It is not an easy question how to build a reliable QSAR (QSPR) model. As an old chemist, which worked in the stone ages of computers, I learnt that the most essential is to think properly: to see intuintively what is the problem, and from

the understanding of problem one has to go further toward its solution. You cannot solve scientific problems mechanically; even if your obtain good correlation, you have to be capable to explain it. If you cannot explain behaviour of your system, the results cannot be generalized, and therefore they are useless.

Now, we come to the crucial point: we teach young scientist all kinds of technics, but nobody teach them how to think. They manipulate instead to contemplate.

Overall goal of feature selection methods is to develop robust model by using minimum number of descriptors. Several possible methods are common in feature selection, correlation is most common and most important method. It's tedious but you can do this job in few minutes by using Weka/Rapidminer. Other popular methods is F-score, Genetic algorithm, random forest. But before using any other feature selection method you must remove highly correlated molecular descriptors.

There are lots of feature selection methods are available in Rapidminer/Weka. Just try Rapidminer/weka.

This discussion here might help: https://www.researchgate.net/post/What_are_criteria_for_good_definitions_of_chemical_space_and_descriptors

Coming from the atomistic simulation community I feel that good descriptors should most importantly be unique, i.e. complete or overcomplete. It's not necessary, but desirable, that they are invariant with respect to atom indexing, and molecular translation and rotation.

It is not an easy question how to build a reliable QSAR (QSPR) model. As an old chemist, which worked in the stone ages of computers, I learnt that the most essential is to think properly: to see intuintively what is the problem, and from

the understanding of problem one has to go further toward its solution. You cannot solve scientific problems mechanically; even if your obtain good correlation, you have to be capable to explain it. If you cannot explain behaviour of your system, the results cannot be generalized, and therefore they are useless.

Now, we come to the crucial point: we teach young scientist all kinds of technics, but nobody teach them how to think. They manipulate instead to contemplate.

Really there is no criteria to choose descriptors. However, you can analyze your protein and guess the physical properties. I suggest you to do a bulk then analyze it based on efficient statistical values.

I agree strongly with Nitish on the use of Weka - it will give you graphs comparing all of your descriptors against each other in a pairwise fashion from a single option - the result a presented graphically so strong correlation is really obvious (even when it's not linear).

The main issue with having lots of decriptors is that it increases chance correlation so validation of your QSAR model is essential.

You can narrow down your choice of descriptors by following Nenad's suggestion - understand the problem you are trying to model as well as you possibly can and use that understanding to guide you in choosing properties that are likely to be relevant to the problem. Also it helps to understand the nature of your descriptors - what is it that they are telling you about your molecules. One problem with computers and efficient coding is that it is too easy just to press buttons and get numbers out with no understanding of what is going on or what the number mean.

Finally, take a look at what others are doing and either ask yourself why? or look in their discussion for where they explain why?

You can find some info about feature selection here http://www.moleculardescriptors.eu/tutorials/tutorials.htm

I perfectly agree with what has been posted. Descriptors must be chosen in order to be as few as possible (generally for QSAR you should have maximum 1 descriptor every 5 molecules, better if more, of the training set,) and need to be uncorrelated between them. Some descriptors might be chosen (or pruned) by critical thinking (i.e. if you are studying a model to predict how a molecule would cross the blood brain barrier, logP should be mandatory, while LUMO energy would probably be not...) then PCA analysis will help in reducing the number. In MOE there also is a nice tool to build correlation matrices, which evidence the internal correlation between descriptors compared with the correlation of each descriptor with the experimental property you are considering (ok for comparing batches of about 15-20 descriptors).

“Molecular descriptors map the structure of the compound into a set of numerical or binary values representing various molecular properties that are important for explaining the activity or property of the molecule. Molecular descriptors are at the very core of the QSAR analysis, because they establish a causal link between structure and property. Molecular structure perceived as an electron–nuclear configuration is extremely complex even for simple compounds, with many degrees of freedom involved (nuclear, electronic, vibrational, and rotational to name just the most well-known and important). But molecular structure can be reduced to or decomposed into myriads of much simpler molecular features, each of which describes some particular aspect of the molecular structure and ignores all other aspects. In turn, the simplicity or complexity of description can vary in a rather wide range: some of the developed descriptors are powerful enough to grasp a substantial part of molecule’s topology, or geometry, or electronic/vibronic structure. Tentatively, two broad families of descriptors can be distinguished, based on the dependence on the information about three-dimensional (3D) structure and conformation of the molecule. In 2D-QSAR, chemical descriptors are calculated from chemical graphs and no information about 3D configuration of molecules is used. A fragment-based QSAR paradigm relates sub- structural fragments present in the dataset compounds to their biological activity. Based on such a fragment-based QSAR model obtained, a fragment weights may be assigned according to their individual coefficient values and descriptor values.

In 3D-QSAR, methods such as the popular comparative molecular field analysis (CoMFA), conformational analysis, and global 3D structure alignment should take place before de- scriptors are calculated. The multidimensional (nD)-QSAR methods such as 4D-QSAR, 5D-QSAR, etc., tackle known 3D- QSAR problems such as subjective molecular alignment and bioactive conformation problems. The success of QSAR depends on the ability to reproduce the bioactive conformers, namely, those conformers which are expected to dominate a protein–ligand complex ensemble. The predictive power, robustness, and reliability of the QSAR/QSPR models depends critically on the use of appropriate molecular descriptors.”

