Generating 3D molecules conditional on receptor binding sites with deep generative models

The goal of structure-based drug discovery is to find small molecules that bind to a given target protein. Deep learning has been used to generate drug-like molecules with certain cheminformatic properties, but has not yet been applied to generating 3D molecules predicted to bind to proteins by sampling the conditional distribution of protein–ligand binding interactions. In this work, we describe for the first time a deep learning system for generating 3D molecular structures conditioned on a receptor binding site. We approach the problem using a conditional variational autoencoder trained on an atomic density grid representation of cross-docked protein–ligand structures. We apply atom fitting and bond inference procedures to construct valid molecular conformations from generated atomic densities. We evaluate the properties of the generated molecules and demonstrate that they change significantly when conditioned on mutated receptors. We also explore the latent space learned by our generative model using sampling and interpolation techniques. This work opens the door for end-to-end prediction of stable bioactive molecules from protein structures with deep learning.

We generate 3D molecules conditioned on receptor binding sites by training a deep generative model on protein–ligand complexes. Our model uses the conditional receptor information to make chemically relevant changes to the generated molecules.     

Deep generative design with 3D pharmacophoric constraints

Generative models have increasingly been proposed as a solution to the molecular design problem. However, it has proved challenging to control the design process or incorporate prior knowledge, limiting their practical use in drug discovery. In particular, generative methods have made limited use of three-dimensional (3D) structural information even though this is critical to binding. This work describes a method to incorporate such information and demonstrates the benefit of doing so. We combine an existing graph-based deep generative model, DeLinker, with a convolutional neural network to utilise physically-meaningful 3D representations of molecules and target pharmacophores. We apply our model, DEVELOP, to both linker and R-group design, demonstrating its suitability for both hit-to-lead and lead optimisation. The 3D pharmacophoric information results in improved generation and allows greater control of the design process. In multiple large-scale evaluations, we show that including 3D pharmacophoric constraints results in substantial improvements in the quality of generated molecules. On a challenging test set derived from PDBbind, our model improves the proportion of generated molecules with high 3D similarity to the original molecule by over 300%. In addition, DEVELOP recovers 10× more of the original molecules compared to the baseline DeLinker method. Our approach is general-purpose, readily modifiable to alternate 3D representations, and can be incorporated into other generative frameworks. Code is available at https://github.com/oxpig/DEVELOP.

A novel deep generative model combines convolution and graph neural networks to allow 3D-aware molecular design. We show how 3D pharmacophoric information can be incorporated into generative models and apply our model to both linker and R-group design.     

De novo design based identification of potential HIV-1 integrase inhibitors: A pharmacoinformatics study

In the present study, pharmacoinformatics paradigms include receptor-based de novo design, virtual screening through molecular docking and molecular dynamics (MD) simulation are implemented to identify novel and promising HIV-1 integrase inhibitors. The de novodrug/ligand/molecule design is a powerful and effective approach to design a large number of novel and structurally diverse compounds with the required pharmacological profiles. A crystal structure of HIV-1 integrase bound with standard inhibitor BI-224436 is used and a set of 80,000 compounds through the de novo approach in LigBuilder is designed. Initially, a number of criteria including molecular docking, in-silico toxicity and pharmacokinetics profile assessments are implied to reduce the chemical space. Finally, four de novo designed molecules are proposed as potential HIV-1 integrase inhibitors based on comparative analyses. Notably, strong binding interactions have been identified between a few newly identified catalytic amino acid residues and proposed HIV-1 integrase inhibitors. For evaluation of the dynamic stability of the protein-ligand complexes, a number of parameters are explored from the 100 ns MD simulation study. The MD simulation study suggested that proposed molecules efficiently retained their molecular interaction and structural integrity inside the HIV-1 integrase. The binding free energy is calculated through the Molecular Mechanics Poisson-Boltzmann Surface Area (MM-PBSA) approach for all complexes and it also explains their thermodynamic stability. Hence, proposed molecules through de novo design might be critical to inhibiting the HIV-1 integrase.     

Automated Docking and the Search for HIV Protease Inhibitors

Abstract

This article will discuss the motivations, technologies, and future directions of computational automated docking in the context of the structure-based rational design of HIV-1 protease inhibitors. Docking simulations are widely used for screening of compound libraries to identify new drug leads, employing a simple model for rapid testing of thousands of compounds. Docking simulations are also useful for lead enhancement, using more detailed models to analyze the atomic interactions between inhibitors and target macromolecules. Major advances have been reported in the development of empirical force fields, which now allow assessment of relative binding strength and drug specificity, and extensions of automated docking techniques allow de novo drug design.     

