The LOTUS Initiative for Open Natural Products Research: Knowledge Management through Wikidata

Contemporary bioinformatic and chemoinformatic capabilities hold promise to reshape knowledge management, analysis and interpretation of data in natural products research. Currently, reliance on a disparate set of non-standardized, insular, and specialized databases presents a series of challenges to data access, either within the discipline or to integration and interoperability between related domains. The fundamental elements of exchange are referenced structure-organism pairs that establish relationships between distinct molecular structures and the living organisms from which they were identified. Consolidating and sharing such information via an open platform has strong transformative potential for natural products research and beyond. This is the ultimate goal of the newly established LOTUS initiative, which has now completed the first steps toward the harmonization, curation, validation and open dissemination of 700,000+ referenced structure-organism pairs. LOTUS data is hosted on Wikidata and regularly mirrored on https://lotus.naturalproducts.net. Data sharing within the Wikidata framework broadens data access and interoperability, opening new possibilities for community curation and evolving publication models. Furthermore, embedding LOTUS data into the vast Wikidata knowledge graph will facilitate new biological and chemical insights. The LOTUS initiative represents an important advancement in the design and deployment of a comprehensive and collaborative natural products knowledge base.


Introduction Evolution of Electronic Natural Products Resources
Natural Products (NP) research is a transdisciplinary eld with wide-ranging interests: from fundamental structural aspects of naturally-occurring molecular entities to their e ects on living organisms and extending to the study of chemically-mediated interactions within entire ecosystems. Despite the ambiguous de nition of "natural" ("All natural," 2007), the basis of our de nition of a NP as a chemical entity found in a living organism is predicated on the identi cation of the explicit relationship between a naturally-occurring chemical entity and its source organism. A third fundamental element of a structure-organism pair is a reference to the experimental evidence that establishes the linkages between a chemical structure and a biologicl organism and a future-oriented electronic NP resource should contain only fully-referenced structure-organism pairs.
Reliance on data from the NP literature presents many challenges. The assembly and integration of NP occurrences into an inter-operative platform relies primarily on access to a heterogeneous set of databases (DB) whose content and maintenance status are critical factors in this dependency (Tsugawa, 2018). A tertiary inter-operative NP platform is thus dependent on a secondary set of data that has been selectively annotated into a DB from primary literature sources. The experimental data itself re ects a complex process involving collection or sourcing of natural material (and establishment of its identity), a series of material transformation and separation steps and ultimately the chemical or spectral elucidation of isolates. The specter of human error and the potential for the introduction of biases are present at every phase of this journey. These include publication biases (Lee et al., 2013), such as emphasis on novel and/or bioactive structures in the review process, or, in DB assembly stages, with selective focus on a speci c compound class or a given taxonomic range, or disregard for annotation of other relevant evidence that may have been presented in primary sources. Temporal biases also exist: a technological "state-of-the-art" when published can eventually be recast as anachronistic.
The advancement of NP research has always relied on the development of new technologies. In the past century alone, the rate at which unambiguous identi cation of new NP entities from biological matrices can be achieved has been reduced from years to days and in the past few decades, the scale at which new NP discoveries are being reported has increased exponentially. Without a means to access and process these disparate NP data points, information is fragmented and scienti c progress is impaired (Balietti et al., 2015). To this extent, contemporary bioinformatic tools enable the (re-)interpretation and (re-)annotation of (existing) datasets documenting molecular aspects of biodiversity (Jarmusch et al., 2020;Mongia and Mohimani, 2021).
While large, well-structured and freely accessible DB exist, they are often concerned primarily with chemical structures (e.g. PubChem (Kim et al., 2019), with over 100M entries) or biological organisms (e.g. GBIF ("GBIF.org," 2020), with over 1,400M entries), but scarce interlinkages limit their application for documentation of NP occurrence(s). Currently, no open, cross-kingdom, comprehensive, computer-interpretable electronic NP resource links NP and their producing organisms, along with referral to the underlying experimental work. This shortcoming breaks the crucial evidentiary link required for tracing information back to the original data and assessing its quality. Even valuable commercially available e orts for compiling NP data, such as the Dictionary of Natural Products (DNP), can lack proper documentation of these critical links.
Pioneering e orts to address such challenges led to the establishment of KNApSAck (Shinbo et al., 2006), which is likely the rst public, curated electronic NP resource of referenced structure-organism pairs. KNApSAck currently contains 50,000+ structures and 100,000+ structure-organism pairs. However, the organism eld is not standardized and access to the data is not straightforward. One of the earliest-established electronic NP resources is the NAPRALERT dataset (Graham and Farnsworth, 2010), which was compiled over ve decades from the NP literature, gathering and annotating data derived from over 200,000 primary literature sources. The dataset contains 200,000+ distinct compound names and structural elements, along with 500,000+ records of distinct, fully-cited structure-organism pairs. In total, NAPRALERT contains over 900,000 such records, due to equivalent structure-organism pairs reported in di erent citations. NAPRALERT is not an open platform, employing an access model that provides only limited free searches of the dataset. Finally, the NPAtlas (van Santen et al., 2019) is a more recent project aimed at complying with the FAIR (Findability, Accessibility, Interoperability and Reuse) guidelines for digital assets (Wilkinson et al., 2016) and o ering web access. While the NPAtlas encourages submission of new compounds with their biological source, it focuses on microbial NP and ignores a wide range of biosynthetically active organisms found in the Archaeplastida.
Building on experience with the recently published COlleCtion of Open NatUral producTs (COCONUT) (Sorokina et al., 2021), the LOTUS initiative seeks to address the aforementioned shortcomings. At its current stage of development, LOTUS disseminates 700,000+ referenced structure-organism pairs. After extensive data curation and harmonization, each pair was standardized at the chemical, biological and reference levels. These e orts and experiences represent an intensive preliminary curatorial phase and the rst major step towards providing a high quality, computer-interpretable knowledge base capable of transforming NP research data management from a classical (siloed) database approach to an optimally-shared resource.

