Deep Learning Model for Efficient Protein-Ligand Docking with Implicit Side-Chain Flexibility.

Protein-ligand docking is an essential tool in structure-based drug design with applications ranging from virtual high-throughput screening to pose prediction for lead optimization. Most docking programs for pose prediction are optimized for redocking to an existing cocrystallized protein structure, ignoring protein flexibility. In real-world drug design applications, however, protein flexibility is an essential feature of the ligand-binding process. Flexible protein-ligand docking still remains a significant challenge to computational drug design. To target this challenge, we present a deep learning (DL) model for flexible protein-ligand docking based on the prediction of an intermolecular Euclidean distance matrix (EDM), making the typical use of iterative search algorithms obsolete. The model was trained on a large-scale data set of protein-ligand complexes and evaluated on independent test sets. Our model generates high quality poses for a diverse set of protein and ligand structures and outperforms comparable docking methods.

Efficient virtual high-content screening using a distance-aware transformer model.

Molecular similarity search is an often-used method in drug discovery, especially in virtual screening studies. While simple one- or two-dimensional similarity metrics can be applied to search databases containing billions of molecules in a reasonable amount of time, this is not the case for complex three-dimensional methods. In this work, we trained a transformer model to autoencode tokenized SMILES strings using a custom loss function developed to conserve similarities in latent space. This allows the direct sampling of molecules in the generated latent space based on their Euclidian distance. Reducing the similarity between molecules to their Euclidian distance in latent space allows the model to perform independent of the similarity metric it was trained on. While we test the method here using 2D similarity as proof-of-concept study, the algorithm will enable also high-content screening with time-consuming 3D similarity metrics. We show that the presence of a specific loss function for similarity conservation greatly improved the model's ability to predict highly similar molecules. When applying the model to a database containing 1.5 billion molecules, our model managed to reduce the relevant search space by 5 orders of magnitude. We also show that our model was able to generalize adequately when trained on a relatively small dataset of representative structures. The herein presented method thereby provides new means of substantially reducing the relevant search space in virtual screening approaches, thus highly increasing their throughput. Additionally, the distance awareness of the model causes the efficiency of this method to be independent of the underlying similarity metric.

Ligand pathways in estrogen-related receptors.

The three subtypes of estrogen-related receptors ERRα, ERRß, and ERRγ are nuclear receptors mediating metabolic processes in various tissues such as the skeletal muscle, fat tissue, bone, and liver. Although the knowledge on their physiological ligands is limited, they have been implicated as drug targets for important indications including diabetes, cardiovascular diseases, and osteoporosis. As in other nuclear receptors, their ligand binding pocket is buried within the core of the receptor and connected to its surrounding by ligand pathways. Here, we investigated these pathways with conventional molecular dynamics as well as metadynamics simulations to reveal their distribution and their capability to facilitate ligand translocation. Dependent on the ERR subtype and the conformational state of the receptor, we could detect different pathways to be favored. Overall, the results suggested pathways IIIa and IIIb to be favored in the agonistic conformation, while antagonists preferred pathways I, II, and V. Along the pathways, the ligands passed different gating mechanisms of the receptor, including groups of protein residues as well as whole secondary structure elements, to leave the binding site. Even though these pathways are suggested to influence ligand specificity of the receptors and their elucidation might advance rational drug design, they have not yet been studied in ERRs.Communicated by Ramaswamy H. Sarma.

Accurate Sampling of Macromolecular Conformations Using Adaptive Deep Learning and Coarse-Grained Representation.

Conformational sampling of protein structures is essential for understanding biochemical functions and for predicting thermodynamic properties such as free energies. Where previous approaches rely on sequential sampling procedures, recent developments in generative deep neural networks rendered possible the parallel, statistically independent sampling of molecular configurations. To be able to accurately generate samples of large molecular systems from a high-dimensional multimodal equilibrium distribution function, we developed a hierarchical approach based on expressive normalizing flows with rational quadratic neural splines and coarse-grained representation. Furthermore, system specific priors and adaptive and property-based controlled learning was designed to diminish the likelihood for the generation of high-energy structures during sampling. Finally, backmapping from a coarse-grained to fully atomistic representation is performed through an equivariant transformer model. We demonstrate the applicability of the method on the one-shot configurational sampling of a protein system with more than a hundred amino acids. The results show enhanced expressivity that diminish the invertibility constraints inherent in the normalizing flow framework. Moreover, the capacity of the hierarchical normalizing flow model was tested on a challenging case study of the folding/unfolding dynamics of the peptide chignolin.

