Review: SI of mGluR LTD

http://www.mitpressjournals.org/doi/abs/10.1162/NECO_a_00408

Ever since I started my PhD program I always wanted to do a weekly review, post on a science paper to keep me up to date, focused, and refine my scientific knowledge, reading, and writing skills.


Well, that never happened.

But this time I did write something up. I hope I can have enough motivation to do this again some time, though. Review/Summary of the paper is as follows.


=====

Long term effects on synapses are currently a hot topic in the neuroscience community as it is associated with learning and memory. Two of the major forms of long-term plasticity, known as Long Term Potentiation (LTP) and Long Term Depression (LTD), are studied extensively to understand its underlying mechanisms. In computational neuroscience, we try to use models to further our understanding of LTP and LTD; this paper is just one example of how one may try to go about this.

The schematic of this paper seems rather simple: conduct experiments which will induce LTD, then use System Identification to model it. Nevertheless the work required for such a paper is extensive. This model in particular focuses on mGluR-dependent LTD, in contrast to NMDA-dependent LTD (as is most other models currently available out there). For the experimental portion of this paper they use a chemical known as dihydroxyphenylglycine (DHPG) to stimulate mGluR-dependent LTD... I’ll assume that the references support that this is indeed mGluR-dependent.
For modeling they use a Systems Identification approach, a top-down method where they attempt to extract the dynamics of LTD using transfer functions. They used four different datasets with differing DHPG concentration, duration of DHPG application and sampling rates; the input was defined as the DHPG concentration while the output was the fEPSP (post synaptic current) slope percent change as a result of application.

The transfer function used is of particular interest to me as it is more abstract and thus difficult to grasp than the experimental portion. From the paper, it seems the focus of the transfer function is the polynomial expansion using a “backward shift operator” z^-1. Basic definition of a backward shift operator is:

z^-1 * Y(t) = Y(t-1)

essentially, then, z^-1 = Y(t-1)/Y(t), a ratio of the past inputs compared with the current input. The polynomial coefficients are then estimated depending on a range on order (z^-n represents nth order) and time delays (1-10). The best model was determined through 3 statistical criteria (R^2, Akaike, Young) which each have their own criterion and reason for use.

The results of this paper are simple yet significant. For lower sampling rate (0.0033 Hz)Applying DHPG for 5 minutes at 15 uM gives about 20% reduction in slope size; application for 15 min give about 30% reduction; and applying 30 uM DHPG for 25 min also gives, on average, 30% reduction, but the effects take longer to stabilize. The authors state that the models approximate an integrator, meaning that time of application for input (5, 15, or 25 minutes) changes the response of the system.
The paper then discusses the results of the higher sampling rate (0.033 Hz). It gives a reasoning on why higher sampling rate should be used, but raises the question as to why they did not use this higher sampling rate for their previous experiments, as well, and instead have the higher sampling rate to be separate from the results of the others.
In all their results the highest order with the best fit turns out to be 2nd order. Judging from how the pattern of LTD looks, this seems accurate since there does not appear to be complex nonlinearities in LTD. The paper goes on to describe LTD as a certain subprocesses (serial, parallel, feedback), second order meaning there are two subprocesses, one with fast time constants and one with slow time constants. The results they give suggest there could be either parallel and/or feedback structures for their model.
In terms of the subprocesses, there is still some lack of understanding on behalf of myself to completely know how the coefficients and time constants relate to them, therefore I may need to determine how to work with them better.

As the experiment itself was rather conceptually simple, the discussion doesn’t seem to add much more to what is already known. Dynamics of the system is briefly mentioned. Additionally, they present evidence to support the experimental protocols which they used and explain a bit more about the modeling results with how it compares to the currently know physiological processes.

Overall the paper is a useful reference to see how System Identification can be used to model physiological data and how to interpret the model. As for its significance and usage, it seems a bit too rudimentary to have such a model in a larger scale modeling scheme such as EONS. The conditions are too specific and the experimental data to support the model is not enough. Nevertheless such a model can prove useful, not only through analyzing the dynamics, but as more is learned about mGluR related LTD, it can serve as one of the stepping stones to bring the entire process of synaptic plasticity together, which would then serve use to EONS in the future.