Value of p-Value.

The goal of this manuscript is to discuss important aspects of external validation of classification and category Quantitative Structure - Activity/Property/Toxicity Relationship QS/A/P/T/R models that to the best of author's knowledge are not addressed in publications. Statistical significance (in terms of p-value) and accuracy of prediction (in terms of Correct Classification Rate (CCR)) of external validation set compounds are among most important characteristics of the models. We assert that in most cases the models built for classification or category response variable should be statistically significant and predictive for each class or category. We show that three thresholds of the number of compounds in each class or category of the external validation sets should be satisfied. 1) The p-value criterion can never be satisfied, if the number of compounds is below the first threshold. 2) If the number of compounds is between the first and the second thresholds, p-value criterion should be used. 3) If it is higher than the third threshold, classification or category accuracy criterion should be used. 4) If the number of compounds is between second and third thresholds, either one or the other criterion should be used depending on the value of p-value. 5) When the number of compounds in the class approaches infinity, the maximum relative error of prediction approaches the relative expected error. The results are of interest in other areas of multidimensional data analysis.

Multi-Descriptor Read Across (MuDRA): A Simple and Transparent Approach for Developing Accurate Quantitative Structure-Activity Relationship Models.

Multiple approaches to quantitative structure-activity relationship (QSAR) modeling using various statistical or machine learning techniques and different types of chemical descriptors have been developed over the years. Oftentimes models are used in consensus to make more accurate predictions at the expense of model interpretation. We propose a simple, fast, and reliable method termed Multi-Descriptor Read Across (MuDRA) for developing both accurate and interpretable models. The method is conceptually related to the well-known kNN approach but uses different types of chemical descriptors simultaneously for similarity assessment. To benchmark the new method, we have built MuDRA models for six different end points (Ames mutagenicity, aquatic toxicity, hepatotoxicity, hERG liability, skin sensitization, and endocrine disruption) and compared the results with those generated with conventional consensus QSAR modeling. We find that models built with MuDRA show consistently high external accuracy similar to that of conventional QSAR models. However, MuDRA models excel in terms of transparency, interpretability, and computational efficiency. We posit that due to its methodological simplicity and reliable predictive accuracy, MuDRA provides a powerful alternative to a much more complex consensus QSAR modeling. MuDRA is implemented and freely available at the Chembench web portal ( https://chembench.mml.unc.edu/mudra ).

Application of quantitative structure-activity relationship models of 5-HT1A receptor binding to virtual screening identifies novel and potent 5-HT1A ligands.

The 5-hydroxytryptamine 1A (5-HT1A) serotonin receptor has been an attractive target for treating mood and anxiety disorders such as schizophrenia. We have developed binary classification quantitative structure-activity relationship (QSAR) models of 5-HT1A receptor binding activity using data retrieved from the PDSP Ki database. The prediction accuracy of these models was estimated by external 5-fold cross-validation as well as using an additional validation set comprising 66 structurally distinct compounds from the World of Molecular Bioactivity database. These validated models were then used to mine three major types of chemical screening libraries, i.e., drug-like libraries, GPCR targeted libraries, and diversity libraries, to identify novel computational hits. The five best hits from each class of libraries were chosen for further experimental testing in radioligand binding assays, and nine of the 15 hits were confirmed to be active experimentally with binding affinity better than 10 µM. The most active compound, Lysergol, from the diversity library showed very high binding affinity (Ki) of 2.3 nM against 5-HT1A receptor. The novel 5-HT1A actives identified with the QSAR-based virtual screening approach could be potentially developed as novel anxiolytics or potential antischizophrenic drugs.

Data set modelability by QSAR.

We introduce a simple MODelability Index (MODI) that estimates the feasibility of obtaining predictive QSAR models (correct classification rate above 0.7) for a binary data set of bioactive compounds. MODI is defined as an activity class-weighted ratio of the number of nearest-neighbor pairs of compounds with the same activity class versus the total number of pairs. The MODI values were calculated for more than 100 data sets, and the threshold of 0.65 was found to separate the nonmodelable and modelable data sets.

Integrative chemical-biological read-across approach for chemical hazard classification.