Evaluation of Xa inhibitors as potential inhibitors of the SARS-CoV-2 Mpro protease.

Based on previous large-scale in silico screening several factor Xa inhibitors were proposed to potentially inhibit SARS-CoV-2 Mpro. In addition to their known anticoagulants activity this potential inhibition could have an additional therapeutic effect on patients with COVID-19 disease. In this study we examined the binding of the Apixaban, Betrixaban and Rivaroxaban to the SARS-CoV-2 Mpro with the use of the MicroScale Thermophoresis technique. Our results indicate that the experimentally measured binding affinity is weak and the therapeutic effect due to the SARS-CoV-2 Mpro inhibition is rather negligible.

Investigation of Thiocarbamates as Potential Inhibitors of the SARS-CoV-2 Mpro.

In the present study we tested, using the microscale thermophoresis technique, a small library of thionocarbamates, thiolocarbamates, sulfide and disulfide as potential lead compounds for SARS-CoV-2 Mpro drug design. The successfully identified binder is a representative of the thionocarbamates group with a high potential for future modifications aiming for higher affinity and solubility. The experimental analysis was extended by computational studies that show insufficient accuracy of the simplest and widely applied approaches and underline the necessity of applying more advanced methods to properly evaluate the affinity of potential SARS-CoV-2 Mpro binders.

Computational Selectivity Assessment of Protease Inhibitors against SARS-CoV-2.

The pandemic of the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) poses a serious global health threat. Since no specific therapeutics are available, researchers around the world screened compounds to inhibit various molecular targets of SARS-CoV-2 including its main protease (Mpro) essential for viral replication. Due to the high urgency of these discovery efforts, off-target binding, which is one of the major reasons for drug-induced toxicity and safety-related drug attrition, was neglected. Here, we used molecular docking, toxicity profiling, and multiple molecular dynamics (MD) protocols to assess the selectivity of 33 reported non-covalent inhibitors of SARS-CoV-2 Mpro against eight proteases and 16 anti-targets. The panel of proteases included SARS-CoV Mpro, cathepsin G, caspase-3, ubiquitin carboxy-terminal hydrolase L1 (UCHL1), thrombin, factor Xa, chymase, and prostasin. Several of the assessed compounds presented considerable off-target binding towards the panel of proteases, as well as the selected anti-targets. Our results further suggest a high risk of off-target binding to chymase and cathepsin G. Thus, in future discovery projects, experimental selectivity assessment should be directed toward these proteases. A systematic selectivity assessment of SARS-CoV-2 Mpro inhibitors, as we report it, was not previously conducted.

Computational Assessment of Combination Therapy of Androgen Receptor-Targeting Compounds.

The ligand-binding domain of the androgen receptor (AR) is a target for drugs against prostate cancer and offers three distinct binding sites for small molecules. Drugs acting on the orthosteric hormone binding site suffer from resistance mechanisms that can, in the worst case, reverse their therapeutic effect. While many allosteric ligands targeting either the activation function-2 (AF-2) or the binding function-3 (BF-3) have been reported, their potential for simultaneous administration with currently prescribed antiandrogens was disregarded. Here, we report results of 60 µs molecular dynamics simulations to investigate combinations of orthosteric and allosteric AR antagonists. Our results suggest BF-3 inhibitors to be more suitable in combination with classical antiandrogens as opposed to AF-2 inhibitors based on binding free energies and binding modes. As a mechanistic explanation for these observations, we deduced a structural adaptation of helix-12 involved in the formation of the AF-2 site by classical AR antagonists. Additionally, the changes were accompanied by an expansion of the orthosteric binding site. Considering our predictions, the selective combination of AR-targeting compounds may improve the treatment of prostate cancer.

Decision Making in Structure-Based Drug Discovery: Visual Inspection of Docking Results.

Molecular docking is a computational method widely used in drug discovery. Due to the inherent inaccuracies of molecular docking, visual inspection of binding modes is a crucial routine in the decision making process of computational medicinal chemists. Despite its apparent importance for medicinal chemistry projects, guidelines for the visual docking pose assessment have been hardly discussed in the literature. Here, we review the medicinal chemistry literature with the aim of identifying consistent principles for visual inspection, highlighting cases of its successful application, and discussing its limitations. In this context, we conducted a survey reaching experts in both academia and the pharmaceutical industry, which also included a challenge to distinguish native from incorrect poses. We were able to collect 93 expert opinions that offer valuable insights into visually supported decision-making processes. This perspective shall motivate discussions among experienced computational medicinal chemists and guide young scientists new to the field to stratify their compounds.

