Performance Evaluation of 2D Fingerprint and 3D Shape Similarity Methods in Virtual Screening

iSIM: instant similarity

The quantification of molecular similarity has been present since the beginning of cheminformatics. Although several similarity indices and molecular representations have been reported, all of them ultimately reduce to the calculation of molecular similarities of only two objects at a time. Hence, to obtain the average similarity of a set of molecules, all the pairwise comparisons need to be computed, which demands a quadratic scaling in the number of computational resources. Here we propose an exact alternative to this problem: iSIM (instant similarity). iSIM performs comparisons of multiple molecules at the same time and yields the same value as the average pairwise comparisons of molecules represented by binary fingerprints and real-value descriptors. In this work, we introduce the mathematical framework and several applications of iSIM in chemical sampling, visualization, diversity selection, and clustering.

Metabolic Connectome and Its Role in the Prediction, Diagnosis, and Treatment of Complex Diseases

The interconnectivity of advanced biological systems is essential for their proper functioning. In modern connectomics, biological entities such as proteins, genes, RNA, DNA, and metabolites are often represented as nodes, while the physical, biochemical, or functional interactions between them are represented as edges. Among these entities, metabolites are particularly significant as they exhibit a closer relationship to an organism's phenotype compared to genes or proteins. Moreover, the metabolome has the ability to amplify small proteomic and transcriptomic changes, even those from minor genomic changes. Metabolic networks, which consist of complex systems comprising hundreds of metabolites and their interactions, play a critical role in biological research by mediating energy conversion and chemical reactions within cells. This review provides an introduction to common metabolic network models and their construction methods. It also explores the diverse applications of metabolic networks in elucidating disease mechanisms, predicting and diagnosing diseases, and facilitating drug development. Additionally, it discusses potential future directions for research in metabolic networks. Ultimately, this review serves as a valuable reference for researchers interested in metabolic network modeling, analysis, and their applications.

Network-Based Methods and Their Applications in Drug Discovery

Drug discovery is time-consuming, expensive, and predominantly follows the "one drug → one target → one disease" paradigm. With the rapid development of systems biology and network pharmacology, a novel drug discovery paradigm, "multidrug → multitarget → multidisease", has emerged. This new holistic paradigm of drug discovery aligns well with the essence of networks, leading to the emergence of network-based methods in the field of drug discovery. In this Perspective, we initially introduce the concept and data sources of networks and highlight classical methodologies employed in network-based methods. Subsequently, we focus on the practical applications of network-based methods across various areas of drug discovery, such as target prediction, virtual screening, prediction of drug therapeutic effects or adverse drug events, and elucidation of molecular mechanisms. In addition, we provide representative web servers for researchers to use network-based methods in specific applications. Finally, we discuss several challenges of network-based methods and the directions for future development. In a word, network-based methods could serve as powerful tools to accelerate drug discovery.

Repurposing Drugs for Inhibition against ALDH2 via a 2D/3D Ligand-Based Similarity Search and Molecular Simulation

Aldehyde dehydrogenase-2 (ALDH2) is a crucial enzyme participating in intracellular aldehyde metabolism and is acknowledged as a potential therapeutic target for the treatment of alcohol use disorder and other addictive behaviors. Using previously reported ALDH2 inhibitors of Daidzin, CVT-10216, and CHEMBL114083 as reference molecules, here we perform a ligand-based virtual screening of world-approved drugs via 2D/3D similarity search methods, followed by the assessments of molecular docking, toxicity prediction, molecular simulation, and the molecular mechanics Poisson-Boltzmann surface area (MM-PBSA) analysis. The 2D molecular fingerprinting of ECFP4 and FCFP4 and 3D molecule-shape-based USRCAT methods show good performances in selecting compounds with a strong binding behavior with ALDH2. Three compounds of Zeaxanthin (q = 0), Troglitazone (q = 0), and Sequinavir (q = +1 e) are singled out as potential inhibitors; Zeaxanthin can only be hit via USRCAT. These drugs displayed a stronger binding strength compared to the reported potent inhibitor CVT-10216. Sarizotan (q = +1 e) and Netarsudil (q = 0/+1 e) displayed a strong binding strength with ALDH2 as well, whereas they displayed a shallow penetration into the substrate-binding tunnel of ALDH2 and could not fully occupy it. This likely left a space for substrate binding, and thus they were not ideal inhibitors. The MM-PBSA results indicate that the selected negatively charged compounds from the similarity search and Vina scoring are thermodynamically unfavorable, mainly due to electrostatic repulsion with the receptor (q = -6 e for ALDH2). The electrostatic attraction with positively charged compounds, however, yielded very strong binding results with ALDH2. These findings reveal a deficiency in the modeling of electrostatic interactions (in particular, between charged moieties) in the virtual screening via the 2D/3D similarity search and molecular docking with the Vina scoring system.

High-Quality Conformer Generation with CONFORGE: Algorithm and Performance Assessment

Knowledge of the putative bound-state conformation of a molecule is an essential prerequisite for the successful application of many computer-aided drug design methods that aim to assess or predict its capability to bind to a particular target receptor. An established approach to predict bioactive conformers in the absence of receptor structure information is to sample the low-energy conformational space of the investigated molecules and derive representative conformer ensembles that can be expected to comprise members closely resembling possible bound-state ligand conformations. The high relevance of such conformer generation functionality led to the development of a wide panel of dedicated commercial and open-source software tools throughout the last decades. Several published benchmarking studies have shown that open-source tools usually lag behind their commercial competitors in many key aspects. In this work, we introduce the open-source conformer ensemble generator CONFORGE, which aims at delivering state-of-the-art performance for all types of organic molecules in drug-like chemical space. The ability of CONFORGE and several well-known commercial and open-source conformer ensemble generators to reproduce experimental 3D structures as well as their computational efficiency and robustness has been assessed thoroughly for both typical drug-like molecules and macrocyclic structures. For small molecules, CONFORGE clearly outperformed all other tested open-source conformer generators and performed at least equally well as the evaluated commercial generators in terms of both processing speed and accuracy. In the case of macrocyclic structures, CONFORGE achieved the best average accuracy among all benchmarked generators, with RDKit's generator coming close in second place.

