Automatic descriptor selection

Gathering of Common Set of Descriptors. The descriptor files of ADAPT typically hold many more descriptors for a set of compounds of interest than can be studied at one time. Therefore, a routine allows the user to select which of the available descriptors to form into a data set for analysis. This allows the user to conveniently study different sets of descriptors for the same set of compounds. When a data set is generated, the descriptors are automatically autoscaled in order to give each descriptor equal weight in the subsequent analysis. 

Bayesian neural networks (BNNs) are an alternative to the more traditional ANNs. The main advantage with BNNs is that they are less prone to overtraining compared to ANNs. BNNs are based on Bayesian probabilistics for the network training. Network weights are determined by Bayesian inference. BNNs have been successfully used together with automatic relevance determination (ARD) for the selection of relevant descriptors to model aqueous solubility [89]. For a good review on BNNs, see Ref. [90]. 

One of the problems when applying classical QSAR techniques is the right choice of the method and the descriptor combination. In principle, two general approaches might be undertaken to overcome this issue, which normally is pursued on a trial and error basis. One is to automatically combine feature selection algorithms with classification and regression tools and the other is to combinatorially explore the descriptor/method space. The latter was recently introduced by the group of Tropsha (combinatorial QSAR) [47]. 

The dimensions for a descriptor are dehned in two groups — Cartesian distance and 2D property — where the minimum, maximum, and resolution of the vector in the hrst dimension of the descriptor can be dehned. The track bars are adapted automatically to changes for example, resolution is calculated and minimum-maximum dependencies are corrected. When the binary checkbox is clicked, only selections are possible that result in dyadic vector length (i.e., the dimension is a factor of 2"). This feature prevents the complicated adjustment of all settings to gain a binary vector that is necessary for transformations. 

It is claimed that CASE differs from other techniques in being completely automatic and by learning directly from the crude data, selecting its own descriptors from the practically infinite number of possible structural assemblies and creating an ad hoc dictionary without human interference [524]. While this statement is against all prior experience with automated approaches, even from a critical point of view it cannot be ruled out that the CASE approach may for the first time be an approximation of artificial intelligence to the medicinal chemist s intuition and skill. 

---