Conformational Changes of Thyroid Receptors in Response to Antagonists.

Thyroid hormone receptors (TRs) play a critical role in human development, growth, and metabolism. Antagonists of TRs offer an attractive strategy to treat hyperthyroidism without the disadvantage of a delayed onset of drug action. While it is challenging to examine the atomistic behavior of TRs in a laboratory setting, computational methods such as molecular dynamics (MD) simulations have proven their value to elucidate ligand-induced conformational changes in nuclear receptors. Here, we performed MD simulations of TRα and TRß complexed to their native ligand triiodothyronine (T3) as well as several antagonists. Based on the examination of 27 µs MD trajectories, we showed how binding of these compounds influences various structural features of the receptors including the helicity of helices 3 and 10 as well as the location of helix-12. Helices 3 and 12 are known to mediate coactivator association required for downstream signaling, suggesting these changes to be the molecular basis for TR antagonism. A mechanistic analysis of the trajectories revealed an allosteric pathway between H3 and H12 to be responsible for the conformational adaptations. Even though a mechanistic understanding of conformational adaptations triggered by TR antagonists is important for the development of novel therapeutics, they have not been previously examined in detail as it was done here.

Potential Inhibitors for Novel Coronavirus Protease Identified by Virtual Screening of 606 Million Compounds.

The rapid outbreak of the novel severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) in China followed by its spread around the world poses a serious global concern for public health. To this date, no specific drugs or vaccines are available to treat SARS-CoV-2 despite its close relation to the SARS-CoV virus that caused a similar epidemic in 2003. Thus, there remains an urgent need for the identification and development of specific antiviral therapeutics against SARS-CoV-2. To conquer viral infections, the inhibition of proteases essential for proteolytic processing of viral polyproteins is a conventional therapeutic strategy. In order to find novel inhibitors, we computationally screened a compound library of over 606 million compounds for binding at the recently solved crystal structure of the main protease (Mpro) of SARS-CoV-2. A screening of such a vast chemical space for SARS-CoV-2 Mpro inhibitors has not been reported before. After shape screening, two docking protocols were applied followed by the determination of molecular descriptors relevant for pharmacokinetics to narrow down the number of initial hits. Next, molecular dynamics simulations were conducted to validate the stability of docked binding modes and comprehensively quantify ligand binding energies. After evaluation of potential off-target binding, we report a list of 12 purchasable compounds, with binding affinity to the target protease that is predicted to be more favorable than that of the cocrystallized peptidomimetic compound. In order to quickly advise ongoing therapeutic intervention for patients, we evaluated approved antiviral drugs and other protease inhibitors to provide a list of nine compounds for drug repurposing. Furthermore, we identified the natural compounds (-)-taxifolin and rhamnetin as potential inhibitors of Mpro. Rhamnetin is already commercially available in pharmacies.

Elucidating the multiple roles of hydration for accurate protein-ligand binding prediction via deep learning.

Accurate and efficient prediction of protein-ligand interactions has been a long-lasting dream of practitioners in drug discovery. The insufficient treatment of hydration is widely recognized to be a major limitation for accurate protein-ligand scoring. Using an integration of molecular dynamics simulations on thousands of protein structures with novel big-data analytics based on convolutional neural networks and deep Taylor decomposition, we consistently identify here three different patterns of hydration to be essential for protein-ligand interactions. In addition to desolvation and water-mediated interactions, the formation of enthalpically favorable networks of first-shell water molecules around solvent-exposed ligand moieties is identified to be essential for protein-ligand binding. Despite being currently neglected in drug discovery, this hydration phenomenon could lead to new avenues in optimizing the free energy of ligand binding. Application of deep neural networks incorporating hydration to docking provides 89% accuracy in binding pose ranking, an essential step for rational structure-based drug design.

Instantaneous generation of protein hydration properties from static structures.

