An open source chemical structure curation pipeline using RDKit — TEI-tagged view

A. Patrícia BentopersName, Anne HerseypersName, Eloy FélixpersName, Greg LandrumpersName, Anna GaultonpersName, Francis AtkinsonpersName, Louisa J. BellispersName, Marleen De VeijpersName, Andrew R. LeachpersName (2020date). “An open source chemical structure curation pipeline using RDKit.” Journal of Cheminformaticstitle·j.

Object of study — automated validation, standardization, and parent generation for chemical structuresterm entering ChEMBL

Research question — whether a maintainable open-source RDKitterm pipeline can curate diverse chemical-structure inputs robustly and replace ChEMBL's commercial legacy workflow

Contribution — a public three-component ChEMBL structure pipeline, revised curation rules, production deployment, and comparative tests on curated and uncurated datasets

Full text — tagged

An open source chemical structureterm curation pipeline using RDKitterm Abstract The ChEMBL databaseterm is one of a number of public databases that contain bioactivity dataterm on small molecule compoundsterm curated from diverse sources. Incoming compounds are typically not standardised according to consistent rules. In order to maintain the quality of the final database and to easily compare and integrate data on the same compound from different sources it is necessary for the chemical structuresterm in the database to be appropriately standardised.

A chemical curation pipelineterm has been developed using the open source toolkit RDKitterm. It comprises three components: a Checkerterm to test the validity of chemical structuresterm and flag any serious errors; a Standardizerterm which formats compounds according to defined rules and conventions and a GetParentterm component that removes any saltsterm and solventsterm from the compound to create its parent. This pipeline has been applied to the latest version of the ChEMBL databaseterm as well as uncurated datasets from other sources to test the robustness of the process and to identify common issues in database molecular structures.

All the components of the structure pipeline have been made freely available for other researchers to use and adapt for their own use. The code is available in a GitHub repository and it can also be accessed via the ChEMBL Beaker webservices. It has been used successfully to standardise the nearly 2 million compounds in the ChEMBL databaseterm and the compound validity checkerterm has been used to identify compounds with the most serious issues so that they can be prioritised for manual curationterm.

Keywords: Chemistry, Curation, ChEMBL, RDKitterm, Open source, Standardisation Introduction The ChEMBL databaseterm is a freely available bioactivity database containing close to 2.5 million compound records on nearly 2 million unique chemical structuresterm [ 1 ]. The compound structures and associated bioactivity dataterm are extracted on a regular basis primarily from the medicinal chemistry literature. A growing number of researchers are also depositing experimental data directly in order to make these available in the public domain. Furthermore, ChEMBL contains a set of manually curated marketed drugs and clinical candidates as well as selected bioactivity dataterm from other public databases such as BindingDB [ 2 ] and PubChemterm [ 3 ]. Bioactivity dataterm on the same compound from all ChEMBL sources (scientific articles, deposited datasets and curated drug sources) are aggregated according to chemical structureterm. Compounds may be physically tested in bioassays as the so-called parent moleculeterm·def or as one of a number of different salt formsterm. Scientists commonly wish to aggregate the data on these different forms on the basis of the common underlying parent structureterm, and so it is necessary to link these various forms of the “same” underlying parent moleculeterm·def.

In order to facilitate the use of the database, a key objective of the ChEMBL compound curation process is to standardise the chemical structuresterm stored in the database and to assign a unique identifierterm to each distinct chemical structureterm regardless of the source. It is worth noting that there are over 5000 unique compounds in the ChEMBL databaseterm with data from ten or more different sources, and four compounds (doxorubicin, ciprofloxacin, chloroquine and paclitaxel) each with data from over 1000 sources. For each ChEMBL release, more than 50,000 new structures are added to the database, which makes manual curationterm and standardisation of the chemical structuresterm impracticable. Hence an automated procedureterm is required.

The chemical structuresterm submitted to the ChEMBL databaseterm are generally received as molfilesterm [ 4 ] but can also be in SMILESterm format [ 5 ]. There is no universal standard for these formats and the challenges of converting between chemical structureterm formats is well documented [ 6 ]. Even the simple process of loading molecules into and out of different cheminformatics packages can subtly change the structure, particularly if it was not well drawn in the first place. Chemical structuresterm from the primary scientific literature are mostly manually drawn from the structural information in the papers prior to loading into ChEMBL. These structures are often represented in the publication as Markush structures with different R-groups shown in SAR (structure–activity relationship) tables. Compounds may also be shown with charges on acidic or basic groups, to indicate the form in which they are likely to interact with amino acid residues in a binding pocket. In some articles the compounds synthesised and reported are isotopically labelled. The ChEMBL compound curation procedure therefore needs to process molecules represented in all these ways (and more) to determine which compounds are the same. Examples of these situations can be found in recent articles where data was extracted for the ChEMBL databaseterm [ 7 – 9 ].

