[Comp-neuro] Re: good models versus bad models versus realistic models

[james bower]

First -- with respect to Purkinje cells -- we have been engaged for  
the last 4 years in a difficult series of experiments coupled to  
modeling that takes a fundamentally new approach to figuring out what  
are (perhaps if you can rank) the most important electrical (passive)  
features of Purkinje cells.  Most neurobiologists, including me, spend  
most of their lives working on one part of the brain in one species  
(rat Purkinje cells in the upper lip patch of cerebellar crus IIA).    
In fact, the Purkinje cell model referred to previously, is actually  
based on an Israeli Guinea Pig Purkinje cell from Idan Segev's  
laboratory (approaching the most graphically reproduced neuronal of  
all time).

So, 3 years ago I decided to take a different approach -- why not let  
evolution tell us something.  So, for 3 years we have been recording  
and injecting Purkinje cells in different species (at present, fish,  
turtles, frogs, mice, guinea pigs, rats and soon, lamprey, birds, and  
more fish), to ask a fairly simple question -- what anatomical  
(electrical) features of the Purkinje cell have been preserved through  
evolutionary time.  Might as well let evolution work for  you.  This  
is of course a hard study because anytime you switch biological preps,  
you have to start everything from scratch.  The study will take 4  
years to complete - which is also tough in the current "how many  
papers have you published this year" environment.  Never-the-less the  
results are interesting.

We are also currently using imaging techniques to get anatomical  
reconstructions of several 10s perhaps hundreds of individual mouse  
Purkinje cells -- to specifically look at the anatomical variability  
and what it potentially means for function.  Finally, with respect to  
the circuit, we have a collaborator in Brazil who is counting EVERY  
Purkinje cell and Granule cell (150,000 of which provide inputs to  
each PC) in each of the brains we are studying.  This will allow us to  
look at the relationship between the ratio of input cells to the  
electrical properties of their target neurons across multiple  
species.  Of course, all of this data would be impossible to interpret  
or even analyze coherently without modeling.  And it is because of  
modeling that we were driven to collect this data.


The second point your comment raises involves the issue of levels of  
scale.  Physicists almost always point out when talking about detailed  
realistic models that physics has made great progress by considering  
behavior at different scales (see several recent postings).  However,  
as I believe I may have also said before, the issue of scales in  
biology is interesting and likely different.  It seems very likely to  
me that one of the ways that biology pulls off what it does, is by  
linking scales together.  Thus we predicted a number of  years ago  
that gamma and theta frequences recorded at the level of the surface  
of cerebral cortex, were likely to also be manifest at the level of  
the resonance of interactions between molecules associated with  
synapses.  This seems to be the case.  However, I agree completely  
that figuring out how to cross scales is one of the technical /  
computational feats we need to figure out how to do.

Your question about where one chooses to work and simple vs complex  
models is related to deep question number 4 for the day - what detail  
in the nervous system is irrelevant -- or the inverse -- could you  
build a functioning nervous system with less complexity than the one  
we have -- I suspect not -- therefore, eventually, to understand how  
it works, we are going to have to have tools to at least represent its  
full complexity.

Finally, in all things there are degrees -- yes, realistic modelers  
make assumptions as well (cell morphology matters, channels matter,  
the position of synaptic inputs mater, ephaptic field effects don't  
matter as much) -- however, it is different in kind to make these  
assumptions and to assume what a cell does, what it cares about, what  
computation it performs, and what computation the circuit it is in  
performs.  This is the horse way way way ahead of the cart.  I  
actually believe that models should be tuned on the most artificial  
possible biological responses (responses of neurons to current  
injected in their somas -- which is a TOTALLY artificial stimulous for  
almost all neurons).  Then parameters should be frozen and synapses  
added.

Finally final -- all models are wrong -- by definition -- we don't  
really have any idea how brains work -- so it is not a matter of right  
or wrong -- it is a question for me, of what tools have the greatest  
chance of letting biology (the brain) tell us what it knows and how it  
works -- seems to me those models by definition are those that are  
built, on first principles, from the structure of the brain itself.


