Knowledge-based fragment binding prediction

Streamlining Computational Fragment-Based Drug Discovery through Evolutionary Optimization Informed by Ligand-Based Virtual Prescreening

Recent advances in computational methods provide the promise of dramatically accelerating drug discovery. While mathematical modeling and machine learning have become vital in predicting drug-target interactions and properties, there is untapped potential in computational drug discovery due to the vast and complex chemical space. This paper builds on our recently published computational fragment-based drug discovery (FBDD) method called fragment databases from screened ligand drug discovery (FDSL-DD). FDSL-DD uses in silico screening to identify ligands from a vast library, fragmenting them while attaching specific attributes based on predicted binding affinity and interaction with the target subdomain. In this paper, we further propose a two-stage optimization method that utilizes the information from prescreening to optimize computational ligand synthesis. We hypothesize that using prescreening information for optimization shrinks the search space and focuses on promising regions, thereby improving the optimization for candidate ligands. The first optimization stage assembles these fragments into larger compounds using genetic algorithms, followed by a second stage of iterative refinement to produce compounds with enhanced bioactivity. To demonstrate broad applicability, the methodology is demonstrated on three diverse protein targets found in human solid cancers, bacterial antimicrobial resistance, and the SARS-CoV-2 virus. Combined, the proposed FDSL-DD and a two-stage optimization approach yield high-affinity ligand candidates more efficiently than other state-of-the-art computational FBDD methods. We further show that a multiobjective optimization method accounting for drug-likeness can still produce potential candidate ligands with a high binding affinity. Overall, the results demonstrate that integrating detailed chemical information with a constrained search framework can markedly optimize the initial drug discovery process, offering a more precise and efficient route to developing new therapeutics.

Fragment databases from screened ligands for drug discovery (FDSL-DD)

Fragment-based drug design (FBDD) is one major drug discovery method employed in computer-aided drug discovery. Due to its inherent limitations, this process experiences long processing times and limited success rates. Here we present a new Fragment Databases from Screened Ligands Drug Design method (FDSL-DD) that intelligently incorporates information about fragment characteristics into a fragment-based design approach to the drug development process. The initial step of the FDSL-DD is the creation of a fragment database from a library of docked, drug-like ligands for a specific target, which deviates from the traditional in silico FBDD strategy, incorporating structure-based design screening techniques to combine the advantages of both approaches. Three different protein targets have been tested in this study to demonstrate the potential of the created fragment library and FDSL-DD. Utilizing the FDSL-DD led to an increase in binding affinity for each protein target. The most substantial increase was exhibited by the ligand designed for TIPE2, with a 3.6 kcalmol-1 difference between the top ligand from the FDSL-DD and top ligand from the high throughput virtual screening (HTVS). Using drug-like ligands in the initial HTVS allows for a greater search of chemical space, with higher efficiency in fragments selection, less grid boxes, and potentially identifying more interactions.

BMaps: A Web Application for Fragment-Based Drug Design and Compound Binding Evaluation

Fragment-based drug design uses data about where, and how strongly, small chemical fragments bind to proteins, to assemble new drug molecules. Over the past decade, we have been successfully using fragment data, derived from thermodynamically rigorous Monte Carlo fragment-protein binding simulations, in dozens of preclinical drug programs. However, this approach has not been available to the broader research community because of the cost and complexity of doing simulations and using design tools. We have developed a web application, called BMaps, to make fragment-based drug design widely available with greatly simplified user interfaces. BMaps provides access to a large repository (>550) of proteins with 100s of precomputed fragment maps, druggable hot spots, and high-quality water maps. Users can also employ their own structures or those from the Protein Data Bank and AlphaFold DB. Multigigabyte data sets are searched to find fragments in bondable orientations, ranked by a binding-free energy metric. The designers use this to select modifications that improve affinity and other properties. BMaps is unique in combining conventional tools such as docking and energy minimization with fragment-based design, in a very easy to use and automated web application. The service is available at https://www.boltzmannmaps.com.