A standardised V2000 molfileterm was chosen as the primary chemical structureterm representation in the database. The Standard InChIterm and the corresponding hashed InChIKeyterm [ 10 , 11 ] are used in ChEMBL as the measure of uniqueness for a chemical structureterm and are calculated from the molfileterm using software provided by the InChI Trust, currently version 1.05 [ 12 ]. Thus, when compounds from different scientific articles have the same Standard InChIterm and InChIKeyterm they are considered to be the same compound and are assigned the same ChEMBL identifierterm (CHEMBL_ID). The use of the Standard InChIterm has a number of advantages in the context of the ChEMBL databaseterm. It was developed as an IUPAC open standard to enable information exchange and interoperability between large databases. The simple format of Standard InChIterm is also used by many other database providers and hence it is an ideal choice for an open database such as ChEMBL. However, unlike molfilesterm and SMILESterm, it was designed as an identifierterm and not as a structure format suitable for cheminformatics applications. For example, the Standard InChIterm is independent of the tautomeric form of a compound and hence different tautomers of a compound will have the same standard InChIterm. As a consequence, they are considered in ChEMBL as being the same compound and hence have the same ChEMBL_ID. It should however be remembered that there are a few disadvantages to using Standard InChIterm as an identifierterm, including: its inability to recognise some 1,5 keto-enol tautomers as being the same compound; its inability to recognise cis/trans isomerism in organometallic compounds (e.g. cisplatin and transplatin) and it does not support the use of relative stereochemistryterm·var, only absolute or no stereochemistryterm·var. Despite these limitations, it is a good compromise for a structural identifierterm for a public database such as ChEMBL. To aid interoperability, a canonical SMILESterm is also generated from each primary molfileterm, but this is considered to be a secondary structure.

The challenges of registering compounds in a database and determining chemical uniqueness are not specific to ChEMBL. PubChemterm, the open chemistry database of over 100 million compounds maintained by the National Institutes of Health (NIH) has developed an automated and publicly available standardisation process based on OpenEye toolkits [ 13 ]. They utilise an isomeric canonical SMILESterm to identify unique structures and show that their method generates more unique structures than when Standard InChIterm is used to determine uniqueness. They also generate a canonical tautomerterm for their compounds. The ChEMBL group decided against the use of a canonical tautomerterm, for a number of reasons. Firstly, it is anticipated that the authors of an article in the medicinal chemistry literature will have worked on their compound set for some considerable time, performed docking and molecular modelling studies and will be well placed to assign the most appropriate tautomeric form, for example based on specific interactions with the target protein. Secondly, it is possible for tautomers to interconvert under experimental conditions. Thirdly, changing the tautomeric form of a compound can alter its stereochemistryterm·var by removing, or introducing, a chiral centre from a molecule, requiring often difficult decisions to be made about how to “merge” compounds/data with different chirality or which stereo form to create. Thalidomide is a well-known example of the second phenomenon which interconverts between the therapeutic R-enantiomer and the teratogenic S-enantiomer via the enol tautomer [ 14 ].

The Royal Society of Chemistry (RSC) developed the Chemical Validation and Standardization Platform (CVSP) [ 15 ] to support compound deposition into their ChemSpider chemical database [ 16 ] and as a contribution to the Innovative Medicines Initiative (IMI) funded OpenPHACTS project [ 17 ] which aimed to standardise chemical structuresterm from multiple databases. The CVSP methodology uses sets of SMARTS-based rules that can be tailored by the user. The code is available on GitHub [ 18 ] but currently has no obvious documentation and the interface is no longer available [ 19 ] although it is still possible for depositors to ChemSpider to use the rules to validate their structures.

