Solubility, neural network prediction

Recently, several QSPR solubility prediction models based on a fairly large and diverse data set were generated. Huuskonen developed the models using MLRA and back-propagation neural networks (BPG) on a data set of 1297 diverse compoimds [22]. The compounds were described by 24 atom-type E-state indices and six other topological indices. For the 413 compoimds in the test set, MLRA gave = 0.88 and s = 0.71 and neural network provided... 

Figure 10.1-3. Predicted versus experimental solubility values of 552 compounds in the test set by a back-propagation neural network with 18 topological descriptors.

Huuskonen, J., Salo, M., Taskinen, J., Aqueous solubility prediction of drugs based on molecular topology and neural network modeling, J. 

From a modeling standpoint, the prediction of a molecule s solubility is a very difficult task because of the issues listed above [13-15]. The problem of predicting solubility has been attacked with reasonable success with complex neural network models. While not interpretable, neural networks can function as an in silico assay. Other techniques which are more interpretable have also been applied to the problem. 

Kiss IZ, Mandi G, Beck MT (2000) Artificial neural network approach to predict the solubility of C60 in various solvents. J. Phys. Chem. Sect A 104 8081-8088. 

In addition to methods that rely on experimental inputs, there are prediction tools that are based on structure alone. A number of different techniques have been used to correlate a variety of structure-derived descriptors with observed solubilities, including linear regression [20-25] and neural networks [26-36]. Interestingly, despite... 

Bodor, et al. [42] compare the use of artificial neural networks with regression analysis techniques for the development of predictive solubility models. They report that the performance of the neural network model is superior to the regression-based model. Their study is based on a training set of 331 compounds. The model requires a diverse set of molecular descriptors to account for the structural variety in the training compounds. 

Although various computational approaches for the prediction of intestinal drug permeability and solubility have been reported [219], recent computer-based absorption models utilize a large number of topological, electronic, and geometric descriptors in an effort to take both aqueous drug solubility and permeability into account. Thus, descriptors of partitioned total surface areas [168], Abraham molecular descriptors [220,221], and a variety of structural descriptors in combination with neural networks [222] have been shown to be determinants of oral drug absorption. 

Duffythe group contribution approach of Klopman and the neural network model of Huuskonen. " The PCCHEM program used at the US Environmental Protection Agency (EPA) incorporates three different equations. All of them are similar to GSE but have different coefficients to predict aqueous solubility depending on the range of log P values. " Meylan and Howard used a database of 817 (RMSE = 0.62) compounds to derive a heuristic equation ... 

China used these polymers in designed controlled release tablets of nifedipine and nimodipine. Using the drug solubility, polymer hydration rate, quantity of polymer, and quantity of an added surfactant as the inputs, they were able to use neural networks to predict drug release successfully. 

The previous considered methods usually depend on linear methods (MLR, PLS) to establish structure-solubility correlations for prediction of solubility of molecules. The work of Goller et al. [51] used a neural network ensemble to predict the apparent solubility of Bayer in-house organic compounds. The solubility was measured in buffer at pH 6.5, which mimics the medium in the human gastrointestinal tract. The authors used the calculated distribution coefficient log/1 (at several pH values), a number of 3D COSMO-derived parameters and some 2D descriptors. The final model was developed using 4806 compounds (RMSE = 0.72) and provided a similar accuracy (RMSE = 0.73) for the prediction of 7222 compounds that were not used to develop the model. The method, however, is quite slow, and it takes about 15 seconds to screen one molecule on an Intel Xeon 2.8 GHz CPU. 

Methods Using 2D and ID Descriptors A good number of articles on aqueous solubility used a nonlinear method of data analysis, in particular, for methods developed with ID and 2D descriptors. Huuskonen [16] used E-state indexes [52,53] and several other topological indexes, with a total of 30 indexes, to develop his models. The predicted results for the 413 test set, SE = 0.71, calculated with MLRA were improved with a neural network, resulting in SE = 0.6. Tetko [17] noticed that E-state indexes represent a complete system of descriptors for molecules, and thus only these descriptors are sufficient to develop the aqueous solubility model. Indeed the model developed by the authors using exclusively E-state indexes provides similar results when compared to the model of Huuskonen [16]. Later on, the model was redeveloped using the Associative Neural Network (ASNN) method [54],... 

Tabaraki, R., Khayamian, T. and Ensafi, A.A. (2006) Wavelet neural network modeling in QSPR for prediction of solubility of 25 anthraquinone dyes at different temperatures and pressures in supercritical carbon dioxide./. Md. Graph. Model., 25, 46-54. 

Next to ADME phenomena, recent data mining studies also focused on the development or improvement of models predicting physicochemical properties relevant to the field of ADME. Examples are Henry s law constant [92], polar surface area [93], and log P [94]. These models try to overcome limitations of already existing models, see for example SlogP [94] vs. Clogp [95], or aqueous solubility [96], The latter study used more than 2000 compounds selected from the AQUASOL [97] and PHYSOPROP [98] databases. Comparison with a multilinear regression showed clear preference for the neural network. 

Although a majority of the published ADMET models are based on linear multivariate methods as discussed in Section 16.3.3.1, other nonlinear methods have also been employed. The most commonly used nonlinear method in ADMET modeling is neural networks (NNs). Backpropagation NNs have been used to model absorption, permeation, as well as solubility and toxicological effects. A particular problem for many NNs is the tendency for these networks to overtrain (see further discussions on model validation in Section 16.3.3.4), which needs to be closely monitored to avoid the situation where the derived model becomes an encyclopedia , that is, the model can perfectly explain the variance of the investigated property of the compounds used to derive the model but have quite poor predictive ability with respect to new compounds. 

---