SwissADME: a free web tool to evaluate pharmacokinetics, drug-likeness and medicinal chemistry friendliness of small molecules — TEI-tagged view

Antoine DainapersName, Olivier MichielinpersName, Vincent ZoetepersName (2017date). “SwissADME: a free web tool to evaluate pharmacokinetics, drug-likeness and medicinal chemistry friendliness of small molecules.” Scientific Reportstitle·j.

Object of study — SwissADMEterm's property models, molecule interface, classifier validation, drug-likenessterm filters, and interoperable outputs

Research question — how to provide fast, robust, interpretable, and freely accessible multi-property prediction for small-molecule drug discovery

Contribution — a login-free web tool pooling predictive modelsterm for physicochemical propertiesterm, pharmacokineticsterm, drug-likenessterm, and medicinal-chemistry friendliness

Full text — tagged

The reference list is not part of the text we hold for this paper. Our copy was extracted from the publisher's page and the bibliography did not survive that extraction, so the article below ends before its references. Nothing has been removed by us.

SwissADMEterm: a free web tool to evaluate pharmacokineticsterm, drug-likenessterm and medicinal chemistry friendlinessterm of small molecules Abstract To be effective as a drug, a potent molecule must reach its target in the body in sufficient concentration, and stay there in a bioactive form long enough for the expected biologic events to occur. Drug development involves assessment of absorption, distribution, metabolism and excretionterm (ADMEterm) increasingly earlier in the discovery process, at a stage when considered compounds are numerous but access to the physical samples is limited. In that context, computer modelsterm constitute valid alternatives to experimentsterm. Here, we present the new SwissADME web toolterm that gives free access to a pool of fast yet robust predictive modelsterm for physicochemical propertiesterm, pharmacokineticsterm, drug-likenessterm and medicinal chemistry friendlinessterm, among which in-house proficient methods such as the BOILED-Eggterm, iLOGPterm and Bioavailability Radarterm. Easy efficient input and interpretation are ensured thanks to a user-friendly interface through the login-free website http://www.swissadmeterm.ch . Specialists, but also nonexpert in cheminformatics or computational chemistry can predict rapidly key parameters for a collection of molecules to support their drug discovery endeavours.

During the time- and resource-consuming processes of drug discovery and development, a large number of molecular structures are evaluated according to very diverse parameters in order to steer the selection of which chemicals to synthetize, test and promote, with the final goal to identify those with the best chance to become an effective medicine for the patients. The molecules must show high biological activity together with low toxicity. Equally important is the access to and concentration at the therapeutic target in the organism. The traditional way to consider pharmacokineticsterm (i.e. the fate of a therapeutic compound in the organism) is to break down the various effects that impact the access to the target into individual parameters. In turn, these ADMEterm parameters (for Absorption, Distribution, Metabolism and Excretionterm) can be evaluated separately by dedicated methods. It has been demonstrated that early estimation of ADMEterm in the discovery phase reduces drastically the fraction of pharmacokinetics-related failure in the clinical phases 1 . Computer modelsterm have been fostered as a valid alternative to experimental procedures for prediction of ADMEterm, especially at initial steps, when investigated chemical structures are numerous but the availability of compounds is scarce 2 .

A large variety of in silico methods share the objective of predicting ADMEterm parameters from molecular structure 3 . Noteworthy, the pioneer work of Lipinski et al . examined orally active compounds to define physicochemical ranges for high probability to be an oral drug (i.e. the drug-likenessterm) 4 . This so-called Rule-of-fiveterm delineated the relationship between pharmacokineticterm and physicochemical parametersterm.

Whereas physicochemical parametersterm give a global description of the structure, molecules can be directly described by substructure searchesterm. These techniques are at the root of Structural Alertterm 5 , the PAINSterm 6 or the Lilly MedChem 7 filters applied to cleanse chemical libraries from compounds most likely unstable, reactive, toxic, or prone to interfere with biological assays because unspecific frequent hitters, dyes or aggregators 8 .

Cheminformaticians developed different molecular descriptorsterm mined from chemical structures. One of the most popular examples is the molecular fingerprintterm (FPterm), which consists in a sequence of bitsterm defining the presence or absence of chemical features in a molecule. The FP2 methodterm 9 is one archetype of topological (or path-based) FPterm, which considers all fragments of the molecular structure following a linear path up to a given number of bonds. Every possible path is hashed to generate the bit string (i.e. the FPterm). A major advantage of FPterm is the efficiency by which computers handle such bit strings 10 , allowing for instance large-scale virtual screeningterm or the rapid estimation of synthetic accessibilityterm of molecules 11 . FPterm are also employed in classification models for ADMEterm behaviours built by support vector machineterm (SVMterm) or Bayesian techniques 12 . Remarkably, computer-aided drug designterm (CADDterm) has been a pioneer field for the application of such machine learning technologies 13 .

Because most freely available in silico ADME toolsterm focus on one specific property or model only, while generalist ADME packagesterm are commercial software, we felt the need of gathering what we consider the most relevant computational methods to provide a global appraisal of the pharmacokineticsterm profile of small molecules. The methods were selected for robustness, speed, but also and importantly for ease of interpretation to enable efficient translation to medicinal chemistry through molecular design. Some of them were adapted with open-source algorithms to ensure freedom to operate for the global scientific community. Others are our own models developed and tested on purpose. When applicable, multiple predictionsterm of the same parameter are provided to allow for a consensus viewterm of a given property.

