Optibrium

Partial Least Squares - This is a robust technique for generation of linear models based on multiple descriptors

Radial Basis Functions - This unique approach is an efficient numerical technique to generate a non-linear function that ‘maps’ the property variations in a descriptor space. A more computationally intensive version of this approach can also be used where a genetic algorithm initially searches for the optimal descriptor space in which to describe those variations.

Gaussian Processes - This powerful ‘machine learning’ technique generates robust non-linear models using a Bayesian statistical approach1.

Decision Trees - This is a recursive partitioning approach to building classification models in cases where good ‘continuous’ data are not available.

All appropriate methods are automatically applied simultaneously to each modeling problem and the results are rigorously compared to identify the best model of your data.

Yes, once a validated model has been built, you can easily share it or ‘publish’ it on a central server for all StarDrop users to access. This enables new models to be rapidly applied to the design of new chemistry.

The data set of molecules and associated property values that have been experimentally determined are split into three sub-sets; a training set used to train individual models using each modeling technique, a validation set used to compare the performance of each model in order to select the best and an independent test set used to confirm the predictive power of the final model. This provides for the development of a robust model.