The United States Environmental Protection Agency (EPA) also strive to have well curated structures in their DSSTox database [ 20 ]. They have described, in detail, the complexity of the task and have undertaken extensive curation of their chemical structuresterm using a combination of manual and automated methods. Much of their focus has been on resolving the mismatches between names, identifiersterm and structures between their compound set and those in the US National Library of Medicine databases ChemID [ 21 ] and PubChemterm. Due to the inability of the V2000 molfilesterm to distinguish relative and absolute stereochemistryterm·var they have chosen to use V3000 molfilesterm as their preferred structure format. They also use an InChIKeyterm calculated using the ChemAxon JChem toolkit to determine uniqueness, but this differs from the Standard InChIKeyterm discussed above. SMILESterm are provided for users not wanting to use V3000 molfilesterm, but they make the point that this results in a less rich representation of the structures. Their methods are based on a combination of commercial toolkits and their own manual curationterm tool [ 22 ].

The Global Substance Registration System (GSRS), developed by the regulatory authorities in the USA and Europe, creates a Unique Ingredient Identifierterm (UNII) for components of medicinal products which includes small molecules but also more complex molecules such as proteins and polymers [ 23 ]. This is a particularly useful resource for identifying the structures of specific pharmaceutically relevant compounds, but it is less useful for the bulk curation of the larger set of bioactive molecules in the ChEMBL databaseterm.

Many commercial software providers also provide toolkits for standardising compound structures. These are widely used by large pharmaceutical and agrochemical companies where the precise annotation of a chemical structureterm is crucial for intellectual property protection. Those produced by ChemAxon and 3DS are examples of these toolkits [ 24 , 25 ].

The legacy pipelineterm for processing compound structures prior to deposition into ChEMBL was based on commercial software and had evolved over more than 10 years. During this time, more and more incorrect, unusual, and exceptional situations were identified and integrated into the pipeline. Unfortunately, these refinements were increasingly making the code difficult to maintain and modify. In reviewing options for a sustainable future solution that also removed the dependence on commercial software it became apparent that none of the existing toolkits fitted the ChEMBL group’s requirements. Therefore, the decision was made to build a curation pipeline around the widely used open-source RDKitterm toolkit and its implementation of the MolVSterm molecule validation and standardisation tool [ 26 , 27 ]. The resulting pipeline is now available as an open-source solution, freely available for the wider scientific community [ 28 ]. As part of this project the current rules used for curating ChEMBL structures have also been revised and rewritten to make them more sustainable. The resulting processes are designed and developed to suit the specific situation of ChEMBL, but the code can also be used by others “as is” or can be modified to suit other requirements.

It is important to note that this curation pipeline is not intended as a replacement for a similarly named standardiser tool [ 29 ] which was previously developed by the ChEMBL group as a contribution to the IMI eTOX project [ 30 ] and was specifically designed to standardise molecules in preparation for molecular modelling applications.

The newly developed curation pipeline will now be described in more detail.

Methods Three new components have been developed using the RDKitterm toolkit. Two of these components ( Standardizerterm and GetParentterm ) have been rewritten and adapted from rules originally implemented using a commercial software toolkit. In contrast, the Checkerterm component was developed more recently in an attempt to identify problem structures before they were added to the ChEMBL databaseterm.

Hence the new ChEMBL curation pipeline comprises three processes: Checkerterm : identifies and validates structures and identifies problems before they are added to the database Standardizerterm : processes (standardises) chemical structuresterm according to a set of predefined rules GetParentterm : generates parent structuresterm based on a set of rules and defined lists of saltsterm and solventsterm Checkerterm component The Checkerterm component validates structures prior to the compounds being loaded into ChEMBL. If an error or problem is detected in the structure a score is reported for the molecule; the score increases with the severity of the perceived problemterm. In the majority of cases compounds are loaded into the database even if a warning flagterm is identified. The scores are recorded but at this point errors are not corrected. Instead, they are prioritised and subjected to subsequent manual curationterm, as time and degree of seriousness permits. A summary of the structure checks performed, and the resultant penalty scoresterm assigned are shown in Table 1 .