Drug search for leishmaniasis: a virtual screening approach by grid computing

The trypanosomatid protozoa Leishmania is endemic in ~100 countries, with infections causing ~2 million new cases of leishmaniasis annually. Disease symptoms can include severe skin and mucosal ulcers, fever, anemia, splenomegaly, and death. Unfortunately, therapeutics approved to treat leishmaniasis are associated with potentially severe side effects, including death. Furthermore, drug-resistant Leishmania parasites have developed in most endemic countries. To address an urgent need for new, safe and inexpensive anti-leishmanial drugs, we utilized the IBM World Community Grid to complete computer-based drug discovery screens (Drug Search for Leishmaniasis) using unique leishmanial proteins and a database of 600,000 drug-like small molecules. Protein structures from different Leishmania species were selected for molecular dynamics (MD) simulations, and a series of conformational “snapshots” were chosen from each MD trajectory to simulate the protein’s flexibility. A Relaxed Complex Scheme methodology was used to screen ~2000 MD conformations against the small molecule database, producing >1 billion protein-ligand structures. For each protein target, a binding spectrum was calculated to identify compounds predicted to bind with highest average affinity to all protein conformations. Significantly, four different Leishmania protein targets were predicted to strongly bind small molecules, with the strongest binding interactions predicted to occur for dihydroorotate dehydrogenase (LmDHODH; PDB:3MJY). A number of predicted tight-binding LmDHODH inhibitors were tested in vitro and potent selective inhibitors of Leishmania panamensis were identified. These promising small molecules are suitable for further development using iterative structure-based optimization and in vitro/in vivo validation assays.     

Automated affinity selection for rapid discovery of peptide binders

In-solution affinity selection (AS) of large synthetic peptide libraries affords identification of binders to protein targets through access to an expanded chemical space. Standard affinity selection methods, however, can be time-consuming, low-throughput, or provide hits that display low selectivity to the target. Here we report an automated bio-layer interferometry (BLI)-assisted affinity selection platform. When coupled with tandem mass spectrometry (MS), this method enables both rapid de novo discovery and affinity maturation of known peptide binders with high selectivity. The BLI-assisted AS-MS technology also features real-time monitoring of the peptide binding during the library selection process, a feature unattainable by current selection approaches. We show the utility of the BLI AS-MS platform toward rapid identification of novel nanomolar (dissociation constant, K_D < 50 nM) non-canonical binders to the leukemia-associated oncogenic protein menin. To our knowledge, this is the first application of BLI to the affinity selection of synthetic peptide libraries. We believe our approach can significantly accelerate the use of synthetic peptidomimetic libraries in drug discovery.

This work reports an automated affinity selection-mass spectrometry (AS-MS) approach amenable to both de novo peptide binder discovery and affinity maturation of known binders in a high-throughput and selective manner.     

Identification of the druggable concavity in homology models using the PLB index

Identification of the druggable concavity, in which drug-like molecules are highly inclined to bind, is an important step in structure-based drug design. We previously proposed an index named PLB (propensity for ligand binding), which is based on the amino acid composition characteristically observed at the small molecule binding sites in the X-ray structures of the complexes between proteins and drug-like small molecules. The PLB index was proven to be useful in identifying the druggable concavities in the quality X-ray structures of proteins. Here, we apply the PLB to predicting the druggable concavity in target proteins using the structures of homologous proteins constructed by homology modeling. In this study, we assembled a set of reference proteins that were accurately determined by X-ray analysis in forms of complexes with drug-like small molecules. Homology models for the reference protein were constructed using multiple homologous proteins as templates. The PLB index was then used to predict the druggable concavity. If the template protein in a complex with a drug-like small molecule was used, the druggable concavity was predicted well, with a prediction rate of 78%. When only the apo protein was available as the template, the practical prediction rate was 71%. Interestingly, even when the percent sequence identity between the reference and template proteins was lower than 30, the PLB index could successfully identify the druggable concavity in some cases. This study demonstrates the practical value of applying the PLB index to identifying the drugabble concavity in the homology model.     