Traditional read-across approaches typically rely on the chemical similarity principle to predict chemical toxicity; however, the accuracy of such predictions is often inadequate due to the underlying complex mechanisms of toxicity. Here, we report on the development of a hazard classification and visualization method that draws upon both chemical structural similarity and comparisons of biological responses to chemicals measured in multiple short-term assays ("biological" similarity). The Chemical-Biological Read-Across (CBRA) approach infers each compound's toxicity from both chemical and biological analogues whose similarities are determined by the Tanimoto coefficient. Classification accuracy of CBRA was compared to that of classical RA and other methods using chemical descriptors alone or in combination with biological data. Different types of adverse effects (hepatotoxicity, hepatocarcinogenicity, mutagenicity, and acute lethality) were classified using several biological data types (gene expression profiling and cytotoxicity screening). CBRA-based hazard classification exhibited consistently high external classification accuracy and applicability to diverse chemicals. Transparency of the CBRA approach is aided by the use of radial plots that show the relative contribution of analogous chemical and biological neighbors. Identification of both chemical and biological features that give rise to the high accuracy of CBRA-based toxicity prediction facilitates mechanistic interpretation of the models.

Discovery of novel antimalarial compounds enabled by QSAR-based virtual screening.

Quantitative structure-activity relationship (QSAR) models have been developed for a data set of 3133 compounds defined as either active or inactive against P. falciparum. Because the data set was strongly biased toward inactive compounds, different sampling approaches were employed to balance the ratio of actives versus inactives, and models were rigorously validated using both internal and external validation approaches. The balanced accuracy for assessing the antimalarial activities of 70 external compounds was between 87% and 100% depending on the approach used to balance the data set. Virtual screening of the ChemBridge database using QSAR models identified 176 putative antimalarial compounds that were submitted for experimental validation, along with 42 putative inactives as negative controls. Twenty five (14.2%) computational hits were found to have antimalarial activities with minimal cytotoxicity to mammalian cells, while all 42 putative inactives were confirmed experimentally. Structural inspection of confirmed active hits revealed novel chemical scaffolds, which could be employed as starting points to discover novel antimalarial agents.

Does rational selection of training and test sets improve the outcome of QSAR modeling?

Prior to using a quantitative structure activity relationship (QSAR) model for external predictions, its predictive power should be established and validated. In the absence of a true external data set, the best way to validate the predictive ability of a model is to perform its statistical external validation. In statistical external validation, the overall data set is divided into training and test sets. Commonly, this splitting is performed using random division. Rational splitting methods can divide data sets into training and test sets in an intelligent fashion. The purpose of this study was to determine whether rational division methods lead to more predictive models compared to random division. A special data splitting procedure was used to facilitate the comparison between random and rational division methods. For each toxicity end point, the overall data set was divided into a modeling set (80% of the overall set) and an external evaluation set (20% of the overall set) using random division. The modeling set was then subdivided into a training set (80% of the modeling set) and a test set (20% of the modeling set) using rational division methods and by using random division. The Kennard-Stone, minimal test set dissimilarity, and sphere exclusion algorithms were used as the rational division methods. The hierarchical clustering, random forest, and k-nearest neighbor (kNN) methods were used to develop QSAR models based on the training sets. For kNN QSAR, multiple training and test sets were generated, and multiple QSAR models were built. The results of this study indicate that models based on rational division methods generate better statistical results for the test sets than models based on random division, but the predictive power of both types of models are comparable.

Quantitative structure-property relationship modeling of remote liposome loading of drugs.

Remote loading of liposomes by trans-membrane gradients is used to achieve therapeutically efficacious intra-liposome concentrations of drugs. We have developed Quantitative Structure Property Relationship (QSPR) models of remote liposome loading for a data set including 60 drugs studied in 366 loading experiments internally or elsewhere. Both experimental conditions and computed chemical descriptors were employed as independent variables to predict the initial drug/lipid ratio (D/L) required to achieve high loading efficiency. Both binary (to distinguish high vs. low initial D/L) and continuous (to predict real D/L values) models were generated using advanced machine learning approaches and 5-fold external validation. The external prediction accuracy for binary models was as high as 91-96%; for continuous models the mean coefficient R(2) for regression between predicted versus observed values was 0.76-0.79. We conclude that QSPR models can be used to identify candidate drugs expected to have high remote loading capacity while simultaneously optimizing the design of formulation experiments.

Development, validation, and use of quantitative structure-activity relationship models of 5-hydroxytryptamine (2B) receptor ligands to identify novel receptor binders and putative valvulopathic compounds among common drugs.