Pharmacophore-Based Machine Learning Model To Predict Ligand Selectivity for E3 Ligase Binders

E3 ligases are enzymes that play a critical role in ubiquitin-mediated protein degradation and are involved in various cellular processes. Pharmacophore analysis is a useful approach for predicting E3 ligase binding selectivity, which involves identifying key chemical features necessary for a ligand to interact with a specific protein target cavity. While pharmacophore analysis is not always sufficient to accurately predict ligand binding affinity, it can be a valuable tool for filtering and/or designing focused libraries for screening campaigns. In this study, we present a fast and an inexpensive approach using a pharmacophore fingerprinting scheme known as ErG, which is used in a multi-class machine learning classification model. This model can assign the correct E3 ligase binder to its known E3 ligase and predict the probability of each molecule to bind to different E3 ligases. Practical applications of this approach are demonstrated on commercial libraries such as Asinex for the rational design of E3 ligase binders. The scripts and data associated with this study can be found on GitHub at https://github.com/Fraunhofer-ITMP/E3_binder_Model.

Employing Molecular Conformations for Ligand-Based Virtual Screening with Equivariant Graph Neural Network and Deep Multiple Instance Learning

Ligand-based virtual screening (LBVS) is a promising approach for rapid and low-cost screening of potentially bioactive molecules in the early stage of drug discovery. Compared with traditional similarity-based machine learning methods, deep learning frameworks for LBVS can more effectively extract high-order molecule structure representations from molecular fingerprints or structures. However, the 3D conformation of a molecule largely influences its bioactivity and physical properties, and has rarely been considered in previous deep learning-based LBVS methods. Moreover, the relative bioactivity benchmark dataset is still lacking. To address these issues, we introduce a novel end-to-end deep learning architecture trained from molecular conformers for LBVS. We first extracted molecule conformers from multiple public molecular bioactivity data and consolidated them into a large-scale bioactivity benchmark dataset, which totally includes millions of endpoints and molecules corresponding to 954 targets. Then, we devised a deep learning-based LBVS called EquiVS to learn molecule representations from conformers for bioactivity prediction. Specifically, graph convolutional network (GCN) and equivariant graph neural network (EGNN) are sequentially stacked to learn high-order molecule-level and conformer-level representations, followed with attention-based deep multiple-instance learning (MIL) to aggregate these representations and then predict the potential bioactivity for the query molecule on a given target. We conducted various experiments to validate the data quality of our benchmark dataset, and confirmed EquiVS achieved better performance compared with 10 traditional machine learning or deep learning-based LBVS methods. Further ablation studies demonstrate the significant contribution of molecular conformation for bioactivity prediction, as well as the reasonability and non-redundancy of deep learning architecture in EquiVS. Finally, a model interpretation case study on CDK2 shows the potential of EquiVS in optimal conformer discovery. The overall study shows that our proposed benchmark dataset and EquiVS method have promising prospects in virtual screening applications.

Updates on drug designing approach through computational strategies: a review

The drug discovery and development (DDD) process in pursuit of novel drug candidates is a challenging procedure requiring lots of time and resources. Therefore, computer-aided drug design (CADD) methodologies are used extensively to promote proficiency in drug development in a systematic and time-effective manner. The point in reference is SARS-CoV-2 which has emerged as a global pandemic. In the absence of any confirmed drug moiety to treat the infection, the science fraternity adopted hit and trial methods to come up with a lead drug compound. This article is an overview of the virtual methodologies, which assist in finding novel hits and help in the progression of drug development in a short period with a specific medicinal solution.

Virtual Screening-Based Drug Development for the Treatment of Nervous System Diseases

The incidence rate of nervous system diseases has increased in recent years. Nerve injury or neurodegenerative diseases usually cause neuronal loss and neuronal circuit damage, which seriously affect motor nerve and autonomic nervous function. Therefore, safe and effective treatment is needed. As traditional drug research becomes slower and more expensive, it is vital to enlist the help of cutting- edge technology. Virtual screening (VS) is an attractive option for the identification and development of promising new compounds with high efficiency and low cost. With the assistance of computer- aided drug design (CADD), VS is becoming more and more popular in new drug development and research. In recent years, it has become a reality to transform non-neuronal cells into functional neurons through small molecular compounds, which provides a broader application prospect than transcription factor-mediated neuronal reprogramming. This review mainly summarizes related theory and technology of VS and the drug research and development using VS technology in nervous system diseases in recent years, and focuses more on the potential application of VS technology in neuronal reprogramming, thus facilitating new drug design for both prevention and treatment of nervous system diseases.

Random-forest model for drug-target interaction prediction via Kullbeck-Leibler divergence

Virtual screening has significantly improved the success rate of early stage drug discovery. Recent virtual screening methods have improved owing to advances in machine learning and chemical information. Among these advances, the creative extraction of drug features is important for predicting drug–target interaction (DTI), which is a large-scale virtual screening of known drugs. Herein, we report Kullbeck–Leibler divergence (KLD) as a DTI feature and the feature-driven classification model applicable to DTI prediction. For the purpose, E3FP three-dimensional (3D) molecular fingerprints of drugs as a molecular representation allow the computation of 3D similarities between ligands within each target (Q–Q matrix) to identify the uniqueness of pharmacological targets and those between a query and a ligand (Q–L vector) in DTIs. The 3D similarity matrices are transformed into probability density functions via kernel density estimation as a nonparametric estimation. Each density model can exploit the characteristics of each pharmacological target and measure the quasi-distance between the ligands. Furthermore, we developed a random forest model from the KLD feature vectors to successfully predict DTIs for representative 17 targets (mean accuracy: 0.882, out-of-bag score estimate: 0.876, ROC AUC: 0.990). The method is applicable for 2D chemical similarity.

Multi-Level High-Throughput Screening for Discovery of Ligands That Inhibit Insulin Aggregation

We have developed a multi-level virtual screening protocol to identify lead molecules from the FDA inactives database that can inhibit insulin aggregation. The method is based on the presence of structural and interaction specificity in non-native aggregation pathway protein-protein interactions. Some key challenges specific to the present problem, when compared with native protein association, include structural heterogeneity of the protein species involved, multiple association pathways, and relatively higher probability of conformational rearrangement of the association complex. In this multi-step method, the inactives database was first screened using the dominant pharmacophore features of previously identified molecules shown to significantly inhibit insulin aggregation nucleation by binding to its aggregation-prone conformers. We then performed ensemble docking of several low-energy ligand conformations on these aggregation-prone conformers followed by molecular dynamics simulations and binding affinity calculations on a subset of docked complexes to identify a final set of five potential lead molecules to inhibit insulin aggregation nucleation. Their effect on aggregation inhibition was extensively investigated by incubating insulin under aggregation-prone aqueous buffer conditions (low pH, high temperature). Aggregation kinetics were characterized using size exclusion chromatography and Thioflavin T fluorescence assay, and the secondary structure was determined using circular dichroism spectroscopy. Riboflavin provided the best aggregation inhibition, with 85% native monomer retention after 48 h incubation under aggregation-prone conditions, whereas the no-ligand formulation showed complete monomer loss after 36 h. Further, insulin incubated with two of the screened inactives (aspartame, riboflavin) had the characteristic α-helical dip in CD spectra, while the no-ligand formulation showed a change to β-sheet rich conformations.

