Complete databases

Figure 2. Distribution of predicted scores for the training sets (solid lines) and for the complete databases (dashed lines).

contained in public databases. The World Drug Index [13] and the Available Chemicals Directory [14] were used as collections of drugs and non-drugs, respectively. Both databases were preprocessed [4] in order to remove compounds with certain reactive or otherwise unwanted substructures and duplicates. In addition, exact matches of WDI compounds (i.e., drugs) were removed from the ACD. Clearly, there are much more drug-like compounds in the ACD and there are certain compound classes in the WDI which are not typical drugs as, e.g., cytostatics. This is the remaining noise in such data collections. The WDI and ACD compounds were assigned drug-likeness scores of 1 (drug) and 0 (non-drug).

The fingerprint descriptor consists simply of the counts of the Ghose/Crip-pen atom types [6-11] within a molecule.

A neural network approach was chosen due to the advantages such a nonlinear approach has over comparable linear methods like linear regression or partial least squares (the main disadvantage of a neural network is its limited use in the interpretation of the data). The neural network was trained with a randomly drawn training set of 5000 drugs from the WDI and 5000 non-drugs from the ACD by using the SNNS program [12]. The parameters for the network training and a suited network architecture were determined by a number of empirical tests. It turned out that a feed-forward network with 92 input neurons (the Ghose/Crippen fingerprint), five hidden units, and one output neuron (the drug-likeness score) achieved sufficient predictivity. For details of this approach see the original publication [4]. The trained network is now available for predicting the drug-likeness of chemical compounds. Figure 2 shows the quality of this prediction.

The score distributions for the training sets (closed lines) and the complete databases (dashed lines) are compared for training and test data. First, there is no significant difference between the predictions for the training data (5000 drugs and 5000 non-drugs) and the complete databases (38 416 drugs from the WDI and 169 331 non-drugs from the ACD). Thus, the network is highly predictive for data that were not shown to it during the training process. Secondly, about 80% of each dataset were predicted correctly. This approach can be further used to prioritize compounds for synthesis, purchase, or biological testing. At the moment we use at BASF a threshold value of 0.3 for the discrimination between drugs and non-drugs. This shifts the number of correctly classified drugs to 90% while keeping the correctly predicted non-drugs still at a level of 70%.

Retrospective analysis of HTS data

The approach for predicting the drug-likeness of compounds was applied to the analysis of high-throughput screening data. For two receptor assays and three enzyme assays, the percentage of drug-like molecules with a score greater than 0.3 was estimated for four different data sets in the screening cascade:

1. The total amount of compounds that went into screening - usually several hundred thousands.

2. The subset of compounds passing the primary screen at a certain level of percent inhibition - several thousands.

3. The subset passing the secondary screen, a certain IC50 level - several hundred compounds.

4. Manual selection by medicinal chemists for further development in chemistry projects - about ten compounds. The criteria for this selection are somehow the intuitions of experienced medicinal chemists for drug-likeness, e.g., bioavailability, toxicity, stability, or synthesizability. This is score > 0.3 [%] 100 90 80 70 60 50

0 0

Post a comment