Gradient
Search

The integration of pharmacophore-based 3D QSAR modeling and virtual screening in safety profiling

The integration of pharmacophore-based 3D QSAR modeling and virtual screening in safety profiling: A case study to identify antagonistic activities against adenosine receptor, A2A, using 1,897 known drugs


Safety pharmacology screening against a wide range of unintended vital targets using in vitro assays is crucial to understand off-target interactions with drug candidates. With the increasing demand for in vitro assays, ligand- and structure-based virtual screening approaches have been evaluated for potential utilization in safety profiling. Although ligand based approaches have been actively applied in retrospective analysis or prospectively within well-defined chemical space during the early discovery stage (i.e., HTS screening and lead optimization), virtual screening is rarely implemented in later stage of drug discovery (i.e., safety). Here they present a case study to evaluate ligand-based 3D QSAR models built based on in vitro antagonistic activity data against adenosine receptor 2A (A2A). The resulting models, obtained from 268 chemically diverse compounds, were used to test a set of 1,897 chemically distinct drugs, simulating the real-world challenge of safety screening when presented with novel chemistry and a limited training set. Due to the unique requirements of safety screening versus discovery screening, the limitations of 3D QSAR methods (i.e., chemotypes, dependence on large training set, and prone to false positives) are less critical than early discovery screen.


They demonstrated that 3D QSAR modeling can be effectively applied in safety assessment prior to in vitro assays, even with chemotypes that are drastically different from training compounds. It is also worth noting that their model is able to adequately make the mechanistic distinction between agonists and antagonists, which is important to inform subsequent in vivo studies. Overall, they present an in-depth analysis of the appropriate utilization and interpretation of pharmacophore-based 3D QSAR models for safety screening.


Workflow illustration for pharmacophore-based 3D QSAR modeling and virtual screening to identify compounds with antagonistic activities against A2A.


Safety profiling against a wide range of molecular off-targets, prior to in vivo toxicity testing with animal models, has been widely implemented across the pharmaceutical industry. Such a “bottom-up approach” reflects a continuous effort for a paradigm shift in early safety evaluations.

Besides preventing hazardous chemicals from entering animals, systematic screening is a necessary step to realize the vision of predicting human adverse events from mechanisms of action and the molecular targets involved. Safety profiling utilizes in vitro high throughput screens (HTS) against a broad array of unintended and vital targets. However as a safety screening panel typically includes a large number of targets, i.e., up to ~200, developing each liability target into a reliable HTS assay is resource demanding. As complementary approaches to help improve the utilization of in vitro HTS assays, tools such as ligand- and structure-based virtual screening have been evaluated. One advantage of in silico approaches is that it can be used to examine new compounds before they are synthesized, providing an attractive possibility for early hazard identification. If molecules with undesirable properties can be ruled out using in silico approaches, such as virtual screening, significant resources can be saved where only “prescreened” molecules are advanced to more costly in vitro screens.


For liability targets with little or no structural information, a ligand-based approach using quantitative structure-activity relationship (QSAR) models may provide value. QSAR is a machine learning process to develop meaningful correlations (model) between independent variables (e.g., structural features of compounds, molecular descriptors) and a dependent variable which is typically the value one wishes to predict. The conceptual basis of such modeling is based on the hypotheses that compounds of similar structural features may exhibit similar biological activities. A QSAR model is determined by factors such as activity data, molecule descriptors, and statistical algorithms. Due to the advantages in throughput, cost-saving(labor and reagents), turn-around-time, and the possibility to test compounds even before they are made, QSAR has been frequently used in various aspects of drug discovery such as lead optimization. However, it has not been widely used in safety profiling, especially the 3D (i.e., pharmacophore) QSAR models, as most of commonly used QSAR models used in safety were built based on 2D features or molecular descriptors, such as the OECD QSAR toolbox, SEA, Toxmatch, ToxTree, and DSSTox. It is important to bear in mind the unique aspects for safety profiling. In an efficacy screening (one target against many compounds), only a small amount of positives was considered. Quantitative determination of potency is crucial for lead optimization and ranking compounds. The negatives were of less value. Whereas in safety profiling (often one compound against many targets), every data point counts including all negatives. In fact, a negative result against a liability target for a drug candidate would be regarded as “good news”. As such, a false negative (contributing to sensitivity) result would be of greater concern in the safety space in comparison to a false positive (contributing to specificity), because it would mean advancing a potentially hazardous compound into further development. The quantitative value of potency is of less importance than efficacy screening. Due to these unique features and mindset, questions regarding QSAR applications remain in data interpretation as well as how to best incorporate these tools.


They present here a case study to evaluate the utilization of 3D QSAR modeling as a part of an integrated approach to support safety profiling. Adenosine receptor 2a (A2A) is one of the four class A GPCRs that regulate the activity of adenosine’s biological actions as a signaling molecule. Due to its presence in both the central nervous system and peripheral tissues, A2A plays important roles in a wide range of biological processes such as locomotion, anxiety, memory, cognition, sleep regulation, angiogenesis, coronary blood flow, inflammation, and anti-tumoral immunity. Disruption of A2A activities, consequentially, may result in undesired side effects in behavioral, vascular, respiratory, inflammatory, and central nervous systems. Indeed A2A is a well-established liability target, as demonstrated in an industrial survey across four pharmaceutical companies.


