BCL::Mol2D—a robust atom environment descriptor for QSAR modeling and lead optimization

Abstract

Comparing fragment based molecular fingerprints of drug-like molecules is one of the most robust and frequently used approaches in computer-assisted drug discovery (CADD). Molprint2D, a popular atom environment (AE) descriptor, yielded the best enrichment of active compounds across a diverse set of targets in a recent large-scale study. We present here BCL::Mol2D descriptors that outperformed Molprint2D on nine PubChem datasets spanning a wide range of protein classes. Because BCL::Mol2D records the number of AEs from a universal AE library, a novel aspect of BCL::Mol2D over the Molprint2D is its reversibility. This property enables decomposition of prediction from machine learning models to particular molecular substructures. Artificial neural networks (ANNs) with dropout, when trained on BCL::Mol2D descriptors outperform those trained on Molprint2D descriptors by up to 26% in logAUC metric. When combined with the Reduced Short Range (RSR) descriptor set, our previously published set of descriptors optimized for QSARs, BCL::Mol2D yields a modest improvement. Finally, we demonstrate how the reversibility of BCL::Mol2D enables visualization of a 'pharmacophore map' that could guide lead optimization for serine/threonine kinase 33 (STK33) inhibitors.

Advancements in Ligand-Based Virtual Screening through the Synergistic Integration of Graph Neural Networks and Expert-Crafted Descriptors

“…We used the BioChemical Library [27] to generate traditional QSAR descriptors. Following previous examples [2,28], we use the optimal descriptors where 391-element molecular-level features are generated.…”

Structure–Activity Relationship Study of the High-Affinity Neuropeptide Y4 Receptor Positive Allosteric Modulator VU0506013

“…Hence, ANNs can drive the virtual screening of QSAR models by providing predictions on the activities of new compounds and by identifying novel chemical structures with desired properties. , In this study, ANN-QSAR models were trained on a set of 19 confirmed active and 71,765 inactive compounds that were discovered through previous HTS efforts . Two QSAR models trained on two different types of two-dimensional molecular descriptor configurations, BCL::Mol2D and BCL::RSR, , were used to calculate locally accurate positive predictive values (LocalPPV) of 174,440 compounds from the Vanderbilt Discovery Collection. Higher LocalPPV values mean that the compounds are predicted to be more likely to be active.…”

“…Generation of Molecular Descriptors. Two types of descriptor configurations, the BCL::Mol2D descriptors 29 and the reduced short range (BCL::RSR) descriptor set, 28 were generated for both the training compound set and external compound set. The atom encoding schemes and atom environment height of one were set for the BCL::Mol2D descriptors (Atom harsh = Atom, AE height = 1, number of descriptors: 574).…”

Machine Learning-Based Quantitative Structure-Activity Relationship and ADMET Prediction Models for ERα Activity of Anti-Breast Cancer Drug Candidates

“…To hunt for characteristic compound descriptors and construct quantitative prediction models for QSAR, several studies used machine learning approaches such as regression analysis [12][13][14] , decision tree algorithm [15] , artificial neural network [16][17][18] , support vector machine [19] , and so on. The specific technique is as follows: given a disease-related target (ERα), gather a series of compounds that operate on the target and their biological activity data, and then utilize a series of molecular structure descriptors as independent variables and the biological activity values of the compounds as a dependent variable in the construction of a QSAR model of the compound, which is subsequently used to forecast novel compound molecules with improved biological activity [20][21][22][23] .…”