id
stringlengths 24
24
| idx
int64 0
402
| paragraph
stringlengths 106
17.2k
|
|---|---|---|
66f5321812ff75c3a18553fd
| 21 |
A fresh strategy in multi-target drug development, polypharmacology-regulating numerous targets with one or more drugs ( ) has attracted significant attention. This method mainly targets specific targets that particular medications can reasonably target ( ). Thanks to substantial developments in the field of in silico pharmacology, several practical approaches for exploring multi-target medicines have emerged. Among these methods are pharmacophore modelling, network pharmacology, molecular docking, multi-omics-based system biology, Multi-target Quantitative Structure-Activity Relationship (mt-QSAR), machine learning, and perturbation model combined with machine learning (PLMT ( )). Thus, it is expected that including different biological data and improving machine learning methods will help to optimise multi-target drug prediction models.
|
66f5321812ff75c3a18553fd
| 22 |
Conducting late-stage clinical trials is a time-consuming and expensive process that spans several years and costs millions ( )). Therefore, developing, validating, and utilising predictive models earlier would be most advantageous, using preclinical and earlystage clinical trial data. Translational biomarkers can be forecasted by employing ma-chine learning techniques on preclinical datasets. Once the model and its accompanying biomarker have been validated using separate data sets from preclinical or clinical studies, they can be used to categorise patients, determine possible uses for medicine, and propose how the treatment works ([28])). For example, ML model used in COVID-19 to identify biomarkers ( )) (Figure ) Figure : The representation of ML application in biomarker identification
|
66f5321812ff75c3a18553fd
| 23 |
The plasma proteome of patients with septic acute kidney injury (AKI) and COVID-19 was collected and examined at a different location. The data was retrieved and subjected to further analysis, which yielded datasets specific to each disease. The method commences by creating a BINN (Binary Neural Network) for each dataset. This is achieved by extracting a subset of the route database, such as Reactome, based on the proteomic content of the dataset in question. The extracted data is then organised into a sequential structure resembling a neural network. The protein amounts of each sample are utilised to train the corresponding BINNs to distinguish between two subphenotypes. Subsequently, the networks are analysed using SHAP, which provides feature importance values that can be used to identify biomarkers and do route analysis ( )).
|
66f5321812ff75c3a18553fd
| 24 |
IntelliGenes, another ML application in biomarker identification, is a revolutionary ML pipeline for multi-genomics biomarker research with excellent accuracy for illness prediction ( ). Using multi-genomic, clinical, and demographic data, this unique approach combines statistical methods with ML algorithms ( ). At the heart of Intelli-Genes is the Intelligent Gene (I-Gene) score, a new metric that assesses biomarkers' ability to predict complex features. IntelliGenes uses Recursive Feature Elimination, Random Forest, Support Vector Machine, XGBoost, k-Nearest Neighbours, Multi-Layer Perceptron, and soft and hard voting classifiers with Pearson correlation, Chi-square test, and ANOVA. IntelliGenes can find disease-associated biomarkers that traditional approaches miss with this combo. The method starts with AI/ML-ready Clinically Integrated Ge-nomics and Transcriptomics (CIGT) data, including patient age, gender, race, ethnicity, diagnoses, and RNA-seq gene expression data. IntelliGenes ranks the top percentage of multi-genomic data-based profiles to predict diagnoses accurately. Also, IntelliGenes' personalised early diagnosis of common and unusual diseases is a significant benefit. Biomarkers' relevance in disease prediction is categorised by the novel I-Gene score. The I-Gene score measures biomarkers' weighted utility and biological system expression using SHAP and HHI. This directionality lets researchers establish whether biomarker overexpression or underexpression causes disease, revealing disease processes. Intel-liGenes' versatility and adaptability allow it to be used on personal devices and highperformance computing platforms ( ). This makes it an effective tool for research and clinical applications, enabling personalised interventions and novel therapy targets ( ).
|
66f5321812ff75c3a18553fd
| 25 |
Computational pathology predictions are a breakthrough approach in pathology using advanced algorithms and machine-learning techniques to increase diagnosis accuracy and efficiency. These computer methods use vast histological data to produce reliable, consistent results above accepted manual interpretations. For instance, Lee et al. have demonstrated that using computational techniques to analyse the surrounding non-cancerous tissue in prostate cancer can yield important information on the chance of cancer returning ( ). In oral malignancies and early-stage oestrogen receptor-positive breast tumours, Lu et al. showed a strong association between survival rates and characteristics defining the shape ( ) and orientation of the nucleus ( ). Such a combination of computational prediction in pathology reduces human error and accelerates the diagnosis process, thereby improving patient outcomes. The continuous evolution of computational pathology has significant power to change tailored treatment and enable the discovery of fresh biomarkers. The combination of technology and pathology (Figure ) can transform healthcare by improving diagnosis accuracy, allowing targeted therapies, and raising the general quality of medical treatment. The graphic shows a thorough, pathologist-centered medical system combining many data sources for improved diagnosis and patient care. Aggregated into the Laboratory Information System (LIS), clinical data from electronic health records (EHR), omics data from molecular pathology, whole slide images (WSIs) from digital pathology, and clinical laboratory results are included. This integration is the foundation of "Algorithm 1," a diagnostic tool synthesising several kinds of data to raise diagnostic accuracy. Updated disease-related data gathered during patient follow-up over time is merged with the original dataset to develop "Algorithm 2." Through constant improvement of diagnostic and prognostic capacities made possible by this iterative process, more exact and individualised patient treatment finally results ( ).
|
66f5321812ff75c3a18553fd
| 26 |
Forecasting the three-dimensional configuration of a particular protein is challenging due to the vast array of possible conformations that the amino acid sequence could adopt ( ). However, AlphaFold ( ) initiates its computational exploration for the probable structure by utilising a template, a pre-existing structure for similar proteins. Alternatively, AlphaFold can utilise data on the biological evolution of amino-acid sequences within the protein family, which consists of proteins with similar activities and presumably possess identical folds ( ). The information used in AlphaFold is valuable because consistent associated evolutionary changes in pairs of amino acids suggest that these amino acids have a direct interaction despite being located far apart in the sequence ( ). Information of this nature can be derived from protein families' multiple sequence alignments (MSAs), obtained by analysing the evolutionary differences in sequences across different biological species. Nevertheless, the dependence on MSAs is limiting due to the unavailability of evolutionary knowledge for all proteins ( ). 13
|
66f5321812ff75c3a18553fd
| 27 |
Discovering and developing medications, regulatory review, and authorisation depend critically on pharmacokinetic (PK) and pharmacokinetic/pharmacodynamic (PK/PD) models. The design of clinical trials, data interpretation, and dose ( ) computation depends critically on pharmacokinetic (PK) and pharmacokinetic/pharmacodynamic (PK/PD) models. Applying machine learning technology for PK and PK/PD prediction offers not only better results but also faster forecasts for this critical chore (Figure ). A drug-drug interaction (DDI) is the condition wherein another drug ingested or supplied concurrently modulates the way a drug is metabolised in the body (pharmacokinetics) or its effects on the body (pharmacodynamics) ( ). Drug-drug interactions (DDIs) can cause products to be removed from the market; for example, astemizole, a medication used to treat allergy symptoms, was taken off the market because it can cause the QT interval to become longer and lead to irregular heart rhythms when combined with substances that inhibit cytochrome P450 3A4 (CYP3A4), such grapefruit juice and erythromycin ( ). Taken off the market due to a sluggish heart rate and muscle breakdown when used with some cardiovascular drugs such as beta-blockers or statins, Mibefradil was used to treat excessive blood pressure and chronic chest discomfort ( ). Thus, after time-consuming and expensive drug research and development phases, it is imperative to effectively ascertain drug-drug interactions to prevent withdrawals by employing promising ML models subsequently. For example, a deep learning framework called Deepdrug has investigated drug-drug interaction in Figure ). Besides Deepdrug Figure ( ), several machine-learning methods have been created to forecast drug-drug interactions (DDIs) to address the shortage of known DDI combinations ( ). Prior research has yielded numerous models for predicting the occurrence or non-occurrence of drug-drug interactions (DDIs). These models have successfully identified DDI combinations that result in adverse effects and categorised the various forms of DDIs by utilising publicly available databases. However, the models showed limited success in identifying DDI combinations ( ).
|
66f5321812ff75c3a18553fd
| 28 |
These limitations often stem from incomplete data, variability in patient responses, and the complexity of accurately modelling the factors influencing DDIs ( ). Also, most models have solely offered a fundamental forecast for the presence or categorization of DIs ( ). These models do not help make complex therapeutic decisions, such as accurately adjusting the dosage or selecting alternative drugs. Physicians and chemists require predictions on the fold change of PK parameters. However, thus far, no models have successfully forecasted this. Also, there is currently no available dataset that has been methodically developed to contain true-negative instances exclusively. For example, the DDI DB, such as DrugBank, is commonly utilised for predicting drug-drug interactions (DDIs). It provides information indicating the presence of a DDI between drugs A and B but does not include information indicating the absence of a DDI. Therefore, researchers have unavoidably chosen random sets of medication combinations under the assumption that there were no interactions ( ). Indeed, the lack of proof does not serve as evidence for the absence of something. Utilising a model lacking in quality input poses challenges in acquiring dependable results. When the negative set is randomly generated, it becomes challenging to pinpoint the specific reason for unanticipated problematic output ( ). As a result, there is a continuous need for more sophisticated and robust ML models to better account for these variables and improve the reliability of DDI predictions.
|
66f5321812ff75c3a18553fd
| 29 |
Clinical research includes both interventional and observational studies that utilise volunteers to assess the efficacy and safety of novel drugs, treatments, or medical devices ( ). The design of clinical trials is a crucial element in interventional studies, as it enables a thorough evaluation of the effectiveness and safety of therapies ( ). Clini-calTrials.gov houses more than 400,000 records of clinical studies conducted in 220 countries. The sponsor or primary investigator of the clinical research is responsible for registering the studies on the website. In 2022, 14.5% of interventional trials listed in the clinicaltrials.gov registry ended prematurely ( ).
|
66f5321812ff75c3a18553fd
| 30 |
The usual methods to replicate clinical trials primarily relied on deterministic models and simple probabilistic approaches to replicate trial circumstances and evaluate prospective outcomes ( ). More complex models that consider patient demographics, illness development, and therapeutic effects can be produced as computer capability develops (Figure ). By predicting anticipated difficulties and improving trial procedures, clinical trial simulations have helped to raise the ethical standards and efficiency of trial design. These approaches created the basis for the present advanced simulations, which today incorporate sophisticated algorithms and machine learning techniques to produce even more exact and dynamic forecasts ( ), even if they have some constraints. Model A suggests using in-brace corrections for patients in group B, while Model B predicts the impact of these suggestions. The predicted results were compared to the actual outcomes recorded in the charts of patients in group B ( ).
|
66f5321812ff75c3a18553fd
| 31 |
Multiple studies in the literature have concentrated on utilising machine learning algorithms to forecast the success of clinical trials ( ). These algorithms are based on structured and unstructured information obtained from clinical studies. These investigations aim to predict the premature termination of clinical trials by utilising trial characteristic data in conjunction with unstructured data. Follett et al. utilised both structured and unstructured data to forecast the discontinuation of clinical trials ( ). The study employed text mining techniques to extract vector features from the studies' description fields and the studies' fundamental aspects. A dataset of 130,000 investigations was compiled by selecting completed and terminated trials using accessible information before the study commenced. Subsequently, they employed machine learning methodologies to forecast results ( ). Also, The study done by Elkin et al. ( ) employed machine learning techniques to predict the termination of clinical trials based on data taken from clinicalTrials.gov ( ). The study produced a dataset consisting of 68,999 studies. A total of 640 features were created by employing document embedding, keyword features, and trial characteristics. Their study aimed to comprehensively understand clinical trial termination and produce accurate prediction outcomes ( ). Machine learning (ML) in clinical trial simulations demonstrates its potential to enhance performance by reducing the time required.
|
66f5321812ff75c3a18553fd
| 32 |
In pharmacological research and development, applying machine learning (ML) in drug delivery simulations marks a radical change ( ). Often, depending on empirical data and crude computational models, conventional drug delivery systems can be time-consuming and less exact. Still, machine learning offers a vital substitute by using large databases to predict drug behaviour, streamline delivery systems, and raise treatment outcome accuracy ( ). Using cutting-edge algorithms and data-driven insights, MLdriven drug delivery simulations faithfully replicate complex biological systems, identify the most efficient delivery paths, and project tailored patient reactions. This approach not only accelerates the development of new pharmaceuticals but also provides opportunities for tailored medicine, ensuring that treatments are adjusted to fit every patient's particular requirements and produce the best possible effectiveness and the fewest possible side effects. The advent of machine learning (ML) in medication delivery models has the power to revolutionise the whole industry. Increasing speed, efficiency, and accuracy above past capabilities ( ) can improve drug development (Figure ). The pharmaceutical industry is increasingly adopting AI techniques, including machine learning, to enhance the effectiveness of drugs in terms of their bioavailability, stability, and ability to target specific organs. Computational pharmaceutics transforms medication delivery paradigms by utilising big data and multiscale modelling techniques. Machine learning allows for identifying relevant characteristics, the discovery of unexpected patterns, and immediately adjusting to pathogen evolution, hence allowing effective treatment development, selection, and dosage optimisation. Machine learning integrated into clinical devices can significantly transform clinical decision-making in anti-infective drug delivery strategies, thanks to its impressive speed and practicality ( ).
