Scott D. Bembenek

The recent outbreak of the novel coronavirus (SARS-CoV-2) poses a significant challenge to the scientific and medical communities to find immediate treatments. The usual process of identifying viable molecules and transforming them into a safe and effective drug takes 10-15 years, with around 5 years of that time spent in preclinical research and development alone. The fastest strategy is to identify existing drugs or late-stage clinical molecules (originally intended for other therapeutic targets) that already have some level of efficacy. To this end, we tasked our novel molecular modeling-AI hybrid computational platform with finding potential inhibitors of the SARS-CoV-2 main protease (Mpro, 3CLpro). Over 13,000 FDA-approved drugs and clinical candidates (represented by just under 30,000 protomers) were examined. This effort resulted in the identification of several promising molecules. Moreover, it provided insight into key chemical motifs surely to be beneficial in the design of future inhibitors. Finally, it facilitated a unique perspective into other potentially therapeutic targets and pathways for SARS-CoV-2.

The recent outbreak of the novel coronavirus (SARS-CoV-2) poses a significant challenge to the scientific and medical communities to find immediate treatments. The usual process of identifying viable molecules and transforming them into a safe and effective drug takes 10-15 years, with around 5 years of that time spent in preclinical research and development alone. The fastest strategy is to identify existing drugs or late-stage clinical molecules (originally intended for other therapeutic targets) that already have some level of efficacy. To this end, we tasked our novel molecular modeling-AI hybrid computational platform with finding potential inhibitors of the SARS-CoV-2 main protease (M ^pro , 3CL ^pro ). Over 13,000 FDA-approved drugs and clinical candidates (represented by just under 30,000 protomers) were examined. This effort resulted in the identification of several promising molecules. Moreover, it provided insight into key chemical motifs surely to be beneficial in the design of future inhibitors. Finally, it facilitated a unique perspective into other potentially therapeutic targets and pathways for SARS-CoV-2.

An electronic density model was developed and used to identify a novel pyrrolotriazinone replacement for quinazolinone, a commonly used moiety to impart selectivity in inhibitors for PI3Kγ and PI3Kδ. Guided by molecular docking, this new specificity piece was then linked to the hinge portion of the inhibitor using a novel cyclic moiety. Further SAR optimization around the hinge region led to the discovery of candidate 26, a highly potent and selective PI3Kγ−PI3Kδ dual inhibitor with favorable DMPK properties in preclinical species.

A combination of structure-based drug design and medicinal chemistry efforts led us from benzimidazole-2-carboxamide with modestly active hypoxia-inducible factor prolyl hydroxylase 2 inhibition to certain benzimidazole-2-pyrazole carboxylic acids that were more potent as well as orally efficacious stimulators of erythropoietin secretion in our in vivo mouse model. To better understand the structure–activity relationship, it was necessary to account for (i) the complexation of the ligand with the active site Fe²⁺, (ii) the strain incurred by the ligand upon binding, and (iii) certain key water interactions identified by a crystal structure analysis. With this more complete computational model, we arrived at an overarching paradigm that accounted for the potency differences between benzimidazole-2-carboxamide and benzimidazole-2-pyrazole carboxylic acid enzyme inhibitors. Moreover, the computational paradigm allowed us to anticipate that the bioisostere replacement strategy (amide → pyrazole), which had shown success in the benzimidazole series, was not generally applicable to other series. This illustrates that to fully reconcile the important ligand–active site interactions for certain targets, one often needs to move beyond traditional structure-based drug design (such as crystallographic analysis, docking, etc.) and appeal to a higher level of computational theory.

We analyzed an extensive data set of 3000 Janssen kinase inhibitors (spanning some 40 therapeutic projects) profiled at 414 kinases in the DiscoverX KINOMEscan to better understand the necessity of using such a full kinase panel versus simply profiling one’s compound at a much smaller number of kinases, or mini kinase panel (MKP), to assess its selectivity. To this end, we generated a series of MKPs over a range of sizes, and of varying kinase membership using Monte Carlo simulations. By defining the kinase hit index (KHI), we quantified a compound’s selectivity based on the number of kinases it hits. We find that certain combinations (rather than a random selection) of kinases can result in a much lower average error. Indeed, we identified a focused MKP with a 45.1% improvement in the average error (compared to random) that yields an overall correlation of R2 = 0.786 – 0.826 for the KHI compared to the full kinase panel value. Unlike using a full kinase panel, which is both time and cost restrictive, a focused MKP is amenable to the triaging of all early-stage compounds. In this way, promiscuous compounds are filtered out early on, leaving the most selective compounds for lead optimization.

Here, we report a high-throughput virtual screening (HTVS) study using phosphoinositide 3-kinase (both PI3Kγ and PI3Kδ). Our initial HTVS results of the Janssen corporate database identified small focused libraries with hit rates at 50% inhibition showing a 50-fold increase over those from a HTS (high-throughput screen). Further, applying constraints based on “chemically intuitive” hydrogen bonds and/or positional requirements resulted in a substantial improvement in the hit rates (versus no constraints) and reduced docking time. While we find that docking scoring functions are not capable of providing a reliable relative ranking of a set of compounds, a prioritization of groups of compounds (e.g., low, medium, and high) does emerge, which allows for the chemistry efforts to be quickly focused on the most viable candidates. Thus, this illustrates that it is not always necessary to have a high correlation between a computational score and the experimental data to impact the drug discovery process.

Recently, we introduced a novel computer simulation technique to determine the optimal set of parameters of an interaction potential for a simple monatomic liquid. This technique was used to obtain interaction potentials of the Lennard-Jones form that accurately describe argon over its entire liquid phase at a fixed pressure [S. D. Bembenek and B. M. Rice, Mol. Phys. 97, 1085 (1999)]. Here, we extend this technique to a homonuclear diatomic molecular system, liquid oxygen. This technique was first applied to a system in which the oxygen molecules were treated as point masses interacting through a modified Lennard-Jones potential. Simulations using the resulting optimal set of potential parameters of this system predict densities that are within 0.25% of experiment over the entire liquid range of oxygen at a fixed pressure. However, the errors in the internal energy and the enthalpy were as large as 9.8%. The technique was then used to determine the optimal parameters for a system of harmonic molecules, in which each molecule has two interaction sites centered at the atomic nuclei. The intermolecular interaction is the sum of all site–site interactions described by a modified Lennard-Jones potential. Simulations using these parameters reproduce experimental densities with an error no greater than 0.80%. The predictions of the internal energy and enthalpy differ from experiment by no more than 3.0% for temperatures below 90 K; predictions at 90 K differ from experiment by no more than 4.11%. These results seem to suggest that our method for determining parameters for an interaction potential is also applicable to simple molecular systems.