Penalty scoresterm and annotation that are output from the Checkerterm module It is an individual user’s choice what they decide to do with molecules that return specific penalty scoresterm. For ChEMBL, a penalty scoreterm of 7 is considered to be a fatal errorterm and the molfileterm is not loaded into the database. Examples of illegal input are, for example, unknown elements in the molfileterm, or molfilesterm that cannot be read in RDKitterm due to the inability to process their aromatic bonds. Molecules with a penalty scoreterm of 6 are loaded into the ChEMBL databaseterm but without a molfileterm, as it is considered that these have a significant issue with the structure, and it is preferable to fix the problem than have a badly formed molecule in the database. Most of the issues that give rise to a penalty scoreterm of 6 are self-explanatory and are described in Table 1 . If the penalty scoreterm is 5 or 2 the molecule is loaded but the compounds are also prioritised for manual curationterm. Again, many of the 5 and 2 penalty scoresterm are self-explanatory, but the stereo mismatchterm errors perhaps need further explanation. These are reported when the number of stereocentresterm perceived by the following calculation methods differ: Mol: number of atoms where a wedged bond starts InChI: number of tetrahedral stereocentresterm RDKitterm: number of atomic stereocentresterm remaining after calling Chem.AssignStereochemistry() Hence the “InChi_RDKitterm/Mol stereo mismatchterm” warning message indicates that the InChI and RDKitterm algorithms perceive the number of stereocentresterm to be the same but different from the molfileterm. “Mol/Inchi/RDKitterm stereo mismatchterm” means that all three methods perceive different stereocentre countsterm·var. The majority of these issues occur in complex molecules such as bridged bicyclic molecules that are badly drawn. As Standard InChIterm is derived from the molfileterm, the errors where the molfileterm and InChI differ in their stereocentre countsterm·var are given a higher penalty scoreterm [ 5 ] than when they are the same but different from the RDKitterm stereo count [ 2 ].

The InChI software may give a number of warnings. These are also reported by the ChEMBL Checkerterm module. Some of these are considered important, but others such as “InChI: Omitted undefined stereo”, “InChI: Charges were rearranged”, “InChI: Ambiguous stereo”, “InChI: Proton(s) added/removed” and “InChI: Not chiral” are generated for large numbers of molecules. These either reflect the reorganisation of atoms in order to generate the InChI or are related to stereochemical ambiguity arising, for example, from the fact that the compound is a racemate; these are not considered issues for a database such as ChEMBL. Therefore, these are given a low penalty scoreterm [ 2 ]. However, in other contexts they might be more relevant and so are reported in the Checkerterm output.

Standardizerterm component The standardisation rules implemented in the ChEMBL databaseterm are based largely on the FDA/IUPAC guidelines [ 31 , 32 ]. Whilst the aim is to adhere to these rules as closely as possible, the practical reality is that submitted compounds are sometimes drawn imperfectly or the structures are ambiguously defined in the original publication or by the depositor. An automated standardiser can only safely correct some of the potential issues and the standardisation rules, currently encoded in the Standardizerterm component, are outlined here.

For certain compound types, particularly organometallic and those with a large number of boron atoms, a flag is set (exclude flagterm) and no attempt is made to standardise them. This is largely due to the V2000 molfileterm format used by ChEMBL being unable to accurately represent coordination bondsterm. For this reason, although the bioactivity dataterm on these compounds is available in ChEMBL, the chemical structuresterm are not curated nor provided in the release version of the database.

The first step in the standardisation process is therefore to exclude molecules if they contain more than 7 boron atoms or any of the following atoms: [Sc], [Ti], [V], [Cr], [Mn], [Fe], [Co], [Ni], [Cu], [Ga], [Y], [Zr], [Nb], [Mo], [Tc], [Ru], [Rh], [Pd], [Cd], [In], [Sn], [La], [Hf], [Ta], [W], [Re], [Os], [Ir], [Pt], [Au], [Hg], [Tl], [Pb], [Bi], [Po], [Ac], [Ce], [Pr], [Nd], [Pm], [Sm], [Eu], [Gd], [Tb], [Dy], [Ho], [Er], [Tm], [Yb], [Lu], [Th], [Pa], [U], [Np], [Pu], [Am], [Cm], [Bk], [Cf], [Es], [Fm], [Md], [No], [Lr], [Ge], [Sb].

The following standardisations are then made to the molecule (where they occur): Standardise unknown stereochemistryterm·var Clear S Group data from the molfileterm Generate a kekulé form of the structure Remove explicit H atomsterm except: Hs where an isotopeterm of hydrogen has been specifically set Hs that have a wedged or dashed bond to them Hs bonded to atoms with tetrahedral stereochemistryterm·var set (“Chiral Hs”). This is an example: Hs bonded to atoms in a non-default valence state that are not simply protonated. An example is phosphinic acid: Normalise structure: Fix hypervalent nitro groupsterm Convert covalently drawn alkaline metals connected to O or N to ionic forms (e.g. NaO to Na+ O−) Fix incorrect amide tautomers, e.g. N=COH to HNC(=O) Standardise sulphoxides to charge-separated form Standardise diazonium N to N+ Ensure quaternary N is charged Ensure trivalent O is charged Ensure trivalent S is charged Ensure halogen not bonded to a neighbouring atom is charged Ensure molecule is neutralisedterm, if possible, by: Adding or removing Hs Moving Hs from one atom to another (including between components) Note that if the Hs could be added to more than one atom an arbitrary choiceterm is made but this is done canonically so the result will always be the same for a given molecule 7. Normalise (straighten) triple bonds and allenes In the context of the ChEMBL databaseterm, it is the molfilesterm standardised according to these rules that are stored in the database and which are in turn the structures made available to the database users.

