Automating biomedical literature review for rapid drug discovery: Leveraging GPT-4 to expedite pandemic response.

OBJECTIVE: The rapid expansion of the biomedical literature challenges traditional review methods, especially during outbreaks of emerging infectious diseases when quick action is critical. Our study aims to explore the potential of ChatGPT to automate the biomedical literature review for rapid drug discovery. MATERIALS AND METHODS: We introduce a novel automated pipeline helping to identify drugs for a given virus in response to a potential future global health threat. Our approach can be used to select PubMed articles identifying a drug target for the given virus. We tested our approach on two known pathogens: SARS-CoV-2, where the literature is vast, and Nipah, where the literature is sparse. Specifically, a panel of three experts reviewed a set of PubMed articles and labeled them as either describing a drug target for the given virus or not. The same task was given to the automated pipeline and its performance was based on whether it labeled the articles similarly to the human experts. We applied a number of prompt engineering techniques to improve the performance of ChatGPT. RESULTS: Our best configuration used GPT-4 by OpenAI and achieved an out-of-sample validation performance with accuracy/F1-score/sensitivity/specificity of 92.87%/88.43%/83.38%/97.82% for SARS-CoV-2 and 87.40%/73.90%/74.72%/91.36% for Nipah. CONCLUSION: These results highlight the utility of ChatGPT in drug discovery and development and reveal their potential to enable rapid drug target identification during a pandemic-level health emergency.

Expanding FTMap for Fragment-Based Identification of Pharmacophore Regions in Ligand Binding Sites.

The knowledge of ligand binding hot spots and of the important interactions within such hot spots is crucial for the design of lead compounds in the early stages of structure-based drug discovery. The computational solvent mapping server FTMap can reliably identify binding hot spots as consensus clusters, free energy minima that bind a variety of organic probe molecules. However, in its current implementation, FTMap provides limited information on regions within the hot spots that tend to interact with specific pharmacophoric features of potential ligands. E-FTMap is a new server that expands on the original FTMap protocol. E-FTMap uses 119 organic probes, rather than the 16 in the original FTMap, to exhaustively map binding sites, and identifies pharmacophore features as atomic consensus sites where similar chemical groups bind. We validate E-FTMap against a set of 109 experimentally derived structures of fragment-lead pairs, finding that highly ranked pharmacophore features overlap with the corresponding atoms in both fragments and lead compounds. Additionally, comparisons of mapping results to ensembles of bound ligands reveal that pharmacophores generated with E-FTMap tend to sample highly conserved protein-ligand interactions. E-FTMap is available as a web server at https://eftmap.bu.edu.

Computational identification of antibody-binding epitopes from mimotope datasets.

Introduction: A fundamental challenge in computational vaccinology is that most B-cell epitopes are conformational and therefore hard to predict from sequence alone. Another significant challenge is that a great deal of the amino acid sequence of a viral surface protein might not in fact be antigenic. Thus, identifying the regions of a protein that are most promising for vaccine design based on the degree of surface exposure may not lead to a clinically relevant immune response. Methods: Linear peptides selected by phage display experiments that have high affinity to the monoclonal antibody of interest ("mimotopes") usually have similar physicochemical properties to the antigen epitope corresponding to that antibody. The sequences of these linear peptides can be used to find possible epitopes on the surface of the antigen structure or a homology model of the antigen in the absence of an antigen-antibody complex structure. Results and Discussion: Herein we describe two novel methods for mapping mimotopes to epitopes. The first is a novel algorithm named MimoTree that allows for gaps in the mimotopes and epitopes on the antigen. More specifically, a mimotope may have a gap that does not match to the epitope to allow it to adopt a conformation relevant for binding to an antibody, and residues may similarly be discontinuous in conformational epitopes. MimoTree is a fully automated epitope detection algorithm suitable for the identification of conformational as well as linear epitopes. The second is an ensemble approach, which combines the prediction results from MimoTree and two existing methods.

Identification of a druggable site on GRP78 at the GRP78-SARS-CoV-2 interface and virtual screening of compounds to disrupt that interface.

