Gufodotto would like you to read these:

Friday, July 14, 2006

The Power of Models

I'm going to break non-discolusre agreement, here... may be... or may be not. but haven't talked about my work for a while. I'm currently developing predictive models for generic cardio toxicity
to be used by chemist to decide which way to go when they are pondering about new syntheses. usually, they make the molecule, then send it to the biologists to test it against some protein target. ion channels in the case of cardio toxicity, the same proteins which lets ion in and out of the cardiac cells when your heart beats - if you stop them, goodye heart - and life too.
in vitro inhibition of the channels is good indication that something nasty may happen in vivo. so if the molecules come back with a red flag, well, they've had it, and it's back to the drawing table (and the synthesys bench).

but now there's another way. with my models, they do not have to synthesize the molecule any more - they can use the models I created (based on previous experience - i.e. on the data accumulated in years of trial and errors procedures) to see whether the molecule is active or not just from its expected chemical structure.

and the incredible thing is, it works! as good as the best ion channel inhibition test they've used until now.

here's a picture:

the three bars on the left represent the percentage of success of the three commonly used ion channel assays. Sodium, Calcium and Potassium. this percentage tells you that for example the socdium channel assay is consistent with cardiotoxicity for 80% of the time. whether there's or there is no toxycity, that is.
on the right, instead, the success percentage of two models I built, one based on bayesian statistics and arcane entities called fingrprints - the other one on chemical properties and the exotheric practice of Projection to Latent Structures by means of Partial Least Squares - Discriminant Analysis (PLS-DA for friends).
as you can see, both are slightly below the sodium accuracy, but the good thing is, you need to make the molecule for the assay - with my test, it's all done in one second or less within a computer... and this rocks!!!

see you soon! I'm going to enjoy a well deserved week end, now!!!

3 comments:

PAC said...

Models? I still remember the course i had about System Theory by the great prof.ing. Giancarlo Zini: a symple dynamic matrix e^At, some stability/controllability criteria and the world is yours B-)

Unknown said...

uhm, I think you have quite a skewed idea of what a model is...

PAC said...

That was only the beginning with models for quite simple dynamic systems. The story goes on with "Modellistica & Identificazione", where a model is any possible way to describe the system (from descriptive (i.e. words) , to stochastic and lie algebra based models). My comment was mainly focused on your enthusiasm B-) and yes, my point of view is a little different from your one.