The SwissADME web toolterm presented here is freely accessible at http://www.swissadmeterm.ch and meant for user-friendly submission and easy analysis of the results, also for nonexpert in CADDterm. Compared to the state-of-the art of free web-based tools for ADMEterm and pharmacokineticsterm (e.g. pk-CSM 14 and admetSAR 15 ) and apart from unique access to proficient methods (e.g. iLOGPterm 16 or the BOILED-Eggterm 17 ), SwissADMEterm strong points are, non-exhaustively: different input methods, computation for multiple molecules, and the possibility to display, save and share results per individual molecule or through global intuitive and interactive graphs. Finally, SwissADMEterm is integrated in the SwissDrugDesign workspace. One-click interoperabilityterm gives access to various CADDterm tools developed by the Molecular Modeling Group of the SIB Swiss Institute of Bioinformatics, e.g. ligand-based virtual screeningterm (SwissSimilarity 18 ), biotarget prediction (SwissTargetPrediction 19 ), molecular docking (SwissDock 20 ), bioisosteric design (SwissBioisostere 21 ), or molecular mechanics (SwissParam 22 ).

Submission Web page Accessing http://www.swissadmeterm.ch in a web browser displays directly the submission page of SwissADMEterm, where molecules to be estimated for ADMEterm, physicochemistry, drug-likenessterm, pharmacokineticsterm and medicinal chemistry friendlinessterm properties can be input. As shown in Fig. 1 , a black toolbar at the top of the Web page allows the user to navigate within the different SwissDrugDesign tools. A second bar gives access to different information regarding SwissADMEterm, among which the FAQ and Help pages as well as legal disclaimer and contacts.

The input zone itself comprises a molecular sketcherterm based on ChemAxon’s Marvin JS ( http://www.chemaxon.com ) that enables the user to import (from a file or an external database), draw, and edit a 2D chemical structure, and to transfer it to a list of molecules. This list, on the right-hand side of the submission page, is the actual input for computation. It can be edited as a standard text, allowing for typing or pasting SMILESterm. The list is made to contain one input molecule per line, defined by a SMILESterm and optionally a name separated by a space. If name is omitted, SwissADMEterm will automatically provide an identifier.

Noteworthy, both buttons for transferring the sketch to SMILESterm list and for running the computation are dynamic, in the sense that they are active only if the action is possible. At the time of writing, one can expect a result in 1 to 5 seconds for a drug-like moleculeterm.

Examples can be loaded in the SMILESterm list by clicking on the “Fill with an example” button.

One-panel-per-moleculeterm Output The output panels are loaded in the same Web page. There is one panel compiling all values for each molecule. It is filled immediately after calculation completion, one molecule after the other. This way it is possible to inspect the results for the first compounds without waiting for the whole list to be treated. This one-panel-per-moleculeterm ( Fig. 2 ) is headed by the molecule name and divided into different sections.

Chemical Structure and Bioavailability Radarterm The first section, including two-dimensional chemical structure and canonical SMILESterm, is located below the title ( Fig. 2 ). It shows on which chemical form the predictions were calculated (refer to Computational Methods ). Moreover, our Bioavailability Radarterm is displayed for a rapid appraisal of drug-likenessterm (refer to Fig. 3 ). Six physicochemical propertiesterm are taken into account: lipophilicityterm·var, sizeterm·var, polarityterm·var, solubilityterm·var, flexibilityterm·var and saturationterm·var. A physicochemical range on each axis was defined by descriptors adapted from refs 23 and 24 and depicted as a pink area in which the radar plot of the molecule has to fall entirely to be considered drug-like. Leaving the mouse over the radar gives further information about the descriptors (see also Physicochemical Propertiesterm and Computational Methods ).

Physicochemical Propertiesterm Simple molecular and physicochemical descriptorsterm like molecular weightterm (MW), molecular refractivityterm (MR), count of specific atom types and polar surface areaterm (PSAterm) are compiled in this section. The values are computed with OpenBabelterm 9 , version 2.3.0. The PSAterm is calculated using the fragmental technique called topological polar surface areaterm (TPSAterm), considering sulfur and phosphorus as polar atoms 25 . This has proven a useful descriptor in many models and rules to quickly estimate some ADMEterm properties, especially with regards to biological barrier crossing such as absorption and brain access 17 .

Lipophilicityterm·var The partition coefficient between n -octanol and waterterm (log P o/wterm ) is the classical descriptor for Lipophilicityterm·var . It has a dedicated section in SwissADMEterm due to the critical importance of this physicochemical property for pharmacokineticsterm drug discovery 26 27 . Many computational methods for log P o/wterm estimation were developed with diverse performance on various chemical sets. Common practice is to use multiple predictors either to select the most accurate methods for a given chemical series or to generate consensus estimation. The models behind the predictors should be as diverse as possible to increase the prediction accuracy through consensus log P o/wterm 28 . In that regard, SwissADMEterm gives access to five freely available predictive modelsterm; i.e. XLOGP3, an atomistic method including corrective factors and knowledge-based library 29 ; WLOGPterm·var, our own implementation of a purely atomistic method based on the fragmental system of Wildman and Crippen 30 ; MLOGP, an archetype of topological method relying on a linear relationship with 13 molecular descriptorsterm implemented from refs 31 and 32 ; SILICOS-IT, an hybrid method relying on 27 fragments and 7 topological descriptors ( http://silicos-it.be.s3-website-eu-west-1.amazonaws.com/software/filter-it/1.0.2/filter-it.html , accessed June 2016); and finally iLOGPterm, our in-house physics-based methodterm relying on free energies of solvation in n -octanol and water calculated by the Generalized-Born and solvent accessible surface area (GB/SA) model. iLOGPterm was benchmarked on two drug or drug-like external sets and performed equally as or better than six well-established predictors 16 . The consensus log P o/wterm is the arithmetic mean of the values predicted by the five proposed methods.