Accommodating Principles of FAIRness and TRUSTworthiness for Natural Products Knowledge Management
In awareness of the multi-faceted pitfalls associated with implementing, using and maintaining classical scienti c DBs (Helmy et al., 2016), and to enhance current and future sharing options, the LOTUS initiative selected the Wikidata platform for disseminating its resources. Since its creation, Wikidata has focused on cross-disciplinary and multilingual support. Wikidata is curated and governed collaboratively by a global community of volunteers, about 20,000 of which are contributing monthly. Wikidata currently contains more than 1 billion statements in the form of subject-predicate-object triples. Triples are machine-interpretable and can be enriched with quali ers and references. Within Wikidata, data triples correspond to approximately 90 million entries, which can be grouped into classes as diverse as countries, songs, disasters, or chemical compounds. The statements are closely integrated with Wikipedia and serve as the source for its infoboxes. Various work ows have been established for reporting such classes, particularly those of interest to life sciences, such as genes, proteins, diseases, drugs, or biological taxa (Waagmeester et al., 2020).
Building on the principles and experiences described above, the present report introduces the development and implementation of the LOTUS work ow for NP occurrences' curation and dissemination, which applies both FAIR and TRUST (Transparency, Responsibility, User focus, Sustainability and Technology) principles (Lin et al., 2020). LOTUS' data upload and retrieval procedures ensure optimal accessibility by the research community, allowing any researcher to contribute, edit and reuse the data with a clear and open CC0 license (Creative Commons 0). Despite many advantages, Wikidata hosting has some notable, yet manageable drawbacks. While its SPARQL query language o ers a powerful way to query available data, it can also appear intimidating to the less experienced user. Furthermore, some typical queries of molecular electronic NP resources such as structural or spectral searches are not yet available in Wikidata. To bridge this gap, LOTUS is hosted in parallel at https://lotus.naturalproducts.net (LNPN) within the naturalproducts.net ecosystem. The Natural Products Online is a portal for open-source and open-data resources for NP research. In addition to the generalistic COCONUT and LNPN databases, the portal will enable hosting of arbitrary and skinned collections, themed in particular by species or taxonomic clade, by geographic location or by institution, together with a range of cheminformatics tools for NP research. LNPN is periodically updated with the latest LOTUS data. This dual hosting provides an integrated, community-curated and vast knowledge base (via Wikidata), as well as a NP community-oriented product with tailored search modes (via LNPN).
The LOTUS initiative and its multiple data interaction options establish the basis for transparent and sustainable access, sharing and creation of knowledge on NP occurrence. LOTUS represents an important advancement in the design and deployment of a comprehensive and collaborative NP knowledge base. More broadly, the LOTUS initiative fosters cross-fertilization of the elds of chemistry, biology and associated disciplines.

