Current and Future Roles of Artificial Intelligence in Medicinal Chemistry Synthesis.

Artificial intelligence and machine learning have demonstrated their potential role in predictive chemistry and synthetic planning of small molecules; there are at least a few reports of companies employing in silico synthetic planning into their overall approach to accessing target molecules. A data-driven synthesis planning program is one component being developed and evaluated by the Machine Learning for Pharmaceutical Discovery and Synthesis (MLPDS) consortium, comprising MIT and 13 chemical and pharmaceutical company members. Together, we wrote this perspective to share how we think predictive models can be integrated into medicinal chemistry synthesis workflows, how they are currently used within MLPDS member companies, and the outlook for this field.

The evolution of library design: crafting smart compound collections for phenotypic screens.

The (re)emergence of phenotypic drug discovery has been marked by a growing interest in screening campaigns that utilize phenotypic assays. The key objectives of phenotypic screens are different from those of target-based screens and can require alternate library-design strategies. Designing a library that is appropriate to the selected assay increases the likelihood of identifying better quality hits, which can reduce both timelines and overall cost of the drug-discovery process. Here, we provide an overview of small-molecule library design principles as applied to phenotypic screening.

POSIT: Flexible Shape-Guided Docking For Pose Prediction.

We present a new approach to structure-based drug design (POSIT) rigorously built on the simple concept that pose prediction is intimately coupled to the quality and availability of experimental structural data. We demonstrate the feasibility of the approach by performing retrospective analyses on three data sets designed to explore the strengths and weaknesses of POSIT relative to existing methods. We then present results documenting 2.5 years of prospective use of POSIT across a variety of structure-based industrial drug-discovery research projects. We find that POSIT is well-suited to guiding research decision making for structure-based design and, in particular, excels at enabling lead-optimization campaigns. We show that the POSIT framework can drive superior pose-prediction performance and generate results that naturally lend themselves to prospective decision making during lead optimization. We believe the results presented here are (1) the largest prospective validation of a pose prediction method reported to date (71 crystal structures); (2) provide an unprecedented look at the scope of impact of a computational tool; and (3) represent a first-of-its-kind analysis. We hope that this work inspires additional studies that look at the real impact and performance of computational research tools on prospective drug design.

Novel human D-amino acid oxidase inhibitors stabilize an active-site lid-open conformation.

The NMDAR (N-methyl-D-aspartate receptor) is a central regulator of synaptic plasticity and learning and memory. hDAAO (human D-amino acid oxidase) indirectly reduces NMDAR activity by degrading the NMDAR co-agonist D-serine. Since NMDAR hypofunction is thought to be a foundational defect in schizophrenia, hDAAO inhibitors have potential as treatments for schizophrenia and other nervous system disorders. Here, we sought to identify novel chemicals that inhibit hDAAO activity. We used computational tools to design a focused, purchasable library of compounds. After screening this library for hDAAO inhibition, we identified the structurally novel compound, 'compound 2' [3-(7-hydroxy-2-oxo-4-phenyl-2H-chromen-6-yl)propanoic acid], which displayed low nM hDAAO inhibitory potency (Ki=7 nM). Although the library was expected to enrich for compounds that were competitive for both D-serine and FAD, compound 2 actually was FAD uncompetitive, much like canonical hDAAO inhibitors such as benzoic acid. Compound 2 and an analog were independently co-crystalized with hDAAO. These compounds stabilized a novel conformation of hDAAO in which the active-site lid was in an open position. These results confirm previous hypotheses regarding active-site lid flexibility of mammalian D-amino acid oxidases and could assist in the design of the next generation of hDAAO inhibitors.

A unified, probabilistic framework for structure- and ligand-based virtual screening.

We present a probabilistic framework for interpreting structure-based virtual screening that returns a quantitative likelihood of observing bioactivity and can be quantitatively combined with ligand-based screening methods to yield a cumulative prediction that consistently outperforms any single screening metric. The approach has been developed and validated on more than 30 different protein targets. Transforming structure-based in silico screening results into robust probabilities of activity enables the general fusion of multiple structure- and ligand-based approaches and returns a quantitative expectation of success that can be used to prioritize (or deprioritize) further discovery activities. This unified probabilistic framework offers a paradigm shift in how docking and scoring results are interpreted, which can enhance early lead-finding efforts by maximizing the value of in silico computational tools.

Molecular shape and medicinal chemistry: a perspective.