Maximizing the Performance of Similarity-Based Virtual Screening Methods by Generating Synergy from the Integration of 2D and 3D Approaches

Methods for the pairwise comparison of 2D and 3D molecular structures are established approaches in virtual screening. In this work, we explored three strategies for maximizing the virtual screening performance of these methods: (i) the merging of hit lists obtained from multi-compound screening using a single screening method, (ii) the merging of the hit lists obtained from 2D and 3D screening by parallel selection, and (iii) the combination of both of these strategies in an integrated approach. We found that any of these strategies led to a boost in virtual screening performance, with the clearest advantages observed for the integrated approach. On test sets for virtual screening, covering 50 pharmaceutically relevant proteins, the integrated approach, using sets of five query molecules, yielded, on average, an area under the receiver operating characteristic curve (AUC) of 0.84, an early enrichment among the top 1% of ranked compounds (EF1%) of 53.82 and a scaffold recovery rate among the top 1% of ranked compounds (SRR1%) of 0.50. In comparison, the 2D and 3D methods on their own (when using a single query molecule) yielded AUC values of 0.68 and 0.54, EF1% values of 19.96 and 17.52, and SRR1% values of 0.20 and 0.17, respectively. In conclusion, based on these results, the integration of 2D and 3D methods, via a (balanced) parallel selection strategy, is recommended, and, in particular, when combined with multi-query screening.

Discovery of New Estrogen-Related Receptor α Agonists via a Combination Strategy Based on Shape Screening and Ensemble Docking

Estrogen-related receptor α (ERRα), a member of nuclear receptors (NRs), plays a role in the regulation of cellular energy metabolism and is reported to be a novel potential target for type 2 diabetes therapy. To date, only a few agonists of ERRα have been identified to improve insulin sensitivity and decrease blood glucose levels. Herein, the discovery of novel potent agonists of ERRα determined using a combined virtual screening approach is described. Molecular dynamics (MD) simulations were used to obtain structural ensembles that can consider receptor flexibility. Then, an efficient virtual screening strategy with a combination of similarity search and ensemble docking was performed against the Enamine, SPECS, and Drugbank databases to identify potent ERRα agonists. Finally, a total of 66 compounds were purchased for experimental testing. Biological investigation of promising candidates identified seven compounds that have activity against ERRα with EC50 values ranging from 1.11 to 21.70 μM, with novel scaffolds different from known ERRα agonists until now. Additionally, the molecule GX66 showed micromolar inverse activity against ERRα with an IC50 of 0.82 μM. The predicted binding modes showed that these compounds were anchored in ERRα-LBP via interactions with several residues of ERRα. Overall, this study not only identified the novel potent ERRα agonists or an inverse agonist that would be the promising starting point for further exploration but also demonstrated a successful molecular dynamics-guided approach applicable in virtual screening for ERRα agonists.

Natural product drug discovery in the artificial intelligence era

Natural products (NPs) are primarily recognized as privileged structures to interact with protein drug targets. Their unique characteristics and structural diversity continue to marvel scientists for developing NP-inspired medicines, even though the pharmaceutical industry has largely given up. High-performance computer hardware, extensive storage, accessible software and affordable online education have democratized the use of artificial intelligence (AI) in many sectors and research areas. The last decades have introduced natural language processing and machine learning algorithms, two subfields of AI, to tackle NP drug discovery challenges and open up opportunities. In this article, we review and discuss the rational applications of AI approaches developed to assist in discovering bioactive NPs and capturing the molecular "patterns" of these privileged structures for combinatorial design or target selectivity.

PyRMD: A New Fully Automated AI-Powered Ligand-Based Virtual Screening Tool

Artificial intelligence (AI) algorithms are dramatically redefining the current drug discovery landscape by boosting the efficiency of its various steps. Still, their implementation often requires a certain level of expertise in AI paradigms and coding. This often prevents the use of these powerful methodologies by non-expert users involved in the design of new biologically active compounds. Here, the random matrix discriminant (RMD) algorithm, a high-performance AI method specifically tailored for the identification of new ligands, was implemented in a new fully automated tool, PyRMD. This ligand-based virtual screening tool can be trained using target bioactivity data directly downloaded from the ChEMBL repository without manual intervention. The software automatically splits the available training compounds into active and inactive sets and learns the distinctive chemical features responsible for the compounds' activity/inactivity. PyRMD was designed to easily screen millions of compounds in hours through an automated workflow and intuitive input files, allowing fine tuning of each parameter of the calculation. Additionally, PyRMD features a wealth of benchmark metrics, to accurately probe the model performance, which were used here to gauge its predictive potential and limitations. PyRMD is freely available on GitHub (https://github.com/cosconatilab/PyRMD) as an open-source tool.

De novo design with deep generative models based on 3D similarity scoring

We have demonstrated the utility of a 3D shape and pharmacophore similarity scoring component in molecular design with a deep generative model trained with reinforcement learning. Using Dopamine receptor type 2 (DRD2) as an example and its antagonist haloperidol 1 as a starting point in a ligand based design context, we have shown in a retrospective study that a 3D similarity enabled generative model can discover new leads in the absence of any other information. It can be efficiently used for scaffold hopping and generation of novel series. 3D similarity based models were compared against 2D QSAR based, indicating a significant degree of orthogonality of the generated outputs and with the former having a more diverse output. In addition, when the two scoring components are combined together for training of the generative model, it results in more efficient exploration of desirable chemical space compared to the individual components.

