Prediction of intrinsic aqueous solubility (logS₀): The prediction server calculates intrinsic solubility (solubility of the non-ionized compound) values based on the chemical structure information. For the prediction a neural network model trained on a 4548  compound drug-like data set from the PHYSPROP database, and tested on two different of external validation sets, is used. The model is built on nine 2D-MOE (Molecular Operational Environment) descriptors, and the training was done in a three-fold cross validation using a fully connected feed-forward neural network with one hidden layer and nine hidden neurons. (The chemical structures can be fed to the server in SMILES or sdf format, or sketched in the JME applet, the structures are transformed to 2D sdf-format with the Molconvert program from ChemAxon and the descriptors at generated with MOE and fed to the NN-predictor.)

 

Confidence estimate for the intrinsic solubility prediction: A simple confidence index is assigned to each predicted solubility value, indicating how well the compound matches the chemical space of the training set. Compounds for which the predicted target value or the most important descriptor fall within two standard deviations of the corresponding values of the training set are considered to have high accuracy. For compounds that fall within three standard deviation of either property the prediction accuracy is evaluated to be moderate, and compounds that fall outside this range are evaluated to have low prediction accuracy.

 

Prediction of pH-dependent aqueous solubility: The predicted pH-solubility profiles (logS) are calculated using the Henderson-Hasselbalch (HH) equation, using the logS₀ values predicted with the server described above and acid-base dissociation coefficients (pKa values) computed with the Marvin program from ChemAxon.

 

In case of a monoprotic acid the HH equation has the form

log S = log S₀ + log (1 + 10^pH-pKa)

 

in the case of a monoprotic base becomes

log S = log S₀ + log (1 + 10^pKa-pH)

 

and for an ampholyte, the above two equations are combined to give

log S = log S₀ + log (1 + 10^pH-pKa(acid) + 10^pKa(base)-pH)

 

The salt solubility limit of the compounds are not implemented in the present model, but will be included in the next version, and for this purpose a new data set has been measured by colleges at Warsaw University of Technology in Poland.

# Compound                   p_logS0    p_pKa          p_pKb          Rel.
# ========================================================================
Aspirin                      -1.389     3.41                          high
# ========================================================================