SARS-CoV-2, the virus that causes COVID-19, led to a global health emergency that claimed the lives of millions. Despite the widespread availability of vaccines, the virus continues to exist in the population in an endemic state which allows for the continued emergence of new variants. Most of the current vaccines target the spike glycoprotein interface of SARS-CoV-2, creating a selection pressure favoring viral immune evasion. Antivirals targeting other molecular interactions of SARS-CoV-2 can help slow viral evolution by providing orthogonal selection pressures on the virus. GRP78 is a host auxiliary factor that mediates binding of the SARS-CoV-2 spike protein to human cellular ACE2, the primary pathway of cell infection. As GRP78 forms a ternary complex with SARS-CoV-2 spike protein and ACE2, disrupting the formation of this complex is expected to hinder viral entry into host cells. Here, we developed a model of the GRP78-Spike RBD-ACE2 complex. We then used that model together with hot spot mapping of the GRP78 structure to identify the putative binding site for spike protein on GRP78. Next, we performed structure-based virtual screening of known drug/candidate drug libraries to identify binders to GRP78 that are expected to disrupt spike protein binding to the GRP78, and thereby preventing viral entry to the host cell. A subset of these compounds has previously been shown to have some activity against SARS-CoV-2. The identified hits are starting points for the further development of novel SARS-CoV-2 therapeutics, potentially serving as proof-of-concept for GRP78 as a potential drug target for other viruses.

Conservation of Hot Spots and Ligand Binding Sites in Protein Models by AlphaFold2.

The neural network-based program AlphaFold2 (AF2) provides high accuracy structure prediction for a large fraction of globular proteins. An important question is whether these models are accurate enough for reliably docking small ligands. Several recent papers and the results of CASP15 reveal that local conformational errors reduce the success rates of direct ligand docking. Here, we focus on the ability of the models to conserve the location of binding hot spots, regions on the protein surface that significantly contribute to the binding free energy of the protein-ligand interaction. Clusters of hot spots predict the location and even the druggability of binding sites, and hence are important for computational drug discovery. The hot spots are determined by protein mapping that is based on the distribution of small fragment-sized probes on the protein surface and is less sensitive to local conformation than docking. Mapping models taken from the AlphaFold Protein Structure Database show that identifying binding sites is more reliable than docking, but the success rates are still 5% to 10% lower than based on mapping X-ray structures. The drop in accuracy is particularly large for models of multidomain proteins. However, both the model binding sites and the mapping results can be substantially improved by generating AF2 models for the ligand binding domains of interest rather than the entire proteins and even more if using forced sampling with multiple initial seeds. The mapping of such models tends to reach the accuracy of results obtained by mapping the X-ray structures.

Vaccines Alone Cannot Slow the Evolution of SARS-CoV-2.

The rapid emergence of immune-evading viral variants of SARS-CoV-2 calls into question the practicality of a vaccine-only public-health strategy for managing the ongoing COVID-19 pandemic. It has been suggested that widespread vaccination is necessary to prevent the emergence of future immune-evading mutants. Here, we examined that proposition using stochastic computational models of viral transmission and mutation. Specifically, we looked at the likelihood of emergence of immune escape variants requiring multiple mutations and the impact of vaccination on this process. Our results suggest that the transmission rate of intermediate SARS-CoV-2 mutants will impact the rate at which novel immune-evading variants appear. While vaccination can lower the rate at which new variants appear, other interventions that reduce transmission can also have the same effect. Crucially, relying solely on widespread and repeated vaccination (vaccinating the entire population multiple times a year) is not sufficient to prevent the emergence of novel immune-evading strains, if transmission rates remain high within the population. Thus, vaccines alone are incapable of slowing the pace of evolution of immune evasion, and vaccinal protection against severe and fatal outcomes for COVID-19 patients is therefore not assured.

Machine learning for the identification of respiratory viral attachment machinery from sequences data.

At the outset of an emergent viral respiratory pandemic, sequence data is among the first molecular information available. As viral attachment machinery is a key target for therapeutic and prophylactic interventions, rapid identification of viral "spike" proteins from sequence can significantly accelerate the development of medical countermeasures. For six families of respiratory viruses, covering the vast majority of airborne and droplet-transmitted diseases, host cell entry is mediated by the binding of viral surface glycoproteins that interact with a host cell receptor. In this report it is shown that sequence data for an unknown virus belonging to one of the six families above provides sufficient information to identify the protein(s) responsible for viral attachment. Random forest models that take as input a set of respiratory viral sequences can classify the protein as "spike" vs. non-spike based on predicted secondary structure elements alone (with 97.3% correctly classified) or in combination with N-glycosylation related features (with 97.0% correctly classified). Models were validated through 10-fold cross-validation, bootstrapping on a class-balanced set, and an out-of-sample extra-familial validation set. Surprisingly, we showed that secondary structural elements and N-glycosylation features were sufficient for model generation. The ability to rapidly identify viral attachment machinery directly from sequence data holds the potential to accelerate the design of medical countermeasures for future pandemics. Furthermore, this approach may be extendable for the identification of other potential viral targets and for viral sequence annotation in general in the future.

No magic bullet: Limiting in-school transmission in the face of variable SARS-CoV-2 viral loads.