A comprehensive comparative assessment of 3D molecular similarity tools in ligand-based virtual screening

Three-dimensional (3D) molecular similarity, one major ligand-based virtual screening (VS) method, has been widely used in the drug discovery process. A variety of 3D molecular similarity tools have been developed in recent decades. In this study, we assessed a panel of 15 3D molecular similarity programs against the DUD-E and LIT-PCBA datasets, including commercial ROCS and Phase, in terms of screening power and scaffold-hopping power. The results revealed that (1) SHAFTS, LS-align, Phase Shape_Pharm and LIGSIFT showed the best VS capability in terms of screening power. Some 3D similarity tools available to academia can yield relatively better VS performance than commercial ROCS and Phase software. (2) Current 3D similarity VS tools exhibit a considerable ability to capture actives with new chemotypes in terms of scaffold hopping. (3) Multiple conformers relative to single conformations will generally improve VS performance for most 3D similarity tools, with marginal improvement observed in area under the receiving operator characteristic curve values, enrichment factor in the top 1% and hit rate in the top 1% values showed larger improvement. Moreover, redundancy and complementarity analyses of hit lists from different query seeds and different 3D similarity VS tools showed that the combination of different query seeds and/or different 3D similarity tools in VS campaigns retrieved more (and more diverse) active molecules. These findings provide useful information for guiding choices of the optimal 3D molecular similarity tools for VS practices and designing possible combination strategies to discover more diverse active compounds.

Molecular docking of alpha-enolase to elucidate the promising candidates against Streptococcus pneumoniae infection

PURPOSE

To predict potential inhibitors of alpha-enolase to reduce plasminogen binding of Streptococcus pneumoniae (S. pneumoniae) that may lead as an orally active drug. S. pneumoniae remains dominant in causing invasive diseases. Fibrinolytic pathway is a critical factor of S. pneumoniae to invade and progression of disease in the host body. Besides the low mass on the cell surface, alpha-enolase possesses significant plasminogen binding among all exposed proteins.

METHODS

In-silico based drug designing approach was implemented for evaluating potential inhibitors against alpha-enolase based on their binding affinities, energy score and pharmacokinetics. Lipinski's rule of five (LRo5) and Egan's (Brain Or IntestinaL EstimateD) BOILED-Egg methods were executed to predict the best ligand for biological systems.

RESULTS

Molecular docking analysis revealed, Sodium (1,5-dihydroxy-2-oxopyrrolidin-3-yl)-hydroxy-dioxidophosphanium (SF-2312) as a promising inhibitor that fabricates finest attractive charges and conventional hydrogen bonds with S. pneumoniae alpha-enolase. Moreover, the pharmacokinetics of SF-2312 predict it as a therapeutic inhibitor for clinical trials. Like SF-2312, phosphono-acetohydroxamate (PhAH) also constructed adequate interactions at the active site of alpha-enolase, but it predicted less favourable than SF-2312 based on binding affinity.

CONCLUSION

Briefly, SF-2312 and PhAH ligands could inhibit the role of alpha-enolase to restrain plasminogen binding, invasion and progression of S. pneumoniae. As per our investigation and analysis, SF-2312 is the most potent naturally existing inhibitor of S. pneumoniae alpha-enolase in current time.

Discovery of Natural Products Targeting NQO1 via an Approach Combining Network-Based Inference and Identification of Privileged Substructures

NAD(P)H:quinone oxidoreductase 1 (NQO1) has been shown to be a potential therapeutic target for various human diseases, such as cancer and neurodegenerative disorders. Recent advances in computational methods, especially network-based methods, have made it possible to identify novel compounds for a target with high efficiency and low cost. In this study, we designed a workflow combining network-based methods and identification of privileged substructures to discover new compounds targeting NQO1 from a natural product library. According to the prediction results, we purchased 56 compounds for experimental validation. Without the assistance of privileged substructures, 31 compounds (31/56 = 55.4%) showed IC₅₀ < 100 μM, and 11 compounds (11/56 = 19.6%) showed IC₅₀ < 10 μM. With the assistance of privileged substructures, the two success rates were increased to 61.8 and 26.5%, respectively. Seven natural products further showed inhibitory activity on NQO1 at the cellular level, which may serve as lead compounds for further development. Moreover, network analysis revealed that osthole may exert anticancer effects against multiple cancer types by inhibiting not only carbonic anhydrases IX and XII but also NQO1. Our workflow and computational methods can be easily applied in other targets and become useful tools in drug discovery and development.

Establishing an Analogue Based In Silico Pipeline in the Pursuit of Novel Inhibitory Scaffolds against the SARS Coronavirus 2 Papain-Like Protease

The ongoing coronavirus pandemic has been a burden on the worldwide population, with mass fatalities and devastating socioeconomic consequences. It has particularly drawn attention to the lack of approved small-molecule drugs to inhibit SARS coronaviruses. Importantly, lessons learned from the SARS outbreak of 2002-2004, caused by severe acute respiratory syndrome coronavirus 1 (SARS-CoV-1), can be applied to current drug discovery ventures. SARS-CoV-1 and SARS-CoV-2 both possess two cysteine proteases, the main protease (Mpro) and the papain-like protease (PLpro), which play a significant role in facilitating viral replication, and are important drug targets. The non-covalent inhibitor, GRL-0617, which was found to inhibit replication of SARS-CoV-1, and more recently SARS-CoV-2, is the only PLpro inhibitor co-crystallised with the recently solved SARS-CoV-2 PLpro crystal structure. Therefore, the GRL-0617 structural template and pharmacophore features are instrumental in the design and development of more potent PLpro inhibitors. In this work, we conducted scaffold hopping using GRL-0617 as a reference to screen over 339,000 ligands in the chemical space using the ChemDiv, MayBridge, and Enamine screening libraries. Twenty-four distinct scaffolds with structural and electrostatic similarity to GRL-0617 were obtained. These proceeded to molecular docking against PLpro using the AutoDock tools. Of two compounds that showed the most favourable predicted binding affinities to the target site, as well as comparable protein-ligand interactions to GRL-0617, one was chosen for further analogue-based work. Twenty-seven analogues of this compound were further docked against the PLpro, which resulted in two additional hits with promising docking profiles. Our in silico pipeline consisted of an integrative four-step approach: (1) ligand-based virtual screening (scaffold-hopping), (2) molecular docking, (3) an analogue search, and, (4) evaluation of scaffold drug-likeness, to identify promising scaffolds and eliminate those with undesirable properties. Overall, we present four novel, and lipophilic, scaffolds obtained from an exhaustive search of diverse and uncharted regions of chemical space, which may be further explored in vitro through structure-activity relationship (SAR) studies in the search for more potent inhibitors. Furthermore, these scaffolds were predicted to have fewer off-target interactions than GRL-0617. Lastly, to our knowledge, this work contains the largest ligand-based virtual screen performed against GRL-0617.