Some antipsychotic drugs are known to cause valvular heart disease by activating serotonin 5-HT(2B) receptors. We have developed and validated binary classification QSAR models capable of predicting potential 5-HT(2B) actives. The classification accuracies of the models built to discriminate 5-HT(2B) actives from the inactives were as high as 80% for the external test set. These models were used to screen in silico 59,000 compounds included in the World Drug Index, and 122 compounds were predicted as actives with high confidence. Ten of them were tested in radioligand binding assays and nine were found active, suggesting a success rate of 90%. All validated actives were then tested in functional assays, and one compound was identified as a true 5-HT(2B) agonist. We suggest that the QSAR models developed in this study could be used as reliable predictors to flag drug candidates that are likely to cause valvulopathy.

A novel two-step hierarchical quantitative structure-activity relationship modeling work flow for predicting acute toxicity of chemicals in rodents.

BACKGROUND: Accurate prediction of in vivo toxicity from in vitro testing is a challenging problem. Large public-private consortia have been formed with the goal of improving chemical safety assessment by the means of high-throughput screening. OBJECTIVE: A wealth of available biological data requires new computational approaches to link chemical structure, in vitro data, and potential adverse health effects. METHODS AND RESULTS: A database containing experimental cytotoxicity values for in vitro half-maximal inhibitory concentration (IC(50)) and in vivo rodent median lethal dose (LD(50)) for more than 300 chemicals was compiled by Zentralstelle zur Erfassung und Bewertung von Ersatz- und Ergaenzungsmethoden zum Tierversuch (ZEBET; National Center for Documentation and Evaluation of Alternative Methods to Animal Experiments). The application of conventional quantitative structure-activity relationship (QSAR) modeling approaches to predict mouse or rat acute LD(50) values from chemical descriptors of ZEBET compounds yielded no statistically significant models. The analysis of these data showed no significant correlation between IC(50) and LD(50). However, a linear IC(50) versus LD(50) correlation could be established for a fraction of compounds. To capitalize on this observation, we developed a novel two-step modeling approach as follows. First, all chemicals are partitioned into two groups based on the relationship between IC(50) and LD(50) values: One group comprises compounds with linear IC(50) versus LD(50) relationships, and another group comprises the remaining compounds. Second, we built conventional binary classification QSAR models to predict the group affiliation based on chemical descriptors only. Third, we developed k-nearest neighbor continuous QSAR models for each subclass to predict LD(50) values from chemical descriptors. All models were extensively validated using special protocols. CONCLUSIONS: The novelty of this modeling approach is that it uses the relationships between in vivo and in vitro data only to inform the initial construction of the hierarchical two-step QSAR models. Models resulting from this approach employ chemical descriptors only for external prediction of acute rodent toxicity.

QSAR modeling of the blood-brain barrier permeability for diverse organic compounds.

PURPOSE: Development of externally predictive Quantitative Structure-Activity Relationship (QSAR) models for Blood-Brain Barrier (BBB) permeability. METHODS: Combinatorial QSAR analysis was carried out for a set of 159 compounds with known BBB permeability data. All six possible combinations of three collections of descriptors derived from two-dimensional representations of molecules as chemical graphs and two QSAR methodologies have been explored. Descriptors were calculated by MolconnZ, MOE, and Dragon software. QSAR methodologies included k-Nearest Neighbors and Support Vector Machine approaches. All models have been rigorously validated using both internal and external validation methods. RESULTS: The consensus prediction for the external evaluation set afforded high predictive power (R2 = 0.80 for 10 compounds within the applicability domain after excluding one activity outlier). Classification accuracies for two additional external datasets, including 99 drugs and 267 organic compounds, classified as permeable (BBB+) or non-permeable (BBB-) were 82.5% and 59.0%, respectively. The use of a fairly conservative model applicability domain increased the prediction accuracy to 100% and 83%, respectively (while naturally reducing the dataset coverage to 60% and 43%, respectively). Important descriptors that affect BBB permeability are discussed. CONCLUSION: Models developed in these studies can be used to estimate the BBB permeability of drug candidates at early stages of drug development.

Combinatorial QSAR modeling of specificity and subtype selectivity of ligands binding to serotonin receptors 5HT1E and 5HT1F.