Workalemahu M. Berhanu, Girinath G. Pillai, Alexander A. Oliferenko, and Alan R. Katritzky (2012) Quantitative Structure–Activity/Property Relationships: The Ubiquitous Links between Cause and Effect. ChemPlusChem 2012, 77, 507 – 517

The difficulties in selecting descriptors were most eloquently described above and many useful suggestions proposed. There are a few things that I can offer my comments, though. One is a warning that there is always an urge to replace the firm understanding of the "real" causes with correlation based reasoning (like in medicine).

We are looking for a "complete" and "finite" coverage of the infinite and unknown space, therefore, the task is not algorithmic and it has natural limitations coded into it. As Nitish nicely put it we are looking for a minimal description which is equivalent for maximal abstraction (informational load) in an individual parameter. This view nicely intersects with what Nenad described as experience and intuition selections driven process.

Therefore there are at least two observations (recommendations):

1) The set of descriptors should be as minimal as possible and enhanced only when needed.

2) The particular model will always, by design, be "local" which means fitting better one set of problems than the other, so we should not overextend our expectations.

So, the optimal strategy would be to go from small to large number of parameters not the other way around. Even though, QSAR strives to be a theory, by design, it is a simple correlation method, therefore, correlation replaces causal reasoning (basic science). For instance in entropically driven binding any QSAR model will be somewhat faulty. Also, any choice of parameters is by definition static while the fitting space is dynamic therefore every choice will be somewhat defective.

As a final remark I can only say that we probably need to design the optimization technique (a theory) how to derive (and define) the optimal number of parameters similar to physical minimal action principle. Many individual fields are using Bayesian reasoning for this particular task but the definition of weights would still be an arbitrary task.

In addition to the above comments, the following points are worth considering while choosing descriptors

what is the ultimate goal of your final model ?

a) to predict the activity of future structurally similar analogs (compounds)? and not to worry about structural characteristics of the binding site (as well as compounds)

(when you use too many molecular descriptors and statistical methods like PCA or PLS, tracing back to descriptors for further improvement/modification of compounds would be very difficult if not impossible)

b) to create theoretical binding site model ? (absence of target protein structure or absence of knowledge about the target protein itself ?)

Consider using 3D descriptors and linear or non-linear regression analysis

Just like to add something while performing test and training maintain a balance of 1:5 ratio of descriptors to the number of compounds.Otherwise there could be high bias in you data.

My work is on EEG processing and somehow off-topic but based on our research the redundancy among features is also an important issue. Let me suggest two papers:

Guyon et al., An Introduction to Variable and Feature Selection, Journal of Machine Learning Research 3, 2003

http://jmlr.csail.mit.edu/papers/volume3/guyon03a/guyon03a.pdf

and

"Feature selection based on mutual information: criteria of max-dependency, max-relevance, and min-redundancy,"

Hanchuan Peng, Fuhui Long, and Chris Ding

IEEE Transactions on Pattern Analysis and Machine Intelligence,

Vol. 27, No. 8, pp.1226-1238, 2005.

dear nutan, as such there is no fix criteria for the selection of descriptors but we can find out the probability of selecting the descriptor based on the chemical constitution of the molecule. Instead of MOE if you go for dragon ,it gives correlation with all the descriptor calculated ,you dont have to check correlation for each descriptors. in dragon you can select most correlated descriptor to build QSAR model using suitable method of regression like PLS, MLR, GFA etc.

Frist of all you have to formulate problem properly, i.e. to put the right question.

Is the studied effect determined by electronic or sterical factors? Which parts of molecules are most influential? What the molecules in the studied set have in common? Did the previous studies point to a specific effect?

After that you have to choose descriptors which are presumed best for the particular effect. If there is two or more effects, you have to choose unrelated

(and uncorrelated) descriptors for each effect.

This lead to the old wisdom: if you want to find good answer, you have first to find a good question!

Dear Prof Dr Hamdy M. Abdel-Rahman,

Salam and Eid adha moubarek.

You can use POM theory instead MOE because most pertinent descriptors are there...

As it was fixed by Prof Dr Nenad raos, you should take carefull of both electronic and/or sterical factors...

What is exactly your biotarget? Bacteria, Virus, cancer or Fungus? For each biotarget you can use POM theory...

Try it..then we can talk details..

ALLAH HAFIZ

​Prof Dr Taibi Ben Hadda

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

Mohammed First University, Faculty of Sciences

Materials Chemistry Laboratory, Oujda-60000. Morocco.

Phone: (+212) (0) 6 66 13 41 78; Fax: (+212) (0) 5 36 50 06 03

E-mail: [email protected]; My Skype: Taibi Ben Hadda

​======================================

URL-01: http://scholar.google.co.in/citations?user=KQDRRUMAAAAJ&hl=en

URL-02: http://www.researchgate.net/profile/Taibi_BEN_HADDA/publications/

URL-03: http://www.mp3quran.net/rashad.html