What is a Predictor?
Predictor is another name for QSAR or predictive model. It is a statistical function that maps molecules to a given property or objective (activity, ADMET, etc.). It needs an initial dataset to be fitted on. Once fitted, it can be used on new molecules to predict their value.
Example of a toxicity QSAR model
What is Makya's QSAR Module?
Various components of the Makya’s AI framework (molecular representation, models, etc.) have parameters associated with them. For each individual objective that the user wants to model (activity, ADMET properties...), the optimal combination of these parameters to get the most effective predictor might differ. Therefore, it is crucial for the predictor training framework to be flexible and choose the best combination of parameters for every different objective.
In Makya, this is achieved via a proprietary module called QSAR that performs AutoML (Auto Machine Learning). AutoML automatically tests different combinations of representations and models for each objective and optimizes the parameters to select an optimal solution for each objective.
This optimization is the reason why training predictors in Makya takes some time, since multiple parameter combinations need to be benchmarked. It also explains why the parameters differ between objectives.