A comprehensive comparison of molecular feature representations for use in predictive modeling

Machine learning methods are commonly used for predicting molecular properties to accelerate material and drug design. An important part of this process is deciding how to represent the molecules. Typically, machine learning methods expect examples represented by vectors of values, and many methods for calculating molecular feature representations have been proposed. In this paper, we perform a comprehensive comparison of different molecular features, including traditional methods such as fingerprints and molecular descriptors, and recently proposed learnable representations based on neural networks. Feature representations are evaluated on 11 benchmark datasets, used for predicting properties and measures such as mutagenicity, melting points, activity, solubility, and IC50. Our experiments show that several molecular features work similarly well over all benchmark datasets. The ones that stand out most are Spectrophores, which give significantly worse performance than other features on most datasets. Molecular descriptors from the PaDEL library seem very well suited for predicting physical properties of molecules. Despite their simplicity, MACCS fingerprints performed very well overall. The results show that learnable representations achieve competitive performance compared to expert based representations. However, task-specific representations (graph convolutions and Weave methods) rarely offer any benefits, even though they are computationally more demanding. Lastly, combining different molecular feature representations typically does not give a noticeable improvement in performance compared to individual feature representations.

Comparing predictive ability of QSAR/QSPR models using 2D and 3D molecular representations

Quantitative structure-activity relationship (QSAR) and quantitative structure-property relationship (QSPR) models predict biological activity and molecular property based on the numerical relationship between chemical structures and activity (property) values. Molecular representations are of importance in QSAR/QSPR analysis. Topological information of molecular structures is usually utilized (2D representations) for this purpose. However, conformational information seems important because molecules are in the three-dimensional space. As a three-dimensional molecular representation applicable to diverse compounds, similarity between a test molecule and a set of reference molecules has been previously proposed. This 3D representation was found to be effective on virtual screening for early enrichment of active compounds. In this study, we introduced the 3D representation into QSAR/QSPR modeling (regression tasks). Furthermore, we investigated relative merits of 3D representations over 2D in terms of the diversity of training data sets. For the prediction task of quantum mechanics-based properties, the 3D representations were superior to 2D. For predicting activity of small molecules against specific biological targets, no consistent trend was observed in the difference of performance using the two types of representations, irrespective of the diversity of training data sets.

Neuraldecipher - reverse-engineering extended-connectivity fingerprints (ECFPs) to their molecular structures

Protecting molecular structures from disclosure against external parties is of great relevance for industrial and private associations, such as pharmaceutical companies. Within the framework of external collaborations, it is common to exchange datasets by encoding the molecular structures into descriptors. Molecular fingerprints such as the extended-connectivity fingerprints (ECFPs) are frequently used for such an exchange, because they typically perform well on quantitative structure-activity relationship tasks. ECFPs are often considered to be non-invertible due to the way they are computed. In this paper, we present a fast reverse-engineering method to deduce the molecular structure given revealed ECFPs. Our method includes the Neuraldecipher, a neural network model that predicts a compact vector representation of compounds, given ECFPs. We then utilize another pre-trained model to retrieve the molecular structure as SMILES representation. We demonstrate that our method is able to reconstruct molecular structures to some extent, and improves, when ECFPs with larger fingerprint sizes are revealed. For example, given ECFP count vectors of length 4096, we are able to correctly deduce up to 69% of molecular structures on a validation set (112 K unique samples) with our method.

DeepGraphMolGen, a multi-objective, computational strategy for generating molecules with desirable properties: a graph convolution and reinforcement learning approach

We address the problem of generating novel molecules with desired interaction properties as a multi-objective optimization problem. Interaction binding models are learned from binding data using graph convolution networks (GCNs). Since the experimentally obtained property scores are recognised as having potentially gross errors, we adopted a robust loss for the model. Combinations of these terms, including drug likeness and synthetic accessibility, are then optimized using reinforcement learning based on a graph convolution policy approach. Some of the molecules generated, while legitimate chemically, can have excellent drug-likeness scores but appear unusual. We provide an example based on the binding potency of small molecules to dopamine transporters. We extend our method successfully to use a multi-objective reward function, in this case for generating novel molecules that bind with dopamine transporters but not with those for norepinephrine. Our method should be generally applicable to the generation in silico of molecules with desirable properties.

Machine learning and AI-based approaches for bioactive ligand discovery and GPCR-ligand recognition

In the last decade, machine learning and artificial intelligence applications have received a significant boost in performance and attention in both academic research and industry. The success behind most of the recent state-of-the-art methods can be attributed to the latest developments in deep learning. When applied to various scientific domains that are concerned with the processing of non-tabular data, for example, image or text, deep learning has been shown to outperform not only conventional machine learning but also highly specialized tools developed by domain experts. This review aims to summarize AI-based research for GPCR bioactive ligand discovery with a particular focus on the most recent achievements and research trends. To make this article accessible to a broad audience of computational scientists, we provide instructive explanations of the underlying methodology, including overviews of the most commonly used deep learning architectures and feature representations of molecular data. We highlight the latest AI-based research that has led to the successful discovery of GPCR bioactive ligands. However, an equal focus of this review is on the discussion of machine learning-based technology that has been applied to ligand discovery in general and has the potential to pave the way for successful GPCR bioactive ligand discovery in the future. This review concludes with a brief outlook highlighting the recent research trends in deep learning, such as active learning and semi-supervised learning, which have great potential for advancing bioactive ligand discovery.

Interfering with S100B-effector protein interactions for cancer therapy

S100 calcium-binding protein B (S100B) is overexpressed in various malignant tumors, where it regulates cancer cell proliferation and metabolism by physical interactions with other molecules. Interfering with S100B-effector protein interactions is a potential strategy to treat malignant tumors. Although some S100B inhibitors have been discovered by virtual screening (VS), most target the S100B-p53 interaction. Hence, there is scope for the discovery of other S100B-effector protein interaction modulators for malignant tumors. In this review, we provide an overview of S100B-effector protein interaction inhibitor discovery using VS and discuss promising S100B-effector protein interaction targets that permit in silico analysis for drug discovery.

