4-((3-bromophenyl)amino)-6,7-dimethoxyquinazoline

QSAR models establish quantitative relationships between chemical structures characterized by chemical descriptors and a target property, e.g., biological activity of chemicals in specific biological assays. Validated and externally predictive models29 (link) can be applied to screen virtual chemical libraries to retrieve compounds with desired properties.32 (link), 37 (link), 38 (link) QSAR modeling employs complex machine leaning algorithms such as Support Vector Machines (SVM) or the k Nearest Neighbors (kNN) that take the descriptor matrix of compounds as inputs and output a predicted value for the modeled property.
The QSAR modeling workflow can be divided into three major steps: (i) data preparation/analysis39 (link) (selection of compounds and descriptors), (ii) model building, and (iii) model validation/selection (including the evaluation of its Applicability Domain – AD). A set of compounds with known experimental activity is randomly split into several training and test sets. Models are built using compounds of each training set and then applied to test set compounds to assess their properties. After application of rigorous tests (such as leave one out, n-fold cross-validation, and Y-randomization), and calculation of model accuracy metrics described below, certain models are selected if and only if they can reasonably predict both the training set as assessed by cross-validation procedures and the test set.40 (link) Models obtained for the modeling set with randomized activities (Y-randomization) should have significantly lower predictive capabilities than models built using modeling set with real activities. Finally, the selected models are applied to the external validation set compounds.
Chemical structures are represented by molecular descriptors.41 In Case Study 2, we used the following two-dimensional MOE descriptors (commercial software distributed by Chemical Computing Group): physical properties, surface areas, atom and bond counts, Kier & Hall connectivity indices, kappa shape indices, adjacency and distance matrix descriptors, pharmacophore feature descriptors, and molecular charges.
The clustering of a chemical dataset consists of merging compounds into independent clusters that include chemically similar molecules42 based on any similarity metrics (e.g., compounds can be clustered based on their biological activity profiles). In this study, we have employed the ISIDA/Cluster program31 implementing the Sequential Agglomerative Hierarchical Non-overlapping (SAHN) method. The parent-child relationships between clusters result in a hierarchical data representation, or dendrogram. In particular, we used ISIDA/Cluster to obtain the heat map of the proximity matrix and the dynamic dendrogram (Fig. 2).
The kNN QSAR method43 (link), 44 (link) is based on the idea that the activity of a given compound can be predicted by averaging the activities of k compounds from the modeling set, which are most chemically similar to this compound. Briefly, our algorithm employs the kNN classification principle and variable selection procedure (simulated annealing with the Metropolis-like acceptance criteria): it generates both an optimum k value, typically from one to five, and an optimal nvar subset of descriptors that maximize the QSAR model’s training set accuracy as estimated by the Q²_abs statistical parameter. The Euclidean distance between compounds is used as a metric that characterizes compounds’ dissimilarity in multidimensional descriptor space. Additional details of the method can be found elsewhere.29 (link) For SVM classification, we used the WinSVM program (version 1.1.8)37 (link) developed in our group at UNC, which implements the opensource libsvm package (http://www.csie.ntu.edu.tw/~cjlin/libsvm/).
The Applicability Domain (AD) of a model is defined in order to determine if a given model is capable of predicting the activity of a query compound29 (link), 38 (link) within a reasonable error. In this study, we defined the AD as a threshold distance DT between a query compound and its nearest neighbors in the training set, calculated as follows: D_T = ȳ + Zσ where ȳ is the average Euclidean distance between each compound and its k nearest neighbors in the training set, σ is the standard deviation of the Euclidean distances, and Z is an arbitrary parameter to control the significance level; k is the parameter optimized in the course of QSAR modeling. We set the default value of Z at 0.5, which formally places the allowed distance threshold at the mean plus one-half of the standard deviation. If the distance of the test compound from any of its k nearest neighbors in the training set exceeds the threshold, the prediction is considered unreliable. In this study, we used this same approach for both Case Studies 1 and 2.
