Spline-Fitting with a Genetic Algorithm:  A Method for Developing Classification Structure−Activity Relationships

Abstract

Classification methods allow for the development of structure-activity relationship models when the target property is categorical rather than continuous. We describe a classification method which fits descriptor splines to activities, with descriptors selected using a genetic algorithm. This method, which we identify as SFGA, is compared to the well-established techniques of recursive partitioning (RP) and soft independent modeling by class analogy (SIMCA) using five series of compounds: cyclooxygenase-2 (COX-2) inhibitors, benzodiazepine receptor (BZR) ligands, estrogen receptor (ER) ligands, dihydrofolate reductase (DHFR) inhibitors, and monoamine oxidase (MAO) inhibitors. Only 1-D and 2-D descriptors were used. Approximately 40% of compounds in each series were assigned to a test set, "cherry-picked" from the complete set such that they lie outside the training set as much as possible. SFGA produced models that were more predictive for all but the DHFR set, for which SIMCA was most predictive. RP gave the least predictive models for all but the MAO set. A similar trend was observed when using training and test sets to which compounds were randomly assigned and when gradually eliminating compounds from the (designed) training set. The stability of models was examined for the random and reduced sets, where stability means that classification statistics and the selected descriptors are similar for models derived from different sets. Here, SIMCA produced the most stable models, followed by SFGA and RP. We show that a consensus approach that combines all three methods outperforms the single best model for all data sets.

Mixed learning algorithms and features ensemble in hepatotoxicity prediction

“…Many SVM, kNN or naive Bayes models were generated from the same number of compounds but with differing molecular descriptor set referred to as vary(MDes) in the figure of its efficiency and ease of use. In addition, the random method had shown to be effective in methods such as Random Forest and Random Decision Trees [50,80]. This process was repeated for 50 times, that is, there were 50 ensemble models built from each combination of 5 base classifiers, 9 base classifiers, etc.…”

“…The method uses a fixed amount of training data on the basis that a model should learn from as many sample as possible to exploit all available information. To the best of our knowledge, there were eight other QSAR studies on ensemble of mixed features and mixed algorithms, applied on training set size of 48-816 compounds [33,[50][51][52][53][54][55][56]. The application of ensemble method had improved the final performances in a majority of these studies when compared with the best performing individual model [52,55].…”

Using Local Models to Improve (Q)SAR Predictivity

“…[22] 4QSAR DHFR 362 0.529 regression 4 QSAR database [22] CPD MOUSE 442 0.152 regression ACD DSSTox databases [23] CPD RAT 580 0.158 regression ACD DSSTox databases [23] ISS MOUSE 316 0.153 regression Benigni/Vari Carcinogenicity datasets [24] ISS RAT 375 0.160 regression Benigni/Vari Carcinogenicity datasets [24] Suth COX2 414 0.493 regression Sutherland dataset [25] Suth DHFR 672 0.501 regression Sutherland dataset [25] Suth ER TOX 410 0.360 regression Sutherland dataset [25] FDAMDD 1216 0.270 regression ACD DSSTox databases [26] Fontaine 435 0.495 classification Fontaine et al [27] CYP INH 2C9 700 0.298 classification Yap and Chen [28] CYP SUB 2C9 700 0.298 classification Yap and Chen [28] NCI AIDS 1000 0.331 classification DTP AIDS Antiviral Screen [29] Values smaller than the black horizontal line indicate that more query molecules are predicted with local models than with the global model. As it can be seen from Figure 3, for the majority of datasets local models are more often applied when it comes to prediction.…”

Feature-map vectors: a new class of informative descriptors for computational drug discovery

“…A detailed description of these two [46]. An alignment reference and feature map were created by applying conformation mining to three DHFR inhibitors with available crystal structures (PDB codes 1HFQ [47], 1S3U [48], and 1KMS [49]).…”