Water Solubilityterm Having a soluble molecule greatly facilitates many drug development activities, primarily the ease of handling and formulation 33 . Moreover, for discovery projects targeting oral administration, solubilityterm·var is one major property influencing absorption 34 . As well, a drug meant for parenteral usage has to be highly soluble in water to deliver a sufficient quantity of active ingredient in the small volume of such pharmaceutical dosage 35 . Two topological methods to predict Water Solubilityterm are included in SwissADMEterm. The first one is an implementation of the ESOL modelterm 36 and the second one is adapted from Ali et al . 37 . Both differ from the seminal general solubilityterm·var equation 38 since they avoid the melting point parameter; the latter being challenging to predict. Moreover they demonstrate strong linear correlation between predicted and experimental values (R 2 = 0.69 and 0.81, respectively). SwissADMEterm third predictor for solubilityterm·var was developed by SILICOS-IT. The linear correlation coefficient of this fragmental method corrected by molecular weightterm is R 2 = 0.75 ( http://silicos-it.be.s3-website-eu-west-1.amazonaws.com/software/filter-it/1.0.2/filter-it.html , accessed June 2016).

All predicted values are the decimal logarithm of the molar solubilityterm·var in water (log Sterm ). SwissADMEterm also provides solubilityterm·var in mol/l and mg/ml along with qualitative solubilityterm·var classes (please refer to Computational Methods ).

Pharmacokineticsterm Specialized models, whose predictions are compiled in the Pharmacokineticsterm section, evaluate individual ADMEterm behaviours of the molecule under investigation.

One model is a multiple linear regression, which aims at predicting the skin permeability coefficientterm ( K p ). It is adapted from Potts and Guy 39 , who found K p linearly correlated with molecular sizeterm·var and lipophilicityterm·var (R 2 = 0.67). The more negative the log K pterm (with K p in cm/s), the less skin permeant is the molecule.

The predictions for passive human gastrointestinal absorptionterm (HIAterm) and blood-brain barrierterm (BBBterm) permeation both consist in the readout of the BOILED-Eggterm model 17 , an intuitive graphical classification model, which can be displayed in the SwissADMEterm result page by clicking the red button appearing below the sketcher when all input molecules have been processed (refer to Graphical Output ). Other binary classification models are included, which focus on the propensity for a given small molecule to be substrate or inhibitor of proteins governing important pharmacokineticterm behaviours.

The knowledge about compounds being substrate or non-substrate of the permeability glycoproteinterm (P-gpterm, suggested the most important member among ATP-binding cassette transporters or ABC-transporters) is key to appraise active efflux through biological membranes, for instance from the gastrointestinal wall to the lumen or from the brain 40 . One major role of P-gpterm is to protect the central nervous systemterm (CNS) from xenobiotics 41 . Importantly as well, P-gpterm is overexpressed in some tumour cells and leadsterm to multidrug-resistant cancers 42 .

Also essential is the knowledge about interaction of molecules with cytochromes P450term (CYPterm). This superfamily of isoenzymes is a key player in drug elimination through metabolic biotransformation 43 . It has been suggested that CYPterm and P-gpterm can process small molecules synergistically to improve protection of tissues and organisms 44 . One can estimate that 50 to 90% (depending on the authors) of therapeutic molecules are substrate of five major isoforms (CYP1A2, CYP2C19, CYP2C9, CYP2D6, CYP3A4) 45 46 . Inhibition of these isoenzymesterm is certainly one major cause of pharmacokinetics-related drug-drug interactionsterm 47 48 leading to toxic or other unwanted adverse effectsterm due to the lower clearance and accumulation of the drug or its metabolites 49 . Numerous inhibitors of the CYPterm isoforms have been identified. Some are affecting different CYPterm isoforms, while other compounds show selectivity for specific isoenzymes 50 . It is therefore of great importance for drug discovery to predict the propensity with which the molecule will cause significant drug interactions through inhibition of CYPsterm, and to determine which isoforms are affected.

SwissADMEterm enables the estimation for a chemical to be substrate of P-gpterm or inhibitor of the most important CYPterm isoenzymes. We applied the support vector machineterm algorithm (SVMterm) 51 on meticulously cleansed large datasets of known substrates/non-substrates or inhibitors/non-inhibitors (for details, see Computational Methods ). In similar contexts, SVMterm was found to perform better than other machine-learning algorithms for binary classification 40 52 . The models return “Yes” or “No” if the molecule under investigation has higher probability to be substrate or non-substrate of P-gpterm (respectively inhibitor or non-inhibitor of a given CYPterm). The statistical performance of the classification models is given in Table 1 , in comparison with previous SVMterm models on the same targets. We restricted the benchmark to state-of-the-art methods, published after 2010.

