Quantitative Structure–Activity Relationship Models That Stand the Test of Time

Exploring Deep Learning of Quantum Chemical Properties for Absorption, Distribution, Metabolism, and Excretion Predictions

Quantum mechanical (QM) descriptors of small molecules have wide applicability in understanding organic reactivity and molecular properties, but the substantial compute cost required for ab initio QM calculations limits their broad usage. Here, we investigate the use of deep learning for predicting QM descriptors, with the goal of enabling usage of near-QM accuracy electronic properties on large molecular data sets such as those seen in drug discovery. Several deep learning approaches have previously been benchmarked on a published data set called QM9, where 12 ground-state properties have been calculated for molecules with up to nine heavy atoms, limited to C, H, N, O, and F elements. To advance the work beyond the QM9 chemical space and enable application to molecules encountered in drug discovery, we extend the QM9 data set by creating a QM9-extended data set covering an additional ∼20,000 molecules containing S and Cl atoms. Using this extended set, we generate new deep learning models as well as leverage ANI-2x models to provide predictions on larger, more diverse molecules common in drug discovery, and we find the models estimate 11 of 12 ground-state properties reasonably. We use the predicted QM descriptors to augment graph convolutional neural network (GCNN) models for selected ADME end points (rat microsomal clearance, hepatic clearance, total clearance, and P-glycoprotein efflux) and found varying degrees of performance improvement compared to nonaugmented GCNN models, including pronounced improvement in P-glycoprotein efflux prediction.

Stability of Prediction in Production ADMET Models as a Function of Version: Why and When Predictions Change

As with other pharma companies, we maintain production QSAR models of ADMET end points and update them regularly. Here, for six ADMET end points, we examine the predictions of test set molecules on multiple versions of random forest models spanning a period of 10 years. For any given end point, the predictions for the majority of molecules are similar for all model versions. However, for a small minority of molecules, the prediction shifts substantially over the span of a few versions. For most molecules that shift, the prediction becomes more accurate at later times. This Perspective investigates metrics that can help indicate which molecules will shift substantially in prediction and when the shift will occur.

Prediction Accuracy of Production ADMET Models as a Function of Version: Activity Cliffs Rule

As with many other institutions, our company maintains many quantitative structure-activity relationship (QSAR) models of absorption, distribution, metabolism, excretion, and toxicity (ADMET) end points and updates the models regularly. We recently examined version-to-version predictivity for these models over a period of 10 years. In this approach we monitor the goodness of prediction of new molecules relative to the training set of model version V before they are incorporated in the updated model V+1. Using a cell-based permeability assay (Papp) as an example, we illustrate how the QSAR models made from this data are generally predictive and can be utilized to enrich chemical designs and synthesis. Despite the obvious utility of these models, we turned up unexpected behavior in Papp and other ADMET activities for which the explanation is not obvious. One such behavior is that the apparent predictivity of the models as measured by root-mean-square-error can vary greatly from version to version and is sometimes very poor. One intuitively appealing explanation is that the observed activities of the new molecules fall outside the bulk of activities in the training set. Alternatively, one may think that the new molecules are exploring different regions of chemical space than the training set. However, the real explanation has to do with activity cliffs. If the observed activities of the new molecules are different than expected based on similar molecules in the training set, the predictions will be less accurate. This is true for all our ADMET end points.

Nearest Neighbor Gaussian Process for Quantitative Structure-Activity Relationships

While Gaussian process models are typically restricted to smaller data sets, we propose a variation which extends its applicability to the larger data sets common in the industrial drug discovery space, making it relatively novel in the quantitative structure-activity relationship (QSAR) field. By incorporating locality-sensitive hashing for fast nearest neighbor searches, the nearest neighbor Gaussian process model makes predictions with time complexity that is sub-linear with the sample size. The model can be efficiently built, permitting rapid updates to prevent degradation as new data is collected. Given its small number of hyperparameters, it is robust against overfitting and generalizes about as well as other common QSAR models. Like the usual Gaussian process model, it natively produces principled and well-calibrated uncertainty estimates on its predictions. We compare this new model with implementations of random forest, light gradient boosting, and k-nearest neighbors to highlight these promising advantages. The code for the nearest neighbor Gaussian process is available at https://github.com/Merck/nngp.

Experimental Error, Kurtosis, Activity Cliffs, and Methodology: What Limits the Predictivity of Quantitative Structure-Activity Relationship Models?

Given a particular descriptor/method combination, some quantitative structure-activity relationship (QSAR) datasets are very predictive by random-split cross-validation while others are not. Recent literature in modelability suggests that the limiting issue for predictivity is in the data, not the QSAR methodology, and the limits are due to activity cliffs. Here, we investigate, on in-house data, the relative usefulness of experimental error, distribution of the activities, and activity cliff metrics in determining how predictive a dataset is likely to be. We include unmodified in-house datasets, datasets that should be perfectly predictive based only on the chemical structure, datasets where the distribution of activities is manipulated, and datasets that include a known amount of added noise. We find that activity cliff metrics determine predictivity better than the other metrics we investigated, whatever the type of dataset, consistent with the modelability literature. However, such metrics cannot distinguish real activity cliffs due to large uncertainties in the activities. We also show that a number of modern QSAR methods, and some alternative descriptors, are equally bad at predicting the activities of compounds on activity cliffs, consistent with the assumptions behind "modelability." Finally, we relate time-split predictivity with random-split predictivity and show that different coverages of chemical space are at least as important as uncertainty in activity and/or activity cliffs in limiting predictivity.