|
66f5321812ff75c3a18553fd
| 33 |
One of the main advantages of applying artificial intelligence in medical delivery systems is its capacity to handle and evaluate large amounts of data effectively. Artificial intelligence algorithms can use large amounts of data to identify trends in medication formulations that might not be readily apparent using more traditional approaches ( ). Artificial intelligence methods can examine data-driven sources such as chemical structures, physicochemical properties, and biological activity data to expose hidden relationships. Furthermore, enabling the forecast of medication durability in different storage conditions, the solubility of pharmaceuticals in various formulations, and the release of kinetics from other delivery systems is the ability of machine learning models to grasp complex connections and non-linear interdependencies ( ). Using pattern and trend analysis, AI can thus enhance medicine delivery systems, influencing ) greater effectiveness, safety, stability, solubility, and release kinetics.
|
66f5321812ff75c3a18553fd
| 34 |
Although machine learning offers excellent possibilities to change drug discovery processes, it is not free from challenges and restrictions. The accessibility and quality of data create a significant challenge since the effectiveness of machine learning models depends much on large, diversified, carefully kept datasets ( ). Furthermore, a considerable challenge of machine learning models is their comprehensibility since complex algorithms usually produce predictions without explaining underlying mechanisms. Furthermore, the opaque character of many machine learning models could hinder regulatory approval and acceptance inside the pharmaceutical company ( ). Moreover, the ability of machine learning models to be applied to various datasets and experimental environments is still tricky since models trained on particular datasets could not attain optimal performance on fresh, undiscovered data ( ). Dealing with these challenges calls for cooperation between several fields, advances data organisation and standardising, and generates interpretable and transferable machine learning models specially tailored for the complex character of drug development.
|
66f5321812ff75c3a18553fd
| 35 |
Over the last ten years, there have been significant advancements in the field of cancer biology. These include major breakthroughs in genomics technology, which have allowed scientists to obtain DNA sequences of thousands of tumours. Additionally, researchers have been able to analyse the gene expression profiles of tumours at the individual cell level. Furthermore, the functional properties of genes have been studied through comprehensive CRISPR screens of the entire genome, leading to the ident. Regrettably, many therapeutic targets arising from these advancements have been classified as "undruggable" based on traditional criteria ( ). These targets often include extensive protein-protein interaction surfaces that do not have readily identifiable pockets for binding ligands. However, the drug discovery community has shown a growing commitment to the challenge of developing drugs for targets that were previously considered impossible to drug. As a result, researchers have focused on developing new methods to target previously considered "undruggable" targets, such as KRAS ( ). The targeting of KRAS with drugs aims to analyse the variables that led to the development of innovative approaches to address this challenging and difficult-to-drug target ( ). Another successful example of targeting "undruggable" targets is PROTAC ( ). PROTAC can degrade the protein of interest using the ubiquitin mechanism in the cell ( ). The last successful example is that Prodrugs have been used in the targeting of "undruggable" targets ).
|
66f5321812ff75c3a18553fd
| 36 |
KRAS is an oncogene that is commonly mutated in human cancer, with mutations occurring in around 30% of all tumours ( ). This makes it a significant focus for drug development. However, progress in this area has been limited because there is no apparent binding pocket, and the nucleotide-binding affinity is considerable. At first, the tactics were centred on posttranslational alterations. However, these strategies faced challenges due to the wide range of substances that the enzymes involved may act upon. In 2013, Kevan Shokat made a significant discovery by devising a technique to target the mutant form of KRASG12C by covalent bonding specifically. This method involved adopting a cysteine-based tethering strategy, which led to the identification of a hidden binding pocket in the switch II domain. The study shows that KRASG12C may be selectively targeted by small compounds that specifically attach to its GDP-bound conformation, thereby offering a novel treatment strategy. Five distinct KRASG12C inhibitors are being tested in clinical trials, namely AMG-510 and MRTX849, which are now in Phase II investigations ).
|
66f5321812ff75c3a18553fd
| 37 |
In cancer treatment, prodrugs like PRIMA-1 and its derivative PRIMA-1MET present a viable method for attacking "undruggable" targets such mutant p53. These molecules restore the active conformation of mutant p53, therefore producing sequence-specific DNA binding and cellular death. With their active metabolite methylene quinuclidine alkylating p53 and re-establishing an oxidative environment inside tumour cells, PRIMA-1 and PRIMA-1MET, most importantly, behave as prodrugs. Small compounds such as nutlins also help to solve the "undruggable" target problem by blocking the interaction between p53 and MDM2, a negative regulating protein that drives p53 breakdown. Nutlins offer great potential in preclinical and clinical contexts and help to stabilise p53, therefore keeping its tumour-suppressing properties. These creative approaches draw attention to the possibilities for efficiently modifying difficult cancer targets by the use of prodrugs and tailored compounds ([132]).
|
66f5321812ff75c3a18553fd
| 38 |
Although KRAS G12C has been successful, numerous recurring oncoproteins and tumour suppressors, including MYC, β-catenin, and p53, have not been targeted with drugs. Additionally, resistance to single-drug treatments remains a significant obstacle. Recent technological advancements and novel strategies have started to tackle these obstacles, converting previously difficult-to-treat targets into ones that can be targeted with drugs. Some examples of these advancements include kinase inhibitors, antiapoptotic BCL2 proteins, and transcription factors. Innovations have dramatically influenced the growth in this field in structure-based design, chemically induced protein degradation, and targeted covalent inhibition ).
|
66f5321812ff75c3a18553fd
| 39 |
The further development of proximity-based platforms will significantly influence the identification of cancer drugs. Proximity-based platforms rely on the principle that molecules or molecular entities, such as proteins or nucleic acids in close physical proximity within a cellular or tissue environment, can interact or affect each other's function, even without direct binding. For example, building upon Shokat's research on KRASG12C using proximity-based platforms, the use of chemoproteomic methods, specifically activitybased protein profiling, has emerged as a practical approach for identifying covalent ligands and potential binding sites for proteins that are difficult to target using traditional drug-discovery methods ( ). Also, Nomura and colleagues recently performed a cysteine-reactive covalent ligand screen to discover chemicals that can interfere with the binding of MYC to its DNA consensus sequence ( . On this screen, EN4 was found as a covalent ligand that specifically targets Cys-171 of MYC within a predicted intrinsically disordered area. Subsequently, this molecule was further refined to establish the first structure-activity relationships. The potential exists for ligandable pockets in numerous traditionally 'undruggable' targets to be discovered by employing mass spectrometry techniques for covalent screening, particularly within the natural cellular environment. This approach could also enable the targeting of intrinsically disordered regions of proteins, which are commonly found in transcription factors ). Discovered by Ebert's team, BI-3802 and CR8 both represent small molecule degraders aiming against BCL6 and CDK, respectively. These compounds enhance already existing protein-protein interactions, hence promoting breakdown. The classic molecular glue, cyclosporin, hooks cyclophilin to block calcineurin, hence modulating IL2 synthesis. Companies such as Revolution Medicines and WarpDrive Bio apply similar ideas for new therapies like KRASG12C inhibitors. By extending the range of druggable targets, these strategies should transform drug development ( ).
|
66f5321812ff75c3a18553fd
| 40 |
Small-molecule treatments have attracted interest in not just targeting proteins but also in RNA regulation. For targets in difficult-to-treat cancer especially, this offers fresh opportunities. The promise of RNA targeting is shown by the effectiveness of PCSK9 inhibition with R-IMPP, a small molecule that binds especially to ribosomes. In clinical studies, experimental data on RNA-targeted drugs, including linezolid and ribocil, has shown their efficacy. This implies that RNAs causing diseases can offer particular sites where medications might bind. As our knowledge of structure develops, we could be able to apply several strategies of action, including translational stalling and splicing control.
|
66f5321812ff75c3a18553fd
| 41 |
To tackle drug resistance in targeted treatments, it is necessary to comprehend both genetic and nongenetic causes. These mechanisms include mutations in the gene targeted by the medication and the adaptive reprogramming of signalling pathways. Novel pharmaceuticals, such as combination therapy and bivalent compounds, have been created to counteract resistance mutations in different cancer-causing sites, prolonging treatment effectiveness. Nevertheless, resistance to specific protein degraders such as PROTACs and molecular glues may develop. This would need the implementation of techniques to counteract resistance caused by mutations. Furthermore, resistance to drugs that do not involve mutations, such as adaptive reprogramming and cell plasticity, poses difficulties but also provides potential for rational medication combinations and treatments that target epigenetic changes to avoid or overcome resistance. Small compounds that target transcription factors are a valuable addition to the range of tools available to combat drug resistance in cancer treatment ( ).
|
66f5321812ff75c3a18553fd
| 42 |
Mainly by the use of genomics technology, single-cell gene expression analysis, and thorough CRISpen screens, significant advances in cancer biology over the past ten years have revolutionised medication development. Many therapeutic targets these developments found were first judged "undruggable" because of sizeable protein-protein interaction surfaces devoid of clear ligand-binding pockets. Still, the scientific community has embraced creating strategies to target these challenging proteins. Among the notable achievements are creating medicines aimed at KRASG12C, covalent bonding techniques to take advantage of latent binding sites, and PROTAC technology-which uses the ubiquitin-proteasome system to break down target proteins. These days, new trends in drug development centre on creative lead-finding tools and fresh drug modalities. For typically "undruggable" proteins, proximity-based platforms, including chemoproteomic techniques, have evolved as valuable tools for locating covalent ligands and possible binding sites. Furthermore, the idea of proximity induction-as shown by IMiDs and PROTACs-has created fresh paths for inducing protein breakdown and thereby targeting less ligandable proteins. Targeting RNA also presents fresh opportunities for challenging-to-treat cancer targets since it seems a complementary approach to protein targeting. These developments are transforming the field of drug discovery and broadening the range of druggable targets when combined with attempts to overcome drug resistance with fresh drugs and combination therapies.
|
66f5321812ff75c3a18553fd
| 43 |
Integrating machine learning (ML) into drug discovery represents a significant advancement, transforming traditional approaches and accelerating the development of new therapeutics. This review has highlighted the diverse ML methodologies currently employed in the field, including supervised learning, neural networks, and reinforcement learning, and their applications across various stages of drug discovery. By leveraging ML, researchers can enhance predictive accuracy, streamline drug design processes, and optimize clinical trials, ultimately leading to more efficient and effective drug development.
|
66f5321812ff75c3a18553fd
| 44 |
Despite the promising advancements, several challenges remain, including the need for high-quality data, model interpretability, and integration with existing workflows. Addressing these challenges will be crucial for realizing ML's full potential in drug discovery. Future developments in ML techniques and their applications promise to further improve the drug discovery pipeline, offering new opportunities for innovation and discovery.
|
6319d1fb5351a324a7f31463
| 0 |
Activating mutations within the epidermal growth factor receptor (EGFR) kinase domain, most prevalently L858R and exon19del, are common causes of non-small cell lung cancer (NSCLC) and often serve as predictive markers for the selection of EGFR tyrosine kinase inhibitors (TKIs) as effective targeted therapies. Prolonged efficacy of first-generation TKIs (gefitinib and erlotinib) is eventually made limited due to drug resistance as a result of patients acquiring a second T790M "gatekeeper" mutation. To produce a viable treatment option for T790M-positive NSCLC tumors, drug development efforts have yielded the clinically-approved drug AZD9291 (osimertinib), which is selective for T790M-containing EGFR and made potent by forming an irreversible covalent bond to C797. Despite promising indications, drug resistance to osimertinib is inevitable and caused in part by the acquisition of a third kinase domain mutation C797S that prevents formation of the potency-enabling covalent bond. More recently, osimertinib has been shown effective, and clinically-approved, as a front-line therapy in untreated patients harboring EGFR L858R and exon19del activating mutations. 6 As osimertinib is the only approved third-generation EGFR TKI for L858R and exon19del EGFR mutant tumors, development efforts from Yuhan and Janssen biotech sought to produce a drug with improved medicinal chemistry properties. These efforts resulted in YH25448 (lazertinib), which is structurally related to osimertinib comprising an aminopyrimidine core and acrylamide warhead but is distinct with respect to the substituted pyrazole as well as morpholine groups (Figure ). Preclinical head-to-head or in combination with the antibody Amivantamab (NCT04077463), and is currently approved to treat T790M-containing NSCLC in the Republic of Korea. Regardless of these improvements, drug resistance to lazertinib has been shown to be due to the acquisition of C797S mutation. 9-10 Despite the improved properties and positive clinical outlook, no crystal structures have been reported detailing the molecular basis for lazertinib inhibition of mutant EGFR.
|
6319d1fb5351a324a7f31463
| 1 |
Results and Discussion. To characterize the binding mode of lazertinib in complex with the EGFR kinase domain, we determined a 2.4 Å resolution X-ray co-crystal structure of lazertinib soaked into WT EGFR crystals (Figure & S1, PDB ID 7UKV). WT kinase domains crystallize in the active "ɑC-helix in" conformation due to crystal packing of the kinase domains as asymmetric dimers. As expected, lazertinib binds with the aminopyrimidine anchored to the hinge region by H-bonds to M793 and covalent bond formed at C797 as generally observed for third-generation TKIs. Importantly, the unique pyrazole group extends away from the hinge in a conformation that positions the phenyl ring toward the K745-E762 salt bridge and the N,N-dimethylmethyleneamine (methyleneamine) in a H-bond with the DFG-motif D855 carboxylate. The substituted pyrazole of lazertinib is unique among third-generation EGFR TKIs and most likely the basis for mutant-selectivity and improved medicinal chemistry properties. A comparison of our structures from the "ɑC-helix in" active WT (Figure ) and "ɑC-helix out" inactive T790M (Figure ) show very similar binding positions indicating that lazertinib is anchored to the EGFR kinase domain identically in both active and inactive states and independent of the Thr versus Met 790 gatekeeper residue (Figure ). We expect that this conformation of lazertinib is preferred as compared to the "flipped" conformation (Figure ) in the "ɑC-helix in" active state due to the requirement to anchor the positive methyleneamine near the K745-E762 salt bridge (Figure ).
