Try generating multiple models using multiple random allocation of data points to training and test sets. Then, either select the best few models that represent all of the data or use a suitable consensus method for prediction.

thank you dear chandrika

do you have any suggestion in  set this protocol for make a good model?

Another point : we have often found that 2D QSAR models have lower error than 3D QSAR models for the same data. This would be due to fewer assumptions in 2D QSAR. Developing 2D models may work better for you.