Results & Discussion
This section is structured as follows: rst, we present an overview of the LOTUS initiative at its current stage of development. The central curation and dissemination elements of the LOTUS initiative are then explained in detail. The third section addresses the interaction modes between LOTUS and its end-users, including data retrieval, addition and editing. The nal section is dedicated to the interpretation of LOTUS data and illustrates the dimensions and qualities of the current LOTUS dataset from chemical and biological perspectives.

Blueprint of the LOTUS Initiative
Building on the standards established by the related Wikidata project (Chemistry, Taxonomy and Source Metadata), a NP chemistry-oriented subproject was created (Chemistry/Natural products). Its central data consists of three minimal su cient objects: A chemical structure object, with associated Simpli ed Molecular Input Line Entry System (SMILES) (Weininger, 1988), International Chemical Identi er (InChI) (Heller et al., 2013) and InChIKey (a hashed version of the InChI). A biological organism object, with associated taxon name, the taxonomic DB where it was described and the taxon ID in the respective DB. A reference object describing the structure-organism pair, with the associated article title and a Digital Object Identi er (DOI), a PubMed (PMID), or PubMed Central (PMCID) ID.
As data formats are largely inhomogeneous among existing electronic NP resources, elds related to chemical structure, biological organism and references are variable and essentially not standardized. Therefore, LOTUS implements multiple stages of harmonization, cleaning and validation ( Figure 1, stages 1 to 3). LOTUS employs a Single Source of Truth (SSOT, Single_source_of_truth) to ensure data reliability and continuous availability of the latest curated version of LOTUS data in both Wikidata and LNPN (Figure 1, stage 4). The SSOT approach consists of a PostgreSQL DB that structures links and data schemes such that every data element has a single place. The LOTUS processing pipeline is tailored to e ciently include and di use novel or curated data directly from new sources or at the Wikidata level. This iterative work ow relies both on data addition and retrieval actions as described in the Data Interaction section. The overall process leading to referenced and curated structureorganisms pairs is illustrated in Figure 1 and detailed below.
By design, this iterative process fosters community participation, essential to e ciently document NP occurrences. All stages of the work ow are described on the git sites of the LOTUS initiative at https://gitlab.com/lotus7 and https://github.com/mSorok/LOTUSweb. At the time of writing, 700,000+ LOTUS entries contained a curated chemical structure, biological organism and reference and were available on both Wikidata and LNPN. As the LOTUS data volume is expected to increase over time, a frozen (as of 2021-05-23), tabular version of this dataset with its associated metadata is made available at https://osf.io/eydjs/.
The contacts of the electronic NP resources not explicitly licensed as open were individually reached for permission to access and reuse data. A detailed list of data sources and related information is available as SI-1. All necessary scripts for data gathering and harmonization can be found in the lotusprocessor repository in the src/1_gathering directory. All subsequent and future iterations that include additional data sources, either updated information from the same data sources or new data, will involve a comparison of the new with previously gathered data at the SSOT level to ensure that the data is only curated once.

