[Research] Machine Learning in Drug Discovery (punch line!)

[Jeff Schneider]

Congratulations to everyone attending the meeting today!  You were so 
successful at fulfilling my request to ask lots of questions and 
brainstorm on other topics, that I forgot to give the final punch line.

Here it is in abbreviated form:

Psychogenics has now used this system to screen thousands of compounds.

Doing so, they have found about a dozen lead candidates.  These are 
drugs that:

1. Were predicted by SmartCube to be effective in one of the classes.
2. Had their activity successfully confirmed by running two different 
traditional behavioral tests (in mice) for that class.
3. Are known to work by a novel mechanism because a follow-up in vitro 
test showed that they do not use any of the mechanisms of drugs 
currently on the market.

These lead candidates were found at a fraction of the time and cost that 
a large pharmaceutical using existed methods would need.  And that 
answered what was a multi-million dollar question/bet for Psychogenics. 
  Can you develop this system to efficiently find new drugs that work by 
completely new biological mechanisms.  The answer is yes!!





Jeff Schneider wrote:
> Since I actually have one handy, here's an abstract for today:
> 
> Machine Learning in in vivo CNS Drug Discovery
> 
> Researchers in machine learning and operations research have made great
> strides in modeling and optimization manufacturing and other commercial
> processes.  A more recent trend is to observe that the scientific method 
> is a process that can be modeled and optimized with similar techniques. 
>   In this talk we consider a specific example of that: discovery of 
> central nervous system drugs (e.g. antidepressants antipsychotics, 
> anxiolytics, etc.) using in vivo behavioral testing.  Algorithms will be 
> discussed in the following areas: the use of kernel density estimators 
> to provide improved posterior probabilities in multi-class applications; 
> the use of semi-supervised learning to handle training data with 
> uncertain class labels; and the use of active learning to control 
> experimentation in the discovery process.
> 
> _______________________________________________
> Research mailing list
> Research at autonlab.org
> https://www.autonlab.org/mailman/listinfo/research
>