Thursday, December 23, 2004

More model disenchantment

I happened to spend a few minutes yesterday talking to one of the professors around here about what folks in the biotech industry were looking for in people coming out of grad school. That conversation actually served to further my current skeptical stance on [certain types of] mathematical models in biology, at least in terms of their current practical utility.

It turns out that the folks trying to come up with therapeutics [ie drugs] don't really care very much about mechanistic models of what's going on in a particular biological process. By "mechanistic" models, I mean models that explicitly describe molecule A reacting with molecule B to produce C, which in turn reacts with molecule D etc. Apparently, the reason they don't care about them is that

a. In order to create and refine these computer models, you have to do a lot of quantitative biochemistry that allows you to measure how fast certain reactions are, how much of a particular compound gets produced etc. Obtaining that sort of data is a lot of work because there are no good high-throughput [ie fast], accurate ways of getting it, so you end up having to do a lot of painful grunt work.
b. Even if you get all this data, chances are that you measured it all in your little petri dish ["in vitro"] ie not under the actual physiological conditions that exist inside an animal ["in vivo"]. That makes it questionable how applicable the computer simulation is to what you're really interested in, namely how an animal/person would react to the drug. So, in the end you have to go do the majority of the experimental work [that a computer model was supposed to save you from] anyway.

What industry folks actually do care about is being able to analyze large data sets coming from high-throughput measurement methods and seeing what they can figure out that way.

An analogy based on cars and car engines goes something like this: people building mechanistic models are basically taking the engine apart, figuring out what the carburetor does, how it's connected to the combustion chamber, how the spark plugs work etc, and then building up a picture of how the various components all fit together. In contrast, the folks analyzing large data sets are doing the equivalent of stepping on the brake, the accelerator, changing gears etc and seeing how the car reacts. If all you care about in the short/medium term is driving the car in a particular direction, the second set of actions produces results much more quickly, and is more useful than the first.

Definitely a useful reality check to have early on in my graduate career.


Post a Comment

<< Home