The Quantitative Structure-Activity Relationship (QSAR) approach has been applied to model binding affinity and receptor subtype selectivity of human 5HT1E and 5HT1F receptor-ligands. The experimental data were obtained from the PDSP Ki Database. Several descriptor types and data-mining approaches have been used in the context of combinatorial QSAR modeling. Data mining approaches included k Nearest Neighbor, Automated Lazy Learning (ALL), and PLS; descriptor types included MolConnZ, MOE, DRAGON, Frequent Subgraphs (FSG), and Molecular Hologram Fingerprints (MHFs). Highly predictive QSAR models were generated for all three data sets (i.e., for ligands of both receptor subtypes and for subtype selectivity), and different individual techniques were proved best in each case. For real value activity data available for 5HT1E and 5HT1F ligand binding, models were characterized by leave-one-out cross-validated R(2) (q(2)) for the training sets and predictive R(2) values for the test sets. The best models for 5HT1E ligands were obtained with the kNN approach combined with MolConnZ descriptors (q(2)=0.69, R(2)=0.92); for 5HT1F ligands ALL QSAR method using MolConnZ descriptors gave the best results (R(2)=0.92). Rigorously validated classification models were also developed for the 5HT1E/5HT1F subtype selectivity data set with high correct classification accuracy for both training (CCRtrain=0.88) and test (CCRtest=1.00) sets using kNN with MolConnZ descriptors. The external predictive power of QSAR models was further validated by virtual screening of The Scripps Research Institute (TSRI) screening library to recover 5HT1E agonists and antagonists (not present in the original PDSP data set) with high enrichment factors. The successful development of externally predictive and interpretative QSAR models affords further design and discovery of novel subtype specific GPCR agents.

Differentiation of AmpC beta-lactamase binders vs. decoys using classification kNN QSAR modeling and application of the QSAR classifier to virtual screening.

The use of inaccurate scoring functions in docking algorithms may result in the selection of compounds with high predicted binding affinity that nevertheless are known experimentally not to bind to the target receptor. Such falsely predicted binders have been termed 'binding decoys'. We posed a question as to whether true binders and decoys could be distinguished based only on their structural chemical descriptors using approaches commonly used in ligand based drug design. We have applied the k-Nearest Neighbor (kNN) classification QSAR approach to a dataset of compounds characterized as binders or binding decoys of AmpC beta-lactamase. Models were subjected to rigorous internal and external validation as part of our standard workflow and a special QSAR modeling scheme was employed that took into account the imbalanced ratio of inhibitors to non-binders (1:4) in this dataset. 342 predictive models were obtained with correct classification rate (CCR) for both training and test sets as high as 0.90 or higher. The prediction accuracy was as high as 100% (CCR = 1.00) for the external validation set composed of 10 compounds (5 true binders and 5 decoys) selected randomly from the original dataset. For an additional external set of 50 known non-binders, we have achieved the CCR of 0.87 using very conservative model applicability domain threshold. The validated binary kNN QSAR models were further employed for mining the NCGC AmpC screening dataset (69653 compounds). The consensus prediction of 64 compounds identified as screening hits in the AmpC PubChem assay disagreed with their annotation in PubChem but was in agreement with the results of secondary assays. At the same time, 15 compounds were identified as potential binders contrary to their annotation in PubChem. Five of them were tested experimentally and showed inhibitory activities in millimolar range with the highest binding constant K(i) of 135 microM. Our studies suggest that validated QSAR models could complement structure based docking and scoring approaches in identifying promising hits by virtual screening of molecular libraries.

Predictive QSAR modeling workflow, model applicability domains, and virtual screening.

Quantitative Structure Activity Relationship (QSAR) modeling has been traditionally applied as an evaluative approach, i.e., with the focus on developing retrospective and explanatory models of existing data. Model extrapolation was considered if only in hypothetical sense in terms of potential modifications of known biologically active chemicals that could improve compounds' activity. This critical review re-examines the strategy and the output of the modern QSAR modeling approaches. We provide examples and arguments suggesting that current methodologies may afford robust and validated models capable of accurate prediction of compound properties for molecules not included in the training sets. We discuss a data-analytical modeling workflow developed in our laboratory that incorporates modules for combinatorial QSAR model development (i.e., using all possible binary combinations of available descriptor sets and statistical data modeling techniques), rigorous model validation, and virtual screening of available chemical databases to identify novel biologically active compounds. Our approach places particular emphasis on model validation as well as the need to define model applicability domains in the chemistry space. We present examples of studies where the application of rigorously validated QSAR models to virtual screening identified computational hits that were confirmed by subsequent experimental investigations. The emerging focus of QSAR modeling on target property forecasting brings it forward as predictive, as opposed to evaluative, modeling approach.

A novel automated lazy learning QSAR (ALL-QSAR) approach: method development, applications, and virtual screening of chemical databases using validated ALL-QSAR models.