Data Cleaning & Validation
As shown in Figure 1, data curation consisted of three stages: harmonization, cleaning and validation. Thereby, after the harmonization stage, each of the three central objects -chemical compounds, biological organisms and reference -were cleaned. Given the data size (2.5M+ initial entries), manual validation was unfeasible. Curating the references was a particularly challenging part of the process. Whereas organisms are typically reported by at least their vernacular or scienti c denomination and chemical structures via their SMILES, InChI, InChIKey or image (not covered in this work), references su er from largely insu cient reporting standards. Despite poor standardization of the initial reference eld, proper referencing remains an indispensable way to establish the validity of structureorganism pairs. Better reporting practices, supported by tools such as Scholia (Nielsen et al., 2017;Rasberry et al., 2019) and relying on Wikidata, Fatcat, or Semantic Scholar should improve referencerelated information retrieval in the future.
In addition to curating the entries during data processing, 420 referenced structure-organism pairs were selected for manual validation. An entry was considered as valid if: i) the structure (in the form of any structural descriptor that could be linked to the nal sanitized InChIKey) was described in the reference ii) the containing organism (as any organism descriptor that could be linked to the accepted canonical name) was described in the reference and iii) the reference was describing the occurrence of the chemical structure in the biological organism. This process allowed us to establish rules for automatic ltering and validation of the entries. The ltering was then applied to all entries. To con rm the e cacy of the ltering process, a new subset of 100 diverse, automatically curated and automatically validated entries was manually checked, yielding a rate of 97% of true positives. The detailed results of the two manual validation steps are reported in Supporting Information SI-2. The resulting data is also available in the dataset shared at https://osf.io/eydjs/. Table 1 shows an example of a referenced structure-organism pair before and after curation. This process resolved the structure to an InChIKey, the organism to a valid taxonomic name and the reference to a DOI, thereby completing the essential referenced structure-organism pair. The gure highlights, for example, the essential contribution of the DOI category of references contained in NAPRALERT towards the current set of validated references in LOTUS. The combination of the results of the automated curation pipeline and the manually curated entries led to the establishment of four categories (manually validated, manually rejected, automatically validated and automatically rejected) of the referenced structure-organism pairs that formed the processed part of the SSOT. Out of a total of 2.5M+ pairs, the manual and automatic validation retained 700,000+ pairs (approximately 30%), which were then selected for dissemination on Wikidata. The disseminated data contains 250,000+ unique chemical structures, 30,000+ distinct organisms and 75,000+ references.

Data Dissemination
Research worldwide can bene t the most when all results of published scienti c studies are fully accessible immediately upon publication (Agosti and Johnson, 2002). This concept is considered the foundation of scienti c investigation and a prerequisite for e ectively directing new research e orts based on prior information. To achieve this, research results have to be made publicly available and reusable. As computers are now the main investigation tool for a growing number of scientists, all research data including those in publications should be disseminated in computer-readable format, following the FAIR principles. LOTUS uses Wikidata as a repository for referenced structure-organism pairs, as this allows documented research data to be integrated with a large, pre-existing and extensible body of chemical and biological knowledge. The dynamic nature of Wikidata fosters the continuous curation of deposited data through the user community. Independence from individual and institutional funding represents another major advantage of Wikidata. The Wikidata knowledge base and the option to use elaborate SPARQL queries allow the exploration of the dataset from a sheer unlimited number of angles. The openness of Wikidata also o ers unprecedented opportunities for community curation, which will support, if not guarantee, a dynamic and evolving data repository. At the same time, certain limitations of this approach can be anticipated. Despite (or possibly due to) their power, SPARQL queries are complex and often require a relatively in-depth understanding of the models and data structure. This involves a steep learning curve which tends to discourage end-users. Furthermore, traditional ways to query electronic NP resources such as structural or spectral searches are currently not within the scope of Wikidata and, thus, are addressed in LNPN. Using the preexisting COCONUT template, LNPN hosting allows the user to perform structural searches by drawing a molecule, thereby addressing the current lack of structural search possibilities in Wikidata. Since metabolite pro ling by Liquid Chromatography (LC) -Mass Spectrometry (MS) is now routinely used for the chemical composition assessment of natural extracts, future versions of LOTUS and COCONUT are envisioned to be augmented by predicted MS spectra and hosted at https://naturalproducts.net/ to allow mass and spectral-based queries. To facilitate queries focused on speci c taxa (e.g., "return all molecules found in the Asteraceae family"), a uni ed taxonomy is paramount. As the taxonomy of living organisms is a complex and constantly evolving eld, all the taxon identi ers from all accepted taxonomic DB for a given taxon name were kept. Initiatives such as the Open Tree of Life (OTL) (Rees and Cranston, 2017) will help to gradually reduce these discrepancies, the Wikidata platform should support such developments. OTL also bene ts from regular expert curation and new data. As the taxonomic identi er property for this resource did not exist in Wikidata, its creation was requested and obtained. The property is now available as "Open Tree of Life ID" (P9157).
Following the previously described curation process, all validated entries have been made available through Wikidata and LNPN. LNPN will be regularly mirroring Wikidata LOTUS through the SSOT as described in Figure 1.