TF3P: Three-Dimensional Force Fields Fingerprint Learned by Deep Capsular Network

Molecular fingerprints are the workhorse in ligand-based drug discovery. In recent years, an increasing number of research papers reported fascinating results on using deep neural networks to learn 2D molecular representations as fingerprints. It is anticipated that the integration of deep learning would also contribute to the prosperity of 3D fingerprints. Here, we unprecedentedly introduce deep learning into 3D small molecule fingerprints, presenting a new one we termed as the three-dimensional force fields fingerprint (TF3P). TF3P is learned by a deep capsular network whose training is in no need of labeled data sets for specific predictive tasks. TF3P can encode the 3D force fields information of molecules and demonstrates the stronger ability to capture 3D structural changes, to recognize molecules alike in 3D but not in 2D, and to identify similar targets inaccessible by other 2D or 3D fingerprints based on only ligands similarity. Furthermore, TF3P is compatible with both statistical models (e.g., similarity ensemble approach) and machine learning models. Altogether, we report TF3P as a new 3D small molecule fingerprint with a promising future in ligand-based drug discovery. All codes are written in Python and available at https://github.com/canisw/tf3p.

Comparing a Query Compound with Drug Target Classes Using 3D-Chemical Similarity

3D similarity is useful in predicting the profiles of unprecedented molecular frameworks that are 2D dissimilar to known compounds. When comparing pairs of compounds, 3D similarity of the pairs depends on conformational sampling, the alignment method, the chosen descriptors, and the similarity coefficients. In addition to these four factors, 3D chemocentric target prediction of an unknown compound requires compound-target associations, which replace compound-to-compound comparisons with compound-to-target comparisons. In this study, quantitative comparison of query compounds to target classes (one-to-group) was achieved via two types of 3D similarity distributions for the respective target class with parameter optimization for the fitting models: (1) maximum likelihood (ML) estimation of queries, and (2) the Gaussian mixture model (GMM) of target classes. While Jaccard-Tanimoto similarity of query-to-ligand pairs with 3D structures (sampled multi-conformers) can be transformed into query distribution using ML estimation, the ligand pair similarity within each target class can be transformed into a representative distribution of a target class through GMM, which is hyperparameterized via the expectation-maximization (EM) algorithm. To quantify the discriminativeness of a query ligand against target classes, the Kullback-Leibler (K-L) divergence of each query was calculated and compared between targets. 3D similarity-based K-L divergence together with the probability and the feasibility index, (F_m), showed discriminative power with regard to some query-class associations. The K-L divergence of 3D similarity distributions can be an additional method for (1) the rank of the 3D similarity score or (2) the p-value of one 3D similarity distribution to predict the target of unprecedented drug scaffolds.

Discovering Anti-Cancer Drugs via Computational Methods

New drug discovery has been acknowledged as a complicated, expensive, time-consuming, and challenging project. It has been estimated that around 12 years and 2.7 billion USD, on average, are demanded for a new drug discovery via traditional drug development pipeline. How to reduce the research cost and speed up the development process of new drug discovery has become a challenging, urgent question for the pharmaceutical industry. Computer-aided drug discovery (CADD) has emerged as a powerful, and promising technology for faster, cheaper, and more effective drug design. Recently, the rapid growth of computational tools for drug discovery, including anticancer therapies, has exhibited a significant and outstanding impact on anticancer drug design, and has also provided fruitful insights into the area of cancer therapy. In this work, we discussed the different subareas of the computer-aided drug discovery process with a focus on anticancer drugs.

Predicting or Pretending: Artificial Intelligence for Protein-Ligand Interactions Lack of Sufficiently Large and Unbiased Datasets

Predicting protein-ligand interactions using artificial intelligence (AI) models has attracted great interest in recent years. However, data-driven AI models unequivocally suffer from a lack of sufficiently large and unbiased datasets. Here, we systematically investigated the data biases on the PDBbind and DUD-E datasets. We examined the model performance of atomic convolutional neural network (ACNN) on the PDBbind core set and achieved a Pearson R² of 0.73 between experimental and predicted binding affinities. Strikingly, the ACNN models did not require learning the essential protein-ligand interactions in complex structures and achieved similar performance even on datasets containing only ligand structures or only protein structures, while data splitting based on similarity clustering (protein sequence or ligand scaffold) significantly reduced the model performance. We also identified the property and topology biases in the DUD-E dataset which led to the artificially increased enrichment performance of virtual screening. The property bias in DUD-E was reduced by enforcing the more stringent ligand property matching rules, while the topology bias still exists due to the use of molecular fingerprint similarity as a decoy selection criterion. Therefore, we believe that sufficiently large and unbiased datasets are desirable for training robust AI models to accurately predict protein-ligand interactions.

In Silico Strategies in Tuberculosis Drug Discovery

Tuberculosis (TB) remains a serious threat to global public health, responsible for an estimated 1.5 million mortalities in 2018. While there are available therapeutics for this infection, slow-acting drugs, poor patient compliance, drug toxicity, and drug resistance require the discovery of novel TB drugs. Discovering new and more potent antibiotics that target novel TB protein targets is an attractive strategy towards controlling the global TB epidemic. In silico strategies can be applied at multiple stages of the drug discovery paradigm to expedite the identification of novel anti-TB therapeutics. In this paper, we discuss the current TB treatment, emergence of drug resistance, and the effective application of computational tools to the different stages of TB drug discovery when combined with traditional biochemical methods. We will also highlight the strengths and points of improvement in in silico TB drug discovery research, as well as possible future perspectives in this field.

Virtual Screening in the Cloud: How Big Is Big Enough?

Virtual screening is a standard tool in Computer-Assisted Drug Design (CADD). Early in a project, it is typical to use ligand-based similarity search methods to find suitable hit molecules. However, the number of compounds which can be screened and the time required are usually limited by computational resources. We describe here a high-throughput virtual screening project using 3D similarity (FastROCS) and automated evaluation workflows on Orion, a cloud computing platform. Cloud resources make this approach fully scalable and flexible, allowing the generation and search of billions of virtual molecules, and give access to an explicit 3D virtual chemistry space not available before. We discuss the impact of the size of the search space with respect to finding novel chemical hits and the size of the required hit list, as well as computational and economical aspects of resource scaling.