GetParentterm component Many compound registration systems, including the ChEMBL databaseterm, identify compounds that are related by virtue of being a salt formterm of a common parent structureterm. Therefore, as part of the ChEMBL compound curation pipeline, molecules are identified where the molfileterm contains more than one connected componentterm as well as molecules containing atoms with specified isotopesterm.

The GetParentterm module is applied to just those compounds that match one or both of these criteria. All information about isotopesterm is removed, as are solventsterm and saltsterm present in the molfileterm which match any of the components in the defined saltterm and solventterm lists. Having removed all saltsterm (e.g. Na+ that might be included to neutralise a carboxyl group), the resulting molecule is neutralisedterm and a new molfileterm created as the “parent” molecule. Compounds containing more than one component that are genuine mixturesterm (i.e., all of the components are absent from the saltterm and solventterm lists) has, in the context of the ChEMBL databaseterm, its parent registered as the identical mixture. For cases such as sodium chloride and sodium citrate, where both components are in the saltterm list, the GetParentterm module does not remove any component and the parent remains the same as the saltterm. Here again, the parent is registered as the multicomponent mixture. Compounds containing any of the excluded atoms described above have their isotopesterm and solventsterm removed and then parents created, so that bioactivity dataterm can subsequently be aggregated. For example, the antimony-containing compound sodium stibogluconate has two versions in ChEMBL 26term, both with bioactivity dataterm: CHEMBL3754364 is a version with water of crystallisation and CHEMBL3764926 is a version without. These are annotated as related forms so that the bioactivity dataterm can be seen aggregated in the database. Cyanocobalamin is a cobalt-containing compound which is recorded as a parent and three different isotopesterm in the database (CHEMBL2110563, CHEMBL2104118, CHEMBL2104381 and CHEMBL2096655). Again, the GetParentterm module enables their data aggregation. Organometallic compounds do not however have saltsterm removed due to the complexity of how they are often represented in the deposited molfileterm. For example, this is often achieved by drawing them as disconnected components as is the case for transplatin (CHEMBL1386) which was deposited into the database as N.N.[Cl−].[Cl−].[Pt+2]. Removal of the chloride and ammonia components would incorrectly result in a platinum ion as the parent.

The list of saltsterm used in ChEMBL is based on the USAN Council’s list of pharmacological saltsterm [ 33 ]. Additional entities have been added to this list where a significant number of examples have been present in ChEMBL datasets. The GetParentterm module will remove saltsterm regardless of: (i) the charge statusterm·var (e.g. acetic acid or acetate, Cl− or HCl); (ii) whether or not stereochemistryterm·var is depicted (e.g. tartaric acid); (iii) cis/trans isomers (e.g. maleic and fumaric acid). The saltsterm and solventsterm files are available in the GitHub repository [ 28 ]. Currently, these files contain 162 saltsterm and 9 solventsterm respectively. This list will be maintained and extended if additional saltsterm and solventsterm are identified.

For the avoidance of any doubt, although parents, saltsterm, solventsterm, isotopesterm and mixtures are all identified using the process just described, the bioactivity dataterm recorded in ChEMBL is registered against the form it was measured on. The aggregation by parent structureterm is undertaken to make it easier to identify all the data for saltsterm and isotopesterm of a common parent. For example, paroxetine (CHEMBL490) has bioactivity dataterm determined for the parent moleculeterm·def, two saltsterm, one saltterm/solventterm mixture and two different isotopesterm as well as there being an additional saltterm registered as an FDA approved drug. Another example is amphetamine (CHEMBL405), which has bioactivity dataterm in ChEMBL for eight different saltsterm in addition to the parent amphetamine and an additional two saltsterm that are recorded in drug sources such as the FDA orange book [ 34 ]. The parent aggregation processterm makes this data easily identified and grouped. This is illustrated in Fig. 1 for these two compounds.