The quantification of the models performance is not straightforwardly comparable, because the training sets are different, most of the published models are less than 10-fold cross-validated and some statistical parameters are missing. Nevertheless, the SwissADMEterm classifiers are competitive with previous models in term of robustness, with cross-validation accuracyterm (ACC CV ) grossly at the same level. Furthermore, cross-validated areas under receiver operating characteristic (ROC) curves (AUC CV ) are equal to the corresponding values found in the literature. Likewise, external prediction power is difficult to compare, as each test set includes different molecules. However, the predictive capacity of SwissADMEterm classifiers is grossly equivalent to the related SVMterm methods, both in terms of external accuracy (ACC ext ) and external area under ROC curveterm (AUC ext ). The models for which external validation was not found (P-gpterm of refs 15 and 53 ) have to be taken with extreme caution since they possibly suffer from overfitting biasesterm. Noteworthy, some of the published models (e.g. CYP2C9 and CYP3A4 of ref. 54 or CYP2C9 of ref. 55 ) were built on severely unbalanced training sets and tested on clearly unbalanced external sets. As demonstrated by Carbon-Mangels et al . 56 the relevance of machine-learning classification methods, and especially SVMterm, are negatively impacted by datasets with one significantly more populated class. In that case, accuracy measurements are overestimated and prone to mislead the construction and the evaluation of the model. Moreover, some SVMterm models were published with small training and test setsterm (P-gpterm of refs 15 and 57 ), which imply questionable capacity of generalization and broadness of applicability domainsterm. We emphasize that for SwissADMEterm classifiers, both training and test setsterm were carefully cleansed and checked for sizeterm·var, diversity and balance between classes. Furthermore, our SVMterm models rely merely on molecular and physicochemical descriptorsterm generated by SwissADMEterm. We believe that this improves robustness and sustainability of the underlying methodologies. In particular, not using molecular fingerprintsterm, molecular graphs or other structural descriptions can be an handicap to generate high statistical values but should also limit overfitting biasesterm and yield more generalist predictive modelsterm, not necessarily influenced by specific chemical scaffolds or moieties. In our practice, these well-performing models able to estimate important ADMEterm behaviours are of great support for pharmacokineticsterm optimization and evaluation of small molecules.

Drug-likenessterm As defined earlier, “drug-likenessterm” assesses qualitatively the chance for a molecule to become an oral drug with respect to bioavailability. Drug-likenessterm was established from structural or physicochemical inspections of development compounds advanced enough to be considered oral drug-candidates. This notion is routinely employed to perform filtering of chemical libraries to exclude molecules with properties most probably incompatible with an acceptable pharmacokineticsterm profile. This SwissADMEterm section gives access to five different rule-based filtersterm, with diverse ranges of properties inside of which the molecule is defined as drug-like. These filters often originate from analyses by major pharmaceutical companies aiming to improve the quality of their proprietary chemical collections. The Lipinski (Pfizer) filter is the pioneer rule-of-fiveterm implemented from ref. 4 . The Ghose (Amgen), Veber (GSK), Egan (Pharmacia) and Muegge (Bayer) methods were adapted from refs 58 , 59 , 60 , 61 , respectively. Multiple estimations allow consensus viewsterm or selection of methods best fitting the end-user’s specific needs in terms of chemical space or project-related demands. Any violation of any rule described here appears explicitly in the output panel.

The Abbot Bioavailability Scoreterm 62 is similar but seeks to predict the probability of a compound to have at least 10% oral bioavailability in rat or measurable Caco-2 permeability. This semi-quantitative rule-based score relying on total charge, TPSAterm, and violation to the Lipinski filter defines four classes of compounds with probabilities of 11%, 17%, 56% or 85%. Like the other methods in this section, it primary focuses on the fast screening of chemical libraries, to select the best molecules to be purchased, synthetized or promoted at a further stage of a medicinal chemistry project.

Medicinal Chemistry The purpose of this section is to support medicinal chemists in their daily drug discovery endeavours. Two complementary pattern recognition methods allow for identification of potentially problematic fragments. PAINSterm (for pan assay interference compounds, a.k.a. frequent hitters or promiscuous compounds) are molecules containing substructures showing potent response in assays irrespective of the protein target. Such fragments, yielding false positive biological output, have been identified by Baell et al . 6 in analysing six orthogonal assays and breaking down the molecules active on 2 or more assays into 481 recurrent fragments, considered as potentially leading to promiscuous compounds. SwissADMEterm returns warnings if such moieties are found in the molecule under evaluation.

Besides, we implemented Structural Alertterm, which consists in a list of 105 fragments identified by Brenk et al . 5 to be putatively toxic, chemically reactive, metabolically unstable or to bear properties responsible for poor pharmacokineticsterm. In SwissADMEterm, it is possible to have a chemical description of the problematic fragments found in a given molecule by flying over the “question mark” icon appearing after the fragment list. This is implemented for both PAINSterm and Brenk filters. By applying these and other physicochemical filters to design screening libraries, Brenk et al . 5 observed that most of the remaining compounds satisfy criteria for “leadlikenessterm”. This concept is similar to drug-likenessterm, yet focusing on physicochemical boundaries defining a good leadterm, i.e. a molecular entity suitable for optimization. By definition, leadsterm are subjected to chemical modifications that will most likely increase sizeterm·var and lipophilicityterm·var 63 . As a consequence, leadsterm are required to be smaller and less hydrophobic than drug-like moleculesterm. Since it is crucial for a chemist to judge whether a given molecule is suitable to initiate leadterm optimization, in addition to structural filters, we implemented a rule-based method for leadlikenessterm, which was adapted from ref. 64 .