The eight contributions here provide ample evidence that shape as a volume or as a surface is a vibrant and useful concept when applied to drug discovery. It provides a reliable scaffold for "decoration" with chemical intuition (or bias) for virtual screening and lead optimization but also has its unadorned uses, as in library design, ligand fitting, pose prediction, or active site description. Computing power has facilitated this evolution by allowing shape to be handled precisely without the need to reduce down to point descriptors or approximate metrics, and the diversity of resultant applications argues for this being an important step forward. Certainly, it is encouraging that as computation has enabled our intuition, molecular shape has consistently surprised us in its usefulness and adaptability. The first Aurelius question, "What is the essence of a thing?", seems well answered, however, the third, "What do molecules do?", only partly so. Are the topics covered here exhaustive, or is there more to come? To date, there has been little published on the use of the volumetric definition of shape described here as a QSAR variable, for instance, in the prediction or classification of activity, although other shape definitions have been successful applied, for instance, as embodied in the Compass program described above in "Shape from Surfaces". Crystal packing is a phenomenon much desired to be understood. Although powerful models have been applied to the problem, to what degree is this dominated purely by the shape of a molecule? The shape comparison described here is typically of a global nature, and yet some importance must surely be placed on partial shape matching, just as the substructure matching of chemical graphs has proved useful. The approach of using surfaces, as described here, offers some flavor of this, as does the use of metrics that penalize volume mismatch less than the Tanimoto, e.g., Tversky measures. As yet, there is little to go on as to how useful a paradigm this will be because there is less software and fewer concrete results.Finally, the distance between molecular shapes, or between any shapes defined as volumes or surfaces, is a metric property in the mathematical sense of the word. As yet, there has been little, if any, application of this observation. We cannot know what new application to the design and discovery of pharmaceuticals may yet arise from the simple concept of molecular shape, but it is fair to say that the progress so far is impressive.

Healthy skepticism: assessing realistic model performance.

Although the development of computational models to aid drug discovery has become an integral part of pharmaceutical research, the application of these models often fails to produce the expected impact on productivity. One reason for this may be that the expected performance of many models is simply not supported by the underlying data, because of often neglected effects of assay and prediction errors on the reliability of the predicted outcome. Another significant challenge to realizing the full potential of computational models is their integration into prospective medicinal chemistry campaigns. This article will analyze the impact of assay and prediction error on model quality, and explore scenarios where computational models can expect to have a significant influence on drug discovery research.

Large-scale application of high-throughput molecular mechanics with Poisson-Boltzmann surface area for routine physics-based scoring of protein-ligand complexes.

We apply a high-throughput formulation of the molecular mechanics with Poisson-Boltzmann surface area (htMM-PBSA) to estimate relative binding potencies on a set of 308 small-molecule ligands in complex with the proteins urokinase, PTP-1B, and Chk-1. We observe statistically significant correlation to experimentally measured potencies and report correlation coefficients for the three proteins in the range 0.72-0.83. The htMM-PBSA calculations illustrate the feasibility of procedural automation of physics-based scoring calculations to produce rank-ordered binding-potency estimates for protein-ligand complexes, with sufficient throughput for realization of practical implementation into scientist workflows in an industrial drug discovery research setting.

Synthesis and evaluation of benzothiazole-based analogues as novel, potent, and selective fatty acid amide hydrolase inhibitors.

High-throughput screening (HTS) identified benzothiazole analogue 3 as a potent fatty acid amide hydrolase (FAAH) inhibitor. Structure-activity relationship (SAR) studies indicated that the sulfonyl group, the piperidine ring and benzothiazole were the key components to their activity, with 16j being the most potent analogue in this series. Time-dependent preincubation study of compound 3 was consistent with it being a reversible inhibitor. Activity-based protein-profiling (ABPP) evaluation of 3 in rat tissues revealed that it had exceptional selectivity and no off-target activity with respect to other serine hydrolases. Molecular shape overlay of 3 with a known FAAH inhibitor indicated that these compounds might act as transition-state analogues, forming putative hydrogen bonds with catalytic residues and mimicking the charge distribution of the tetrahedral transition state. The modeling study also indicated that hydrophobic interactions of the benzothiazole ring with the enzyme contributed to its extraordinary potency. These compounds may provide useful tools for the study of FAAH and the endocannabinoid system.

Contrasting disease and nondisease protein aggregation by molecular simulation.