User Interaction with LOTUS Data
The possibilities to interact with the LOTUS data are numerous. The following gives examples of how to retrieve, add and edit LOTUS data.

Data Retrieval
LOTUS data can be queried and retrieved either directly in Wikidata or on LNPN, both of which have distinct advantages. While Wikidata o ers modularity at the cost of potentially complex access to the data, LNPN has a graphical user interface with capabilities of drawing chemical structure, simpli ed structural or biological ltering and advanced chemical descriptors, albeit with a more rigid structure. For bulk download, a frozen version of LOTUS data (timestamp of 2021-05-23) is also available at https://osf.io/eydjs/. More re ned approaches to the direct interrogation of the up-to-date LOTUS data both in Wikidata and LNPN are detailed in the following.

Wikidata
The easiest way to search for NP occurrence information in Wikidata is by typing the name of a chemical structure directly into the "Search Wikidata" eld on the upper right of the Wikidata homepage. For example, by typing "erysodine", the user will land on the page of this compound (Q27265641). Scrolling down to the "found in taxon" statement will allow the user to view the biological organisms reported to contain this NP ( Figure 3). Clicking the reference link under each taxon name links to the publication(s) documenting the occurrence. The typical approach to more elaborated queries consists in writing SPARQL queries using the Wikidata Query Service or a direct connection to a SPARQL endpoint. Below are some examples from simple to more elaborated queries, demonstrating what can be done using this approach. The fulltext queries with explanations are included in SI-3. The queries presented in Table 2 are only a few examples and many other ways of interrogating LOTUS can be formulated. Generic queries can be used, for example, for hypothesis generation when starting a research project. For instance, a generic SPARQL query -listed in Table 2 as "Which are the available referenced structure-organism pairs?" -retrieves all structures, identi ed by their InChIKey (P235), which contain "found in taxon" (P703) statements that are stated in (P248) a bibliographic reference: https://w.wiki/3JpE. Data can then be exported in various formats, such as classical tabular formats, json, or html tables (see Download tab on the lower right of the query frame). At the time of writing (2021-05-05), this query (without the LIMIT 1000) returned 797,123 entries; a frozen query result is available at https://osf.io/thqaw/.
Targeted queries allowing to interrogate LOTUS data from the perspective of one of the three objects forming the referenced structure-organism pairs can be also built. Users can, for example, retrieve a list of all structures reported from a given organism, such as all structures reported from Arabidopsis thaliana (Q158695) (https://w.wiki/3HLn). Alternatively, all organisms containing a given chemical can be queried via its structure, such as in the search for all organisms where β-sitosterol (Q121802) was found in (https://w.wiki/3HLy). For programmatic access, the lotus-wikidata-exporter repository also allows data retrieval in RDF format and as TSV tables.
As indicated, certain types of queries that are customary in existing molecular electronic resources, such as structure or similarity searches, are not directly available in Wikidata as SPARQL does not natively support a simple integration of such queries. To address this issue, Galgonek et al. developed an in-house SPARQL engine that allows utilization of Sachem, a high-performance chemical DB cartridge for PostgreSQL for ngerprint-guided substructure and similarity search (Kratochvíl et al., 2018). The engine is used by the Integrated Database of Small Molecules (IDSM) that operates, among other things, several dedicated endpoints allowing structural search in selected small-molecule datasets via SPARQL (Kratochvíl et al., 2019). To allow substructure and similarity searches via SPARQL also on compounds from Wikidata, a dedicated IDSM/Sachem endpoint was created for the LOTUS project. The endpoint indexes isomeric (P2017) and canonical (P233) SMILES code available in Wikidata. To ensure that data is kept up-to-date, SMILES codes are automatically downloaded from Wikidata daily. The endpoint allows users to run federated queries and, thereby, proceed to structureoriented searches on the LOTUS data hosted at Wikidata. For example, the SPARQL query https://w.wiki/3HMD returns a list of all organisms that produce NP with an indolic sca old. The output is aggregated at the parent taxa level of the containing organisms and ranked by the number of sca old occurrences.