In the face of a long-running pandemic, understanding the drivers of ongoing SARS-CoV-2 transmission is crucial for the rational management of COVID-19 disease burden. Keeping schools open has emerged as a vital societal imperative during the pandemic, but in-school transmission of SARS-CoV-2 can contribute to further prolonging the pandemic. In this context, the role of schools in driving SARS-CoV-2 transmission acquires critical importance. Here we model in-school transmission from first principles to investigate the effectiveness of layered mitigation strategies on limiting in-school spread. We examined the effect of masks and air quality (ventilation, filtration and ionizers) on steady-state viral load in classrooms, as well as on the number of particles inhaled by an uninfected person. The effectiveness of these measures in limiting viral transmission was assessed for variants with different levels of mean viral load (ancestral, Delta, Omicron). Our results suggest that a layered mitigation strategy can be used effectively to limit in-school transmission, with certain limitations. First, poorly designed strategies (insufficient ventilation, no masks, staying open under high levels of community transmission) will permit in-school spread even if some level of mitigation is present. Second, for viral variants that are sufficiently contagious, it may be difficult to construct any set of interventions capable of blocking transmission once an infected individual is present, underscoring the importance of other measures. Our findings provide practical recommendations; in particular, the use of a layered mitigation strategy that is designed to limit transmission, with other measures such as frequent surveillance testing and smaller class sizes (such as by offering remote schooling options to those who prefer it) as needed.

Rapid relaxation of pandemic restrictions after vaccine rollout favors growth of SARS-CoV-2 variants: A model-based analysis.

The development and deployment of several SARS-CoV-2 vaccines in a little over a year is an unprecedented achievement of modern medicine. The high levels of efficacy against transmission for some of these vaccines makes it feasible to use them to suppress SARS-CoV-2 altogether in regions with high vaccine acceptance. However, viral variants with reduced susceptibility to vaccinal and natural immunity threaten the utility of vaccines, particularly in scenarios where a return to pre-pandemic conditions occurs before the suppression of SARS-CoV-2 transmission. In this work we model the situation in the United States in May-June 2021, to demonstrate how pre-existing variants of SARS-CoV-2 may cause a rebound wave of COVID-19 in a matter of months under a certain set of conditions. A high burden of morbidity (and likely mortality) remains possible, even if the vaccines are partially effective against new variants and widely accepted. Our modeling suggests that variants that are already present within the population may be capable of quickly defeating the vaccines as a public health intervention, a serious potential limitation for strategies that emphasize rapid reopening before achieving control of SARS-CoV-2.

Controlling long-term SARS-CoV-2 infections can slow viral evolution and reduce the risk of treatment failure.

The rapid emergence and expansion of novel SARS-CoV-2 variants threatens our ability to achieve herd immunity for COVID-19. These novel SARS-CoV-2 variants often harbor multiple point mutations, conferring one or more evolutionarily advantageous traits, such as increased transmissibility, immune evasion and longer infection duration. In a number of cases, variant emergence has been linked to long-term infections in individuals who were either immunocompromised or treated with convalescent plasma. In this paper, we used a stochastic evolutionary modeling framework to explore the emergence of fitter variants of SARS-CoV-2 during long-term infections. We found that increased viral load and infection duration favor emergence of such variants. While the overall probability of emergence and subsequent transmission from any given infection is low, on a population level these events occur fairly frequently. Targeting these low-probability stochastic events that lead to the establishment of novel advantageous viral variants might allow us to slow the rate at which they emerge in the patient population, and prevent them from spreading deterministically due to natural selection. Our work thus suggests practical ways to achieve control of long-term SARS-CoV-2 infections, which will be critical for slowing the rate of viral evolution.

Risk of rapid evolutionary escape from biomedical interventions targeting SARS-CoV-2 spike protein.

The spike protein receptor-binding domain (RBD) of SARS-CoV-2 is the molecular target for many vaccines and antibody-based prophylactics aimed at bringing COVID-19 under control. Such a narrow molecular focus raises the specter of viral immune evasion as a potential failure mode for these biomedical interventions. With the emergence of new strains of SARS-CoV-2 with altered transmissibility and immune evasion potential, a critical question is this: how easily can the virus escape neutralizing antibodies (nAbs) targeting the spike RBD? To answer this question, we combined an analysis of the RBD structure-function with an evolutionary modeling framework. Our structure-function analysis revealed that epitopes for RBD-targeting nAbs overlap one another substantially and can be evaded by escape mutants with ACE2 affinities comparable to the wild type, that are observed in sequence surveillance data and infect cells in vitro. This suggests that the fitness cost of nAb-evading mutations is low. We then used evolutionary modeling to predict the frequency of immune escape before and after the widespread presence of nAbs due to vaccines, passive immunization or natural immunity. Our modeling suggests that SARS-CoV-2 mutants with one or two mildly deleterious mutations are expected to exist in high numbers due to neutral genetic variation, and consequently resistance to vaccines or other prophylactics that rely on one or two antibodies for protection can develop quickly -and repeatedly- under positive selection. Predicted resistance timelines are comparable to those of the decay kinetics of nAbs raised against vaccinal or natural antigens, raising a second potential mechanism for loss of immunity in the population. Strategies for viral elimination should therefore be diversified across molecular targets and therapeutic modalities.