|
6319d1fb5351a324a7f31463
| 2 |
showcasing that both inhibitors are made selective for T790M-containing EGFR through van der Waals interactions (Figure ). Additionally, versatile H-bonding methyleneamine moiety of lazertinib is distinctive potentially enabling the modest improvement in binding compared to osimertinib. Another informative comparison is the structurally-related imidazole-based covalent inhibitor LN2057 (Figure ) that forms an H-bond with K745 enabling C797S mutant inhibition at the expense of a loss of selectivity, i.e., enhanced binding to WT EGFR. By comparison, the lazertinib pyrazole substituents sterically block K745 potentially diminishing binding to WT EGFR (Figure ). The correlation of these differences in binding mode demonstrates how the pyrazole moiety of lazertinib affords distinct interactions with EGFR to enable T790M mutant selectivity.
|
6319d1fb5351a324a7f31463
| 3 |
To further understand the functional significance of the lazertinib binding mode, we conducted biochemical assays with purified kinase domains. The three inhibitors potently inhibit EGFR mutants L858R and L858R/T790M as consistent with previous studies. Their strong potency for the T790M-containing variant is most likely due to productive binding to the T790M through intermolecular van der Waals interactions (lazertinib and osimertinib, Figure ) or methionine pi-stacking (LN2057, Figure ). LN2057 is found more potent against WT EGFR compared to osimertinib and lazertinib, most likely due to added binding affinity afforded by the imidazole-K745 H-bond (Figure ). It is also likely that enhanced binding from this H-bond direct for the observed higher potency against L858R. To compliment earlier studies, we assessed biochemical potencies for these three inhibitors against HER2. We observe that lazertinib exhibits significantly lower potency against HER2, which is proposed to limit adverse events and improve drug tolerability. For completeness, we confirmed lazertinib forms a covalent bond at Cys-805 within the HER2 purified kinase domain with LC-MS/MS (Figure ). These trends in biochemical assays showcase how structural differences between these inhibitors elicit differential effects on inhibitor potency and selectivity. Since enhanced efficacy against EGFR and diminished HER2 targeting is a proposed advantage for treatment of mutant EGFR NSCLC with lazertinib, 8 we were motivated to assess lazertinib inhibition in cellular contexts compared to osimertinib and LN2057. We first assessed dose-dependence of inhibition of EGFR(L858R/T790M) in H1975 NSCLC cells by osimertinib, lazertinib, and LN2057 by blotting for active EGFR (pY1068). After dosing 5 or 50 nM of these drugs for 2 hours, we observed that lazertinib suppressed pY1068 to a greater extent compared to equivalent dosing of osimertinib and LN2057, as consistent with previous studies (Figure ). Uniquely, lazertinib is observed to be notably less effective at inhibiting active HER2 (pY1221/1222) in BT474 HER2 overexpressing breast cancer cells, confirming the selective targeting of EGFR by lazertinib compared to osimertinib and LN2057 (Figure ). These findings are consistent with our biochemical activity assays (Table ) indicating that lazertinib potently and selectively inhibits mutant EGFR(L858R/T790M) and WT EGFR while simultaneously affording limited activity against HER2.
|
6319d1fb5351a324a7f31463
| 4 |
In conclusion, we have determined the molecular basis of the novel EGFR TKI lazertinib bound to EGFR in X-ray co-crystal structures showcasing that the lazertinib pyrazole ring binds facilitates H-bonds and van der Waals interactions consistent with drug efficacy and T790M selectivity. Structural and functional correlation to osimertinib and LN2057 demonstrate the importance of productive intermolecular interactions with T790M. Additionally, we find that lazertinib does not H-bond with K745, which likely contributes to lower potency for WT EGFR. Another important feature to lazertinib is the lack of potency on HER2 that is often associated with dose-limiting adverse events, as confirmed here in biochemical and cell-based studies. Our present structural analysis, however, does not reveal a discrete molecular origin for the preference of lazertinib for EGFR when compared to HER2. We speculate that differences in the sequence and dynamics of the HER2 kinase domain, as evident from reported crystal structures, negatively impacts lazertinib reversible binding to HER2 and not osimertinib and LN2057
|
6319d1fb5351a324a7f31463
| 5 |
and motivates future structural studies to understand the structural basis for EGFR kinase specificity. Results from these studies define the binding mode of a novel third-generation mutant-selective EGFR TKI lazertinib with improved dose-limiting toxicity as well as ontarget potency and selectivity and serves as a noteworthy example for developing nextgeneration kinase inhibitors.
|
646b7381ccabde9f6e2e8cf8
| 0 |
Intensive research efforts are currently underway to identify novel carbon allotropes with the goal of achieving specific mechanical and thermal properties approaching those of diamond. For this purpose, modern materials research programs based on evolutionary crystallography are used to predict novel allotropes . Regarding the carbon database, SACADA is recognized as a reliable one available to all researchers. The novel allotropes are subjected to topology analysis using TopCryst software and then published in the Cambridge Structural Database (CCDC).
|
646b7381ccabde9f6e2e8cf8
| 1 |
To illustrate this approach, and keeping in mind the C4 tetrahedron building block of diamond, a simple arrangement of 8 C4 tetrahedra at a cube corner and center is shown in Fig. , with both ball-and-stick and tetrahedral representations exhibiting corner-sharing tetrahedra. The resulting body-centered tetragonal (bct) structure C 4 submitted to quantum mechanical calculation within the Density Functional Theory (DFT) to obtain the ground state energy and related properties was found to be another formulation of diamond, and the allotrope is qualified with dia topology (Table ) . Using bct C 4 as a template, we devised novel simple allotropes, such as the latest one: hybrid (sp 3 /sp 2 ) C 6 with tfa topology, which we called "neoglitter" (cf. and the works cited therein).
|
646b7381ccabde9f6e2e8cf8
| 2 |
Within a similar approach, we propose herein two novel carbon allotropes, tetragonal C 9 and C 12 , both characterized by sp 3 carbon in original architectures comprising corner and edge sharing tetrahedra. The purpose of this paper is to present their crystal structures and to demonstrate ultrahardness, while also presenting high phonon optical mode frequencies of > 40 THz like in diamond and C 4 , a large electronic band gap, and a temperature evolution of the heat capacity close to that of diamond.
|
646b7381ccabde9f6e2e8cf8
| 3 |
The computational methodology was developed in previous works . Essentially, the search for the ground state structures with minimal energies was systematically performed using unconstrained geometry optimizations within the DFT-based Vienna Ab initio Simulation Package (VASP) using the projector augmented wave (PAW) method for potentials . The DFT exchange-correlation (XC) effects were treated using a generalized gradient approximation (GGA) . The plane wave energy cut-off was 500 eV. The study of the mechanical properties was based on the calculations of the elastic properties, which were determined by performing finite distortions of the lattice and deriving the elastic constants from the strain-stress relationship. The calculated elastic constants C ij are then used to obtain the bulk (B) and the shear (G) moduli with the Voigt averaging method based on uniform strain. In addition to the mechanical properties, the dynamic stability was determined from the phonons, which are described as vibrational quanta.
|
646b7381ccabde9f6e2e8cf8
| 4 |
They are represented by phonon band structures obtained using the interface code "Phonopy" based on the Python language . The representations of the structures and charge density projections were obtained with the graphical program VESTA . For the evaluation of the electronic properties the band structures were obtained with the all-electrons augmented spherical wave method ASW .
|
646b7381ccabde9f6e2e8cf8
| 5 |
The crystal parameters of the template bct C 4 (I-4m2, No. 119) are given in Table . We adopt an original method to introduce novel allotropes. The protocol consists of removing the central carbon, resulting in a simple tetragonal C 3 (P-4m2, No. 115). The central void can then accommodate the inserted carbon atoms in an enlarged cell.
|
646b7381ccabde9f6e2e8cf8
| 6 |
The insertion of two sets of four-fold (4i) and (4j) occupations in the same space group leads to the C 9 stoichiometry. The geometry optimization down to the ground state energies through successive calculation cycles with increasing precision of the tetragonal Brillouin zone led to the lattice parameters and energies given in Table , 2 nd column. In Fig. , the structure of C 9 in both the ball-and-stick and tetrahedral representations shows similar, though larger, tetragonal C 4 with tetrahedra at corner site (shown as brown and white spheres) at the 8 corners connected by C2(4j) (white spheres). C3(4i) at z = ½ forming the central square (green spheres) are connected to C2 to form two edge-sharing tetrahedra (Table ). As the original occurrence, the structure then consists of both corner and edge sharing tetrahedra. Note that the tetrahedra are significantly distorted, as can be seen from the angles.
|
646b7381ccabde9f6e2e8cf8
| 7 |
Starting from the square configuration of C 9 with C3(4i) at x,x,½ transformed to C1(4h) at x,x,0 (brown spheres), the structure was completed with C2(8i) (white spheres) and removal of the corner atoms, then subjected to full geometry optimization. The crystal data are given in Table . The structure of C 12 shown in Fig. now consists only of edge-sharing C4 tetrahedra.
|
646b7381ccabde9f6e2e8cf8
| 8 |
The CIFs (crystal information files) of two structures were submitted to topology studies performed with the TopCryst program . While C 4 used as template is designated with dia topology, C 9 was assigned with "4,4,4T6902-HZ" topology, and C 12 was identified with "sqc5532" topology. We note here that there are three allotropes with sqc topology in the SACADA database: sqc1427 (space group No. 125), No. sqc3051 (space group No. 114), and sqc6952 (space group No. 133) . The first two contain sp 2 carbon, while the third one shows only tetrahedral C(sp 3 ) with corner sharing tetrahedra. Structures with carbon squares are documented in the SCACADA database as crb topology with bct space groups (cf. No. 60-64 ). The novel C 9 and C 12 , not cataloged in databases, were subsequently deposited, curated, and then deposited in CCSD: C 9 and C 12 .
|
646b7381ccabde9f6e2e8cf8
| 9 |
From the densities, the diamond-like C 4 structure is found to have the highest density, i.e. ρ(C 4 ) = 3.50 g/cm 3 versus ρ(C 9 ) = 2.92 g/cm 3 , and ρ(C 12 ) = 2.97 g/cm 3 . The reason is that in the two novel allotropes the connection between the tetrahedra is not as regular as in diamond or C 4 .
|
646b7381ccabde9f6e2e8cf8
| 10 |
The cohesive energies follow this trend (last row of Table ), showing cohesive energies with smaller magnitudes than C 4 with E coh /atom = -2.49 eV. Note that C 12 with edge sharing tetrahedra is found to be more cohesive than C 9 with a mix of edge-and corner-sharing tetrahedra.
|
646b7381ccabde9f6e2e8cf8
| 11 |
Four modern theoretical models have been used to predict (H V ). It has been previously reported that the thermodynamic (T) model , which is based on thermodynamic properties and crystal structure, shows surprising agreement with available experimental data and is therefore recommended for hardness evaluation of superhard and ultrahard phases . The Lyakhov-Oganov (LO) model takes into account the topology of the crystal structure, the strength of covalent bonding, the degree of ionicity and directionality; however, in the case of ultrahard phases of light elements, this model gives underestimated hardness values . Empirical models, Mazhnik-Oganov (MO) and Chen-Niu (CN) , are based on elastic properties, namely bulk and shear moduli. Fracture toughness (K Ic ) was evaluated using the Mazhnik-Oganov model .
|
646b7381ccabde9f6e2e8cf8
| 12 |
Table shows the crystal parameters, density and Vickers hardness calculated from the thermodynamic model. For template C 4 , all values are close to those of diamond and lonsdaleite, as expected. The hardness of tetragonal C 9 and C 12 is lower, but in both cases it exceeds 80 GPa, which allows us to assign these allotropes to the family of ultrahard phases . We also calculated hardness of two sqc allotropes from the SACADA database, sqc3051 and sqc6952. The hardness values were found to be 51 and 67 GPa, respectively, which is a correlation with the density.
|
646b7381ccabde9f6e2e8cf8
| 13 |
Table shows the hardness values calculated using other models in addition to the thermodynamic model. As in the case of other ultrahard phases , the Lyakhov-Oganov model gives underestimated hardness values, while both empirical models do not work properly in the case of C 9 and C 12 , which are characterized by low (0.82 and 0.71) values of the Pugh's modulus ratio (G/B).
|
646b7381ccabde9f6e2e8cf8
| 14 |
Further stability criteria can be sought from the phonon dispersion relations in the Brillouin zone (BZ), i.e. with the phonon band structures. The phonon band structures obtained for the three carbon allotropes are shown in Fig. . In each panel the bands run along the main lines of the tetragonal BZ.
|
646b7381ccabde9f6e2e8cf8
| 15 |
Along the vertical direction the frequency is given in units of terahertz (THz). Since no negative energy values are observed, the three allotropes are considered to be dynamically stable. There are 3N-3 high-energy optical modes and 3 acoustic modes. The acoustic modes start from zero energy (ω = 0) at the point, BZ-center, up to a few terahertz. These modes correspond to the lattice rigid translation modes of the crystal (two transverse and one longitudinal).
|
646b7381ccabde9f6e2e8cf8
| 16 |
Like C 4 , C 9 and C 12 exhibit high-frequency bands in the range of 40 THz. Such a magnitude is a signature observed for diamond by Raman spectroscopy . It can be suggested that both novel C 9 and C 12 are expected to be dynamically close to diamond.
|
646b7381ccabde9f6e2e8cf8
| 17 |
The thermodynamic properties were calculated from the phonon frequencies using the statistical thermodynamic expressions on a high precision sampling mesh in the Brillouin zone . For all three allotropes, C 4 , C 9 and C 12 , the temperature dependencies of the heat capacity at constant volume (C V ) are shown in Fig. in comparison with experimental values for diamond . The curves corresponding to the temperature change of the entropy S are also shown. The nearly linear increase in entropy with temperature is expected as a signature of increased disorder within the lattice. The C V = f(T) curve of C 4 in Fig. exactly follows the evolution of the diamond experimental points, but a relatively small deviation in magnitude can be observed for the calculated C V = f(T) curves of C 9 and C 12 . However, a better agreement with the diamond experimental points can be observed for C 12 . It can be proposed that although the novel allotropes are structurally different from diamond, especially with respect to the stacking of tetrahedra, they remain close to it dynamically and thermally.