A novel automated lazy learning quantitative structure-activity relationship (ALL-QSAR) modeling approach has been developed on the basis of the lazy learning theory. The activity of a test compound is predicted from a locally weighted linear regression model using chemical descriptors and the biological activity of the training set compounds most chemically similar to this test compound. The weights with which training set compounds are included in the regression depend on the similarity of those compounds to a test compound. We have applied the ALL-QSAR method to several experimental chemical data sets including 48 anticonvulsant agents with known ED50 values, 48 dopamine D1-receptor antagonists with known competitive binding affinities (Ki), and a Tetrahymena pyriformis data set containing 250 phenolic compounds with toxicity IGC50 values. When applied to database screening, models developed for anticonvulsant agents identified several known anticonvulsant compounds that were not only absent in the training set but highly chemically dissimilar to the training set compounds. This initial success indicates that ALL-QSAR can be further exploited as a general tool for accurate bioactivity prediction and database screening in drug design and discovery. Because of its local nature, the ALL-QSAR approach appears to be especially well-suited for the development of highly predictive models for the sparse or unevenly distributed data sets.

Novel ligands for the human histamine H1 receptor: synthesis, pharmacology, and comparative molecular field analysis studies of 2-dimethylamino-5-(6)-phenyl-1,2,3,4-tetrahydronaphthalenes.

This paper reports the synthesis of a novel series of (+/-)-2-dimethylamino- 5- and 6-phenyl-1,2,3,4-tetrahydronaphthalene derivatives (5- and 6-APTs), and, corresponding affinity, functional activity, and, molecular modeling studies with regard to drug design targeting the human histamine H1 receptor. The 5-APTs have 2- to 4-fold higher H1 receptor affinity than the endogenous agonist histamine. The chemical nature of a meta-substituent on the 5-APT pendant phenyl moiety does not significantly affect H1 affinity. In contrast, analogous meta-substitution for the 6-APTs increases H1 affinity up to 100-fold. The new APTs do not activate H1 receptor-linked intracellular signaling and apparently are competitive H1 antagonists. A new model that establishes structural parameters for binding to the human H1 receptor by APTs and other ligands was developed using 3-D QSAR (CoMFA). The model predicts H1 ligand binding with a higher degree of external predictability compared to a previously reported model. The APTs also were examined for activity at human serotonin 5-HT2A and 5-HT2C receptors, which are phylogenetically closely related to the H1 receptor. 5-APT and m-Cl-6-APT were identified as novel agonists that selectively activate 5-HT2C receptors. It is concluded that the lipophilic (brain-penetrating) APT molecular scaffold may have pharmacotherapeutic potential in neuropsychiatric diseases.

Combinatorial QSAR modeling of P-glycoprotein substrates.

Quantitative structure-activity (property) relationship (QSAR/QSPR) models are typically generated with a single modeling technique using one type of molecular descriptors. Recently, we have begun to explore a combinatorial QSAR approach which employs various combinations of optimization methods and descriptor types and includes rigorous and consistent model validation (Kovatcheva, A.; Golbraikh, A.; Oloff, S.; Xiao, Y.; Zheng, W.; Wolschann, P.; Buchbauer, G.; Tropsha, A. Combinatorial QSAR of Ambergris Fragrance Compounds. J. Chem. Inf. Comput. Sci. 2004, 44, 582-95). Herein, we have applied this approach to a data set of 195 diverse substrates and nonsubstrates of P-glycoprotein (P-gp) that plays a crucial role in drug resistance. Modeling methods included k-nearest neighbors classification, decision tree, binary QSAR, and support vector machines (SVM). Descriptor sets included molecular connectivity indices, atom pair (AP) descriptors, VolSurf descriptors, and molecular operation environment descriptors. Each descriptor type was used with every QSAR modeling technique; so, in total, 16 combinations of techniques and descriptor types have been considered. Although all combinations resulted in models with a high correct classification rate for the training set (CCR(train)), not all of them had high classification accuracy for the test set (CCR(test)). Thus, predictive models have been generated only for some combinations of the methods and descriptor types, and the best models were obtained using SVM classification with either AP or VolSurf descriptors; they were characterized by CCR(train) = 0.94 and 0.88 and CCR(test) = 0.81 and 0.81, respectively. The combinatorial QSAR approach identified models with higher predictive accuracy than those reported previously for the same data set. We suggest that, in the absence of any universally applicable "one-for-all" QSAR methodology, the combinatorial QSAR approach should become the standard practice in QSPR/QSAR modeling.