COLLAPSE: A representation learning framework for identification and characterization of protein structural sites

The identification and characterization of the structural sites which contribute to protein function are crucial for understanding biological mechanisms, evaluating disease risk, and developing targeted therapies. However, the quantity of known protein structures is rapidly outpacing our ability to functionally annotate them. Existing methods for function prediction either do not operate on local sites, suffer from high false positive or false negative rates, or require large site-specific training datasets, necessitating the development of new computational methods for annotating functional sites at scale. We present COLLAPSE (Compressed Latents Learned from Aligned Protein Structural Environments), a framework for learning deep representations of protein sites. COLLAPSE operates directly on the 3D positions of atoms surrounding a site and uses evolutionary relationships between homologous proteins as a self-supervision signal, enabling learned embeddings to implicitly capture structure-function relationships within each site. Our representations generalize across disparate tasks in a transfer learning context, achieving state-of-the-art performance on standardized benchmarks (protein-protein interactions and mutation stability) and on the prediction of functional sites from the Prosite database. We use COLLAPSE to search for similar sites across large protein datasets and to annotate proteins based on a database of known functional sites. These methods demonstrate that COLLAPSE is computationally efficient, tunable, and interpretable, providing a general-purpose platform for computational protein analysis.

Defining A Global Map of Functional Group-based 3D Ligand-binding Motifs

Uncovering conserved 3D protein-ligand binding patterns on the basis of functional groups (FGs) shared by a variety of small molecules can greatly expand our knowledge of protein-ligand interactions. Despite that conserved binding patterns for a few commonly used FGs have been reported in the literature, large-scale identification and evaluation of FG-based 3D binding motifs are still lacking. Here, we propose a computational method, Automatic FG-based Three-dimensional Motif Extractor (AFTME), for automatic mapping of 3D motifs to different FGs of a specific ligand. Applying our method to 233 naturally-occurring ligands, we define 481 FG-binding motifs that are highly conserved across different ligand-binding pockets. Systematic analysis further reveals four main classes of binding motifs corresponding to distinct sets of FGs. Combinations of FG-binding motifs facilitate the binding of proteins to a wide spectrum of ligands with various binding affinities. Finally, we show that our FG-motif map can be used to nominate FGs that potentially bind to specific drug targets, thus providing useful insights and guidance for rational design of small-molecule drugs.

CAVIAR: a method for automatic cavity detection, description and decomposition into subcavities

The accurate description of protein binding sites is essential to the determination of similarity and the application of machine learning methods to relate the binding sites to observed functions. This work describes CAVIAR, a new open source tool for generating descriptors for binding sites, using protein structures in PDB and mmCIF format as well as trajectory frames from molecular dynamics simulations as input. The applicability of CAVIAR descriptors is showcased by computing machine learning predictions of binding site ligandability. The method can also automatically assign subcavities, even in the absence of a bound ligand. The defined subpockets mimic the empirical definitions used in medicinal chemistry projects. It is shown that the experimental binding affinity scales relatively well with the number of subcavities filled by the ligand, with compounds binding to more than three subcavities having nanomolar or better affinities to the target. The CAVIAR descriptors and methods can be used in any machine learning-based investigations of problems involving binding sites, from protein engineering to hit identification. The full software code is available on GitHub and a conda package is hosted on Anaconda cloud.

High precision protein functional site detection using 3D convolutional neural networks

Motivation

Accurate annotation of protein functions is fundamental for understanding molecular and cellular physiology. Data-driven methods hold promise for systematically deriving rules underlying the relationship between protein structure and function. However, the choice of protein structural representation is critical. Pre-defined biochemical features emphasize certain aspects of protein properties while ignoring others, and therefore may fail to capture critical information in complex protein sites.

Results

In this paper, we present a general framework that applies 3D convolutional neural networks (3DCNNs) to structure-based protein functional site detection. The framework can extract task-dependent features automatically from the raw atom distributions. We benchmarked our method against other methods and demonstrate better or comparable performance for site detection. Our deep 3DCNNs achieved an average recall of 0.955 at a precision threshold of 0.99 on PROSITE families, detected 98.89 and 92.88% of nitric oxide synthase and TRYPSIN-like enzyme sites in Catalytic Site Atlas, and showed good performance on challenging cases where sequence motifs are absent but a function is known to exist. Finally, we inspected the individual contributions of each atom to the classification decisions and show that our models successfully recapitulate known 3D features within protein functional sites.