Examples of the multicomponent forms of paroxetine and amphetamine and how they have been aggregated by use of the GetParentterm component Availability of structure curation pipeline The code for the pipeline has all been developed using the RDKitterm toolkit (version 2019.09.2.0). It is open source and publicly available in GitHub [ 28 ], currently as version 1.0.0. A conda package is also available to facilitate installation [ 35 ]. The Standardizerterm , Checkerterm and GetParentterm functions are also integrated in the ChEMBL Beaker webservices and can be used in this way via the ‘check’, ‘getParentterm’ and ‘standardize’ endpoints [ 36 ]. Any new features developed by the ChEMBL group will be added to the repository and comments and suggestions from others are welcomed.

Results and discussion One of the objectives in developing the new curation pipeline was to improve the quality of the chemical structuresterm in the ChEMBL databaseterm. As the rules implemented in the new Standardizerterm and GetParentterm modules were largely based on the original ChEMBL rules developed over the last 10 years it is difficult to unambiguously quantify this improvement. To provide a consistent starting point, therefore, the current 1.9 million compounds in the latest release of ChEMBL (ChEMBL 26term) have been completely re-standardised and the parents regenerated for the whole database using the new pipeline.

Checkerterm component To investigate the utility of the Checkerterm module three very different compound sets were used. These were: a set of compounds recently extracted (by manual drawing) from the medicinal chemistry literature by the ChEMBL data extractors; a set of compounds from SureChEMBLterm, the patent database where the chemical structuresterm are obtained by entity recognition followed by name or image to structure conversion [ 37 ] and a randomly selected compound set from PubChemterm (SIDs) as initially submitted by the depositors and prior to any compound standardisation. These three datasets are provided as structure data files (sd files) in the Supplementary Information S1, S2, S3. The results from these three sets are shown in Table 2 .

Checkerterm total number of the different penalty scoresterm output from subjecting the ChEMBL Literature setterm, the SureChEMBLterm set and the PubChemterm Set to the Checkerterm process As expected, the ChEMBL literature setterm, being derived via manual extraction and curation, generates the fewest serious penalty scoresterm, although it is clear there are still thousands of molecules with an undesirable representation (e.g. stereo bonds in rings, atoms with multiple stereo bonds and molecules with stereo bonds to stereocentresterm). The PubChemterm deposited set includes a large number of empty molfilesterm and unknown elements (often represented as an asterix, “*”) in addition to molecules containing overlapping atoms. The SureChEMBLterm set has a large number of compounds with V3000 formatted molfilesterm, radical entities as well as bonds with illegal stereo flags (usually molfilesterm with a 4 in the bond stereo field of the bond block). The difference between these sets is not surprising given the differences in the data sourcesterm·var, confirming that the Checkerterm is suitable for identifying a range of different issues of differing severity for datasets of diverse origin.

A breakdown of the number of compounds from each set which returned penalty scoresterm is shown in Table 3 . When more than one score is returned for a compound the highest score is reported. It is worth noting that 75% of the ChEMBL literature setterm compounds and 60% of both the SureChEMBLterm and PubChemterm compound sets showed no issues in the Checkerterm tests. When structures with a penalty scoreterm of 2 associated with InChI Warnings such as “omitted undefined stereo” and “InChI: Charges were rearranged” are excluded (as described above) then the proportion of compounds with undesirable features that would benefit from correction falls to between 4 and 11%.

Percentages of the compounds in each of the SureChEMBLterm, ChEMBL and PubChemterm sets returning each value as their maximum penalty scoreterm From a ChEMBL perspective the compounds in the latest released version of the database (ChEMBL 26term) have also been subjected to the Checkerterm analysis and this has enabled the identification of small sets of compounds with particular problems that were not previously identified. This dataset is available as an sd file on the ChEMBL FTP site [ 38 ]. These have been prioritised for curation for a subsequent ChEMBL release. The numbers of compounds with penalty scoresterm 5 and 6 are summarised in Table 4 . Those considered to be the highest priority for curation are compounds with a penalty scoreterm of 6 (69 compounds). In future the structures of molecules with penalty scoresterm of 6 will not be added to the database until they have been manually checked and corrected.