Current computational methods for predicting protein interactions of natural products

Natural products (NPs) are an indispensable source of drugs and they have a better coverage of the pharmacological space than synthetic compounds, owing to their high structural diversity. The prediction of their interaction profiles with druggable protein targets remains a major challenge in modern drug discovery. Experimental (off-)target predictions of NPs are cost- and time-consuming, whereas computational methods, on the other hand, are much faster and cheaper. As a result, computational predictions are preferentially used in the first instance for NP profiling, prior to experimental validations. This review covers recent advances in computational approaches which have been developed to aid the annotation of unknown drug-target interactions (DTIs), by focusing on three broad classes, namely: ligand-based, target-based, and target-ligand-based (hybrid) approaches. Computational DTI prediction methods have the potential to significantly advance the discovery and development of novel selective drugs exhibiting minimal side effects. We highlight some inherent caveats of these methods which must be overcome to enable them to realize their full potential, and a future outlook is given.

Development and application of a comprehensive machine learning program for predicting molecular biochemical and pharmacological properties

We establish a comprehensive quantitative structure-activity relationship (QSAR) model termed AlphaQ through the machine learning algorithm to associate the fully quantum mechanical molecular descriptors with various biochemical and pharmacological properties. Preliminarily, a novel method for molecular structural alignments was developed in such a way to maximize the quantum mechanical cross correlations among the molecules. Besides the improvement of structural alignments, three-dimensional (3D) distribution of the molecular electrostatic potential was introduced as the unique numerical descriptor for individual molecules. These dual modifications lead to a substantial accuracy enhancement in multifarious 3D-QSAR prediction models of AlphaQ. Most remarkably, AlphaQ has been proven to be applicable to structurally diverse molecules to the extent that it outperforms the conventional QSAR methods in estimating the inhibitory activity against thrombin, the water-cyclohexane distribution coefficient, the permeability across the membrane of the Caco-2 cell, and the metabolic stability in human liver microsomes. Due to the simplicity in model building and the high predictive capability for varying biochemical and pharmacological properties, AlphaQ is anticipated to serve as a valuable screening tool at both early and late stages of drug discovery.

Learning continuous and data-driven molecular descriptors by translating equivalent chemical representations

There has been a recent surge of interest in using machine learning across chemical space in order to predict properties of molecules or design molecules and materials with the desired properties. Most of this work relies on defining clever feature representations, in which the chemical graph structure is encoded in a uniform way such that predictions across chemical space can be made. In this work, we propose to exploit the powerful ability of deep neural networks to learn a feature representation from low-level encodings of a huge corpus of chemical structures. Our model borrows ideas from neural machine translation: it translates between two semantically equivalent but syntactically different representations of molecular structures, compressing the meaningful information both representations have in common in a low-dimensional representation vector. Once the model is trained, this representation can be extracted for any new molecule and utilized as a descriptor. In fair benchmarks with respect to various human-engineered molecular fingerprints and graph-convolution models, our method shows competitive performance in modelling quantitative structure-activity relationships in all analysed datasets. Additionally, we show that our descriptor significantly outperforms all baseline molecular fingerprints in two ligand-based virtual screening tasks. Overall, our descriptors show the most consistent performances in all experiments. The continuity of the descriptor space and the existence of the decoder that permits deducing a chemical structure from an embedding vector allow for exploration of the space and open up new opportunities for compound optimization and idea generation.

e-Sweet: A Machine-Learning Based Platform for the Prediction of Sweetener and Its Relative Sweetness

Artificial sweeteners (AS) can elicit the strong sweet sensation with the low or zero calorie, and are widely used to replace the nutritive sugar in the food and beverage industry. However, the safety issue of current AS is still controversial. Thus, it is imperative to develop more safe and potent AS. Due to the costly and laborious experimental-screening of AS, in-silico sweetener/sweetness prediction could provide a good avenue to identify the potential sweetener candidates before experiment. In this work, we curate the largest dataset of 530 sweeteners and 850 non-sweeteners, and collect the second largest dataset of 352 sweeteners with the relative sweetness (RS) from the literature. In light of these experimental datasets, we adopt five machine-learning methods and conformational-independent molecular fingerprints to derive the classification and regression models for the prediction of sweetener and its RS, respectively via the consensus strategy. Our best classification model achieves the 95% confidence intervals for the accuracy (0.91 ± 0.01), precision (0.90 ± 0.01), specificity (0.94 ± 0.01), sensitivity (0.86 ± 0.01), F1-score (0.88 ± 0.01), and NER (Non-error Rate: 0.90 ± 0.01) on the test set, which outperforms the model (NER = 0.85) of Rojas et al. in terms of NER, and our best regression model gives the 95% confidence intervals for the R²(test set) and ΔR² [referring to |R²(test set)- R²(cross-validation)|] of 0.77 ± 0.01 and 0.03 ± 0.01, respectively, which is also better than the other works based on the conformation-independent 2D descriptors (e.g., 2D Dragon) according to R²(test set) and ΔR². Our models are obtained by averaging over nineteen data-splitting schemes, and fully comply with the guidelines of Organization for Economic Cooperation and Development (OECD), which are not completely followed by the previous relevant works that are all on the basis of only one random data-splitting scheme for the cross-validation set and test set. Finally, we develop a user-friendly platform “e-Sweet” for the automatic prediction of sweetener and its corresponding RS. To our best knowledge, it is a first and free platform that can enable the experimental food scientists to exploit the current machine-learning methods to boost the discovery of more AS with the low or zero calorie content.

Computational Prediction of a New ADMET Endpoint for Small Molecules: Anticommensal Effect on Human Gut Microbiota

The human gut microbiota (HGM), which are evolutionarily commensal in the human gastrointestinal system, are crucial to our health. However, HGM can be broadly shaped by multifaceted factors such as intake of drugs. About one-quarter of the existing drugs for humans, which are designed to target human cells rather than HGM, can notably alter the composition of HGM. Therefore, the anticommensal effect of human drugs should be avoided to the maximum extent possible in the drug discovery and development process. Nevertheless, the anticommensal effect of small molecules is a new ADMET (absorption, distribution, metabolism, excretion, and toxicity) end point, which was never predicted with the computational method before. In this work, we present the first machine-learning based consensus classification model with the accuracy (0.811 ± 0.012), precision (0.759 ± 0.032), specificity (0.901 ± 0.019), sensitivity (0.628 ± 0.036), F1-score (0.687 ± 0.023), and AUC (0.814 ± 0.030) respectively on the test set. Furthermore, we develop an easy-to-use "e-Commensal" program for the automatic prediction. Based on this program, virtual-screening of the food-constituent database (FooDB) indicates that 5888 of 23 202 food-relevant compounds are forecasted to possess an anticommensal effect on HGM. Several top-ranked anticommensal compounds in our prediction are further scrutinized and confirmed by experiments in the existing literature. To the best of our knowledge, this is the first classification model and stand-alone software for the prediction of commensal or anticommensal compounds impacting HGM.