Availability and implementation

The 3DCNN models described in this paper are available at https://simtk.org/projects/fscnn.

Supplementary information

Supplementary data are available at Bioinformatics online.

A deep learning framework to predict binding preference of RNA constituents on protein surface

Protein-RNA interaction plays important roles in post-transcriptional regulation. However, the task of predicting these interactions given a protein structure is difficult. Here we show that, by leveraging a deep learning model NucleicNet, attributes such as binding preference of RNA backbone constituents and different bases can be predicted from local physicochemical characteristics of protein structure surface. On a diverse set of challenging RNA-binding proteins, including Fem-3-binding-factor 2, Argonaute 2 and Ribonuclease III, NucleicNet can accurately recover interaction modes discovered by structural biology experiments. Furthermore, we show that, without seeing any in vitro or in vivo assay data, NucleicNet can still achieve consistency with experiments, including RNAcompete, Immunoprecipitation Assay, and siRNA Knockdown Benchmark. NucleicNet can thus serve to provide quantitative fitness of RNA sequences for given binding pockets or to predict potential binding pockets and binding RNAs for previously unknown RNA binding proteins.

In silico fragment-based drug design with SEED

We report on two fragment-based drug design protocols, SEED2XR and ALTA, which start by high-throughput docking. SEED2XR is a two-stage protocol for fragment-based drug design. The first stage is in silico and consists of the automatic docking of 10³-10⁴ fragments using SEED, which requires about 1 s per fragment. SEED is a docking software developed specifically for fragment docking and binding energy evaluation by a force field with implicit solvent. In the second stage of SEED2XR, the 10-10² fragments with the most favorable predicted binding energies are validated by protein X-ray crystallography. The recent applications of SEED2XR to bromodomains demonstrate that the whole SEED2XR protocol can be carried out in about a week of working time, with hit rates ranging from 10% to 40%. Information on fragment-target interactions generated by the SEED2XR protocol or directly from SEED docking has been used for the discovery of hundreds of hits. ALTA is a computational protocol for screening which identifies candidate ligands that preserve the interactions between the optimal SEED fragments and the protein target. Medicinal chemistry optimization of ligands predicted by ALTA has resulted in pre-clinical candidates for protein kinases and bromodomains. The high-throughput, very low cost, sustainability, and high hit rate of the SEED-based protocols, unreachable by purely experimental techniques, make them perfectly suitable for both academic and industrial drug discovery research.

Ranking Enzyme Structures in the PDB by Bound Ligand Similarity to Biological Substrates

There are numerous applications that use the structures of protein-ligand complexes from the PDB, such as 3D pharmacophore identification, virtual screening, and fragment-based drug design. The structures underlying these applications are potentially much more informative if they contain biologically relevant bound ligands, with high similarity to the cognate ligands. We present a study of ligand-enzyme complexes that compares the similarity of bound and cognate ligands, enabling the best matches to be identified. We calculate the molecular similarity scores using a method called PARITY (proportion of atoms residing in identical topology), which can conveniently be combined to give a similarity score for all cognate reactants or products in the reaction. Thus, we generate a rank-ordered list of related PDB structures, according to the biological similarity of the ligands bound in the structures.

•

We present PARITY, matching atoms in identical topology to gauge ligand similarity

•

Bound-cognate ligand similarity is a useful metric for ranking PDB structures

•

Only 26% of enzyme structures in the PDB have bound-cognate ligand similarity ≥0.7

•

We provide rank-ordered lists of PDBs with the most biologically relevant ligands

Identification of Histamine H3 Receptor Ligands Using a New Crystal Structure Fragment-based Method