|
646b7381ccabde9f6e2e8cf8
| 18 |
The electronic band structures are shown in Fig. . Along the x-axis, the bands develop along the main lines of the tetragonal BZ. Along the y-axis, the zero energy is with respect to the top of the valence band (VB), i.e. E V . Like C 4 , both C 9 and C 12 allotropes are insulating with large band gaps, close to 5 eV. This behavior is an indication of the closeness of these two allotropes to diamond.
|
646b7381ccabde9f6e2e8cf8
| 19 |
In this work we have proposed novel carbon allotropes with tetragonal symmetry: C 9 and C 12 based on square carbon units and leading to mixed arrangements of corner and edge C4 tetrahedra. The initial structures, which were subjected to unconstrained geometry optimization within density functional theory, resulted in cohesive stable phases. C 9 and C 12 were found to be mechanically and dynamically close to diamond and to have similar thermal properties. The electronic band structures show insulating behavior with band gaps close to 5 eV, like diamond.
|
6765e65a81d2151a026d17a6
| 0 |
The question of dihydrogen, how it is produced, and how it is used is central to sustainability issues and the decarbonization of the industry. The reversibility of catalytic hydrogenation and dehydrogenation processes is at the heart of this question, as is the transport of H 2 by liquid organic carriers (LOHC). This is particularly true in research on Sustainable Aviation Fuels (SAF) and shipping. Ammonia has been proposed as a desirable hydrogen carrier for the latter because of its hydrogen density and also as a direct fuel. Still, direct ammonia combustion poses the question of N 2 O formation, a potent greenhouse gas. Catalytic processes that combine the cleavage of H 2 and selective transfer to nitrogen are desirable. Using low-cost heterogeneous catalytic systems, the Haber-Bosch process is currently the most efficient industrial process for producing ammonia by nitrogen hydrogenation. While it is difficult to question the massive production of ammonia via this process, smaller-scale units in less industrial environments, where H 2 is produced from green methods, could make way for catalysts capable of generating ammonia at lower energy costs and with greater versatility. Developing processes that can split the H 2 molecule and selectively functionalize simple substrates, such as nitrogen (for ammonia) or CO 2 (for formic acid or methanol), 7 remains a major challenge in modern organometallic chemistry.
|
6765e65a81d2151a026d17a6
| 1 |
Homogeneous pathways for converting N 2 to NH 3 have been made recently, principally with transition metals and uranium. From a mechanistic point of view, the conversion of N 2 and H 2 to ammonia involves different possible scenarios. In the first one, H 2 cleavage occurs first, forming hydride species, which then transfer the hydrogen atoms to dinitrogen. In the second one, N 2 cleavage occurs first, and the resulting nitride is hydrogenated. In the last scenario, synergetic hydrogenation of a reduced N 2 ligand occurs. Among these possibilities, the role of multi-metallic hydride species in the hydrogen atom transfer to dinitrogen has been documented by Hou and co-workers. Hydrogenation of bridged nitrido units, formed after cleavage of dinitrogen, has been recently reported by the groups of Walter and It is noteworthy that the direct hydrogenation of reduced dinitrogen adducts has been only described with an organometallic Zr In the case of rare-earth (RE) elements, which are constituted by the lanthanides along with the closely related Y and Sc, very rare examples of N 2 functionalization have been disclosed.
|
6765e65a81d2151a026d17a6
| 2 |
Corresponding complexes have been obtained by reaction of complexes bearing side-on triply-reduced (N 2 ) 3-ligands with electrophiles or proton donors (Figure ), but direct reaction with H 2 has not been witnessed to date. Recently, Arnold and co-workers showed that dinuclear Sm and Ce complexes could generate bis-and tris(silyl)amine with a turnover of up to 4 using a large excess of metallic reductant in the presence of excess Me 3 SiCl or proton sources. Additionally, following seminal reports on photochemical assistance in producing ammonia from N 2 , 13 the group of Borbas recently reported photocatalytic ammonia formation by a Sm complex bearing a built-in coumarin chromophore. Historically, the chemistry of divalent lanthanide complexes has been dominated by Sm II , Eu II , and Yb II species, which feature relatively stable 4f n Ln II electronic configurations. The corresponding metallocene chemistry has flourished, especially in the context of single-electron transfer (SET) reductions, typically leading to Ln III end products. Although cyclopentadienyl (Cp) complexes of all the lanthanides, except radioactive Pr, have been isolated in the +II oxidation state, 17 fewer studies have focused on the so-called "non-classical divalent lanthanides", consisting of the remainder of the 4f series. This relative scarcity can be traced back to experimental difficulties in handling these very reactive and reducing species. Among them, Tm II is the most accessible in terms of reduction potential, yet displays strong reducing properties beneficial for molecular activation via SET reactions. 11d,20 In this context, our group recently showed that the Tm II complex [Tm(Cp ttt ) 2 ] (Cp ttt = 1,2,4tris(tert-butyl)cyclopentadienyl) displays unique reactivity in terms of CO reductive coupling and subsequent functionalization reactions. The choice of the Cp ttt ligand was crucial to kinetically stabilize the reactive metal center while still allowing the coordination of an exogenous ligand. The three bulky tert-butyl substituents confer high thermal stability but slightly hamper reactivity, as [Tm(Cp ttt ) 2 ] does not react with N 2 , contrary to the less sterically protected analogues. An extreme case of stabilization was recently reported by Long, Harvey and co-workers through the use of the pentasubstituted Cp i Pr5 (Cp i Pr5 = pentaisopropylcyclopentadienyl) ligand. The redox potential of the corresponding Tm II complex was evaluated as high as -1.57 V, more than 1 V less reducing than other Tm II Cp-type complexes. In the present study, we were interested in investigating whether the Cp ttt ligand would allow the stabilization of even more reducing Lu II complexes. In addition to advantageous properties in terms of quantum applications, Lu II species appear particularly attractive to uncover novel reactivities, especially in terms of small molecule activation via SET reductions. Unlike Tm III species that are strongly paramagnetic, Lu III end products are diamagnetic, allowing precise monitoring of reactions by multinuclear NMR spectroscopy. Surprisingly, only three molecular Lu II complexes have been reported to date (Figure ), the two separated ion-pair complexes [K(crypt)][LuCp' 3 ] (Cp' = C 5 H 4 SiMe 3 ) ][Lu(OAr*) 3 ] (OAr* = 2,6-bis(adamantly)-4-t Bu-C 6 H 2 O), and the neutral [Lu(Cp i Pr5 ) 2 ] complex. 17d In this work, we show how the fluxional behavior of the Cp ttt ligand in [Lu(Cp ttt ) 2 ] allows the stabilization of a highly reactive Lu II complex while still offering an open coordination site for original reactivity (Figure ). It is worth noting that, during this work, Evans and co-workers reported the synthesis of the related [Sc(Cp ttt ) 2 ] and its coordination chemistry towards CO and isocyanides. Herein, we show that [Lu(Cp ttt ) 2 ] induces unprecedented reactivities in lanthanide chemistry, such as direct H 2 splitting and N 2 hydrogenation under very smooth conditions (Figure ).
|
6765e65a81d2151a026d17a6
| 3 |
The mother liquor, which was found to be composed of 1 along with the corresponding half-sandwich complex [Lu(Cp ttt )(BH 4 ) 2 (THF)], can be conveniently reused in subsequent synthesis batches of 1 to improve the yield. As lutetium is the last member in the 4f series, its ionic radius may explain the increased difficulty in accommodating two bulky Cp ttt ligands around the metal center.
|
6765e65a81d2151a026d17a6
| 4 |
The X-ray structure of 1 (Figure ) confirmed the coordination of two 5 -Cp ttt and one 2bound borohydride to the Lu III center. In comparison with the isomorphous Tm III complex, 20c the Lu-Cp(ctr) (ctr = ring centroid) separation is slightly shorter (2.322 Å for Lu vs. 2.356 Å for Tm), in agreement with the smaller ionic radius of Lu. The 1 H NMR spectrum of 1 displays three resonances for the coordinated Cp ttt ligands, consistent with an overall C 2v symmetric species in solution. One notable feature in the IR spectrum of 1 is the presence of three moderately intense bands in the region 2000-2500 cm -1 , more precisely one doublet band at 2459 and 2413 cm -1 associated with one additional band at 2112 cm -1 , arising from the terminal 2 -bound BH These bands are blue-shifted by ca. 50 cm -1 with respect to those in [LuCp 2 (BH 4 )(THF)] initially reported by To access [Lu(Cp ttt ) 2 ] (2), reduction of 1 was performed by treatment with excess KC 8 in pentane at room temperature under an argon atmosphere for 5 days (Figure ). As 2 is extremely sensitive towards N 2 , special precautions must be taken to avoid adventitious reaction with traces of N 2 , even when working in an argon-filled glovebox (see Supplementary Information for details). It should be noted that previous reports on the attempted isolation of non-classical divalent lanthanide [Ln(Cp ttt ) 2 ] complexes highlighted the difficulty in isolating these species owing to their high reactivity and solubility. Building on our recent success in crystallizing the related Tm II complex [Tm(Cp ttt ) 2 ], orangered crystals of 2 suitable for X-ray diffraction (XRD) studies were isolated in 18% yield after crystallization from pentane at low temperature (-40 °C). Although the isolated crystalline yield is relatively low, which can be traced back to the very high solubility of 2 in hydrocarbon solvents, the mother liquor was found to be composed of 2 along with variable amounts of clean thermolysis products (see below), as evidenced by 1 H NMR analysis.
|
6765e65a81d2151a026d17a6
| 5 |
The molecular structure of 2 in the solid state (Figure ) displays two independent molecules of 2 in the asymmetric unit, featuring very similar metric data, and confirms the formation of a basefree bent metallocene complex. The two Cp ttt ligands are in an eclipsed arrangement, providing efficient kinetic stabilization of the Lu II metal center. A comparison of the steric shielding in 2 vs. that in [Lu(Cp i Pr5 ) 2 ] 17d by the Guzei method 36 (see Supporting Information for details) revealed very similar G values (83.5 and 85.5%, respectively). The bent arrangement in 2, which displays an average Cp(ctr)-Lu-Cp(ctr) (ctr = ring centroid) angle of 167.2°, is typical for Cp-based Ln II sandwich complexes. Linear coordination has only been observed when using the Cp i Pr5 ligand, which presents a higher symmetry and different steric profile than the Cp ttt ligand. In contrast to the previously reported [Ln(Cp ttt ) 2 ] (Ln = Sm, Eu, Yb, Tm) complexes, for which a monotonic contraction in the Ln-Cp(ctr) separation of ca. 0.018 Å is observed per atomic number increment, the average Lu-Cp(ctr) distance of 2.305 Å in 2 does not follow this trend (Figure ). Although there have been discussions regarding the model that should be best used to fit the variation in bond distances due to the lanthanide contraction, 37 the Lu-Cp(ctr) in 2 appears ca. 0.056 Å shorter as would be expected from a simple linear fit. The apparent decrease in ionic radius can be traced back to the 4f n 5d 1 (n = 14) configuration for Lu II , compared to the 4f n+1 configuration adopted by the divalent Sm, Eu, Yb and Tm congeners, and covalent interactions involving the populated 5d orbital. The 1 H NMR spectrum of 2 at room temperature displays two hardly identifiable broad resonances. Upon heating to 80 °C, a better-resolved spectrum can be obtained with two main resonances at 2.0 ( 1/2 ≈ 70 Hz) and 0.9 ppm ( 1/2 ≈ 200 Hz) in a respective 1:2 ratio, corresponding to the t Bu groups of freely rotating Cp ttt ligands. Continuous-wave X-band EPR measurements performed on a solution of 2 in pentane at room temperature only revealed one broad isotropic signal at g = 2.274 with unresolved hyperfine coupling, which disappeared upon addition of N 2 (Figures S34-35). In contrast, an eight-line EPR pattern was reported for [K(crypt)][LuCp' 3 ] ( 175 Lu, I = 7/2, 97.4% natural abundance), whereas [Lu(Cp i Pr5 ) 2 ] did not give rise to an observable signal. The geometry of complex 2 was optimized at the DFT level (B3W91 functional) and the optimized geometry compares well with the experimental one. In particular, the Lu-Cp(ctr) distance is well reproduced (2.335 vs. 2.305 Å experimentally) as well as the Cp(ctr)-Lu-Cp(ctr) angle (166.6° vs. 167.2° experimentally). The SOMO (Figure ) of 2 appears to be the metal d z 2 orbital, in line with a Lu II 4f 14 5d 1 configuration, and agrees with the isotropic g value of the EPR signal.
|
6765e65a81d2151a026d17a6
| 6 |
The UV-vis spectrum of 2 in pentane (Figure ) shows a major absorption band at 435 nm ( = 2800 L•mol -1 •cm -1 ), along with the onset of a strong band below 350 nm. It contrasts with the spectrum of the related neutral [Lu(Cp i Pr5 ) 2 ] complex, which features minimal absorption in the visible In comparison, in the [K(crypt)][LuCp' 3 ] separated ion-pair complex, an intense band was observed at 518 nm and assigned to a metal-to-ligand charge-transfer (MLCT) transition. TDDFT calculations were carried out in n-pentane to obtain the UV-vis spectrum of 2 (Figure ). An absorption band is found around 400 nm, in agreement with the experimental spectrum, and consists of an MLCT excitation from the SOMO (d z 2 ) to the Lu-Cp ttt antibonding orbital.