One of the key aspects of CADDterm activities is to help the selection of the most promising virtual molecules that will be synthetized and submitted to biological assays or other experimentsterm. Synthetic accessibilityterm (SA) is a major factor to consider in this selection process. Obviously, for a reasonable number of molecules, medicinal chemists are the best able to determine SA. However, when too many molecular structures prevent an expert evaluation, in silico estimation can be used for pre-filtering. Ertl & Schuffenhauer 11 proposed a fingerprint-based approach for SA estimation but including closed-source information about fingerprint definition that prevents a straightforward implementation in our tool open to the scientific community. As a consequence, we have built our own fragmental method by analysing more than 13 millions compounds immediately deliverable by vendors. We assumed that the most frequent molecular fragmentsterm (technically, FP2term·def bits, refer to Computational Methods ) in this large collection indicates a probably high SA, while rare fragments imply a difficult synthesis. For a given molecule, the fragmental contributions to SA are summed and corrected by the terms describing sizeterm·var and complexity, such as macrocycles, chiral centres, or spiro functions as defined by Ertl & Schuffenhauer 11 . After normalization, the SA Scoreterm ranges from 1 (very easy) to 10 (very difficult). To assess the performance of the method developed for SwissADMEterm, we retrieved two test sets of SA previously published. Both sets involved external molecules, whose difficulty of synthesis was marked from 1 to 10 by nine 11 and four 65 medicinal chemists, respectively. The averaged expert score can then be compared to an in silico SA Scoreterm. As seen in Table 2 , the predictive capacity of all three methods appears very dependent of the test set. Indeed the SAs of set 1, smaller and evaluated by more chemists, turned out to be much more robustly predictable than set 2. Human evaluation of synthetic complexity is undeniably subjective and relies on individual chemist’s training and experience. However, significant linear correlation and small errors — especially with SwissADMEterm SA Scoreterm that outperformed the reference methods on both sets with smaller errors, and equal or higher linear correlation coefficients — demonstrate how this simple and fast methodology can help prioritizing molecules to synthetize.

Graphical output After all calculations completed, the “Show BOILED-Eggterm” red button appears below the sketcher to display the graphical output on the same page (as exemplify in Fig. 4 ). This consists primarily in the BOILED-Eggterm, an intuitive method to predict simultaneously two key ADMEterm parameters, i.e. the passive gastrointestinal absorption (HIAterm) and brain access (BBBterm). Although conceptually very simple as it relies on two physicochemical descriptorsterm only (WLOGPterm·var and TPSAterm, for lipophilicityterm·var and apparent polarityterm·var), this classification model was built with extreme care regarding statistical significance and robustness 17 . As shown in Fig. 4 , the egg-shaped classification plot includes the yolk (i.e. the physicochemical space for highly probable BBBterm permeation) and the white (i.e. the physicochemical space for highly probable HIAterm absorption). Both compartments are not mutually exclusive and the outside grey region stands for molecules with properties implying predicted low absorption and limited brain penetration. In practice, the BOILED-Eggterm has proven straightforward interpretation and efficient translation to molecular design in a variety of drug discovery settings. Whereas the predictive power of the BOILED-Eggterm is broad in term of chemical space, it is restricted to passive penetration through gastro-intestinal wall and BBBterm. We took benefit of its implementation within SwissADMEterm to enrich the graphical output with the prediction of P-gpterm substrate, which is the most important active efflux mechanism involved in those biological barriers 40 (refer to the SVMterm model described in Pharmacokineticsterm ). As a result, the user conveniently obtains on the same graph a global evaluation about passive absorption (inside/outside the white), passive brain access (inside/outside the yolk) and active efflux from the CNS or to the gastrointestinal lumen by colour-coding: blue dots for P-gpterm substrates (PGPterm+) and red dots for P-gpterm non-substrate (PGPterm−).

Contrary to the one-panel-per-moleculeterm concept for other parameters, the graphical output includes prediction for all molecules submitted to SwissADMEterm, thus additional capacities were implemented to enable interactive navigation and easy evaluation. Flying over a specific point makes a semi-transparent box appear including the name and structure of the molecule. Clicking on a specific point makes the page scroll to the corresponding output panel including all predictions for the molecule. Getting back to the graphical output is achieved by clicking on the red up-arrow on the top-right corner of the panel. Moreover, on the right of the plot are displayed possible actions (at the moment, to show the name of the molecules on the graph, only), legends and possible remarks (the number of molecules outside the range of the plot). The user may want to hide the plot by clicking the “Hide BOILED-Eggterm” red button.

Help, saving and interoperabilityterm SwissADMEterm is user-friendly and has been conceived for a variety of users, including nonexpert in cheminformatics or computational chemistry. Efforts were put to insure ease of input and interpretability of output. These aspects are supported by the Help page, which provides technical guidance on the usage of the website and by the FAQ , where we collected not only technical but also scientific and general questions. Contextual help about specific items can be displayed by leaving the mouse over question marks or some boxes on the pages.

Moreover, SwissADMEterm is an integral part of the SwissDrugDesign programterm, an ambitious initiative driven by the Molecular Modeling Group of the SIB Swiss Institute of Bioinformatics that aims at providing a collection of free Web-based tools covering many aspects of CADDterm. As such, SwissADMEterm users can access the different Web sites, through the black toolbar at the top of the page. By clicking on the name of the tool, a new tab opens with the corresponding submission page. We have undertaken the implementation of more advanced interoperabilityterm capabilities. At the time of writing this manuscript, three icons were placed top left of the output panel (below the name of the molecule, see Fig. 2 ). They allow one-click submission of the molecule to SwissTargetPrediction (for reverse screening to predict protein target, by clicking on the red target), SwissSimilarity (for ligand-based virtual screeningterm, by clicking on the red twins) and to re-submit the molecule to SwissADMEterm (for instance to get a single compound graphical output, by clicking the red pill). Reversely, other SwissDrugDesign websites include the SwissADMEterm pill icon for estimation of ADMEterm, pharmacokineticsterm, drug-likenessterm and medicinal chemistry friendlinessterm regarding a small molecule output of any CADDterm process.

Importantly, the user has the opportunity to save or share SwissADMEterm results in files (CSV) or by copying the values in the clipboard of the local computer and pasting in text or spreadsheet applications for further processing.