Checkerterm penalty scoresterm on the current version of ChEMBL (ChEMBL 26term) The ~ 2500 compounds with penalty scoresterm of 5 are placed in a group for second-tier curation with the following penalty annotations being prioritised: molecule has a crossed bond in a ring (632 compounds). This set does however need to be further divided depending on the ring size. It is considered a more serious issue for small rings than large rings where it is not always clear whether cis, trans or either orientation is correct.

molecule has two (or more) atoms with exactly the same coordinates (273 compounds) Those compounds with mismatches in the number of stereocentresterm will then be addressed.

Standardizerterm component Examples of applying the Standardizerterm to some “ChEMBL-like” molecules are shown in Fig. 2 . These examples show some of the specific standardisations described in the methods section for which the Standardizerterm component was designed to correct. Also, as described above, there will be differences between the last two versions of ChEMBL (versions 25 and 26) due to the re-standardisation as well as the manual curationterm that took place between the two releases. Comparing the structures in these two releases does not give any useful statistics on the effectiveness of the new Standardizerterm . Instead, to investigate this, the three sets of molecules previously described have been standardised by the new Standardizerterm module. The critical differences are where the new standardisation results in a different InChI from the one obtained prior to standardisation.

Examples of standardisations that have been applied to a set of compounds. The compound structure before and after standardisation is shown. a Fix hypervalent nitro groupsterm, b remove explicit H atomsterm, c fix covalently drawn alkaline metals connected to O or N to ionic forms, d Standardise sulphoxides to charge separated form, e normalise (straighten) allene bonds The percentage of molecules with changed InChIsterm are low for all datasets and always < 2% in each case. The majority of the changes occur where the Standardizerterm has neutralisedterm a previously charged molecule, for example by removing a proton from a positively charged nitrogen or protonating a carboxyl group. A summary of these changes is shown in Table 5 . This also indicates which layer of the InChI has changed. Whilst initially it may appear that changes to the connectivity layer are unexpected for a standardisation process these examples are due to saltsterm where the overall molfileterm is not neutral and where a proton has been removed to confer neutrality. To further exemplify this, some examples from the ChEMBL literature setterm are shown in Fig. 3 where the effect of standardisation on neutralising the molecules can be observed.

Summary of the number of compounds that have changed InChIKeysterm following standardisation for the SureChEMBLterm, ChEMBL literature and PubChemterm deposited set Examples of compounds from the ChEMBL literature setterm where the InChIKeyterm changed on standardisation due to the rebalancing of the charge on the compound As different databases use different rules to standardise compound structures, it was of interest to compare the new ChEMBL Standardizerterm with the standardiser used by the PubChemterm database [ 39 ]. However, as the PubChemterm PUG service was slow to run via their API, only a small set of compounds were used for this comparison and comprised approved small molecule drugs as annotated in ChEMBL 25 (3071 compounds). This dataset is provided as an sd file in Supplementary Information S4. The ChEMBL 25 structures for these were taken and standardised using the new ChEMBL RDKitterm Standardizerterm and also the PubChem standardiserterm. In total there were 97 compounds that gave different Standard InChIKeysterm after the respective standardisations. The key difference is that the PubChem standardiserterm generates a canonical tautomerterm as part of the process but ChEMBL does not. In many instances this is not reflected by a change to the Standard InChIterm but for some molecules the InChI does change. There are also some differences in the double bond representation (cis or trans (either)). Furthermore, PubChemterm generates all structures with explicit hydrogens which are not present in the ChEMBL structures. Some of these examples are shown in Fig. 4 .

Examples of approved drugs standardised by the ChEMBL RDKitterm Standardizerterm and the PubChem standardiserterm GetParentterm component The GetParentterm module was used to obtain the parent structuresterm for the standardised structures from ChEMBL where these contained more than one component or an isotopeterm. This resulted in over 100,000 compounds where a saltterm, solventterm, isotopeterm or combination thereof could be removed to create a parent structureterm. Molfilesterm that are a mixture of components can contain a number of different combinations such as: a parent and its saltsterm; a parent, its saltsterm and solventterm; a combination of only saltsterm; or a true mixture. A true mixture, for example, would include compounds such as Co-trimoxazole (CHEMBL58061) which is a marketed product comprising trimethoprim and sulfamethoxazole and which has been tested as a combination in a number of bioactivity assays. Additionally, any of these multicomponent compoundsterm may have been tested for bioactivity or used in a clinical setting as a specific isotopic form. A summary of the composition of the multicomponent compoundsterm in ChEMBL 26term are shown in Fig. 5 . Whilst only ~ 6% of the compounds in ChEMBL are multicomponent (mostly saltsterm), these compounds have over a million activity values recorded against them.