Structure-based design technology contour and its application to the design of Renin inhibitors

It is well-known that the structure-based design approach has had a measurable impact on the drug discovery process in identifying novel and efficacious therapeutic agents for a variety of disease targets. The de novo design approach has inherent potential to generate novel molecules that best fit into a protein binding site when compared to all of the computational methods applied to structure-based design. In its initial attempts, this approach did not achieve much success due to technical hurdles. More recently, the algorithmic advancements in the methodologies and clever strategies developed to design drug-like molecules have improved the success rate. We describe a state-of-the-art structure-based design technology called Contour and provide details of the algorithmic enhancements we have implemented. Contour was designed to create novel drug-like molecules by assembling synthetically viable fragments in the protein binding site using a high-resolution crystal structure of the protein. The technology consists of a sophisticated growth algorithm and a novel scoring function based on a directional model. The growth algorithm generates molecules by dynamically selecting only those fragments from the fragment library that are complementary to the binding site, and assembling them by sampling the conformational space for each attached fragment. The scoring function embodying the essential elements of the binding interactions aids in the rank ordering of grown molecules and helps identify those that have high probability of exhibiting activity against the protein target of interest. The application of Contour to identify inhibitors against human renin enzyme eventually leading to the clinical candidate VTP-27,999 will be discussed here.     

Coiled coils 9-to-5: rational de novo design of α-helical barrels with tunable oligomeric states

The rational design of linear peptides that assemble controllably and predictably in water is challenging. Short sequences must encode unique target structures and avoid alternative states. However, the non-covalent forces that stabilize and discriminate between states are weak. Nonetheless, for α-helical coiled-coil assemblies considerable progress has been made in rational de novo design. In these, sequence repeats of nominally hydrophobic (h) and polar (p) residues, hpphppp, direct the assembly of amphipathic helices into dimeric to tetrameric bundles. Expanding this pattern to hpphhph can produce larger α-helical barrels. Here, we show that pentameric to nonameric barrels are accessed by varying the residue at one of the h sites. In peptides with four L/I–K–E–I–A–x–Z repeats, decreasing the size of Z from threonine to serine to alanine to glycine gives progressively larger oligomers. X-ray crystal structures of the resulting α-helical barrels rationalize this: side chains at Z point directly into the helical interfaces, and smaller residues allow closer helix contacts and larger assemblies.

Systematic de novo design of peptides that form α-helical barrels with functionalisable central channels with a range of internal diameters.     

Beyond generative models: superfast traversal,optimization, novelty,exploration and discovery (STONED) algorithm for molecules using SELFIES

Inverse design allows the generation of molecules with desirable physical quantities using property optimization. Deep generative models have recently been applied to tackle inverse design, as they possess the ability to optimize molecular properties directly through structure modification using gradients. While the ability to carry out direct property optimizations is promising, the use of generative deep learning models to solve practical problems requires large amounts of data and is very time-consuming. In this work, we propose STONED – a simple and efficient algorithm to perform interpolation and exploration in the chemical space, comparable to deep generative models. STONED bypasses the need for large amounts of data and training times by using string modifications in the SELFIES molecular representation. First, we achieve non-trivial performance on typical benchmarks for generative models without any training. Additionally, we demonstrate applications in high-throughput virtual screening for the design of drugs, photovoltaics, and the construction of chemical paths, allowing for both property and structure-based interpolation in the chemical space. Overall, we anticipate our results to be a stepping stone for developing more sophisticated inverse design models and benchmarking tools, ultimately helping generative models achieve wider adoption.

Interpolation and exploration within the chemical space for inverse design.     

Approximating protein flexibility through dynamic pharmacophore models: application to fatty acid amide hydrolase (FAAH)

A structure-based drug discovery method is described that incorporates target flexibility through the use of an ensemble of protein conformations. The approach was applied to fatty acid amide hydrolase (FAAH), a key deactivating enzyme in the endocannabinoid system. The resultant dynamic pharmacophore models are rapidly able to identify known FAAH inhibitors over drug-like decoys. Different sources of FAAH conformational ensembles were explored, with both snapshots from molecular dynamics simulations and a group of X-ray structures performing well. Results were compared to those from docking and pharmacophore models generated from a single X-ray structure. Increasing conformational sampling consistently improved the pharmacophore models, emphasizing the importance of incorporating target flexibility in structure-based drug design.     