|
6765e65a81d2151a026d17a6
| 7 |
The thermal stability of 2 was investigated at different concentrations by 1 H NMR and UV-vis spectroscopy. For a dilute solution of 2 in pentane (C = 4.5 mM), a decrease in intensity of ca. 6% was observed in the UV-vis spectrum for the most intense absorption band at 435 nm over 2 h (Figure ).
|
6765e65a81d2151a026d17a6
| 8 |
More concentrated solutions were found to be more stable over time as the 1 H NMR spectrum of 2 in C 6 D 6 (C = 30-50 mM) revealed no significant formation of thermolysis products over several days. The thermal stability of Lu II complexes appears to be highly dependent on the ligand environment: at room temperature and in hydrocarbon solvents, [K(crypt)][LuCp' 3 ] features a half-life of ca. 20 min, whereas [K(crypt)][Lu(OAr*) 3 ] slowly decomposes over the course of several days. An exceptional thermal stability was reported for [Lu(Cp i Pr5 ) 2 ], synthesized upon extensive heating at elevated temperatures (160-180 °C). 17d
|
6765e65a81d2151a026d17a6
| 9 |
Heating a solution of 2 in C 6 D 6 at 80 °C for several days led to an unexpectedly clean and welldefined 1 H NMR spectrum, showing the formation of the Lu III hydride complex 3-H and the cyclometallated complex 4 (Figure ) in close to equimolar amounts (see Figures ). The identity of 4 was testified by its reactivity towards hydrogenolysis, as complete and immediate conversion into 3-H occurred upon addition of H 2 (Figure ). A similar hydrogenolysis reactivity had been earlier reported by Andersen and co-workers on the cerium analogue of 4. The 1 H NMR spectrum of 4 displays, notably, two doublets at 0.65 and -0.18 ppm, with a 2 J HH coupling constant of 13.5 Hz, for the diastereotopic protons of the cyclometallated methylene group. The characterization data of 3-H will be discussed in detail in the next section. The hydrogenolysis reaction is reversible as heating a degassed solution of 3-H under static vacuum for several days led to complete conversion back to 4 (Figures S11-12). The formation of 4 and its reactivity with H 2 to form 3-H were also investigated computationally. The formation of 4 implies a bimetallic transition state (TS) on the triplet spin potential energy surface (PES) where the hydrogen from one t Bu group of one molecule of 2 is transferred to the Lu center of a second molecule of 2 (Figure ). The associated barrier is 39.0 kcal•mol -1 , in line with the relative stability of 2 at room temperature and the experimental observation of a slow thermal degradation at 80 °C. Yet, the reaction is exothermic by 20.8 kcal•mol -1 due to the formation of one equivalent of 4 and one of 3-H. The hydrogenolysis of complex 4 to form 3-H (Figure ) is both kinetically (barrier of 7.2 kcal•mol -1 ) and thermodynamically favorable (-15.8 kcal•mol -1 ). The activation barrier of 22.9 kcal•mol -1 for the reverse transformation is fully consistent with the dehydrogenation of 3-H into 4 accessible upon heating.
|
6765e65a81d2151a026d17a6
| 10 |
The clean access to 2 as a rare base-free, room-temperature stable and highly soluble divalent lanthanide complex gave us the opportunity to study its reactivity towards industrially relevant small molecules, such as H 2 , which usually does not bind strongly with f-elements. Herein, its coordination to the lanthanide center in 2 would not suffer from the presence of a better ligand. Due to its unique properties, 2 reacts with H 2 under very smooth conditions (1.2-1.5 bar of H 2 , ambient temperature), leading to quantitative conversion into the hydride complex 3-H over the course of 1-3 days (Figure ). Reaction with D 2 proceeded similarly, yielding the Lu III deuteride 3-D, without any appreciable deuteration of the ligand backbone. It should be stressed that the reactivity of 2 with H 2 is much faster than its thermal decomposition into 3-H and 4 (see Figure ), meaning that another direct pathway must be active to form 3-H (see below for mechanistic insights by DFT calculations).
|
6765e65a81d2151a026d17a6
| 11 |
In the 1 H NMR spectrum of 3-H, the hydride resonance is detected as a downfield singlet at 10.3 ppm while the aromatic Cp ttt protons give rise to a singlet at 6.13 ppm. Unrestricted rotation of the Cp ttt ligands leads to an overall C 2v symmetry in solution, as evidenced by the presence of two singlets for the t Bu groups in a 2:1 ratio. In the IR spectrum of 3-H, the Lu-H stretching frequency could not be identified, even upon direct comparison with the spectrum of 3-D, similarly to what was also reported for the Ce analogue [Ce(Cp ttt ) 2 H]. The molecular structures of 3-H and 3-D were unambiguously established by single-crystal XRD analyses, confirming the formation of rare examples of monomeric Ln III complexes bearing terminal hydride or deuteride ligands (Figure and Figure ). Although the H and D atoms on lutetium were located on the electron density map, the Lu-H/D separations (1.73(8) and 1.81(5) Å, respectively) suffer from low precision and may not be fully representative of the actual bond distances. They appear particularly short compared to the lutetiumhydride distances in related Cp-type complexes. The Lu-Cp(ctr) distances in 3-H/D are identical, within experimental error, to those in 2, whereas the presence of the hydride leads to a more acute Cp(ctr)-Lu-Cp(ctr) angle (in average 167.2° in 2 vs. 153.1° in 3-H), i.e. a larger deviation from linearity.
|
6765e65a81d2151a026d17a6
| 12 |
To the best of our knowledge, direct H 2 splitting by molecular lanthanide complexes is unprecedented to date. This strategy offers a novel route for the synthesis of molecular Ln III -H species, typically obtained by hydrogenolysis of Ln-alkyl bonds. It can be noted that, in rare instances, molecular RE hydrides have been obtained upon treatment of trivalent complexes with strong alkali metal reductants. However, such reactions did not involve H 2 activation but rather undefined hydrogen abstraction from solvents or possible aromatic C-H activation. Although elemental zerovalent lanthanides are known to react with hydrogen at elevated temperatures to afford binary Ln hydrides, no such reaction has been observed with molecular Ln species. Weak coordination of H 2 to a lanthanide complex has only been spectroscopically evidenced for the base-free Eu II complex [Eu(C 5 Me 5 ) 2 ], in which case no metal oxidation and formation of a hydride complex occurred. Very recently, this phenomenon has been extended to 5f elements with spectroscopic and computational evidence of the reversible formation of a U III -H 2 complex. Regarding heterogeneous lanthanide systems, nondissociative dihydrogen binding has also been very recently evidenced by solid-state NMR studies under pressure of H 2 . Finally, it is worth noting that, within the actinide series, some U II and Th II complexes were reported to activate H 2 , leading to higher-valent hydride complexes, but the corresponding mechanisms remain elusive. DFT calculations were performed to shed light on the mechanism of the unusual H 2 activation reactivity of 2. Reaction profiles were calculated on the triplet and singlet Potential Energy Surface (PES) (Figure ), and the most probable profile is discussed (Figure ). The reaction begins with the formation of a weak van der Waals adduct of H 2 in between two molecules of 2 (1.2 kcal•mol -1 ). This adduct yields to the side-on coordination of H 2 to one Lu II center while the second Lu II remains in the vicinity (+5.4 kcal•mol -1 ). It is important to note that these two adducts do not imply oxidation of the Lu center since the most stable one is on the triplet PES and the unpaired spin density is mainly located at the two Lu centers (see Figure ). However, the H-H distance (0.82 Å) is elongated, and the SOMO implies interaction between a d orbital on Lu II and the σ* of H 2 (see Figure ). This coordination has thus decreased the energy of the H 2 σ* orbital (which is now involved in the SOMO), allowing further reduction of H 2 via a bimetallic transition state (TS). At the TS, H 2 is undergoing a two-electron reduction that allows the disruption of the H-H bond with an accessible barrier of 24.2 kcal•mol -1 . The value for this activation barrier is fully consistent with the relatively slow reaction observed at room temperature.
|
6765e65a81d2151a026d17a6
| 13 |
The TS was located on the singlet PES in line with a two-electron reduction and the need of bimetallic activation since each Lu II is only able to do one single electron transfer. The H 2 molecule is becoming end-on coordinated to the Lu center (Lu-H distance of 2.20 Å) while the H-H distance is only marginally elongated with respect to the adduct (0.88 Å). The second H•••Lu distance remains long (3.47 Å) but the interaction is effective, as highlighted by the HOMO at the TS (Figure ), where the two d z 2 orbitals located on the two Lu atoms overlap with the H 2 σ* orbital. Following the intrinsic reaction coordinate (IRC), it yields two Lu III -H complexes whose formation is highly favorable (-36.4 kcal•mol -1 with respect to the entrance channel).
|
6765e65a81d2151a026d17a6
| 14 |
The exceptional sensitivity of 2 towards N 2 prompted us to examine the nature of the resulting product and study the possibility of synergistic H 2 /N 2 activation reactions. The addition of N 2 to a pentane solution of 2 led to the immediate formation of a deep purple suspension of [{Lu(Cp ttt ) 2 } 2 (- : -N 2 )]
|
6765e65a81d2151a026d17a6
| 15 |
(5) (Figure ). This work parallels the recent synthesis by Layfield and co-workers of the analogous Gd, Tb, and Dy complexes by in situ reaction of the elusive divalent complexes with N Crystals of 5 suitable of X-ray diffraction studies revealed the formation of a centrosymmetric dinuclear Lu III complex featuring an end-on bridging (N 2 ) 2-ligand (Figure ). Although N 2 coordination on low-valent RE complexes is relatively common, the end-on - -N 2 coordination mode is still rare, 35b,50 typically favored by bulky ligands and RE metals of small ionic radii. As expected from the lanthanide contraction, the Lu-N separation of 2.184(5) Å is slightly shorter than the Ln-N separations of 2.325(4), 2.296(4) and 2.268(7) Å in the Gd, Tb and Dy analogues, respectively. The N-N separation of 1.203(11) Å is consistent with a doubly reduced (N 2 ) 2-ligand, which is corroborated by vibrational spectroscopy. Although the IR spectrum of 5 is very similar to that of 2 (Figure ) because the N=N vibration is not IR active in this centrosymmetric structure, the Raman spectrum of 5 revealed a sharp band at 1610 cm -1 for the symmetric N 2 stretch (Figure ). This assignment was confirmed by the shift of this signal to 1559 cm -1 for the 15 N-labeled complex 5-15 N. The experimental ratio of frequencies of 1.033 matches the theoretical value of 1.035 derived from reduced mass considerations. It is worth noting that the extent of N 2 activation in rare-earth metal diazenido complexes seems to particularly depend on the (N 2 ) 2-coordination mode. For the most common side-on coordination mode, the N 2 stretch lies in the range 1371-1473 cm -1 , 11c,51 whereas the reported N 2 stretching vibration in end-on complexes is at higher energy, ranging from 1595 to 1660 cm -1 . 35b,50 Once formed, 5 is only sparingly soluble in hydrocarbon solvents and readily decomposes in ether solvents to unidentified species, which hampered analysis by NMR spectroscopy. The UV-vis spectrum in pentane of in situ formed 5 shows the disappearance of the characteristic band of 2 centered at 435 nm, at the expense of two intense absorption bands at 310 ( = 5900 L•mol -1 •cm -1 ) and 555 nm ( = 5300 L•mol -1 •cm -1 ) (Figure ). The latter band is slightly blue-shifted compared to that in the corresponding Dy complex, following the preliminary trend observed in the [{Ln(Cp ttt ) 2 } 2 (- -N 2 )] (Ln = Gd, Tb, Dy) series. The structure of complex 5 was investigated computationally using the same methodology as before. Both end-on and side-on coordination of N 2 were considered and the end-on coordination is found to be the most favorable. In the same way, the singlet and triplet spin states were considered for complex 5, as a recent report by Jones and co-workers demonstrated that the (N 2 ) 2-ligand can adopt a triplet ground state by occupying the two formally degenerated * orbital of N 2 . A similar triplet ground state was observed by Evans and co-workers for end-on coordinated dinitrogen Sc and Gd ionpair complexes supported by silylamide ligands. In the case of 5, the singlet is found to be slightly more stable than the triplet by , consistent with the recent findings of Layfield and coworkers on the Gd, Tb and Dy analogues. The singlet optimized geometry compares well with the experimental one (Lu-N distance of 2.20 Å and N-N distance of 1.20 Å), in line with a two-electron reduction of N 2 upon coordination to two Lu II fragments. This reduction, evidenced by the N-N distance variation, is also highlighted by the nature of the HOMO where two d xy orbitals overlap with one of the N 2 * while the second * orbital overlaps with the d xz in the LUMO (Figure ).
|
6765e65a81d2151a026d17a6
| 16 |
It should be noted that, although 2 presents a very strong affinity for dinitrogen, the N 2 coordination is reversible with temperature (Figure ). Upon heating a solution of 5 in tol-d 8 to reflux in a sealed NMR tube under an argon atmosphere, the purple color slowly fades away, leading to the orange-red color characteristic of 2. Agitation of the NMR tube immediately returns the original purple colors, indicating re-coordination of the N 2 ligand. A movie about this transformation can be found in the Supporting Information section. Reversible N 2 coordination is unusual in RE dinitrogen complexes, dominated by side-on N 2 coordination. It has historically been observed in some Sm which are less reducing than non-classical Ln II complexes and thus afford weaker N 2 activation. In the case of RE end-on coordinated (N 2 ) 2-complexes, the liberation of N 2 with concomitant generation, even transiently, of Ln II species seems to be more common and has been observed both thermally 50b,50d or under UV light irradiation.