On the virtues of automated quantitative structure-activity relationship: the new kid on the block

Quantitative structure-activity relationship (QSAR) has proved to be an invaluable tool in medicinal chemistry. Data availability at unprecedented levels through various databases have collaborated to a resurgence in the interest for QSAR. In this context, rapid generation of quality predictive models is highly desirable for hit identification and lead optimization. We showcase the application of an automated QSAR approach, which randomly selects multiple training/test sets and utilizes machine-learning algorithms to generate predictive models. Results demonstrate that AutoQSAR produces models of improved or similar quality to those generated by practitioners in the field but in just a fraction of the time. Despite the potential of the concept to the benefit of the community, the AutoQSAR opportunity has been largely undervalued.

Structure-Affinity Relationships and Structure-Kinetic Relationships of 1,2-Diarylimidazol-4-carboxamide Derivatives as Human Cannabinoid 1 Receptor Antagonists

We report on the synthesis and biological evaluation of a series of 1,2-diarylimidazol-4-carboxamide derivatives developed as CB₁ receptor antagonists. These were evaluated in a radioligand displacement binding assay, a [³⁵S]GTPγS binding assay, and in a competition association assay that enables the relatively fast kinetic screening of multiple compounds. The compounds show high affinities and a diverse range of kinetic profiles at the CB₁ receptor and their structure–kinetic relationships (SKRs) were established. Using the recently resolved hCB₁ receptor crystal structures, we also performed a modeling study that sheds light on the crucial interactions for both the affinity and dissociation kinetics of this family of ligands. We provide evidence that, next to affinity, additional knowledge of binding kinetics is useful for selecting new hCB₁ receptor antagonists in the early phases of drug discovery.

Evolution of physicochemical properties of melanin concentrating hormone receptor 1 (MCHr1) antagonists

One pharmacological principle for the treatment of obesity is blockade of the melanin concentrating hormone receptor 1 (MCHr1), which in rodents has been shown to be strongly associated with food intake and energy expenditure. However, discovery of safe and efficacious MCHr1 antagonists has proved to be complex. So far, six compounds have been progressed into clinical trials, but clinical validation of the concept is still lacking. An account of discovery of the three most recent clinical candidates targeting the MCHr1 receptor is given, with an emphasis on their physicochemical properties.

Discovery of (3-(4-(2-Oxa-6-azaspiro[3.3]heptan-6-ylmethyl)phenoxy)azetidin-1-yl)(5-(4-methoxyphenyl)-1,3,4-oxadiazol-2-yl)methanone (AZD1979), a Melanin Concentrating Hormone Receptor 1 (MCHr1) Antagonist with Favorable Physicochemical Properties

A novel series of melanin concentrating hormone receptor 1 (MCHr1) antagonists were the starting point for a drug discovery program that culminated in the discovery of 103 (AZD1979). The lead optimization program was conducted with a focus on reducing lipophilicity and understanding the physicochemical properties governing CNS exposure and undesired off-target pharmacology such as hERG interactions. An integrated approach was taken where the key assay was ex vivo receptor occupancy in mice. The candidate compound 103 displayed appropriate lipophilicity for a CNS indication and showed excellent permeability with no efflux. Preclinical GLP toxicology and safety pharmacology studies were without major findings and 103 was taken into clinical trials.

Role of computer-aided drug design in modern drug discovery

Drug discovery utilizes chemical biology and computational drug design approaches for the efficient identification and optimization of lead compounds. Chemical biology is mostly involved in the elucidation of the biological function of a target and the mechanism of action of a chemical modulator. On the other hand, computer-aided drug design makes use of the structural knowledge of either the target (structure-based) or known ligands with bioactivity (ligand-based) to facilitate the determination of promising candidate drugs. Various virtual screening techniques are now being used by both pharmaceutical companies and academic research groups to reduce the cost and time required for the discovery of a potent drug. Despite the rapid advances in these methods, continuous improvements are critical for future drug discovery tools. Advantages presented by structure-based and ligand-based drug design suggest that their complementary use, as well as their integration with experimental routines, has a powerful impact on rational drug design. In this article, we give an overview of the current computational drug design and their application in integrated rational drug development to aid in the progress of drug discovery research.

QSAR workbench: automating QSAR modeling to drive compound design

We describe the QSAR Workbench, a system for the building and analysis of QSAR models. The system is built around the Pipeline Pilot workflow tool and provides access to a variety of model building algorithms for both continuous and categorical data. Traditionally models are built on a one by one basis and fully exploring the model space of algorithms and descriptor subsets is a time consuming basis. The QSAR Workbench provides a framework to allow for multiple models to be built over a number of modeling algorithms, descriptor combinations and data splits (training and test sets). Methods to analyze and compare models are provided, enabling the user to select the most appropriate model. The Workbench provides a consistent set of routines for data preparation and chemistry normalization that are also applied for predictions. The Workbench provides a large degree of automation with the ability to publish preconfigured model building workflows for a variety of problem domains, whilst providing experienced users full access to the underlying parameterization if required. Methods are provided to allow for publication of selected models as web services, thus providing integration with the chemistry desktop. We describe the design and implementation of the QSAR Workbench and demonstrate its utility through application to two public domain datasets.