Prediction of hERG Potassium Channel Blockade Using kNN‐QSAR and Local Lazy Regression Methods

Abstract

We have collated hERG inhibition data of 165 compounds from literature and employed two regression procedures, namely, Local Lazy Regression (LLR) and k-Nearest Neighbor (kNN)-QSAR regression methods in combination with Genetic Algorithms (GAs) to select significant and independent molecular descriptors and to build robust predictive models. This methodology helped us to derive four, optimal 2D-and 3D-QSPR models, M1 -M4, based on five descriptors. Extensive validation tests using leave-one-out method and 61 compounds that are not used in the model generation strongly suggest that: (i) models M1 and M2, based on LLR, are very stable and robust; (ii) the model, M2 based on 3-D descriptors, performs better than the one based on 2-D descriptors, M1; and (iii) LLR method outperforms kNN regression approach. These results strongly suggest that the combination of GA and LLR method is a promising methodology, to build multiple stable models that are useful in consensus prediction. Further, from the analysis of the physical meaning of the descriptors, used in the best 2-D and 3-D descriptor models, M1 and M2, the significant physico-chemical forces that determine the hERG inhibition profile of small organic compounds are uncovered. Finally, as the models reported herein, are based on computed properties, they appear a valuable tool in virtual screening, where selection and prioritization of candidates is required.

In Silico Prediction of Compounds Binding to Human Plasma Proteins by QSAR Models

“…k ‐Nearest neighbor (kNN) : kNN is a type of instance‐based learning and is thought to be one of the simplest machine learning algorithms. In this case, Manhattan distance was used to calculate the distance matrices, k was set to 28 to reduce the effect of noise, and the final result was an inverse distance weighted average of the k nearest neighbors.…”

Calculations of acute intravenous toxicity in mice based on local regression models in superoverlapping clusters (LRMSC)

“…This approach described in [12,13] is based on the method of k the nearest neighbors (kNN), which employs clusterization and generation of own regres sion model in each cluster.…”

A new strategy to improve the predictive ability of the local lazy regression and its application to the QSAR study of melanin‐concentrating hormone receptor 1 antagonists

“…36,37 Deciding the optimal number k of nearest neighbors to be used for the identification of the best model is of great relevance and vital for the prediction ability of local modeling. 38 When a prediction is required for a query point, LL proceeds by identifying a set of local model candidates with different polynomial degrees and different number of neighbors, by letting k vary among k min and k max (bandwidth, where k min and k max control the minimal and maximal number of neighbors to be used for identifying and validating models, respectively).…”

“…38 When a prediction is required for a query point, LL proceeds by identifying a set of local model candidates with different polynomial degrees and different number of neighbors, by letting k vary among k min and k max (bandwidth, where k min and k max control the minimal and maximal number of neighbors to be used for identifying and validating models, respectively). The prediction ability of each model is commonly assessed through a local leave-one-out cross-validation (LOO-CV) procedure, [36][37][38][39] i.e. the model is built using the k min resultant neighborhood, which is validated using LOO-CV to generate a Q 2 LOO value or prediction error.…”