multiple observations rather than a single observation.

In a cell, a "decision" can be something like upregulating a gene. If this decision is made by a single transcription factor that detects some sort of environment, then the decision (transcription) will be noisy. In contract, if multiple transcription factors that all respond to the same environment are used, then the transcription process will be a function of the sum of multiple stochastic processes. The sum will always have a lower variance. Unfortunately, multiple regulators means the there are multiple association/dissociation events, which would add more noise. The solution would have to be a bit more clever than this, but the general idea holds.

Automatic analysis and construction of feed-forward regulatory networks

Given the assumption that transcriptional regulation is sigmoid shaped, i.e. the steady state diagram of the inducer vs. the target protein has a sigmoid shape (or inverted in the case of a repressor), it might be possible to automatically determine the steady state diagram of a regulatory network with no feedback controls. It might also be possible to generate a feed forward network for a given steady state diagram.


Lets take a simple case, where a transcription factor "a" indirectly upregulates and downregulates the target gene "X", as shown below. Then there are two regulation curves associated with "a". If we assume that a repressor will dominate over an activator, then we can determine what the the final steady state diagram of "X" will be as a function of "a" by looking at the dissociation constants (the center of the sigmoid curves).



The same method can be used to construct networks that have more complex steady state behaviors. For example, suppose "a" and "b" regulate "X" such that the activity of "X" is described by the yellow regions (1,2,3) in the diagram below. Then, it is possible to identify the network that will satisfy each piece and then put the pieces together. The shape of the sigmoid is determined by the dissociation constants.

Inference from topology

Changing parameters of a system can drastically change its behavior in some situations. But even so, it might be possible to deduce certain dynamical properties from topology...

Lets take a simple case:

      The gene product of gene A positively regulates gene B. 

Even if no parameters are known, one can hypothesize the steady state behavior of B will probably be a sigmoid function of A. The exact shape of the curve cannot be infered without additional information. 

Lets take a less specific case:

       Gene A regulates gene B, but the type of regulation is unknown.

Now, there are two relationships that are possible: a sigmoid function (positive regulation) or an inverted sigmoid function (negative regulation).  Again the exact shape of each cannot be known.

Lets extend the situation:  A regulates B and C, and B regulates C.  

Now, there are 2x2x2 possibilities. However, the number of different shapes that the 8 different combinations can make is probably not 8. It can be more in the case that this topology is highly versatile in the types of functions it can realize. It can be less than 8 if many versions of this toplogy produce similar behaviors. It is also possible that a few versions of this general toplogy are very interesting in the variety of functions they can realize, but the other versions are similar to one another. If this last situation is the real situation, then it is possible to make some inferences about the dynamical behavior with just the topological information. Here is how:

For a given general topology, i.e. where the regulation types are not known:

1. Generate all the different "versions" of this topology
2. Analyze each version by varying the parameters. Look for steady state behaviors as well as other interesting qualitative behaviors. 
3. Classify each version of the topology by its qualitative behaviors.
4. Hypothesize possible uses for each qualitative behaviour, especially in the context of where the original toplogy came from. 

The above approach is not specific to genetic networks. If it works for one type of network, it should work for the others.