Jim



On Aug 13, 2008, at 1:39 PM, Paul Miller wrote:

> Hi,
>
> I love this debate and thought I'd chime in here on the idea of  
> assuming "how things work".
>
>> Haven't abused at all -- with one big exception -- realistic  
>> models  are more likely to tell you how things work, than are  
>> models in which  'how things work' is assumed.
>
> But once a realistic model of a single cell has told us "how things  
> work" at the cell level then if one is interested in what arises  
> from the interactions of multiple cells --- "how things work" at the  
> systems level --- it makes sense to use the simplest single cell  
> model that captures how that  cell works (which for those interested  
> in the system I think means "what makes the cell release  
> neurotransmitter or neuromodulator").
>
> Especially since, using your example, every single Purkinje cell in  
> the brain is different with different sets of conductances, so that  
> in all likelihood an "average" Purkinje cell in terms of  
> conductances and morphology may not act like an "average" Purkinje  
> cell in terms of function. So a detailed model of a set of Purkinje  
> cells would need to make each cell different (in which case which  
> model cell is the right one?) but probably have them function in the  
> same way. I would argue that in biology (convergent evolution etc)  
> it appears it is function that matters (for survival and  
> reproduction which is really all that matters) not necessarily the  
> detailed instantiation of that function.
>
> So, agreed, at one level of research, let's ensure with whatever  
> methods available (experimental of computational) that we have an  
> accurate and correct description of single-cell function. However,  
> at the next level (with an aim to understanding behaviour) let's  
> maintain that *correct* function in the simplest possible model to  
> understand its effect on the system.
>
> I think it it is unfair to tar the outcomes of "simple" models with  
> the outcomes of "wrong" models. Or is your claim that we are  
> sufficiently ignorant of what makes any cell in the brain fire a  
> spike, that a simple model of any cell's function is almost  
> certainly wrong (as of now)?
>
> Interestingly, your mentioning of the impossibility for one person  
> to fully understand collaborative software efforts links in here.  
> The key to such programming efforts is to maintain a modular  
> structure. One person understands a module well enough to guarantee  
> that a given set of inputs produces a given set of outputs. Another  
> person need not understand how the module works, but can rely on its  
> input-output relationship to connect it with other modules and  
> produce a larger functioning module and ultimately (through the  
> hierarchy) a functioning system. Of course if one has an incorrect  
> description of the input-output characteristics of a module one is  
> doomed, but that does not imply one needs a detailed description of  
> every module. Furthermore, from the top-down, one can come to  
> realize the system will only work if it has a particular type of  
> component or module -- and can go looking for the existence of such  
> a component.
>
> In summary we all assume "how things work" at one level and based on  
> these assumptions try to explain or predict "how things work" at a  
> higher level. We choose at which level we work (maybe you think some  
> choose to waste their time?) and our goal is to ensure we use the  
> simplest of the correct models of any level we wish to integrate.
>
> Paul.
>




==================================

Dr. James M. Bower Ph.D.

Professor of Computational Neuroscience

Research Imaging Center
University of Texas Health Science Center -
-  San Antonio
8403 Floyd Curl Drive
San Antonio Texas  78284-6240

Main Number:  210- 567-8100
Fax: 210 567-8152
Mobile:  210-382-0553

CONFIDENTIAL NOTICE:
The contents of this email and any attachments to it may be privileged  
or
contain privileged and confidential information. This information is  
only
for the viewing or use of the intended recipient. If you have received  
this
e-mail in error or are not the intended recipient, you are hereby  
notified
that any disclosure, copying, distribution or use of, or the taking of  
any
action in reliance upon, any of the information contained in this e- 
mail, or
any of the attachments to this e-mail, is strictly prohibited and that  
this
e-mail and all of the attachments to this e-mail, if any, must be
immediately returned to the sender or destroyed and, in either case,  
this
e-mail and all attachments to this e-mail must be immediately deleted  
from
your computer without making any copies hereof and any and all hard  
copies
made must be destroyed. If you have received this e-mail in error,  
please
notify the sender by e-mail immediately.