[Figurre: see text]. Protein aggregation can be defined as the sacrifice of stabilizing intrachain contacts of the functional state that are replaced with interchain contacts to form non-functional states. The resulting aggregate morphologies range from amorphous structures without long-range order typical of nondisease proteins involved in inclusion bodies to highly structured fibril assemblies typical of amyloid disease proteins. In this Account, we describe the development and application of computational models for the investigation of nondisease and disease protein aggregation as illustrated for the proteins L and G and the Alzheimer's Abeta systems. In each case, we validate the models against relevant experimental observables and then expand on the experimental window to better elucidate the link between molecular properties and aggregation outcomes. Our studies show that each class of protein exhibits distinct aggregation mechanisms that are dependent on protein sequence, protein concentration, and solution conditions. Nondisease proteins can have native structural elements in the denatured state ensemble or rapidly form early folding intermediates, which offers avenues of protection against aggregation even at relatively high concentrations. The possibility that early folding intermediates may be evolutionarily selected for their protective role against unwanted aggregation could be a useful strategy for reengineering sequences to slow aggregation and increase folding yield in industrial protein production. The observed oligomeric aggregates that we see for nondisease proteins L and G may represent the nuclei for larger aggregates, not just for large amorphous inclusion bodies, but potentially as the seeds of ordered fibrillar aggregates, since most nondisease proteins can form amyloid fibrils under conditions that destabilize the native state. By contrast, amyloidogenic protein sequences such as Abeta 1-40,42 and the familial Alzheimer's disease (FAD) mutants favor aggregation into ordered fibrils once the free-energy barrier for forming a critical nucleus is crossed. However, the structural characteristics and oligomer size of the soluble nucleation species have yet to be determined experimentally for any disease peptide sequence, and the molecular mechanism of polymerization that eventually delineates a mature fibril is unknown. This is in part due to the limited experimental access to very low peptide concentrations that are required to characterize these early aggregation events, providing an opportunity for theoretical studies to bridge the gap between the monomer and fibril end points and to develop testable hypotheses. Our model shows that Abeta 1-40 requires as few as 6-10 monomer chains (depending on sequence) to begin manifesting the cross-beta order that is a signature of formation of amyloid filaments or fibrils assessed in dye-binding kinetic assays. The richness of the oligomeric structures and viable filament and fibril polymorphs that we observe may offer structural clues to disease virulence variations that are seen for the WT and hereditary mutants.

Application of belief theory to similarity data fusion for use in analog searching and lead hopping.

A wide variety of computational algorithms have been developed that strive to capture the chemical similarity between two compounds for use in virtual screening and lead discovery. One limitation of such approaches is that, while a returned similarity value reflects the perceived degree of relatedness between any two compounds, there is no direct correlation between this value and the expectation or confidence that any two molecules will in fact be equally active. A lack of a common framework for interpretation of similarity measures also confounds the reliable fusion of information from different algorithms. Here, we present a probabilistic framework for interpreting similarity measures that directly correlates the similarity value to a quantitative expectation that two molecules will in fact be equipotent. The approach is based on extensive benchmarking of 10 different similarity methods (MACCS keys, Daylight fingerprints, maximum common subgraphs, rapid overlay of chemical structures (ROCS) shape similarity, and six connectivity-based fingerprints) against a database of more than 150,000 compounds with activity data against 23 protein targets. Given this unified and probabilistic framework for interpreting chemical similarity, principles derived from decision theory can then be applied to combine the evidence from different similarity measures in such a way that both capitalizes on the strengths of the individual approaches and maintains a quantitative estimate of the likelihood that any two molecules will exhibit similar biological activity.

Rapid estimation of relative protein-ligand binding affinities using a high-throughput version of MM-PBSA.

By employing a modified protocol of the Molecular Mechanics with Poisson-Boltzmann Surface Area (MM-PBSA) methodology we substantially decrease the required computation time for calculating relative estimates of protein-ligand binding affinities. The modified method uses a generalized Born implicit solvation model during molecular dynamics to enhance conformational sampling as well as a very efficient Poisson-Boltzmann solver and a computational design based on a distributed-computing paradigm. This construction allows for reduction of the computational cost of the calculations by roughly 2 orders of magnitude compared to the traditional formulation of MM-PBSA. With this high-throughput version of MM-PBSA we show that one can produce efficient physics-based estimates of relative binding free energies with reasonable correlation to experimental data and a total computation time that is sufficiently low such that an industrially relevant throughput can be realized given currently accessible computing resources. We demonstrate this approach by performing a comparison of different MM-PBSA implementations on a set of 18 ligands for the protein target urokinase.

High-throughput calculation of protein-ligand binding affinities: modification and adaptation of the MM-PBSA protocol to enterprise grid computing.

We have developed a system for performing computations on an enterprise grid using a freely available package for grid computing that allows us to harvest unused CPU cycles off of employee desktop computers. By modifying the traditional formulation of Molecular Mechanics with Poisson-Boltzmann Surface Area (MM-PBSA) methodology, in combination with a coarse-grain parallelized implementation suitable for deployment onto our enterprise grid, we show that it is possible to produce rapid physics-based estimates of protein-ligand binding affinities that have good correlation to experimental data. This is demonstrated by examining the correlation of our calculated binding affinities to experimental data and also by comparison to the correlation obtained from the binding-affinity calculations using traditional MM-PBSA that are reported in the literature.