The composition and number of the compounds containing more than one component in ChEMBL 26term as identified by the GetParentterm module. The numbers in brackets refer to the number of compounds in each grouping that contain isotopesterm Some examples of the output from the GetParentterm process are shown in Fig. 6 . This illustrates a number of features of the GetParentterm module including the need to re-standardise a parent to some extent once a saltterm is removed, and in particular to re-neutralise the charges on the parent moleculeterm·def. The sodium saltterm of a carboxyl-containing compound is an obvious example of this. Once the Na+ cation is removed this leaves a net negative change on the parent so the GetParentterm module will add a proton to neutralise it. Quaternary nitrogen compounds are the exception, where the counterion (for example a chloride) is removed but, due to the quaternary nature of the nitrogen, it is not possible to neutralise the parent which therefore remains as the positively charged cation. Another important function of the GetParentterm module is that if a compound is a 2:1 complex with two parent moleculesterm·def and one saltterm molecule, when the GetParentterm module is applied it can recognise that the two components remaining after saltterm removal are identical and so only one molecule is returned as the parent. Atorvastatin calcium is an example of this as can be seen in Fig. 6 .

Examples of applying the GetParentterm module to some representative ChEMBL compounds containing varying combinations of saltsterm, isotopesterm and solventsterm. The “Child” is the compound before and “Parent” the compound after the process has been applied Future work Now that the components of the curation pipeline are being used in a production context for curation of compounds being added to the ChEMBL databaseterm, it will provide opportunities to review and refine any of the rules on the basis of a large consistently standardised dataset. The ChEMBL group will continue to evolve the curation pipeline as further needs are identified. The code is also now freely available to the community and other researchers are encouraged to suggest modifications so that the curation pipeline can improve further over time. Comments and issues can be added to the issues section of the ChEMBL Structure Pipeline GitHub repository [ 28 ].

Conclusion The three components of the structure curation pipeline described here have been developed as an open source project and are now available for researchers to use and adapt for their own applications. These components have been used by the ChEMBL group to produce the chemical structuresterm in the latest release of the ChEMBL databaseterm (ChEMBL26). This has resulted in the correction of certain errors in standardisation that were present in previous ChEMBL releases and the identification of other issues that can now be prioritised for future manual curationterm. Additionally, the pipeline has been tested on less well curated datasets and demonstrated to be sufficiently robust to be used in the automatic structure checking and standardisation of such datasets. A comparison between the standardiser used by PubChemterm (another large publicly available database) and the one developed here gave only a small percentage of compounds with non-identical structures (defined as having different Standard InChIsterm). Where such differences were present this was largely due to variations in the business rulesterm of the two database providers.

Acknowledgements The authors would like to explicitly acknowledge the indirect contributions of Matt Swain, the original author of the standalone MolVSterm package, and Susan Leung, who ported the python version of MolVSterm to C++ and integrated it into the RDKitterm core.

Authors’ contributions APB, AH, AG and ARL initiated and organised the work. LJB developed the original curation pipeline, which this work is built upon. APB, AH, AG, LJB, MdV and FA devised the business rulesterm. GL, EF and APB developed and wrote the code. AG and EF implemented the pipeline. LJB and MdV curated the data. AH, APB and AG performed the analysis and wrote the manuscript with contributions from GL, EF and ARL. All authors read and approved the final manuscript.

Funding This work was funded by Strategic Awards from the Wellcome Trust [WT086151/Z/08/Z, WT104104/Z/14/Z] and the Member States of the European Molecular Biology Laboratory (EMBL); Funding for open access charge: Wellcome Trust.

Availability of data and materials The datasets used for the analyses are provided as sd files in Supplementary Files. S1_surechemblterm_set, S2_pubchemterm_set, S3_chembl_literature_set, S4_approved_drug_set and are available from the ChEMBL FTP site: ftp://ftp.ebi.ac.uk/pub/databases/chembl/standardiser_data_sets .

Competing interests The authors have no competing interests.

Footnotes References Associated Data Data Citations Data Availability Statement The datasets used for the analyses are provided as sd files in Supplementary Files. S1_surechemblterm_set, S2_pubchemterm_set, S3_chembl_literature_set, S4_approved_drug_set and are available from the ChEMBL FTP site: ftp://ftp.ebi.ac.uk/pub/databases/chembl/standardiser_data_sets .