An Antipersister Strategy for Treatment of Chronic Pseudomonas aeruginosa Infections.

Bacterial persisters are a quasidormant subpopulation of cells that are tolerant to antibiotic treatment. The combination of the aminoglycoside tobramycin with fumarate as an antibacterial potentiator utilizes an antipersister strategy that is aimed at reducing recurrent Pseudomonas aeruginosa infections by enhancing the killing of P. aeruginosa persisters. Stationary-phase cultures of P. aeruginosa were used to generate persister cells. A range of tobramycin concentrations was tested with a range of metabolite concentrations to determine the potentiation effect of the metabolite under a variety of conditions, including a range of pH values and in the presence of azithromycin or cystic fibrosis (CF) patient sputum. In addition, 96-well dish biofilm and colony biofilm assays were performed, and the cytotoxicity of the tobramycin-fumarate combination was determined utilizing a lactate dehydrogenase (LDH) assay. Enhanced killing of up to 6 orders of magnitude of P. aeruginosa persisters over a range of CF isolates, including mucoid and nonmucoid strains, was observed for the tobramycin-fumarate combination compared to killing with tobramycin alone. Furthermore, significant fumarate-mediated potentiation was seen in the presence of azithromycin or CF patient sputum. Fumarate also reduced the cytotoxicity of tobramycin-treated P. aeruginosa to human epithelial airway cells. Finally, in mucoid and nonmucoid CF isolates, complete eradication of P. aeruginosa biofilm was observed in the colony biofilm assay due to fumarate potentiation. These data suggest that a combination of tobramycin with fumarate as an antibacterial potentiator may be an attractive therapeutic for eliminating recurrent P. aeruginosa infections in CF patients through the eradication of bacterial persisters.

Evaluating Free Energies of Binding and Conservation of Crystallographic Waters Using SZMAP.

The SZMAP method computes binding free energies and the corresponding thermodynamic components for water molecules in the binding site of a protein structure [ SZMAP, 1.0.0 ; OpenEye Scientific Software Inc. : Santa Fe, NM, USA , 2011 ]. In this work, the ability of SZMAP to predict water structure and thermodynamic stability is examined for the X-ray crystal structures of a series of protein-ligand complexes. SZMAP results correlate with higher-level replica exchange thermodynamic integration double decoupling calculations of the absolute free energy of bound waters in the test set complexes. In addition, SZMAP calculations show good agreement with experimental data in terms of water conservation (across multiple crystal structures) and B-factors over a subset of the test set. In particular, the SZMAP neutral entropy difference term calculated at crystallographic water positions within each of the complex structures correlates well with whether that crystallographic water is conserved or displaceable. Furthermore, the calculated entropy of the water probe relative to the continuum shows a significant degree of correlation with the B-factors associated with the oxygen atoms of the water molecules. Taken together, these results indicate that SZMAP is capable of quantitatively predicting water positions and their energetics and is potentially a useful tool for determining which waters to attempt to displace, maintain, or build in through water-mediated interactions when evolving a lead series during a drug discovery program.

Ensemble-based docking using biased molecular dynamics.

Proteins are dynamic molecules, and understanding their movements, especially as they relate to molecular recognition and protein-ligand interactions, poses a significant challenge to structure-based drug discovery. In most instances, protein flexibility is underrepresented in computer-aided drug design due to uncertainties on how it should be accurately modeled as well as the computational cost associated with attempting to incorporate flexibility in the calculations. One approach that aims to address these issues is ensemble-based docking. With this technique, ligands are docked to an ensemble of rigid protein conformations. Molecular dynamics (MD) simulations can be used to generate the ensemble of protein conformations for the subsequent docking. Here we present a novel approach that uses biased-MD simulations to generate the docking ensemble. The MD simulations are biased toward an initial protein-ligand X-ray complex structure. The biasing maintains some of the original crystallographic pocket-ligand information and thereby enhances sampling of the more relevant conformational space of the protein. Resulting trajectories are clustered to select a representative set of protein conformations, and ligands are docked to that reduced set of conformations. Cross-docking to this ensemble and then selecting the lowest scoring pose enables reliable identification of the correct binding mode. Various levels of biasing are investigated, and the method is validated for cyclin-dependent kinase 2 and factor Xa.