Target-templated de novo design of macrocyclic d-/l-peptides: discovery of drug-like inhibitors of PD-1

Peptides are a rapidly growing class of therapeutics with various advantages over traditional small molecules, especially for targeting difficult protein–protein interactions. However, current structure-based methods are largely limited to natural peptides and are not suitable for designing bioactive cyclic topologies that go beyond natural l-amino acids. Here, we report a generalizable framework that exploits the computational power of Rosetta, in terms of large-scale backbone sampling, side-chain composition and energy scoring, to design heterochiral cyclic peptides that bind to a protein surface of interest. To showcase the applicability of our approach, we developed two new inhibitors (PD-i3 and PD-i6) of programmed cell death 1 (PD-1), a key immune checkpoint in oncology. A comprehensive biophysical evaluation was performed to assess their binding to PD-1 as well as their blocking effect on the endogenous PD-1/PD-L1 interaction. Finally, NMR elucidation of their in-solution structures confirmed our de novo design approach.

In silico design of heterochiral cyclic peptides that bind to a specific surface patch on the target protein (PD-1, in this case) and disrupt protein–protein interactions.     

Ligand design by a combinatorial approach based on modeling and experiment: application to HLA-DR4

Combinatorial synthesis and large scale screening methods are being used increasingly in drug discovery, particularly for finding novel lead compounds. Although these "random" methods sample larger areas of chemical space than traditional synthetic approaches, only a relatively small percentage of all possible compounds are practically accessible. It is therefore helpful to select regions of chemical space that have greater likelihood of yielding useful leads. When three-dimensional structural data are available for the target molecule this can be achieved by applying structure-based computational design methods to focus the combinatorial library. This is advantageous over the standard usage of computational methods to design a small number of specific novel ligands, because here computation is employed as part of the combinatorial design process and so is required only to determine a propensity for binding of certain chemical moieties in regions of the target molecule. This paper describes the application of the Multiple Copy Simultaneous Search (MCSS) method, an active site mapping and de novo structure-based design tool, to design a focused combinatorial library for the class II MHC protein HLA-DR4. Methods for the synthesizing and screening the computationally designed library are presented; evidence is provided to show that binding was achieved. Although the structure of the protein-ligand complex could not be determined, experimental results including cross-exclusion of a known HLA-DR4 peptide ligand (HA) by a compound from the library. Computational model building suggest that at least one of the ligands designed and identified by the methods described binds in a mode similar to that of native peptides.     

PIGNet: a physics-informed deep learning model toward generalized drug–target interaction predictions

Recently, deep neural network (DNN)-based drug–target interaction (DTI) models were highlighted for their high accuracy with affordable computational costs. Yet, the models'' insufficient generalization remains a challenging problem in the practice of in silico drug discovery. We propose two key strategies to enhance generalization in the DTI model. The first is to predict the atom–atom pairwise interactions via physics-informed equations parameterized with neural networks and provides the total binding affinity of a protein–ligand complex as their sum. We further improved the model generalization by augmenting a broader range of binding poses and ligands to training data. We validated our model, PIGNet, in the comparative assessment of scoring functions (CASF) 2016, demonstrating the outperforming docking and screening powers than previous methods. Our physics-informing strategy also enables the interpretation of predicted affinities by visualizing the contribution of ligand substructures, providing insights for further ligand optimization.

PIGNet, a deep neural network-based drug–target interaction model guided by physics and extensive data augmentation, shows significantly improved generalization ability and model performance.     

A validation study on the practical use of automated de novo design

The de novo design program Skelgen has been used to design inhibitor structures for four targets of pharmaceutical interest. The designed structures are compared to modeled binding modes of known inhibitors (i) visually and (ii) by means of a novel similarity measure considering the size and spatial proximity of the maximum common substructure of two small molecules. It is shown that the Skelgen algorithm generates representatives of many inhibitor classes within a very short time and that the new similarity measure is useful for comparing and clustering designed structures. The results demonstrate the necessity of properly defining search constraints in practical applications of de novo design.     