|
6765e65a81d2151a026d17a6
| 17 |
With the two components of H 2 splitting and N 2 reduction to a (N 2 ) 2-in hand, we focused on their synergy in relation to N 2 hydrogenation, relevant to the Haber-Bosch process. We were interested in the reactivity of 5 towards H 2 as Ln III dinitrogen complexes, which feature formally oxidized metal centers, have been shown to act as "masked" low-valent synthons. In these examples, twoelectron reduction processes occurred upon addition of reducible substrates (CO, CO 2 , aromatic Nheterocycles), with the systematic release of the N 2 ligand. Similar reductive reactivity has also been recently reported in the case of N 2 complexes of U IV , Mg II and Ca II , where the coordinated and reduced (N 2 ) 2-moiety acted as a redox-active ligand, able to reversibly store two electrons. Initial reactivity attempts in NMR tubes showed that heating a purple suspension of 5 in C 6 D 6 under H 2 atmosphere at 80 °C led to the full consumption of 5 over several hours, as evidenced by the formation of a clear colorless solution. H NMR analysis revealed the formation of 3-H as a major product (ca. 90%) along with a minor species (ca. 10%) assigned to [Lu(Cp ttt ) 2 NH 2 ] (6) (Figure ). The identity of 6 was testified by N labelling experiments as well as independent in situ synthesis of 6 upon exposure of 2 with NH 3 gas (Figures ). In the 1 H NMR spectrum, the singlet at 3.07 ppm for the NH 2 protons in 6 was replaced by a doublet ( 1 J HN = 62.8 Hz) when performing the experiment with the isotopically labelled 5-15 N. Accordingly, the As decoordination of the N 2 ligand in 5 is favored with temperature, we hypothesized that the formation of 3-H might arise from the reaction of in situ liberated 2 with H 2 . We reasoned that lower temperatures should be beneficial for higher conversions into 6 vs. 3-H. Although the reaction appeared kinetically blocked below 0 °C, full conversion was observed upon stirring a toluene or pentane suspension of 5 under H 2 atmosphere over 1-5 days at 0 °C or room temperature. Under these conditions, 6 was consistently formed in 20-32% NMR yields (Figures S19-20).
|
6765e65a81d2151a026d17a6
| 18 |
The exact mechanism of the N 2 hydrogenation reactivity remains elusive to date but control experiments revealed no reaction between 3-H and N 2 at different temperatures (Figure ). Although M-NH 2 moieties have been obtained upon hydrogenation of terminal uranium or iridium nitride complexes, 56 the inaccessibility of oxidation states higher than +III for Lu would prevent the formation of a similar intermediate. As such, the formation of 6 is bound to proceed through direct hydrogenation of the N 2 -ligated complex 5, an unprecedented reactivity in RE chemistry. The observed N 2hydrogenation reactivity is all the more remarkable as it occurs on a doubly-reduced (N 2 ) 2-ligand. So far, the very rare examples of N 2 -functionalization on RE complexes (Figure ) have systematically required further activation of the N 2 ligand via formation of the (N 2 ) 3-radical anion. 11
|
6765e65a81d2151a026d17a6
| 19 |
The present work describes the synthesis and reactivity of a rare example of Lu II neutral complex. The fluxionality of the Cp ttt ligand is essential to provide kinetic stabilization of the divalent metal center while still allowing coordination of exogenous ligands. The Lu II complex, 2, reduces H 2 at room temperature to give a Lu III -hydride complex, which corresponds to the first example of direct H 2 splitting by a molecular rare-earth complex. This unusual reactivity has been fully supported by DFT calculations. In addition, 2 readily reacts with N 2 with reversible formation of an end-on coordinated (N 2 ) 2-complex. The latter reacts with H 2 under very smooth conditions yielding a Lu III -NH 2 via direct
|
6344000b114b7e0e21304d7b
| 0 |
Robust reaction rate enhancement and allosteric regulation are hallmarks of enzyme catalysis, and both aspects may be at least in part underpinned by protein conformational flexibility. The catalytic prowess of enzymes can be significantly ascribed to substrate binding to an electrostatically preorganised active site, which minimises the reorganisation energy required for optimum stabilisation of the charge redistribution as the reaction progresses from the reactant state to the transition state. Yet several lines of evidence also suggest a contribution from protein dynamics, from nonstatistical, femtosecond-timescale vibrations coupled directly to transition-state barrier crossing, to slower, thermally equilibrated motions reshaping the enzyme conformational ensemble towards populations where active-site preorganisation is optimised. Nonetheless, this topic is still controversial possibly due to the inherent flexibility of proteins which makes it difficult to isolate motions that may have evolved to facilitate reaction. Allosteric modulation of enzymes, i.e. the alteration of reaction rate and/or substrate affinity upon ligand binding to, mutation of, or post-translational modification at a site remote from the active site, is a fundamental regulatory mechanism of biochemical reactions. It finds applications in drug discovery to facilitate drug-target selectivity as allosteric sites tend to be less conserved than active sites across homologous proteins, and in enzyme engineering and synthetic biology, where allosteric control may need to be introduced or, more often, eliminated. While enzymes subject to allosteric regulation by ligand binding can be broadly classified as K-type, those where substrate affinity is altered, and V-type, those where the steady-state catalytic rate constant (kcat) is altered, the specific kinetic steps affected can vary depending on the enzyme. For instance, in Mycobacterium tuberculosis -isopropylmalate synthase, the rate-limiting step changes from product release to chemistry upon allosteric inhibition by leucine. The role of protein dynamics in allostery has been much less controversial when discussed in terms of conformational changes to promote physical events such as substrate binding and product release, or the interconversion rate among conformations. However, in systems where allosteric regulation affects the rate of the chemical step itself, the intersection at which local and remote protein motions, active-site electrostatic preorganisation, and ultimately catalysis meet remains challenging to pinpoint, despite recent advances toward this goal with Kemp eliminase. ATP phosphoribosyltransferase (ATPPRT) (EC 2.4.2.17) catalyses the Mg 2+dependent formation of N 1 -(5-phospho-β-D-ribosyl)-ATP (PRATP) and inorganic pyrophosphate (PPi) from ATP and 5-phospho-α-D-ribose 1-pyrophosphate (PRPP) (Fig. ), the first and flux-controlling step of histidine biosynthesis, and is allosterically inhibited by histidine in a negative feedback control loop. ATPPRT is the focus of synthetic biology efforts to enable the production of histidine in bacteria and a promising drug target against some pathogenic bacteria. Short-form ATPPRTs form an intricate allosteric system comprising catalytic (HisGS) and regulatory (HisZ) subunits assembled as a heterooctamer with a tetrameric core of HisZ sandwiched by two dimers of HisGS. HisGS on its own is catalytically active and insensitive to inhibition by histidine. Binding of HisZ, which has no catalytic power of its own, to form the ATPPRT holoenzyme allosterically activates catalysis by HisGS. However, HisZ also harbours the pocket where histidine binds and allosterically inhibits ATPPRT. Thus, the regulatory protein plays a dual role, as allosteric activator of catalysis in the absence of histidine and mediator of allosteric inhibition in the presence of histidine. showed Arg56 of one of the catalytic subunits reaching across the dimer interface to form a salt-bridge with the pyrophosphate moiety of PRPP in the active site of the other subunit (Fig. ) in the PaATPPRT structure but not in the PaHisGS structure. Therefore, allosteric activation was proposed to lead to more efficient leaving group departure at the transition state by stabilisation of the negative charge build-up on the pyrophosphate upon nucleophilic attack of ATP N1 on PRPP C1. Second, no burst in product formation was observed for the reaction catalysed by PaHisGS, and the multiple-turnover, pre-steady-state rate constant was in agreement with kcat, suggesting chemistry is rate limiting in the nonactivated enzyme reaction. In contrast, a burst was observed in the reaction catalysed by PaATPPRT, producing a rate-constant (kburst) much higher than kcat, supporting a mechanism where allosteric activation speeds up the chemical step, making product release rate limiting. Finally, replacement of Mg 2+ by Mn 2+ , which more efficiently offsets the negative charge at the transition state, led to an ~3-fold enhancement of PaHisGS kcat, as would be predicted, qualitatively, for a rate-limiting chemical step, but had no effect on PaATPPRT kcat, where chemistry was already much faster than subsequent steps. As expected, an R56A-PaHisGS mutant had a reduced reaction rate in the nonactivated enzyme as measured at a fixed concentration of substrates, since R56 was posited to be important for leaving group departure in PaHisGS as well, only less efficiently.
|
6344000b114b7e0e21304d7b
| 1 |
Intriguingly, upon PaHisZ binding to R56A-PaHisGS, part of the activity was recovered. Here we employed site-directed mutagenesis, differential scanning fluorimetry (DSF), enzyme kinetics, 31 P-NMR spectroscopy, protein crystallography, and molecular dynamics (MD) simulations to dissect this phenomenon, reveal that other single-and double-mutations at the PaHisGS active site display similar behaviour, and demonstrate how modulation of PaHisGS dynamics by PaHisZ propagates to the active site to affect the chemical step.
|
6344000b114b7e0e21304d7b
| 2 |
Electrospray ionisation/time-of-flight-mass spectrometry (ESI/TOF-MS) confirmed the molecular masses of WT-, C115A-, C115S-, D179A-, R32A-, R56A-, and R56A/K57A-PaHisGS variants were in agreement with the predicted values (Supplementary Fig. ). The introduction of the D179N mutation was confirmed by MS/MS analysis of tryptic fragments (Supplementary Fig. ). In an initial activity screen, reactions were monitored for just under 1 min by the continuous and direct UV/VIS absorbance-based assay for ATPPRT activity at fixed PRPP and ATP concentrations sufficient to saturate WT-PaHisGS. PRATP formation was readily detected with C115A-, D179A-, D179N-, and WT-PaHisGS (Fig. ), and linear regression of the data yielded apparent rate constants shown in Supplementary Table .
|
6344000b114b7e0e21304d7b
| 3 |
Substrate saturation curves for WT-, C115A-, D179A-, and D179N-PaHisGS obeyed Michaelis-Menten kinetics (Fig. ), and data fit to equation (1) produced the apparent steady-state kinetic parameters in Supplementary Table . The Michaelis constant for ATP (KATP) increased less than 2-fold for C115A-PaHisGS, and the Michaelis constant for PRPP (KPRPP) actually decreased between 2-and 3-fold for D179A-and D179N-PaHisGS, suggesting C115 and D179 make negligible contributions to substrate binding. The kcat values for the mutants were reduced only ~4-fold in comparison with the WT-PaHisGS, pointing to these residues' modest importance in catalysis. Over the course of the ATPPRT-catalysed reaction, the 6-NH2 group must donate a proton to yield the 6-NH group of PRATP, and in the case of PaATPPRT, this proton abstraction happens on the enzyme. Based on their respective positions in the active site (Fig. ), both C115 and D179 were candidates to act as general base for this proton abstraction, but the small catalytic effect of their replacements for residues that cannot participate in acid-base catalysis does not support such role, leaving the identity of the general base still elusive.
|
6344000b114b7e0e21304d7b
| 4 |
C115S-, R32A-, R56A-, and R56A/K57A-PaHisGS are catalytically compromised. No PRATP formation could be detected above the background noise of the assay (no-enzyme control) during the initial activity screen when either C115S-, R32A-, R56A-, or R56A/K57A-PaHisGS was used as catalyst (Fig. ), indicating these mutants have impaired catalytic activity. DSF-based thermal denaturation assays showed these mutants display similar thermal unfolding profiles to the WT protein (Fig. ), demonstrating the mutations do not thermally destabilise the tertiary structure of the protein, and data fit to equation (2) yielded melting temperatures (Tm) shown in Supplementary Table . Moreover, as described previously and repeated here for WT-PaHisGS, the presence of PRPP increased the Tm of the mutants (Fig. ; Supplementary Table ), indicating the catalytically impaired PaHisGS variants can bind PRPP, in agreement with the ordered kinetic mechanism proposed for this enzyme. Analytical size-exclusion chromatography produced similar elution profiles for WT-, C115S-, R32A-, R56A-, and R56A/K57A-PaHisGS (Supplementary Fig. ), which includes the expected dimer and a higher oligomeric state. An activity screen with longer reaction times and twice as much enzyme as in the initial screen, to allow more product to accumulate, demonstrated the catalytic ability of C115S-, R32A-, R56A-, and R56A/K57A-PaHisGS is significantly diminished but not abolished (Fig. ). ppm corresponds to the phosphorus in the N 1 -5-phospho-β-D-ribose moiety of PRATP. e Substrate saturation curves for WT and mutant PaATPPRT. Data are mean standard error of two independent measurements. Lines are best fit of the data to equation (1). f Effect of 1 mM histidine on mutant PaATPPRT-catalysed reaction. Two independent measurements were carried out, and all data points are shown.