Novel natural non-nucleoside inhibitors of HIV-1 reverse transcriptase identified by shape- and structure-based virtual screening techniques

In this work we report a parallel application of both docking- and shape-based virtual screening (VS) methods, followed by Molecular Dynamics simulations (MDs), for discovering new compounds able to inhibit the human immunodeficiency virus type 1 (HIV-1) reverse transcriptase (RT) RNA-dependent DNA polymerase activity. Specifically, we screened more than 143000 natural compounds commercially available in the ZINC database against the best five RT crystallographic models, taking into account the five approved NNRTIs as query compounds. As a result, 20 hit molecules were selected and tested on biochemical assays for the inhibition of the RNA dependent DNA polymerase RT function and, among them, an indoline pyrrolidine (hit1), an indonyl piperazine (hit2) and an indolyl indolinone (hit3) derivatives were identified as novel non-nucleoside RT inhibitors in the low micromolar range.

Network-Based Methods for Prediction of Drug-Target Interactions

Drug-target interaction (DTI) is the basis of drug discovery. However, it is time-consuming and costly to determine DTIs experimentally. Over the past decade, various computational methods were proposed to predict potential DTIs with high efficiency and low costs. These methods can be roughly divided into several categories, such as molecular docking-based, pharmacophore-based, similarity-based, machine learning-based, and network-based methods. Among them, network-based methods, which do not rely on three-dimensional structures of targets and negative samples, have shown great advantages over the others. In this article, we focused on network-based methods for DTI prediction, in particular our network-based inference (NBI) methods that were derived from recommendation algorithms. We first introduced the methodologies and evaluation of network-based methods, and then the emphasis was put on their applications in a wide range of fields, including target prediction and elucidation of molecular mechanisms of therapeutic effects or safety problems. Finally, limitations and perspectives of network-based methods were discussed. In a word, network-based methods provide alternative tools for studies in drug repurposing, new drug discovery, systems pharmacology and systems toxicology.

Advances in the Development of Shape Similarity Methods and Their Application in Drug Discovery

Molecular similarity is a key concept in drug discovery. It is based on the assumption that structurally similar molecules frequently have similar properties. Assessment of similarity between small molecules has been highly effective in the discovery and development of various drugs. Especially, two-dimensional (2D) similarity approaches have been quite popular due to their simplicity, accuracy and efficiency. Recently, the focus has been shifted toward the development of methods involving the representation and comparison of three-dimensional (3D) conformation of small molecules. Among the 3D similarity methods, evaluation of shape similarity is now gaining attention for its application not only in virtual screening but also in molecular target prediction, drug repurposing and scaffold hopping. A wide range of methods have been developed to describe molecular shape and to determine the shape similarity between small molecules. The most widely used methods include atom distance-based methods, surface-based approaches such as spherical harmonics and 3D Zernike descriptors, atom-centered Gaussian overlay based representations. Several of these methods demonstrated excellent virtual screening performance not only retrospectively but also prospectively. In addition to methods assessing the similarity between small molecules, shape similarity approaches have been developed to compare shapes of protein structures and binding pockets. Additionally, shape comparisons between atomic models and 3D density maps allowed the fitting of atomic models into cryo-electron microscopy maps. This review aims to summarize the methodological advances in shape similarity assessment highlighting advantages, disadvantages and their application in drug discovery.

Machine learning in chemoinformatics and drug discovery

Chemoinformatics is an established discipline focusing on extracting, processing and extrapolating meaningful data from chemical structures. With the rapid explosion of chemical 'big' data from HTS and combinatorial synthesis, machine learning has become an indispensable tool for drug designers to mine chemical information from large compound databases to design drugs with important biological properties. To process the chemical data, we first reviewed multiple processing layers in the chemoinformatics pipeline followed by the introduction of commonly used machine learning models in drug discovery and QSAR analysis. Here, we present basic principles and recent case studies to demonstrate the utility of machine learning techniques in chemoinformatics analyses; and we discuss limitations and future directions to guide further development in this evolving field.

Eg5 inhibitor YL001 induces mitotic arrest and inhibits tumor proliferation

Eg5 is a kinesin spindle protein that controls chromosomal segregation in mitosis and is thus a critical drug target for cancer therapy. We report the discovery of a potent, selective inhibitor of Eg5 designated YL001. YL001 was obtained through shape similarity based virtual screening, and it bears a 1,5-disubstituted tetrazole scaffold. YL001 exhibits favorable bioactivity in a variety of cancer cell lines, including taxol-resistant ovarian cancer and 6TG-resistant breast cancer cell lines. This compound inhibits tumor growth by 60% and significantly prolongs median survival time by more than 50% in a xenograft mouse model. YL001 blocks the ATPase activity of Eg5 and causes mitotic failure, ultimately resulting in apoptosis of cancer cells through activation of the caspase-3 pathway. Our findings demonstrate that YL001 is a potent antitumor agent that may be developed for cancer therapeutics.

e-Bitter: Bitterant Prediction by the Consensus Voting From the Machine-Learning Methods

In-silico bitterant prediction received the considerable attention due to the expensive and laborious experimental-screening of the bitterant. In this work, we collect the fully experimental dataset containing 707 bitterants and 592 non-bitterants, which is distinct from the fully or partially hypothetical non-bitterant dataset used in the previous works. Based on this experimental dataset, we harness the consensus votes from the multiple machine-learning methods (e.g., deep learning etc.) combined with the molecular fingerprint to build the bitter/bitterless classification models with five-fold cross-validation, which are further inspected by the Y-randomization test and applicability domain analysis. One of the best consensus models affords the accuracy, precision, specificity, sensitivity, F1-score, and Matthews correlation coefficient (MCC) of 0.929, 0.918, 0.898, 0.954, 0.936, and 0.856 respectively on our test set. For the automatic prediction of bitterant, a graphic program “e-Bitter” is developed for the convenience of users via the simple mouse click. To our best knowledge, it is for the first time to adopt the consensus model for the bitterant prediction and develop the first free stand-alone software for the experimental food scientist.