I need to know what the criteria for choosing descriptors shoud be?For example,the use of MOE software can generate more than 100 types of descriptors but it is obvious that not all of them can be used for a QSAR study.Whereas in the case of gaussview we must fix descriptors ourselves.
Really it's a big difference that i didn't find a logical answer.
Which method shoub we use to generate quality descriptors??
For QSAR studies, you can choose descriptors (property) of a molecule, which correlates with your activity (experimental).
Not all descriptors can correlate with activity with a high correlation coefficient.
Most importantly while selecting descriptors, it should be correlated with activity with a good r square value (r2).
regards
Dear Kasmi Rania,
Huge number of descriptors could be derived from the chemical structure, so the descriptor space size needs to be decreased to those which are relevant to the target property. So descriptor filtration and selection step is to be carried out, usually it's done automatically by the used software, this involves the following steps:
1.Removal of correlated (co-linear) descriptors (you can check this by the correlation matrix)
2. Removal of low variance (low standard deviation) descriptors that slightly (or do not) change between the used compounds.
3. Descriptor selection by any of the following methods:
a.Forward selection.
b.Backward selection.
c.Stepwise regression.
d.Genetic algorithm.
e.Manual selection: Based on a previous knowledge about the relation between the molecular structure and the target property and also based on intuition.
In case of MOE, it uses PLS to generate the model, PLS can handle huge amount of descriptors even the colinear ones as it depends on latent variable approach ( You can read more about the PLS https://en.wikipedia.org/wiki/Partial_least_squares_regression)
So my advice to you is that if you don't have previous knowledge about the important related descriptors from for example experimental study, you can calculate a considerable number of descriptors and implement one of the previous descriptor selection methods ( computationally) if you're not gonna use a modeling technique such as the PLS that can handle large number of descriptors, and in all cases you should use the correlation matrix to make sure that your descriptors add meaning and are not redundant.
Hope this can help you :)
Best wishes,
Marwa
- Badges
- Science topic
- Similar topics
- Physical Sciences
- Atomic, Molecular and Optical Physics
- Atoms