Drug characteristics prediction

To predict a drug´s biomedical and biochemical characteristics, such as drug toxicity or efficacy, it often requires large dimensional bio-features as inputs, for example, structure information among frequently used available public data. It is a very challenging problem to apply predictive algorithms when there is a large dimension of descriptive features for each compound while the number of available samples (compounds) is limited. This is also true generally for many bio-informatics applications. We present here an innovative feature clustering method to group, select and reduce dimensionality, therefore is able to produce more accurate prediction. In our example, we studied a sample of1200 drug candidates with 140 having the desired characteristics. There are 3400 bio-features as input features. We used the innovative feature clustering method to intelligently group, select and reduce dimensionality, therefore, significantly increase the predictive accuracy when using the 140 compounds as the training examples to predict the desired characteristics for new candidates.