We used different statistical parameters to evaluate the performance of models. For binary classification problems (like Case Study 1), these are defined as: Accuracy = (TP + TN) / (NA + NI); Sensitivity = TP / NA; Specificity = TN / NI; CCR = 0.5 (Sensitivity + Specificity), where NA is the total number of actives (or class 1), NI is the total number of inactives (or class 0), TP is the number of true positives (experimentally actives predicted as actives), TN is the number of true negatives (experimentally inactives predicted as inactives), and CCR is the Correct Classification Rate.
When activities were represented by a range of values (Case Study 2), we used squared correlation coefficient (R²_abs) for test set compounds, squared leave-one-out cross-validation correlation coefficient (Q²_abs) for training set compounds, and mean absolute error (MAE) for the linear correlation between predicted (Y_pred) and experimental (Y_exp) data. For this study, Y is the Paca2 cellular uptake. These parameters are defined as follows: $R_{\mathit{\text{abs}}}^2=1-{\displaystyle \sum _Y{(Y_{\text{exp}}-Y_{\mathit{\text{pred}}})}^2}/{\displaystyle \sum _Y{(Y_{\text{exp}}-<Y{>}_{\text{exp}})}^2},Q_{\mathit{\text{abs}}}^2=1-{\displaystyle \sum _Y{(Y_{\text{exp}}-Y_{\mathit{\text{LOO}}})}^2}/{\displaystyle \sum _Y{(Y_{\text{exp}}-<Y{>}_{\text{exp}})}^2},\mathit{\text{MAE}}={\displaystyle \sum _Y|Y-Y_{\mathit{\text{pred}}}|}/n$ . In Case Study 1, the classification models were considered acceptable if CCR_CV ≥ 0.6 and CCR_test ≥ 0.6, whereas the regression models were considered acceptable in Case Study 2 if Q²_abs > 0.6 and R²_abs > 0.6.

To a solution of 27 (50 mg, 0.18 mmol) in anhydrous THF, were added triethylamine (26 mg, 0.19 mmol) and hexanoyl chloride. The reaction mixture was stirred at room temperature for 6 hours. The solvent was then removed under vacuum, and the residue was dissolved in ethyl acetate and washed with water. The ethyl acetate fraction was dried using Na₂SO₄ and evaporated under vacuum to obtain the intermediate amide compound. This was dissolved in 0.1 ml of 2M solution of ammonia in MeOH. The sealed reaction vessel was heated in microwave at 150 °C for 1 hour. The solvent was then removed under vacuum and the residue was purified using column chromatography (7% MeOH/dichloromethane) to obtain the compound 32 (8 mg; unoptimized yields). ¹H NMR (500 MHz, CDCl₃) δ 7.81 (d, J = 8.3 Hz, 1H), 7.72 (dd, J = 8.3, 1.0 Hz, 1H), 7.44 (qd, J = 7.0, 3.5 Hz, 1H), 7.38 – 7.29 (m, 3H), 7.18 – 7.10 (m, 1H), 7.06 (d, J = 6.9 Hz, 2H), 5.74 (s, 2H), 5.60 (s, 2H), 2.95 – 2.79 (m, 2H), 1.80 (dt, J = 15.6, 7.6 Hz, 2H), 1.42 – 1.27 (m, 4H), 0.87 (t, J = 7.2 Hz, 3H). ¹³C NMR (126 MHz, CDCl₃) δ 153.50, 150.28, 134.61, 133.41, 128.61, 127.40, 126.51, 125.94, 125.85, 124.85, 121.69, 119.07, 114.42, 48.20, 30.85, 26.97, 26.77, 21.65, 13.23. MS (ESI) calculated for C₂₂H₂₄N₄, m/z 344.20, found 345.21 (M + H)⁺. 2D ¹H COSY and NOESY experiments (included in Supporting Information) provided unambiguous assignment of ¹H resonances.