Fragment-based lead discovery and design.

Fragment-based lead discovery and design has and continues to show increasing promise in drug discovery. In this article, the current state of the art in terms of hot-spot characterization, fragment screening techniques, and fragment-based design is discussed. Three overall fragment-based lead generation strategies are explored and involve the chemical biology characterization of biological targets via fragment screening, fragment screening as a complementary approach to high-throughput screening of drug-like compounds, and direct fragment-based drug discovery, respectively. The evolution and development of fragment libraries is described. With an emphasis on computational approaches and the strategies applied at AstraZeneca, the review illustrates how integration of data from one regime can inform the design of experiments in the other, ultimately leading to the discovery of high quality chemical matter.

Discovery of selective and potent inhibitors of gram-positive bacterial thymidylate kinase (TMK).

Thymidylate kinase (TMK) is an essential enzyme in bacterial DNA synthesis. The deoxythymidine monophosphate (dTMP) substrate binding pocket was targeted in a rational-design, structure-supported effort, yielding a unique series of antibacterial agents showing a novel, induced-fit binding mode. Lead optimization, aided by X-ray crystallography, led to picomolar inhibitors of both Streptococcus pneumoniae and Staphylococcus aureus TMK. MICs < 1 µg/mL were achieved against methicillin-resistant S. aureus (MRSA), S. pneumoniae, and vancomycin-resistant Enterococcus (VRE). Log D adjustments yielded single diastereomers 14 (TK-666) and 46, showing a broad antibacterial spectrum against Gram-positive bacteria and excellent selectivity against the human thymidylate kinase ortholog.

Virtual fragment screening: exploration of MM-PBSA re-scoring.

An NMR fragment screening dataset with known binders and decoys was used to evaluate the ability of docking and re-scoring methods to identify fragment binders. Re-scoring docked poses using the Molecular Mechanics Poisson-Boltzmann Surface Area (MM-PBSA) implicit solvent model identifies additional active fragments relative to either docking or random fragment screening alone. Early enrichment, which is clearly most important in practice for selecting relatively small sets of compounds for experimental testing, is improved by MM-PBSA re-scoring. In addition, the value in MM-PBSA re-scoring of docked poses for virtual screening may be in lessening the effect of the variation in the protein complex structure used.

In vivo validation of thymidylate kinase (TMK) with a rationally designed, selective antibacterial compound.

There is an urgent need for new antibacterials that pinpoint novel targets and thereby avoid existing resistance mechanisms. We have created novel synthetic antibacterials through structure-based drug design that specifically target bacterial thymidylate kinase (TMK), a nucleotide kinase essential in the DNA synthesis pathway. A high-resolution structure shows compound TK-666 binding partly in the thymidine monophosphate substrate site, but also forming new induced-fit interactions that give picomolar affinity. TK-666 has potent, broad-spectrum Gram-positive microbiological activity (including activity against methicillin-resistant Staphylococcus aureus and vancomycin-resistant Enterococcus), bactericidal action with rapid killing kinetics, excellent target selectivity over the human ortholog, and low resistance rates. We demonstrate in vivo efficacy against S. aureus in a murine infected-thigh model. This work presents the first validation of TMK as a compelling antibacterial target and provides a rationale for pursuing novel clinical candidates for treating Gram-positive infections through TMK.

Identification and characterization of acidic mammalian chitinase inhibitors.

Acidic mammalian chitinase (AMCase) is a member of the glycosyl hydrolase 18 family (EC 3.2.1.14) that has been implicated in the pathophysiology of allergic airway disease such as asthma. Small molecule inhibitors of AMCase were identified using a combination of high-throughput screening, fragment screening, and virtual screening techniques and characterized by enzyme inhibition and NMR and Biacore binding experiments. X-ray structures of the inhibitors in complex with AMCase revealed that the larger more potent HTS hits, e.g. 5-(4-(2-(4-bromophenoxy)ethyl)piperazine-1-yl)-1H-1,2,4-triazol-3-amine 1, spanned from the active site pocket to a hydrophobic pocket. Smaller fragments identified by FBS occupy both these pockets independently and suggest potential strategies for linking fragments. Compound 1 is a 200 nM AMCase inhibitor which reduced AMCase enzymatic activity in the bronchoalveolar lavage fluid in allergen-challenged mice after oral dosing.