Conclusions The SwissADME Web toolterm enables the computation of key physicochemical, pharmacokineticterm, drug-like and related parameters for one or multiple molecules. In one hand, efforts were put in the backend to embed free open-access and fast predictive modelsterm showing statistical significance, predictive power, intuitive interpretation, and straightforward translation to molecular design. These models are adapted from published renowned approaches or in-house original methods, specially developed and thoroughly benchmarked. On the other hand, we focused on an ergonomic and user-friendly graphical interface for the cost- and login-free Web site http://www.swissadmeterm.ch . The latter enables easy input and efficient analysis of the output through interactive capabilities, and does not require any prior knowledge in CADDterm. Moreover, interoperabilityterm allows for direct access to other SwissDrugDesign web toolsterm, including SwissSimilarity 66 (virtual screeningterm), SwissBioisostere 21 (ligand-based design), SwissTargetPrediction 19 (prediction of biotargets), SwissDock 20 (molecular docking), SwissSideChain 67 (protein modeling) and SwissParam 22 (molecular mechanics).

As a result, SwissADMEterm has been designed to support the entire community (specialists and nonexperts) in their drug discovery endeavours.

Computational methods Programming and scripting The SwissADMEterm website was written in HTML, PHP5, and JavaScript, whereas the backend of computation was mainly coded in Python 2.7. The use of additional libraries or software for specific tasks is mentioned in the corresponding paragraph.

Submission Page The molecule inputted through the sketcher Marvin JS (version 16.4.18, 2016, www.chemaxon.com ) are converted into SMILESterm by JChem Web Services (version 14.9.29, 2013, www.chemaxon.com ) installed on one of our servers. This on-the-fly conversion allows seamless paste of SMILESterm in the input list. The user has the possibility to edit this list as a standard text, e.g. to modify SMILESterm or add a name to the molecule. Upon calculation submission by clicking the “Run” button, the SMILESterm of each molecule is canonicalised by OpenBabelterm (version 2.3.0, 2012, http://openbabelterm.org ) 9 and processed individually. Several actions are performed through JChem Web Services APIs. First hydrogen atoms are added to the molecular structure, which is dearomatised (i.e. kekulised), neutralised and checked by the Standardizer API. Then a tridimensional conformation is generated though the StringMolExport function with the Clean3D option and stored in MOL2 format. Besides, a two-dimensional image created through the MolConverter API is displayed on demand when scrolling the output web page.

One-panel-per-moleculeterm Output The Bioavailability Radarterm in the first section of the One-panel-per-moleculeterm output complements the two-dimensional image from the JChem webserver and the canonical SMILESterm calculated by OpenBabelterm. We use the JpGraph PHP library (version 3.5.0b1, 2016, http://jpgraph.net ) to produce the radar plot, which bears six axes for six important properties for oral bioavailability. Each property is defined by a descriptor of SwissADMEterm and a range of optimal values is depicted as a pink area. The latter is inspired from commonly accepted bioavailability and drug-likenessterm guidelines 23 24 . For saturationterm·var, the ratio of sp 3 hybridized carbons over the total carbon count of the molecule (Fraction Csp3) should be at least 0.25. For sizeterm·var, the molecular weightterm (MW calculated by OpenBabelterm) should be between 150 and 500 g/mol. For polarityterm·var, the TPSAterm 25 should be between 20 and 130Å 2 . For solubilityterm·var, log Sterm (calculated with the ESOL modelterm 36 ) should not exceed 6. For lipophilicityterm·var, XLOGP3 29 should be in the range from −0.7 to +6.0. For flexibilityterm·var, the molecule should not have more than 9 rotatable bonds. To be estimated as drug-like, the red line of the compound under study must be fully included in the pink area. Any deviation represents a suboptimal physicochemical property for oral bioavailability.

All descriptors and molecular parameters of the Physicochemical Propertiesterm section are computed through the OpenBabelterm API (version 2.3.0, 2012, http://openbabelterm.org ) 9 . Noteworthy, the topological polar surface areaterm (TPSAterm) is strictly based on the fragmental system provided by Ertl et al . 25 including polar sulfur and phosphorus atoms.

Multiple freely available computational methods to predict n -octanol/water partition coefficient (log P o/wterm ) values are made available in the Lipophilicityterm·var section. iLOGPterm (for implicit log P ) is an in-house physics-based methodsterm relying on Gibbs free energy of solvation calculated by GB/SA in water and n -octanol. Generalized-born (GB) parameters are computed through the GBMV2 method 68 and solvent-accessible surface area (SA) is the analytical approximation generated by CHARMM (version c36b1, 2011, https://www.charmm.org ) 69 . The iLOGPterm implemented in SwissADMEterm corresponds to Model9 of the seminal publication 16 , which was trained on 11,993 molecules (r = 0.72, MAE = 0.89, and RMSE = 1.14 against experimental log P ). 5-fold crossvalidation ensured robustness (q 2 CV = 0.52, MAE CV = 0.89, and RMSE CV = 1.14) and external test benchmarks showed the excellent predictive power and extended applicability domainterm compared to well-established methods. XLOGP3 values are obtained through the command-line Linux program (version 3.2.2, courtesy of CCBG, Shanghai Institute of Organic Chemistry) including the knowledge-based corrections 29 . WLOGPterm·var is our own implementation of the atomistic method developed by Wildman and Crippen 30 . MLOGP values are computed through an in-house implementation of Moriguchi’s topological method 31 32 . SILICOS-IT is the log P o/wterm estimation returned by executing the FILTER-IT program (version 1.0.2, 2013, http://silicos-it.be.s3-website-eu-west-1.amazonaws.com/software/filter-it/1.0.2/filter-it.html ). Finally, SwissADMEterm gives a consensus log P o/wterm value, which is the arithmetic mean of the five predictive values mentioned above.

