As noted a few weeks ago, gene regulation is a complicated field, typically with cartoonish views developed from small amounts of data. Mapping out the basic parameters is one thing, but creating quantitative models of how regulation happens in a dynamic environment is something quite different- something still extremely rare. A recent paper uses yeast genetics to develop a more thorough way to model gene regulation, and to decide among and refine such models.
|Binding site preferences for each regulatory protein discussed. One can tell that they are not always very well-defined.|
|Edging towards a model. Individual aspects of the known or hypothesized interactions are encoded in computable form.|
This is where modelling comes into play. The authors set up the known and hypothesized interactions, each into its own equation, whose parameters could vary. Though the number of elements are few, the large number of interactions / equations meant the models, (with 5 interactions, 13 states, and 41 parameters), given a partial set of data, could not be solved analytically, but were rather approximated by Monte Carlo methods, which is to say, by guessing with sample data. Models with various hypothesized interactions were compared with each other in performance over perturbation, where the model is given a change in conditions, such as a switch to low-nitrogen medium, or an inactivating mutation in one component. The model comparison method was Bayesian because it was iterative and took into account well-known data, such as the established interactions and their key parameter levels, wherever known.
Given a model, its ability to match the experimental data from the mRNA expression profiles under various conditions can be measured, adjusted, and re-iterated. Many models can be compared, and eventually a competitive process reveals which models work better. This is informative if the models are sufficiently detailed, and there is enough detailed data to measure them on, which is one of the strong points of this well-studied regulatory system. Whether this method can be extended to other systems with far less data is questionable.
In this case, one hypothesized interaction stood out as always contributing to more succesful models. That was the inhibition of Gzf3 by Dal80, its close relative. Also, in further selections, hypothesis 2 was also strongly supported, which is the auto-activation of Gat1, probably by binding to its own promoter. On the other hand, models that were missing the hypothesized interactions 1,3, and 5 were the top performers, indicating that these (auto-inhibition of Dal80, inhibition of Dal80 by Gzf3, and cooperative binding by Gln3 and Gat1) are probably not real, or at least significant under the measured conditions.
Lastly, the authors do a bit of model validation by creating new experiments against which to measure model predictions. Using their best model, the expression of Dal80 (Y-axis) under various perturbations is reasonably well-fit.
- More on regulatory logic.
- The state can work effectively.
- A little pacifism: "Our government has roughly eight hundred foreign military bases."
- While we have been stagnating, the rest of the world has been catching up and doing better.
- ECB and helicopter money, but not for Greece.
- Pakistan is not the only one playing a double game in Afghanistan.
- Fed, on the wrong track.
- Every day is opposite day. Do gun nuts know anything about Christianity? "Collectivism: humanity's oldest disease."
- Methods of a con artist.
- Abenomics looks a lot more like austerity.