It really depends on your model and the size of test/training set. You can estimate individual contributions to remove unnecessary features and perform cross-validation to detect overfitting.

What fits your model compound!? As for ref is concerned, you will need to search and sift yourself!! Spend some time and effort!!!!

If the activity correlates perfectly with one structural descriptor, the answer is: one.

For an intermediate size data set, one should have at least 5 - 6 data points per variable (or per component) to circumvent chance correlations and overfitting.

Which means, for each 5-6 compounds, there shouldn't be more than one predictor or descriptor.

Lukas Schulig , Marwa Abd El Kader Zaater , Martin Klvana , Riaz A. Khan

thank you very much for the answers and suggestions,

very helpful answers and suggestions

It depends on the selection of small molecules under study...and their activity.

The class of molecules, is it having the same heterocyclic chore or a diverse class of molecules.

The range of activity ( ex IC50 in Cancer studies)can be low to medium.

More the number of molecules in the test set; better is the regression coefficient and viceversa.