Similarly to lipophilicityterm·var, the Water Solubilityterm section includes multiple predictive methods for the user to choose between the most accurate model for a given chemical series and an averaged consensus value. The ESOL modelterm 36 is a QSPR modelterm establishing the linear relationship between log Sterm and five molecular parameters, i.e. MW, the number of rotatable bonds, the fraction of aromatic heavy atoms and Daylight’s CLOGP. Because the lipophilicityterm·var descriptor is not freely available, the implementation of ESOL in SwissADMEterm replaces CLOGP by XLOGP3 as parameter in the linear equation to predict log Sterm . XLOGP3 is known to perform well on external datasets and to return similar predictions as CLOGP 28 . The other three parameters were computed with OpenBabelterm. Likewise, Ali et al . 37 linked log Sterm with log P o/wterm and TPSAterm. The model implemented in SwissADMEterm corresponds to the model 3 of the original publication, with XLOGP3 as lipophilicityterm·var descriptor. The third solubilityterm·var method available in SwissADMEterm is the log Sterm estimated by the FILTER-IT program (version 1.0.2, 2013, http://silicos-it.be.s3-website-eu-west-1.amazonaws.com/software/filter-it/1.0.2/filter-it.html ). This prediction is based on a system of 16 fragmental contributions modulated by the squared root of MW. All three models are predicting log Sterm values, which are also translated within SwissADMEterm into solubilityterm·var in mol/l and mg/ml. Finally a qualitative estimation of the solubilityterm·var class is given according to the following log Sterm scale: insoluble <−10 <poorly <−6 <moderately <−4 <soluble <−2 <very <0 <highly.

The Pharmacokineticsterm section proposes one linear method for skin permeation, which relies on the simple QSPR modelterm by Potts and Guy 39 linking the decimal logarithm of the skin permeability coefficientterm (log K pterm in cm/s) with MW and log P o/wterm . The model implemented in SwissADMEterm uses XLOGP3 as lipophilicityterm·var descriptor. Besides, most of the models in this section are machine-learning binary classifiers for important ADMEterm behaviours. Passive gastro-intestinal (HIAterm) absorption and blood-brain barrierterm (BBBterm) permeation are predicted with the BOILED-Eggterm model, which defines favourable and unfavourable zones in the log P o/wterm versus PSAterm physicochemical space for passive diffusion through both physiological barriers 17 . The classification showed 10-fold cross-validation accuracyterm of 92% and 88% for BBBterm and HIAterm, respectively (refer to Graphical Output ).

Six other classification models are part of the Pharmacokineticsterm section to predict the propensity of the molecule under investigation to be substrate or inhibitor of important pharmacokinetics-related proteins, for which large diverse and balanced datasets were retrieved and meticulously cleansed. For P-glycoprotein1 (P-gpterm), the training set consists of 521 substrates and 512 non-substrates extracted from the Metrabase database 70 ( http://www-metrabase.ch.cam.ac.uk , accessed January 2016), whereas the test set was obtained from ref. 71 . To ensure truly external validation, molecules overlapping with the training set were removed from the test set, which finally includes 215 substrates and 200 non-substrates. For CYPterm major isoforms, all datasets were those of Veith et al . 50 and downloaded from the PubChem database 72 ( http://pubchem.ncbi.nlm.nih.gov , accessed February 2016). In case of unbalanced dataset (all except CYP1A2 and CYP2C19), sufficient chemical diversity was guaranteed by clustering with the Ward method and a reciprocal nearest neighbour (RNN) algorithm 73 , the more populated class to lessen. The number of molecules (described by circular fingerprints) of the large class is reduced by defining clusters with the JKlustor program (version 14.9.29, 2014, http://www.chemaxon.com ). Only the centre of each cluster (i.e. the molecule that has the smallest sum of dissimilarities to the other molecules in the cluster) is included in the training or test set to balance. As a result, the training sets involved respectively 4301 CYP1A2 inhibitors and 4844 CYP1A2 non-inhibitors; 4284 CYP2C19 inhibitors and 4988 CYP2C19 non-inhibitors; 2940 CYP2C9 inhibitors and 3000 CYP2C9 non-inhibitors; 1814 CYP2D6 inhibitors and 1850 CYP2D6 non-inhibitors; and 3758 CYP3A4 inhibitors and 3760 CYP3A4 non-inhibitors. The test sets involved respectively 1412 CYP1A2 inhibitors and 1588 CYP1A2 non-inhibitors; 1386 CYP2C19 inhibitors and 1614 CYP2C19 non-inhibitors; 1020 CYP2C9 inhibitors and 1055 CYP2C9 non-inhibitors; 528 CYP2D6 inhibitors and 540 CYP2D6 non-inhibitors; and 1289 CYP3A4 inhibitors and 1290 CYP3A4 non-inhibitors.

SwissADMEterm’s backend calculations were ran to generate 50 molecular and physicochemical descriptorsterm per molecule (described in the Supplementary Table S1 ). For a given model, a descriptor was rejected if non-zero values for all molecules in the training set are less than 20% or if the coefficient of variation is less than 3%. In case of correlation higher than 0.9 between remaining descriptors, a selection is made based on F-score. The selected descriptors for each model are shown in Supplementary Tables S2–S7 . These tables also include the minimum and maximum values for each descriptor among all molecules used in the training. This enables beholding the broadness of physicochemical space involved and the applicability domainterm of the SVMterm models. The predictive capability of each model can be further appraised on Supplementary Table S8 , where external accuracy was split in sensitivity and specificity to ensure that positive and negative molecules are predicted with the same level of robustness. The final training and test setsterm with selected descriptors were normalized and the respective model ready to be built. First, the libSVM support vector machineterm python library (version 3.20, 2015, https://www.csie.ntu.edu.tw/~cjlin/libsvm/ ) 74 was used for multi-step grid-based optimization of the best coefficients for the above-selected descriptors as well as for the soft-margin permissivity (C) and the hyper-parameter (ϒ) of the RBF Gaussian kernel functionterm. The 10-fold crossvalidated accuracy (ACC CV ) for each model was so maximized and AUC CV was calculated. In a second step, the so-built models were used on the external test sets (normalized according to the training set) in order to evaluate predictive power in terms of external accuracy (ACC ext ) and AUC ext . All final SVMterm models were stored in separate files, which are read through the libSVM API upon SwissADMEterm job submission.