Here, they developed QSAR models to predict compounds’ antagonistic activity against A2A. It is important to note that the crystallographic structure of A2A is known, in contrast to a large number of safety targets (e.g., ion channels and transporters). To make this study generalizable to those targets, however, they chose not to incorporate the structural data for A2A in model building but rather used it to provide additional insights to evaluate the performance of the ligand QSAR model. In their study, they collected 268 in-house and external compounds with IC50 values against A2A available, which were used to build the QSAR models. The majority of these compounds were obtained from early chemistry scaffolds and SAR. Hence, these compounds represented a diverse chemical space but not necessarily with ideal “drug-like features”. Bearing in mind that the goal is the prospective utilization of QSAR in secondary pharmacology profiling, we tailored their study specifically within the setting of drug discovery. First, overtraining the model(s) was avoided. During drug development, it may not be practically possible to obtain many training compounds and assay results, hence the need to implement the QSAR model. Therefore, we did not adhere to the 4:1 or 10:1 ratio. for training and test sets. Second, as new chemotypes are constantly made in pharmaceutical development to drive SAR, a different external set of compounds was used to further challenge the QSAR models. This additional level of the challenge came from 1,897 known drugs. Among these drugs, a subset of 75 known A2A ligands was used as an external set. The 75 ligands in the subset are different in structure compared to the initial 268 training and test compounds. These 75 compounds were thoroughly tested to evaluate prospective utilization of the generated QSAR model(s) before applying them to screen the rest of ~1,800 drugs from the DrugBank. These ~1,800 drugs further differ from the 268 compounds in chemical structure, which created a more realistic challenge. It is important to note that the focus of their study is the repurposing of existing QSAR tools in the realm of drug safety, rather than developing novel QSAR methodology.


They demonstrate that, due to the unique requirements of safety screening, the well-known limitations of QSAR methods (i.e., chemotypes, dependence on the large training set, and prone to false positives) are less critical than in early discovery screening. Overall, what they present is an in-depth case study for the utilization of in silico methods in early safety profiling.

The heat map demonstration for binary fingerprint similarities between training 1 and test 1 (A), training 2 and test 2 (B), the 55 representations of the 268 compounds and the subset of 75 A2A ligands (C), as well as the 55 representations and the 1,832 DrugBank drugs (D). The heat map of 1,832 drugs was truncated due to space limitation. The heat maps were generated using Schrodinger Canvas, as described in details in Materials and Methods. The lowest similarity (0.0) was shown in black, whereas the highest similarity (1.0) was shown in red. See supplementary data for a zoomed in version for each panel.



Pharmacophore hypothesis identified by Phase a



The performance of the 4 models on predicting actives of the test set compounds.

A, model AADR.139 generated from training set 1. B, model AAADR.20 generated from training set 2. C, model AAADR.1 generated from training set 3. D, model AAAR.2 generated from training set 3. In the cases of models AAADR.1 and AAAR.2, there were no test set compounds as all 268 compounds were used as training.



The statistical data of pharmacophore-based 3D QSAR using Phase a,b,c.

The reference ligands used for model AADR.

The reference ligands used for model AADR.139 (A), AAADR.20 (B), AAADR.1 (C), and AAAR.2 (D). The insert of Fig 4D shows that model AADR.5 shared the same reference compound (N-isopropyl-2-((pyridin-3-ylmethyl)amino)thieno[3,2-d]pyrimidine-4-carboxamide) as model AAAR. 2. Hence, N-isopropyl-2-((pyridin-3-ylmethyl)amino)thieno[3,2-d]pyrimidine-4-carboxamide contained all five pharmacophore sites in model AAADR.1. Hydrogen bond acceptor was shown in magenta vector, hydrogen donor was shown in light blue vector, and aromatic residues were shown in brown ring.



The reference ligands used for model AADR.

The reference ligands used for model AADR.139 (A), AAADR.20 (B), AAADR.1 (C), and AAAR.2 (D). The insert of Fig 4D shows that model AADR.5 shared the same reference compound (N-isopropyl-2-((pyridin-3-ylmethyl)amino)thieno[3,2-d]pyrimidine-4-carboxamide) as model AAAR. 2. Hence, N-isopropyl-2-((pyridin-3-ylmethyl)amino)thieno[3,2-d]pyrimidine-4-carboxamide contained all five pharmacophore sites in model AAADR.1. Hydrogen bond acceptor was shown in magenta vector, hydrogen donor was shown in light blue vector, and aromatic residues were shown in brown ring.




Distances are in the unit of Å.



Pictorial representations of the positive (cobalt) and negative (red) coefficients that contribute to A2A antagonist activities, from hydrogen bond donor (A), hydrophobicity (B), electron withdrawing groups (C), and the combined effects (D).



In vitro assay results a.


The performance of pharmacophore-based 3D QSAR modeling results in comparison to in vitro activities, when the similarities of the binary fingerprint between the query compound and the training/test compounds are ≥ 0.29 (A), between 0.22 to 0.29 (B), between 0.14 and 0.22 (C), and < 0.14 (D). In D, red dots indicated that similarity ranges between 0.10 to 0.14; grey dots indicated that similarity was below 0.10.


  1. Fan F, Toledo Warshaviak D, Hamadeh HK, Dunn RT II (2019) The integration of pharmacophore-based 3D QSAR modeling and virtual screening in safety profiling: A case study to identify antagonistic activities against adenosine receptor, A2A, using 1,897 known drugs. PLoS ONE 14(1): e0204378. https://doi.org/10.1371/journal.pone.0204378


Fan-2019-The-integration-of-pharmacophor
.
Download • 3.11MB