|
6344000b114b7e0e21304d7b
| 5 |
Even though C115 is only modestly important for catalysis, its replacement by serine led to a 117-fold reduction in activity (Table ), perhaps due to the introduction of a detrimental interaction. The activities of R32A-and R56A-PaHisGS decreased 25-and 42fold, respectively, in comparison with the WT-PaHisGS (Table ). This demonstrates the importance of these residues in PaHisGS catalysis, possibly because R56 and R32 may contribute to leaving group departure at the transition state. K57 is adjacent to R56 in the PaHisGS primary sequence, but in all PaHisGS and PaATPPRT crystal structures, it points away from the active site. We hypothesised that in the absence of the R56 guanidinium group, the K57 -NH3 + group could move towards the active site and assist in leaving group departure. However, the R56A/K57A-PaHisGS double mutant displayed a 254-fold decrease in activity (Table ), which is only ~6-fold more catalytically impaired than the R56A-PaHisGS, indicating just a modest catalytic importance for K57. R32, R56, K57, and C115 are highly conserved in HisGS across species, and D179 is also conserved but sometimes replaced with a glutamate residue. Nevertheless, out of these five residues in PaHisGS, only the arginine residues seem to be significantly important for catalysis. . No activity was detected when the PaHisGS mutant-catalysed reactions were carried out in the presence of bovine serum albumin (BSA) (Supplementary Fig. ), ruling out that allosteric rescue was due to nonspecific protein binding. To gather orthogonal evidence for the allosteric rescue, the reaction catalysed by each PaHisGS mutant was analysed by P-NMR spectroscopy in the presence and absence of PaHisZ (Supplementary Fig. ) under conditions where product can be detected with WT-PaHisGS. The characteristic chemical shift at ~3.30 ppm (Fig. )
|
6344000b114b7e0e21304d7b
| 6 |
previously assigned to the phosphorus in the N 1 -5-phospho-β-D-ribose moiety of PRATP, was only detected here when PaHisZ was present in the reaction, confirming the rescue of the catalytically compromised mutants by the regulatory protein. Histidine binds to PaHisZ and allosterically inhibits PaATPPRT catalysis, and the suppression of the burst in product formation in the presence of histidine suggests that allosteric inhibition directly affects the chemical step of the reaction. Histidine also inhibits the reaction catalysed by the rescued PaATPPRT mutants (Fig. ; Supplementary Fig. ),
|
6344000b114b7e0e21304d7b
| 7 |
and PaHisGS share 43% and 69% sequence identity with their orthologues from the pathogenic bacterium Acinetobacter baumannii, AbHisZ and AbHisGS, respectively, but PaHisZ has been shown to be a potent allosteric inhibitor of AbHisGS. We thus hypothesised that AbHisZ could inhibit WT-PaHisGS. However, addition of AbHisZ activated catalysis by WT-PaHisGS (Fig. ), and data fit to equation (3) yielded an apparent KD for AbHisZ of 9 1 M. Moreover, AbHisZ also rescued catalysis by R56A-PaHisGS (Fig. ), but their interaction involved positive co-operativity as evidenced by the sigmoidal dependence of the reaction rate on the regulatory protein. The R56A-PaHisGS mutant was chosen due to its significant catalytic impairment and the proposed role in catalysis for R56.
|
6344000b114b7e0e21304d7b
| 8 |
The data were fit to equation ( ), yielding a concentration of AbHisZ at the inflection point (K0.5) and a Hill coefficient (h) of 8.1 0.4 M and 1.68 0.08, respectively. Nonetheless, this fit is intended only to highlight the sigmoidal behaviour of the data, since the experiment could not be carried out under pseudo-first-order conditions, i. e. [R56A-PaHisGS] ~
|
6344000b114b7e0e21304d7b
| 9 |
[AbHisZ] in the experiment; thus, equation ( ) does not hold. WT-PaATPPRT are known to form the Michaelis complex in crystallo, likely due to a highly unfavourable on-enzyme equilibrium for the forward reaction. The active site interactions are very similar between the two structures, except the electron density for Mg 2+ in R56A-PaATPPRT was not well defined, so the metal was not modelled in, and the hydrogen bond between E163 and the PRPP 3ʹ-OH is absent in R56A-PaHisGS (Fig. ). The structures are also very similar to those of the respective WT enzymes. Therefore, any structural or conformational differences that may lead to allosteric rescue of R56A-PaHisGS catalysis are not captured in the static view of the crystal structure. been observed in other allosteric systems, and it can also be seen here that PaHisZ binding increases order in the system.
|
6344000b114b7e0e21304d7b
| 10 |
The conformational behavior of the activated dimer was further explored using the shortest path map (SPM) approach as implemented by Osuna and coworkers, which enables the identification of pairs of residues in both the active site and distal positions with the highest contributions to the communication pathways in nonactivated and activated PaHisGS.
|
6344000b114b7e0e21304d7b
| 11 |
we have performed node-weakening analysis by removing the nodes corresponding to the residues located at the interface between PaHisGS and PaHisZ, where the SPM goes through both proteins (5 nodes on PaHisGS and 7 nodes on PaHisZ), and calculating the change in CPL (characteristic path length) upon removal of each node (Supplementary Table ). We then selected the three residues displaying the highest impact on the CPL and performed molecular dynamic simulations of four selected in-silico mutants of PaATPPRT (Y105A-PaHisGS, Y105F-PaHisGS, N185A-PaHisZ, and K186D-PaHisZ), with the aim of disrupting the communication pathway between PaHisGS and PaHisZ. When comparing the SPM of the WT-PaATPPRT with those of the various mutants we generated in-silico, we see that the pathway is slightly reorganized without displaying critical changes that disrupt the communication signal between the two proteins (Supplementary Fig. ). These results are in line with a proposal that allosteric activation is due to a global "nesting" effect of PaHisZ over PaHisGS, with a preferred but not unique allosteric activation pathway.
|
6344000b114b7e0e21304d7b
| 12 |
This mimics the natural condition in a crowded cellular environment, where the protein is exposed to many different random forces. Thus, the DFI analysis aims to capture the contribution of each position in the protein to the underlying functional dynamics, highlighting the key residues and regions that mediate functionally important dynamical information. Our analysis (Supplementary Fig. ) reveal PaHisZ binding changes the DFI of PaHisGS compared with nonactivated PaHisGS, in particular at the helices inferred to be important for communication at the PaHisGS-PaHisZ interface from the SPM analysis (Fig. ), as well as the region containing R32 and R56. Given the catalytic importance of R32 and R56 and their hypothesized involvement in facilitating leaving group departure, we tracked the distance between Cζ of each side chain and P of the PPi moiety of PRPP during our simulations (Fig. ). It can be gleaned from the data that the R56 side chain displays a bimodal distribution of distances in WT-PaHisGS (Fig. ). This distribution comprises a peak at ~4.4 Å corresponding to a catalytic conformation in which this side chain forms a salt-bridge with the PRPP PPi moiety, and another at ~7.8 Å, corresponding to a non-catalytic rotamer of this residue. Binding of PaHisZ shifts the distance distribution towards the catalytically active rotamer (Fig. ). This furnishes support for PaHisZ binding constraining the conformational dynamics of PaHisGS, fostering a preorganized active site in which the R56 guanidinium group is poised to help stabilize the leaving group at the transition state. increases WT-PaHisGS kcat, and density-functional theory calculations provided a rationale for this effect based on more efficient stabilisation of the negative charges by Mn 2+ via dorbital bonding to the oxygens of the departing PPi at the transition state. This was corroborating evidence that chemistry was the rate-limiting step in PaHisGS catalysis, but not in PaATPPRT where Mn 2+ had no significant impact on kcat. This observation is reproduced here. At saturating concentrations of both substrates, Mn 2+ allows product formation to be detected at a WT-PaHisGS concentration too low to detect reaction with Mg 2+ , but does not increase the WT-PaATPPRT reaction rate (Fig. ). In contrast, when R32A-PaATPPRT and R56A-PaATPPRT reactions were carried out with Mn 2+ instead of Mg 2+ , the rate of product formation increased (Fig. ), and the apparent first-order rate constants increased by ~5-fold and ~3-fold, respectively, in comparison with those with Mg 2+ (Fig. ). This indicates the rates of the rescued mutants reflect the chemical step of the reaction, i. e. unlike WT-PaATPPRT, chemistry is at least partially rate-limiting for R32A-PaATPPRT and R56A-PaATPPRT. Interestingly, R32A-PaATPPRT and R56A-PaATPPRT apparent first-order rate constants with Mn 2+ are even higher than the corresponding one for WT-PaATPPRT.
|
6344000b114b7e0e21304d7b
| 13 |
However, these rate constants do not reflect the same reaction step. It is likely that Mn 2+ enhances the rate of chemistry substantially for WT-PaATPPRT, but the observed rate constant is dominated by product release. On the other hand, with R32A-PaATPPRT and R56A-PaATPPRT, the effect of Mn 2+ is the enhancement of the rate of chemistry itself.
|
6344000b114b7e0e21304d7b
| 14 |
Given the proposed role of R56 in stabilising the departure of the negatively charged leaving group, the possibility that a lysine residue could replace R56 with similar catalytic ability was considered. To evaluate this possibility, R56K-PaHisGS was produced. ESI/TOF-MS analysis resulted in the expected mass (Supplementary Fig. ), and DSF showed the mutation does not change the Tm of the protein (Supplementary Fig. ). At substrate concentrations saturating for WT-PaHisGS, however, the R56K-PaHisGS reaction rate is reduced ~24-fold in comparison with the WT-PaHisGS, and although there is an ~11-fold activation in the presence of PaHisZ, the R56K-PaATPPRT reaction rate is still ~2-fold lower compared with the nonactivated WT variant (Supplementary Fig. ). These observations indicate the amino group cannot substitute for the guanidinium group at position 56 of PaHisGS. Furthermore, allosteric activation by PaHisZ cannot rescue catalysis in this case.
|
6344000b114b7e0e21304d7b
| 15 |
R32A/R56A/K57A-PaHisGS cannot be rescued by PaHisZ. The hypothesis that R32 and R56 can compensate to a certain extent for the absence of the other in the presence of PaHisZ to restore the electrostatic preorganisation of PaHisGS active site predicts that removal of both arginine residues would prevent allosteric rescue of catalysis. To test this prediction, the R32A/R56A/K57A-PaHisGS triple mutant was produced (Supplementary Fig. ) and ESI/TOF-MS analysis resulted in the expected mass (Supplementary Fig. ).
|
6344000b114b7e0e21304d7b
| 16 |
DSF indicated that the additional mutation does not alter the Tm of the protein (Fig. ) as compared with R56A/K57A-PaHisGS or WT-PaHisGS Tm. Furthermore, PRPP led to an increase in Tm, showing the triple mutant can bind this substrate. As expected, PRATP formation could not be detected with R32A/R56A/K57A-PaHisGS as catalyst (Fig. ).
|
6344000b114b7e0e21304d7b
| 17 |
Upon addition of excess PaHisZ, some PRATP formation could be marginally detected above the assay background noise (Fig. ), demonstrating R32A/R56A/K57A-PaATPPRT still retains residual catalytic activity. However, the apparent rate constant is reduced ~777fold in comparison with that of WT-PaATPPRT (Supplementary Fig. ), and ~340-fold and ~222-fold in comparison with those of R32A-PaATPPRT and R56A-PaATPPRT, respectively (Supplementary Fig. ), in accordance with the proposed necessity for at least one of the two arginine residues to aid in leaving group departure for full catalytic power of PaATPPRT.