Development of quantitative structure-binding affinity relationship models based on novel geometrical chemical descriptors of the protein-ligand interfaces.

Novel geometrical chemical descriptors have been derived on the basis of the computational geometry of protein-ligand interfaces and Pauling atomic electronegativities (EN). Delaunay tessellation has been applied to a diverse set of 517 X-ray characterized protein-ligand complexes yielding a unique collection of interfacial nearest neighbor atomic quadruplets for each complex. Each quadruplet composition was characterized by a single descriptor calculated as the sum of the EN values for the four participating atom types. We termed these simple descriptors generated from atomic EN values and derived with the Delaunay Tessellation the ENTess descriptors and used them in the variable selection k-nearest neighbor quantitative structure-binding affinity relationship (QSBR) studies of 264 diverse protein-ligand complexes with known binding constants. Twenty-four complexes with chemically dissimilar ligands were set aside as an independent validation set, and the remaining dataset of 240 complexes was divided into multiple training and test sets. The best models were characterized by the leave-one-out cross-validated correlation coefficient q(2) as high as 0.66 for the training set and the correlation coefficient R(2) as high as 0.83 for the test set. The high predictive power of these models was confirmed independently by applying them to the validation set of 24 complexes yielding R(2) as high as 0.85. We conclude that QSBR models built with the ENTess descriptors can be instrumental for predicting the binding affinity of receptor-ligand complexes.

Quantitative structure-activity relationship analysis of pyridinone HIV-1 reverse transcriptase inhibitors using the k nearest neighbor method and QSAR-based database mining.

We have developed quantitative structure-activity relationship (QSAR) models for 44 non-nucleoside HIV-1 reverse transcriptase inhibitors (NNRTIs) of the pyridinone derivative type. The k nearest neighbor (kNN) variable selection approach was used. This method utilizes multiple descriptors such as molecular connectivity indices, which are derived from two-dimensional molecular topology. The modeling process entailed extensive validation including the randomization of the target property (Y-randomization) test and the division of the dataset into multiple training and test sets to establish the external predictive power of the training set models. QSAR models with high internal and external accuracy were generated, with leave-one-out cross-validated R2 (q2) values ranging between 0.5 and 0.8 for the training sets and R2 values exceeding 0.6 for the test sets. The best models with the highest internal and external predictive power were used to search the National Cancer Institute database. Derivatives of the pyrazolo[3,4-d]pyrimidine and phenothiazine type were identified as promising novel NNRTIs leads. Several candidates were docked into the binding pocket of nevirapine with the AutoDock (version 3.0) software. Docking results suggested that these types of compounds could be binding in the NNRTI binding site in a similar mode to a known non-nucleoside inhibitor nevirapine.

Application of predictive QSAR models to database mining: identification and experimental validation of novel anticonvulsant compounds.

We have developed a drug discovery strategy that employs variable selection quantitative structure-activity relationship (QSAR) models for chemical database mining. The approach starts with the development of rigorously validated QSAR models obtained with the variable selection k nearest neighbor (kNN) method (or, in principle, with any other robust model-building technique). Model validation is based on several statistical criteria, including the randomization of the target property (Y-randomization), independent assessment of the training set model's predictive power using external test sets, and the establishment of the model's applicability domain. All successful models are employed in database mining concurrently; in each case, only variables selected as a result of model building (termed descriptor pharmacophore) are used in chemical similarity searches comparing active compounds of the training set (queries) with those in chemical databases. Specific biological activity (characteristic of the training set compounds) of external database entries found to be within a predefined similarity threshold of the training set molecules is predicted on the basis of the validated QSAR models using the applicability domain criteria. Compounds judged to have high predicted activities by all or the majority of all models are considered as consensus hits. We report on the application of this computational strategy for the first time for the discovery of anticonvulsant agents in the Maybridge and National Cancer Institute (NCI) databases containing ca. 250,000 compounds combined. Forty-eight anticonvulsant agents of the functionalized amino acid (FAA) series were used to build kNN variable selection QSAR models. The 10 best models were applied to mining chemical databases, and 22 compounds were selected as consensus hits. Nine compounds were synthesized and tested at the NIH Epilepsy Branch, Rockville, MD using the same biological test that was employed to assess the anticonvulsant activity of the training set compounds; of these nine, four were exact database hits and five were derived from the hits by minor chemical modifications. Seven of these nine compounds were confirmed to be active, indicating an exceptionally high hit rate. The approach described in this report can be used as a general rational drug discovery tool.