Hi Huy. One of the important things to note is the criteria in which you choose your descriptors. In 2D-QSAR generally descriptors which are more specific for geometrical properties, tend to display higher model qualities compared with descriptors which lack any 3D data (like number of atoms, bond counts, molecular weight and so forth).

Also, generally avoiding the same "kind" of descriptors usually lead to more reliable results. For instance, if you did choose PSA and ALogP as two descriptors in your calculations, then SASA would be an "inaccurate" parameter. It is more logical to cherry-pick from every "type" of descriptors especially geometrically-oriented ones.

Finally, as a matter of fact, generation of large-scale valid 2D-QSAR is not an easy task at all. Diversity in structural profile of the ligands might in fact produce different binding poses in protein binding sites which cannot be conveniently attributed to 2D-parameters (due to such confounding factors).

This article may be productive for better selection of descriptors. Also please check similar papers too.

Article Descriptors and their selection methods in QSAR analysis: Pa...

Thank you Ardavan. Can you please give me name of some softwares to select descriptors? I'm using WEKA, but as we have seen, it's not show a good result. Or maybe I'm using it wrong way?

You're welcome. Sure, you can calculate many descriptors by using ab initio softwares which compute the physicochemical properties of chemicals. For example, Gaussian or Jaguar (one of the Schrodinger's module). Also if ab initio calculation are taking too much time, you can simply use specialized softwares for obtaining theses descriptors like Paddle, MOE or Dragon or Discovery studio. There are some other softwares in this field like CORAL, SARpy and Weka. Perhaps Paddle is one of the easiest and fastest softwares among these packages to handle the data.