Hi peers and experts! I am new to 3D-QSAR and tried to build a model based on 34 molecules, split into 10 test and 24 training sets. In the end, a poor model is evident from r2 and q2 values. Please help to understand the key points, that one should take care of while creating these two sets for model generation. Thanks in advance.
Pratibha Kushwaha It's crucial to ensure that the distribution of compounds in terms of chemical diversity, and activity distribution are roughly balanced between the training and test sets. This balance is essential for preventing bias in model evaluation.
Good morning,
I would also suggest to reduce the size of the test set. Additionally, you can use the Kennard-Stones algorithm for training set/test set splitting. This method has the advantage to consider all the variables. (Article Investigating the influence of data splitting on the predict...
)A python function to perform this operation is available (https://pypi.org/project/kennard-stone/).
Hope this helps!
I suggest you to use NUERAL network in Matlab. The network will split data automatically into training, testing and validation set