Virtual screening offers an efficient alternative to high-throughput screening in the identification of pharmacological tools and lead compounds. Virtual screening is typically based on the matching of target structures or ligand pharmacophores to commercial or in-house compound catalogues. This study provides the first proof-of-concept for our recently reported method where pharmacophores are instead constructed based on the inference of residue-ligand fragments from crystal structures. We demonstrate its unique utility for G protein-coupled receptors, which represent the largest families of human membrane proteins and drug targets. We identified five neutral antagonists and one inverse agonist for the histamine H₃ receptor with potencies of 0.7-8.5 μM in a recombinant receptor cell-based inositol phosphate accumulation assay and validated their activity using a radioligand competition binding assay. H₃ receptor antagonism is of large therapeutic value and our ligands could serve as starting points for further lead optimisation. The six ligands exhibit four chemical scaffolds, whereof three have high novelty in comparison to the known H₃ receptor ligands in the ChEMBL database. The complete pharmacophore fragment library is freely available through the GPCR database, GPCRdb, allowing the successful application herein to be repeated for most of the 285 class A GPCR targets. The method could also easily be adapted to other protein families.

3D deep convolutional neural networks for amino acid environment similarity analysis

Background

Central to protein biology is the understanding of how structural elements give rise to observed function. The surfeit of protein structural data enables development of computational methods to systematically derive rules governing structural-functional relationships. However, performance of these methods depends critically on the choice of protein structural representation. Most current methods rely on features that are manually selected based on knowledge about protein structures. These are often general-purpose but not optimized for the specific application of interest.

In this paper, we present a general framework that applies 3D convolutional neural network (3DCNN) technology to structure-based protein analysis. The framework automatically extracts task-specific features from the raw atom distribution, driven by supervised labels. As a pilot study, we use our network to analyze local protein microenvironments surrounding the 20 amino acids, and predict the amino acids most compatible with environments within a protein structure. To further validate the power of our method, we construct two amino acid substitution matrices from the prediction statistics and use them to predict effects of mutations in T4 lysozyme structures.

Results

Our deep 3DCNN achieves a two-fold increase in prediction accuracy compared to models that employ conventional hand-engineered features and successfully recapitulates known information about similar and different microenvironments. Models built from our predictions and substitution matrices achieve an 85% accuracy predicting outcomes of the T4 lysozyme mutation variants. Our substitution matrices contain rich information relevant to mutation analysis compared to well-established substitution matrices. Finally, we present a visualization method to inspect the individual contributions of each atom to the classification decisions.

Conclusions

End-to-end trained deep learning networks consistently outperform methods using hand-engineered features, suggesting that the 3DCNN framework is well suited for analysis of protein microenvironments and may be useful for other protein structural analyses.

Electronic supplementary material

The online version of this article (doi:10.1186/s12859-017-1702-0) contains supplementary material, which is available to authorized users.

Combination therapeutics in complex diseases

The biological redundancies in molecular networks of complex diseases limit the efficacy of many single drug therapies. Combination therapeutics, as a common therapeutic method, involve pharmacological intervention using several drugs that interact with multiple targets in the molecular networks of diseases and may achieve better efficacy and/or less toxicity than monotherapy in practice. The development of combination therapeutics is complicated by several critical issues, including identifying multiple targets, targeting strategies and the drug combination. This review summarizes the current achievements in combination therapeutics, with a particular emphasis on the efforts to develop combination therapeutics for complex diseases.

IsoMIF Finder: online detection of binding site molecular interaction field similarities

Summary: IsoMIF Finder is an online server for the identification of molecular interaction field (MIF) similarities. User defined binding site MIFs can be compared to datasets of pre-calculated MIFs or against a user-defined list of PDB entries. The interface can be used for the prediction of function, identification of potential cross-reactivity or polypharmacological targets and drug repurposing. Detected similarities can be viewed in a browser or within a PyMOL session.

Availability and Implementation: IsoMIF Finder uses JSMOL (no java plugin required), is cross-browser and freely available at bcb.med.usherbrooke.ca/imfi.

Contact:Rafael.Najmanovich@Usherbrooke.ca

Supplementary information:Supplementary data are available at Bioinformatics online.

Modeling enzyme-ligand binding in drug discovery

Enzymes are one of the most important groups of drug targets, and identifying possible ligand-enzyme interactions is of major importance in many drug discovery processes. Novel computational methods have been developed that can apply the information from the increasing number of resolved and available ligand-enzyme complexes to model new unknown interactions and therefore contribute to answer open questions in the field of drug discovery like the identification of unknown protein functions, off-target binding, ligand 3D homology modeling and induced-fit simulations.