Complex molecular simulation methods are typically required to calculate the thermodynamic properties of biochemical systems. One example thereof is the thermodynamic profiling of (de)solvation of proteins, which is an essential driving force for protein-ligand and protein-protein binding. The thermodynamic state of water molecules depends on its enthalpic and entropic components; the latter is governed by dynamic properties of the molecule. Here, we developed, to the best of our knowledge, two novel machine learning methods based on deep neural networks that are able to generate the converged thermodynamic state of dynamic water molecules in the heterogeneous protein environment based solely on the information of the static protein structure. The applicability of our machine learning methods to predict the hydration information is demonstrated in two different studies, the qualitative analysis and quantitative prediction of structure-activity relationships, and the prediction of protein-ligand binding modes.

Discovery of Inhibitors for Proliferating Cell Nuclear Antigen Using a Computational-Based Linked-Multiple-Fragment Screen.

Proliferating cell nuclear antigen (PCNA) is a central factor in DNA replication and repair pathways that plays an essential role in genome stability. The functional roles of PCNA are mediated through an extensive list of protein-protein interactions, each of which transmits specific information in protein assemblies. The flexibility at the PCNA-protein interaction interfaces offers opportunities for the discovery of functionally selective inhibitors of DNA repair pathways. Current fragment-based drug design methodologies can be limited by the flexibility of protein interfaces. These factors motivated an approach to defining compounds that could leverage previously identified subpockets on PCNA that are suitable for fragment-binding sites. Methodologies for screening multiple connected fragment-binding events in distinct subpockets are deployed to improve the selection of fragment combinations. A flexible backbone based on N-alkyl-glycine amides offers a scaffold to combinatorically link multiple fragments for in silico screening libraries that explore the diversity of subpockets at protein interfaces. This approach was applied to discover new potential inhibitors of DNA replication and repair that target PCNA in a multiprotein recognition site. The screens of the libraries were designed to computationally filter ligands based upon the fragments and positions to <1%, which were synthesized and tested for direct binding to PCNA. Molecular dynamics simulations also revealed distinct features of these novel molecules that block key PCNA-protein interactions. Furthermore, a Bayesian classifier predicted 15 of the 16 new inhibitors to be modulators of protein-protein interactions, demonstrating the method's utility as an effective screening filter. The cellular activities of example ligands with similar affinity for PCNA demonstrate unique properties for novel selective synergy with therapeutic DNA-damaging agents in drug-resistant contexts.

Improving Atom-Type Diversity and Sampling in Cosolvent Simulations Using λ-Dynamics.

Cosolvent molecular dynamics (MD) simulations perform MD simulations of the protein in explicit water mixed with cosolvent molecules that represent functional groups of ligands potentially binding to the protein. The competition between different probes and water molecules allows the identification of the energetic preference of functional groups in different binding site moieties including enthalpic and entropic contributions. Cosolvent MD simulations have recently been applied to a variety of different questions in structure-based drug design but still have significant shortcomings. Among those issues is the limited chemical diversity of probe molecules ignoring the chemical context of the pharmacophoric feature represented by a probe. Here we present a novel cosolvent MD simulation method based on the λ-dynamics simulation concept that significantly increases the chemical diversity of functional groups investigated during cosolvent simulations. Application to four different test cases highlights the utility of the new approach to identify binding preferences of different functional groups and to correctly rank ligand series that differ by their substitution patterns.

Optimization of a 1,3,4-oxadiazole series for inhibition of Ca2+/calmodulin-stimulated activity of adenylyl cyclases 1 and 8 for the treatment of chronic pain.

Adenylyl cyclases type 1 (AC1) and 8 (AC8) are group 1 transmembrane adenylyl cyclases (AC) that are stimulated by Ca2+/calmodulin. Studies have shown that mice depleted of AC1 have attenuated inflammatory pain response, while AC1/AC8 double-knockout mice display both attenuated pain response and opioid dependence. Thus, AC1 has emerged as a promising new target for treating chronic pain and opioid abuse. We discovered that the 1,3,4-oxadiazole scaffold inhibits Ca2+/calmodulin-stimulated cyclic adenosine 3',5'-monophosphate (cAMP) production in cells stably expressing either AC1 or AC8. We then carried out structure-activity relationship studies, in which we designed and synthesized 65 analogs, to modulate potency and selectivity versus each AC isoform in cells. Furthermore, molecular docking of the analogs into an AC1 homology model suggests the molecules may bind at the ATP binding site. Finally, a prioritized analog was tested in a mouse model of inflammatory pain and exhibited modest analgesic properties. In summary, our data indicate the 1,3,4-oxadiazoles represent a novel scaffold for the cellular inhibition of Ca2+/calmodulin-stimulated AC1- and AC8 cAMP and warrant further exploration as potential lead compounds for the treatment of chronic inflammatory pain.