Picomolar FKBP inhibitors enabled by a single water-displacing methyl group in bicyclic [4.3.1] aza-amides

Methyl groups can have profound effects in drug discovery but the underlying mechanisms are diverse and incompletely understood. Here we report the stereospecific effect of a single, solvent-exposed methyl group in bicyclic [4.3.1] aza-amides, robustly leading to a 2 to 10-fold increase in binding affinity for FK506-binding proteins (FKBPs). This resulted in the most potent and efficient FKBP ligands known to date. By a combination of co-crystal structures, isothermal titration calorimetry (ITC), density-functional theory (DFT), and 3D reference interaction site model (3D-RISM) calculations we elucidated the origin of the observed affinity boost, which was purely entropically driven and relied on the displacement of a water molecule at the protein–ligand–bulk solvent interface. The best compounds potently occupied FKBPs in cells and enhanced bone morphogenic protein (BMP) signaling. Our results show how subtle manipulation of the solvent network can be used to design atom-efficient ligands for difficult, solvent-exposed binding pockets.

Enhancement by displacement. A single methyl group displaces a water molecule from the binding site of FKBPs, resulting in the most potent binders known, outperforming the natural products FK506 and rapamycin in biochemical and cellular assays.     

Cooperativity as quantification and optimization paradigm for nuclear receptor modulators

Nuclear Receptors (NRs) are highly relevant drug targets, for which small molecule modulation goes beyond a simple ligand/receptor interaction. NR–ligands modulate Protein–Protein Interactions (PPIs) with coregulator proteins. Here we bring forward a cooperativity mechanism for small molecule modulation of NR PPIs, using the Peroxisome Proliferator Activated Receptor γ (PPARγ), which describes NR–ligands as allosteric molecular glues. The cooperativity framework uses a thermodynamic model based on three-body binding events, to dissect and quantify reciprocal effects of NR–coregulator binding (K^I_D) and NR–ligand binding (K^II_D), jointly recapitulated in the cooperativity factor (α) for each specific ternary ligand·NR·coregulator complex formation. These fundamental thermodynamic parameters allow for a conceptually new way of thinking about structure–activity-relationships for NR–ligands and can steer NR modulator discovery and optimization via a completely novel approach.

A cooperativity framework describes the formation of nuclear receptor ternary complexes and deconvolutes ligand and cofactor binding into intrinsic affinities and a cooperativity factor, providing a conceptually new understanding of NR modulation.     

Incorporating protein flexibility in structure-based drug discovery: using HIV-1 protease as a test case

We have developed a receptor-based pharmacophore method which utilizes a collection of protein structures to account for inherent protein flexibility in structure-based drug design. Several procedures were systematically evaluated to derive the most general protocol for using multiple protein structures. Most notably, incorporating more protein flexibility improved the performance of the method. The pharmacophore models successfully discriminate known inhibitors from drug-like non-inhibitors. Furthermore, the models correctly identify the bound conformations of some ligands. We used unliganded HIV-1 protease to develop and validate this method. Drug design is always initiated with a protein-ligand structure, and such success with unbound protein structures is remarkable - particularly in the case of HIV-1 protease, which has a large conformational change upon binding. This technique holds the promise of successful computer-based drug design before bound crystal structures are even discovered, which can mean a jump-start of 1-3 years in tackling some medically relevant systems with computational methods.     

MGraphDTA: deep multiscale graph neural network for explainable drug–target binding affinity prediction

Predicting drug–target affinity (DTA) is beneficial for accelerating drug discovery. Graph neural networks (GNNs) have been widely used in DTA prediction. However, existing shallow GNNs are insufficient to capture the global structure of compounds. Besides, the interpretability of the graph-based DTA models highly relies on the graph attention mechanism, which can not reveal the global relationship between each atom of a molecule. In this study, we proposed a deep multiscale graph neural network based on chemical intuition for DTA prediction (MGraphDTA). We introduced a dense connection into the GNN and built a super-deep GNN with 27 graph convolutional layers to capture the local and global structure of the compound simultaneously. We also developed a novel visual explanation method, gradient-weighted affinity activation mapping (Grad-AAM), to analyze a deep learning model from the chemical perspective. We evaluated our approach using seven benchmark datasets and compared the proposed method to the state-of-the-art deep learning (DL) models. MGraphDTA outperforms other DL-based approaches significantly on various datasets. Moreover, we show that Grad-AAM creates explanations that are consistent with pharmacologists, which may help us gain chemical insights directly from data beyond human perception. These advantages demonstrate that the proposed method improves the generalization and interpretation capability of DTA prediction modeling.

MGraphDTA is designed to capture the local and global structure of a compound simultaneously for drug–target affinity prediction and can provide explanations that are consistent with pharmacologists.