The Drug-likenessterm section includes six rule-based methods The Lipinski rule-of-fiveterm is exactly as described in ref. 4 including MLOGP <4.15 as lipophilicityterm·var threshold. The Ghose filter is the method detailed in the original publication 58 , where the atomic log P is calculated with WLOGPterm·var. The very simple yet efficient Veber filter is implemented directly from the seminal paper 59 . The Egan filter is yielded from the Egan Egg 60 75 , but the closed-source ALOGP98 was replaced by WLOGPterm·var 17 . The Muegge filter 61 was adapted to fit SwissADMEterm implementation and usage by calculating XLOGP3 as lipophilicityterm·var descriptor. The Bioavailability Scoreterm was implemented without changes from Martin et al . 62 . Similarly, the leadlikenessterm filter included in the Medicinal Chemistry section was adapted from the original rule 64 by using XLOGP3 as lipophilicityterm·var descriptor.

Both methods for identification of problematic fragments within the Medicinal Chemistry section, i.e. PAINSterm 6 and Brenk’s Structural Alertterm 5 , were implemented using the SMARTS recognition capability of OpenBabelterm API. The SMARTS definitions for PAINSterm were retrieved from the Filter-it distribution (version 1.0.2, 2013, http://silicos-it.be.s3-website-eu-west-1.amazonaws.com/software/filter-it/1.0.2/filter-it.html ). Little formatting and cleansing were needed to obtain a screenable collection of 481 fragments. Brenk provided directly the SMARTS descriptions of 105 unwanted chemical groups in the supplementary material of the seminal paper 5 .

SwissADMEterm Synthetic Accessibilityterm (SA) Score is based primarily on the assumption that the frequency of molecular fragmentsterm·var in ‘really’ obtainable molecules correlates with the ease of synthesis. The fragmental contribution to SA should be favourable for frequent chemical moieties and unfavourable for rare moieties. We examined the ‘All Now’ subset of ZINC database 76 (version 12, 2014, http://zinc.docking.org/subsets/all-now accessed April 2015) including 12′782′590 compounds immediately deliverable by vendors. This collection was not submitted to any other filter. FP2term·def fingerprints of all molecules were computed by OpenBabelterm 9 thus generating 12′782′590 bit strings (i.e. the fingerprints). The frequency of occupancy of each of the 1024 bits for the entire library was calculated and its contribution to SA obtained by applying natural logarithm to the normalized bit count. As a result, this fragmental system of bits allows extremely fast evaluation of any input molecule by reading its FP2term·def string. Fragmental contribution of bit occupancy is summed and linearly modulated by corrective factors, which are penalties for sizeterm·var (MW) and complexity. This latter is based on SMARTS recognition of chiral centres, spiro functions, bridged rings and macrocycles (more than 8-membered rings). The coefficients of the corrective terms are those defined by Ertl & Schuffenhauer 11 . Finally the score is normalized to range from 1 (very easy) to 10 (very difficult to synthetize). As crude as it may seem, SwissADMEterm SA Scoreterm is extremely fast and performing slightly better than two similar methods previously published 11 65 (refer to Table 2 ).

Graphical output The graphical output of SwissADMEterm consists of the BOILED-Eggterm directly implemented from ref. 17 to predict passive diffusion through HIAterm and BBBterm by position in a WLOGP- versus -TPSA physicochemical space. The HIAterm model was trained on 660 well-absorbed and non-absorbed molecules and a 10-fold crossvalidation returned an accuracy (ACC CV ) of 92% and a Matthews correlation coefficient (MCC CV ) of 0.65. The BBBterm model was built on a training set of 260 permeant and non-permeant molecules and displayed excellent internal statistics as well: ACC CV = 88% and MCC CV = 0.75. The plot is enriched by color-coding referring to P-gpterm active efflux predicted by a SVMterm classifier (refer to Pharmacokineticsterm ). We aimed at providing a global and efficient analysis of absorption and brain access for multiple molecules at once. For more advanced interactive features, the graph is generated in JavaScript using the Flot plotting library (version 0.8, 2015, http://www.flotcharts.org/ ). The user has the possibility to visualize the position of the molecules between the different BOILED-Eggterm compartments and their propensity of being substrate of P-gpterm by coloured points: blue for substrate (PGPterm+) and red for non-substrate (PGPterm−). The dots are also the mean to visualize name and chemical structure of the molecules (generated on-the-fly by the dedicated Chemaxon web server as described above) by leaving the mouse over the dot. As well, it is possible to access the corresponding output panel summarizing all calculated values for the molecule by clicking on the corresponding dot. Getting back to the BOILED-Eggterm is as simple as clicking on the up arrow at the top-right corner of the panel. All these functionalities, as well as interoperabilityterm with other SwissDrugDesign web toolsterm are coded in PHP and Javascript.

Additional Information How to cite this article : DainapersName, A. et al . SwissADMEterm: a free web tool to evaluate pharmacokineticsterm, drug-likenessterm and medicinal chemistry friendlinessterm of small molecules. Sci. Rep. 7 , 42717; doi: 10.1038/srep42717 (2017).

Publisher's note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary Material