Modeling of Halogen-Protein Interactions in Co-Solvent Molecular Dynamics Simulations.

Co-solvent molecular dynamics (MD) simulations have recently become successful approaches in structure-based drug design but neglect important interactions such as halogen bonding. To be able to successfully model compound libraries containing halogenated ligands using co-solvent simulations, we investigate the use of halogenated benzene probes in co-solvent simulations on the two test systems human cathepsin L (hCatL) and the Y220C mutant of the tumor suppressor p53 (p53-Y220C). Our studies demonstrate that halogenated benzene probes indeed can unambiguously identify halogen-bonding interaction sites in the binding pocket and show superior correlation and ranking performance compared to standard co-solvent approaches.

Rubiscolins are naturally occurring G protein-biased delta opioid receptor peptides.

The impact that ß-arrestin proteins have on G protein-coupled receptor trafficking, signaling and physiological behavior has gained much appreciation over the past decade. A number of studies have attributed the side effects associated with the use of naturally occurring and synthetic opioids, such as respiratory depression and constipation, to excessive recruitment of ß-arrestin. These findings have led to the development of biased opioid small molecule agonists that do not recruit ß-arrestin, activating only the canonical G protein pathway. Similar G protein-biased small molecule opioids have been found to occur in nature, particularly within kratom, and opioids within salvia have served as a template for the synthesis of other G protein-biased opioids. Here, we present the first report of naturally occurring peptides that selectively activate G protein signaling pathways at δ opioid receptors, but with minimal ß-arrestin recruitment. Specifically, we find that rubiscolin peptides, which are produced as cleavage products of the plant protein rubisco, bind to and activate G protein signaling at δ opioid receptors. However, unlike the naturally occurring δ opioid peptides leu-enkephalin and deltorphin II, the rubiscolin peptides only very weakly recruit ß-arrestin 2 and have undetectable recruitment of ß-arrestin 1 at the δ opioid receptor.

Efficient and Accurate Hydration Site Profiling for Enclosed Binding Sites.

Molecular dynamics (MD) simulations allow for accurate prediction of the thermodynamic profile of binding-site water molecules critical for protein-ligand association. Whereas this hydration-site profiling converges rapidly for solvent-exposed sites independent of the initial water placement, an accurate and reliable placement is required for water molecules in occluded binding sites. Here, we present an accurate and efficient hydration-site prediction method for occluded binding sites combining water placement based on 3D-RISM and MD simulations using WATsite.

Molecular Determinants of the Differential Modulation of Cav1.2 and Cav1.3 by Nifedipine and FPL 64176.

Nifedipine and FPL 64176 (FPL), which block and potentiate L-type voltage-gated Ca2+ channels, respectively, modulate Cav1.2 more potently than Cav1.3. To identify potential strategies for developing subtype-selective inhibitors, we investigated the role of divergent amino acid residues in transmembrane domains IIIS5 and the extracellular IIIS5-3P loop region in modulation of these channels by nifedipine and FPL. Insertion of the extracellular IIIS5-3P loop from Cav1.2 into Cav1.3 (Cav1.3+) reduced the IC50 of nifedipine from 289 to 101 nM, and substitution of S1100 with an A residue, as in Cav1.2, accounted for this difference. Substituting M1030 in IIIS5 to V in Cav1.3+ (Cav1.3+V) further reduced the IC50 of nifedipine to 42 nM. FPL increased current amplitude with an EC50 of 854 nM in Cav1.3, 103 nM in Cav1.2, and 99 nM in Cav1.3+V. In contrast to nifedipine block, substitution of M1030 to V in Cav1.3 had no effect on potency of FPL potentiation of current amplitude, but slowed deactivation in the presence and absence of 10 µM FPL. FPL had no effect on deactivation of Cav1.3/dihydropyridine-insensitive (DHPi), a channel with very low sensitivity to nifedipine block (IC50 ∼93 µM), but did shift the voltage-dependence of activation by ∼-10 mV. We conclude that the M/V variation in IIIS5 and the S/A variation in the IIIS5-3P loop of Cav1.2 and Cav1.3 largely determine the difference in nifedipine potency between these two channels, but the difference in FPL potency is determined by divergent amino acids in the IIIS5-3P loop.