Considerations of Protein Subpockets in Fragment-Based Drug Design

While the fragment-based drug design approach continues to gain importance, gaps in the tools and methods available in the identification and accurate utilization of protein subpockets have limited the scope. The importance of these features of small molecule-protein recognition is highlighted with several examples. A generalized solution for the identification of subpockets and corresponding chemical fragments remains elusive, but there are numerous advancements in methods that can be used in combination to address subpockets. Finally, additional examples of approaches that consider the relative importance of small-molecule co-dependence of protein conformations are highlighted to emphasize an increased significance of subpockets, especially at protein interfaces.

Detection of Binding Site Molecular Interaction Field Similarities

Protein binding-site similarity detection methods can be used to predict protein function and understand molecular recognition, as a tool in drug design for drug repurposing and polypharmacology, and for the prediction of the molecular determinants of drug toxicity. Here, we present IsoMIF, a method able to identify binding site molecular interaction field similarities across protein families. IsoMIF utilizes six chemical probes and the detection of subgraph isomorphisms to identify geometrically and chemically equivalent sections of protein cavity pairs. The method is validated using six distinct data sets, four of those previously used in the validation of other methods. The mean area under the receiver operator curve (AUC) obtained across data sets for IsoMIF is higher than those of other methods. Furthermore, while IsoMIF obtains consistently high AUC values across data sets, other methods perform more erratically across data sets. IsoMIF can be used to predict function from structure, to detect potential cross-reactivity or polypharmacology targets, and to help suggest bioisosteric replacements to known binding molecules. Given that IsoMIF detects spatial patterns of molecular interaction field similarities, its predictions are directly related to pharmacophores and may be readily translated into modeling decisions in structure-based drug design. IsoMIF may in principle detect similar binding sites with distinct amino acid arrangements that lead to equivalent interactions within the cavity. The source code to calculate and visualize MIFs and MIF similarities are freely available.

PockDrug-Server: a new web server for predicting pocket druggability on holo and apo proteins

Predicting protein pocket's ability to bind drug-like molecules with high affinity, i.e. druggability, is of major interest in the target identification phase of drug discovery. Therefore, pocket druggability investigations represent a key step of compound clinical progression projects. Currently computational druggability prediction models are attached to one unique pocket estimation method despite pocket estimation uncertainties. In this paper, we propose ‘PockDrug-Server’ to predict pocket druggability, efficient on both (i) estimated pockets guided by the ligand proximity (extracted by proximity to a ligand from a holo protein structure) and (ii) estimated pockets based solely on protein structure information (based on amino atoms that form the surface of potential binding cavities). PockDrug-Server provides consistent druggability results using different pocket estimation methods. It is robust with respect to pocket boundary and estimation uncertainties, thus efficient using apo pockets that are challenging to estimate. It clearly distinguishes druggable from less druggable pockets using different estimation methods and outperformed recent druggability models for apo pockets. It can be carried out from one or a set of apo/holo proteins using different pocket estimation methods proposed by our web server or from any pocket previously estimated by the user. PockDrug-Server is publicly available at: http://pockdrug.rpbs.univ-paris-diderot.fr.

PockDrug: A Model for Predicting Pocket Druggability That Overcomes Pocket Estimation Uncertainties

Predicting protein druggability is a key interest in the target identification phase of drug discovery. Here, we assess the pocket estimation methods' influence on druggability predictions by comparing statistical models constructed from pockets estimated using different pocket estimation methods: a proximity of either 4 or 5.5 Å to a cocrystallized ligand or DoGSite and fpocket estimation methods. We developed PockDrug, a robust pocket druggability model that copes with uncertainties in pocket boundaries. It is based on a linear discriminant analysis from a pool of 52 descriptors combined with a selection of the most stable and efficient models using different pocket estimation methods. PockDrug retains the best combinations of three pocket properties which impact druggability: geometry, hydrophobicity, and aromaticity. It results in an average accuracy of 87.9% ± 4.7% using a test set and exhibits higher accuracy (∼5-10%) than previous studies that used an identical apo set. In conclusion, this study confirms the influence of pocket estimation on pocket druggability prediction and proposes PockDrug as a new model that overcomes pocket estimation variability.