|
6344000b114b7e0e21304d7b
| 18 |
Rescue of catalytically impaired enzyme mutants is well established, but not as observed for PaATPPRT. Chemical rescue by small molecules that mimic missing residue side chains is a useful tool to probe the function of active-site residues in catalysis, and a hyper-nucleophilic cholesterol analogue in which an -OOH group replaces the sterol -OH group could rescue the base-catalysed endoproteolytic activity of a mutant hedgehog protein where the catalytically essential aspartate general base was mutated to an alanine, which rendered the reaction highly impaired with the natural substrate. A catalytically compromised receptor tyrosine kinase carrying mutations in the activation loop tyrosine residues that would otherwise be autophosphorylated could be allosterically rescued by the juxtamembrane segment via autophosphorylation of this segment's Y687, but in this case the effect of the phospho-Y687 is exerted by direct interaction with arginine residues that normally interact with the phosphorylated tyrosine residues in the activation loop. In heterotetrameric tryptophan synthase, where catalysis by the β subunit is allosterically activated by the subunit, the catalytically inactive E109A mutant of the β subunit could not be rescued by the subunit, but the activity of the E109A-2β2 complex could be partially restored by CsCl, possibly by modulation of the conformational ensemble of the complex. In human prolyl isomerase CypA, the second-shell S99T mutation, which is highly detrimental to catalysis, can be partially counteracted by additional mutations outside the active site which rescue the dynamics of interconversion between two essential conformations. PaATPPRT is unique because the rescue of catalytically compromised PaHisGS mutants by PaHisZ, and even by the orthologous AbHisZ, is truly allosteric since the regulatory subunit binds far from the active site where the mutations exert a detrimental effect on transition state stabilisation. The narrowing of the distribution of states sampled by the PaHisGS upon PaHisZ binding has a direct impact on the positioning and orientation of R32 and R56, which are better poised to facilitate leaving group departure by electrostatic stabilisation of the PPi negative charges. This implies allosteric activation of catalysis in PaATPPRT involves modulation of the conformational flexibility of the holoenzyme and electrostatic preorganisation of the active site. The catalytic recruitment of R32 and R56 in concert with Mg 2+ to stabilise PPi is reminiscent of that of arginine residues in adenylate kinase to promote phosphate transfer from ADP to AMP, where thermally equilibrated protein motions were also proposed to help achieve optimal electrostatic preorganisation. Interestingly, replacement of a key arginine residue for a lysine was detrimental to catalysis in adenylate kinase as well. Another important aspect of the allosterically rescued R32A-and R56A-PaATPPRT is that kcat is at least partially limited by the chemical step, a drastic change from WT-PaATPPRT in which kcat is determined by product release. Electrostatic preorganisation exerts its effect on catalysis at the chemical step, i.e. as the reaction progresses from the preorganised Michaelis complex to the transition state. Thus it is paramount that an experimentally measured rate constant purporting to reflect any coupling of protein motions to the preorganisation of the Michaelis complex be limited by the chemical step. This is what is observed with the allosteric rescue of the catalytically compromised mutants of PaHisGS, establishing a direct connection between PaHisZ-modulated rotamers of R32 and R56 and electrostatic preorganisation of the active site, which is required for optimal catalysis. Crystallisation, X-ray data collection and data processing. Crystals of R56A-PaHisGS were grown, soaked in PRPP and ATP and stored as described for WT-PaHisGS, whereas crystals of R56A-PaATPPRT were grown as described for WT-PaATPPRT and soaked in PRPP and ATP and stored as described for WT-PaATPPRT. X-ray diffraction data for R56A-PaHisGS were collected in house as previously reported and processed with iMosflm, Structures were refined using cycles of model building with COOT and refinement with Refmac. ATP was modelled at either 70% or 80% occupancy in R56A-PaATPPRT. Partial charges for the ligand PRPP were calculated in vacuo at the HF/6-31G* level of theory using Gaussian 16 Rev. A.03, and fitted using the standard restrained electrostatic potential (RESP) protocol as implemented in Antechamber (Supplementary Table ). All other force field terms for PRPP were then described using the Amber force field ff14SB together with revised parameters to describe bioorganic phosphates. The parameters for ATP were taken from the literature. We used an octahedral cationic dummy model to describe Mg 2+ , following from previous successful results using this model. All MD simulations were performed using the GPU-accelerated version of Amber16, with the protein and water molecules described using the amber force field ff14SB and the TIP3P water model, respectively. All systems were solvated in an octahedral box of water molecules, extended 8 Å from the closest solute molecule in every direction. Each system was neutralized by adding Na + or Cl -counterions to ensure overall charge neutrality. Counterions were placed using the "addions" approach as implemented in Amber16. The dimeric and hetero-octameric forms of the enzyme were simulated for 10 500 ns and 5 500 ns, respectively, in the NPT ensemble. The solvated systems were first minimized using 5,000 steps of steepest descent minimization with 500 kcal mol -1 Å -2 positional restraints placed on all solute atoms to minimize all hydrogen atoms and solvent molecules, followed by 5,000 steps of conjugate gradient minimization, with the restraint dropped to 5 kcal mol -1 Å -2 . The minimized system was then heated from 0 to 300 K in an NVT ensemble over 250 ps of simulation time using the Berendsen thermostat
|
60c753a5f96a001b412884da
| 0 |
Since the introduction of the term epigenetics by Conrad Waddington in 1942 to denote the mechanisms that relate genotype to phenotype, the term has been used with multiple meanings, going from the classic definition that refers to epigenetics as the study of the alterations in the biological phenotype without underlying changes in the DNA sequence, to one of the most recent and general definitions: "the structural adaptation of chromosomal regions to register, signal, or perpetuate altered activity states." At the molecular level, this adaptation involves the reversible modification of nucleic acids and histones. These modifications are catalyzed by a plethora of proteins, which could be considered as the core epigenetic targets, and that are classified into three main groups: (a) writers -enzymes capable of adding chemical groups to nucleic acids and histones -such as DNA methyltransferases (DNMTs), histone methyltransferases (HMTs) and histone acetyltransferases (HATs), (b) erasers -enzymes capable of removing marks introduced by the writers -such as histone deacetylases (HDACs) and histone demethylases (HDMs), and (c) readers -proteins with specialized domains capable of recognizing these changes -such as the bromodomain and external terminal protein (BET) family. In addition to these core epigenetic targets, a wide range of proteins also play important roles in epigenetic regulation; these proteins include histone chaperones (critical for nucleosome assembly), chromatin remodelers (CHRs -responsible for moving, ejecting, and restructuring the nucleosome), and even some classes of transcription factors. Epigenetics is an essential component in an organism's normal development and responsiveness, so its dysregulation has been associated with altered gene expression patterns related to multiple diseases. This makes epigenetic targets a significant focus for drug discovery research. Successful examples can be found in cancer research, with the approval of eight epigenetic drugs (drugs targeting epigenetic proteins) for clinical use: azacytidine and decitabine targeting DNMT1, vorinostat, belinostat, panobinostat, romidepsin and tucsibinostat targeting HDACs, and tazemostat targeting an HMT (EZH2). The importance of epigenetics in drug discovery is also illustrated by the increasing availability of chemogenomic databases related to epigenetics over the past decade. An example of this is EpiFactors, to the best of our knowledge, the database with the largest number of annotated proteins related to epigenetics reported so far, with a total of 815 different targets.
|
60c753a5f96a001b412884da
| 1 |
In a recent work, we surveyed the status of the compounds tested against these and other epigenetic targets identified from ChEMBL, Therapeutic Target Database, and scientific literature. We found out that for 136 of these targets, there are more than ten reported inhibitors, which meant a considerable increase in comparison with the 52 targets fulfilling the same criteria in 2017. The rich structure-activity relationships (SAR) contained in these large data sets represents an excellent source of information to develop predictive models that have not been developed thus far on a large-scale basis. In a previous work the authors explored the SAR of epigenetic target data sets using the concept of activity landscape.
|
60c753a5f96a001b412884da
| 2 |
Although that work was a quantitative study, it was descriptive. The increase in the publicly available chemogenomic data for all target classes over the years opened up the opportunity for the construction of ligand-based machine learning models to assist target prediction of small molecules. Some of these methods are currently available as easy-to-access web applications, such as Similarity Ensemble Approach (SEA), HitPick, Polypharmacology Browser (PPB), TargetHunter, and SwissTargetPrediction, to name a few examples. These methods usually assign the targets for a given small molecule from the known targets of the most similar ligands in their datasets, employing different descriptions and metrics for the similarity assessment, and often making use of additional statistical models to estimate the significance of the predictions. Despite of the increasing number of chemogenomic databases related to epigenetics, this data still represents a minimal amount when compared to other protein families such as kinases (KINs), ion channels or G protein-coupled receptors. This suggests that epigenetic targets are commonly underrepresented in the current target prediction methods, and that unless the similarity of a known ligand is high enough, they are less likely to be predicted as potential targets of small molecules, which points out the need of developing predictive models focused on epigenetic targets to assist medicinal chemistry efforts in this area.
|
60c753a5f96a001b412884da
| 3 |
Machine learning methods have proven to be useful in multiple areas of drug discovery, one such being target prediction of small molecules. For instance, in a retrospective large-scale comparison of machine learning methods for target prediction on ChEMBL (in the context of biochemical assays), deep neural networks were the best performing method for this task when trained on Extended Connectivity Fingerprints 37 (ECFP) of chemical compounds. However, the application of machine learning models for large-scale epigenetic target prediction has been explored on a limited basis, with most works focused on single targets or protein families such as HDACs or the BET family. Herein, we aimed to develop accurate models for epigenetic target prediction based on stateof-the-art machine learning algorithms trained on different fingerprint representation of compounds. We describe the development of predictive models with high precision for 55 epigenetic targets. Derivation of such predictive models is relevant for medicinal chemistry to develop hypothesis for the discovery of new epigenetic probes and drugs. The best models herein generated are implemented in an easy-to-use web application freely available to support medicinal chemistry projects related to epigenetic drug and probe discovery. It is anticipated that this tool will assist epigenetic drug design and development projects in the design and selection of compounds with potential epigenetic activity.
|
60c753a5f96a001b412884da
| 4 |
This section is organized into three major parts. First, we described the results of the data sets of epigenetic targets used in this work. The second part, entitled "Epigenetic Target Prediction with Machine Learning," presents the results of the development of the machine learning models and their validation using two main strategies. The third main section, "Retrospective Identification of Epigenetic Targets," shows, as a case study, a practical application of the best machine learning model derived in the second part, to identify epigenetic targets for external and recently reported compounds. All the details of the methods used are described in the Experimental Section.
|
60c753a5f96a001b412884da
| 5 |
Chemogenomic Data for Epigenetic Targets. Quantitative compound-protein associations were extracted from ChEMBL 27 and PubChem to build epigenetic target-associated compound datasets meeting the following criteria: (a) containing at least 30 compounds with a quantitative measure of biological activity (IC 50 , EC 50 , K i or K d ) lower or equal to 10 µM ("active") and at least 30 compounds with a quantitative measure of biological activity higher than 10 µM ("inactive"), and (b) modelability index (MODI) higher than 0.7 for at least one of the three molecular fingerprints selected as compound representation (see Experimental Section for further details). As illustrated in Figure The compiled chemogenomic dataset contained 26318 unique compounds and 38129 compound-protein associations, with 28750 of them being labeled as active and 9379 labeled as inactive (due to the natural, although not the best practice of reporting mostly active compounds and not negative -inactive-data in ChEMBL). Consistently with the compound/compound-protein associations ratio, 20318 compounds (77.2%) in the dataset had known associations to a single target, and only 196 compounds (0.7%) had known associations to at least 10 targets, with a maximum of 15 targets for four compounds (Table ). k-nearest neighbors (k-NN) , Random Forest (RF) , Gradient Boosting Trees (GBT) , Support Vector Machines (SVM) , and Feed-Forward Neural Networks (FFNN) , and three molecular fingerprints of different design used as compound representations: Molecular ACCess System (MACCS) Keys (166-bit), Morgan fingerprint with radius 2 (2048-bit), and RDK fingerprint (2048-bit). Each model is denoted as a combination of fingerprint and algorithm (fingerprint::algorithm). For each algorithm and target, hyperparameters were optimized from an exhaustive search detailed in the Experimental Section, using the mean balanced accuracy (BA) over a 10-fold cross-validation as the performance metric to select the best set of hyperparameters. Figure shows the number of targets for which each model was identified as the best performing, considering the mean BA over the ten folds as a point metric. Under this approach, there is no model, fingerprint, nor machine learning algorithm that could be identified as the best performing for all 55 target datasets considered in this work. Figure shows that RDK::GBT had the highest mean BA for 14 out of the 55 targets, making them the most frequent choice. However, in terms of compound representations only, Morgan fingerprint was the best choice for 28 targets, followed by 24 for RDK fingerprint and three for MACCS. Nevertheless, t-tests comparing the sets of BA scores calculated from the ten validation folds revealed that for all the targets, there is at least another model with no significant difference of performance to the one with the highest mean BA (Table in the Supporting Information). Moreover, the t-test comparison revealed that for 35 out of the 55 targets, there are at least 9 other models with no significant difference of performance to the one with the highest BA (at 95% confidence level), a surprising quantity considering the number of algorithms and compound representations included.
|
60c753a5f96a001b412884da
| 6 |
To compare the models herein generated in a more global context, the cross-validated predictions for each optimized model were stored and used to compute single point performance metrics in the context of each target, being BA, F1 score, and Mathews correlation coefficient (MMC). Summary results of the fifteen models' performance are summarized in Table , and their distribution across the 55 epigenetic targets is shown in Figure . Overall, most of the models performed well in the single-target prediction task, having a mean BA and F1 score higher than 0.5 and mean MCC higher than zero. To identify the global best performing model, we applied Wilcoxon signed-rank tests between all pairs of models for the three metrics of performance. Each test involves a comparison between sets of 55 values.
|
60c753a5f96a001b412884da
| 7 |
The Morgan::SVM model showed the highest mean values for the three performance metrics and significantly higher values of BA and MCC when compared to all but the RDK::SVM model (at 95% confidence level). F1 score showed the lower differences between models, with the Morgan::SVM having the highest mean value and significantly higher values when compared to all but five models, being RDK::SVM, RDK::GBT, RDK::FNN, RDK::RF and Morgan::k-NN (at 95% confidence level). These results suggested Morgan and RDK fingerprints and the SVM algorithm as the best combinations to derive binary classifiers for the current sets of studied epigenetic targets.
|
60c753a5f96a001b412884da
| 8 |
It has been pointed out that the combination of predictive models generally has a higher reliability than the individual models. In other to identify the best models combination to construct a consensus model, we performed a hierarchical clustering of the models relying on Morgan and RDK fingerprints by comparing their 38129 crossvalidated predictions obtained in the single target validation strategy (vide supra). Jaccard distance was employed as the metric between models and an average linkage was used for the hierarchical clustering calculation as detailed in the Experimental Section. Figure depicts a dendrogram of the hierarchical clustering. Predictions for all models are closely related, with all average distances between groups being lower than 0.1. It should be noted that models relying on the same fingerprint are clustered together before being grouped with models built on a different fingerprint. In the context of each fingerprint, the clustering follows the same order, with models relying on GBT and RF being grouped at first, followed by those built on SVM, FFNN, and k-NN. Based on these findings, the best performing model built on each fingerprint, Morgan::SVM and RDK::SVM, were combined to derive a consensus model. To prioritize the correct identification of active compounds, the consensus model was constructed by combining the predictions of both models so that a compound was predicted as "active" for a given target only if both models agreed in the prediction and "inactive" otherwise. This consensus model showed a mean BA, F1 score, and MCC of 0.835, 0.851, and 0.676, respectively. Wilcoxon signed-rank tests indicated significantly lower values for F1 score than those obtained by the individual models, and no significant difference for BA and MCC values (at 95% confidence level). Since F1 score is defined as the harmonic mean of precision (PPV) and recall (TPR)
|
60c753a5f96a001b412884da
| 9 |
for the active class, and the consensus model was a priori built to have high precision, the significantly lower values obtained for F1 score are explained by a decrease in the TPR of the model (Table and Figure in the Supporting Information), which is related to the decrease in the number of "active" outcomes for the consensus model.
|
60c753a5f96a001b412884da
| 10 |
Although BA, F1 score, and MCC are well-suited metrics for model performance estimation on imbalanced datasets, in a practical medicinal chemistry application, the correct identification of active compounds is often more important than the correct identification of inactive ones. To this end, the performance of the individual models and the derived consensus model were studied in terms of PPV, TPR, negative predictive value (NPV), and true negative rate (TNR). To estimate the models' applicability domain, these metrics were computed on a distance-to-model (DM) basis as detailed in the Experimental Section. All cross-validated predictions were categorized into four quartiles (Q1-Q4) according to their mean Jaccard distances to the training set in the context of each target (Table in the Supporting Information). Summary results of the three models' performance are presented in Table , and their distribution across the 55 epigenetic targets is shown in Figure . , and their distribution across the 10 samples is shown in Figure . Under this validation strategy, PPV and TPR showed the same trends as in the single target validation: both decreased as the DM increased for all models. Wilcoxon signed-rank tests indicated significantly lower TPR values and higher PPV values for all quartiles when comparing the consensus model to any of the two individual models (at 95% confidence level)
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.