Lotus.NaturalProducts.Net (LNPN)
In the search eld of the LNPN interface (https://lotus.naturalproducts.net/), simple queries can be achieved by typing the molecule name (e.g., protopine) or pasting a SMILES, InChI, InChIKey string, or a Wikidata identi er. All compounds reported from a given organism can be found by entering the organism name at the species or any higher taxa level (e.g. Tabernanthe iboga). Compound search by chemical class is also possible.
Alternatively, a structure can be directly drawn in the structure search interface (https://lotus.naturalproducts.net/search/structure), where the user can also decide on the nature of the structure search (exact, similarity, substructure search). Re ned search mode combining multiple search criteria, in particular physicochemical properties, is available in the advanced search interface (https://lotus.naturalproducts.net/search/advanced).
Within LNPN, LOTUS bulk data can be retrieved as SDF or SMILES les, or as a complete MongoDB dump via https://lotus.naturalproducts.net/download. Extensive documentation describing the search possibilities and data entries is available at https://lotus.naturalproducts.net/documentation. LNPN can also be queried via the application programming interface (API) as described in the documentation.

Data Addition and Evolution
One major advantage of the LOTUS architecture is that every user has the option to contribute to the NP occurrences documentation e ort by adding new or editing existing data. As all LOTUS data applies the SSOT mechanism, reprocessing of previously treated elements is avoided. However, at the moment, the SSOT channels are not open to the public for direct write access to maintain data coherence and evolution of the SSOT scheme. For now, the users can employ the following approaches to add or modify data in LOTUS.

Sources
LOTUS data management involves regular re-importing of both current and new data sources. New and edited information from these electronic NP resources will be checked against the SSOT. If absent or di erent, data will be passed through the curation pipeline and subsequently stored in the SSOT. Accordingly, by contributing to external electronic NP resources, any researcher has a means of providing new data for LOTUS, keeping in mind the inevitable delay between data addition and subsequent inclusion into LOTUS.

Wikidata
The currently favored approach to add new data to LOTUS is to edit Wikidata entries directly. Newly edited data will then be imported into the SSOT repository. There are several ways to interact with Wikidata which depend on the technical skills of the user and the volume of data to be imported/modi ed.

Manual Upload
Any researcher interested in reporting NP occurrences can manually add the data directly in Wikidata, without any particular technical knowledge requirement. The only prerequisite is a Wikidata account and following the general object editing guidelines. Regarding the addition of NP-centered objects (i.e., referenced structure-organisms pairs), users shall refer to the WikiProject Chemistry/Natural products group page.
A tutorial for the manual creation and upload of a referenced structure-organism pair to Wikidata is available in SI-4. While direct Wikidata upload is possible, contributors are encouraged to use the LOTUS curation pipeline as a preliminary step to strengthen the initial data quality. The added data will therefore bene t from the curation and validation stages implemented in the LOTUS processing pipeline.

Batch and Automated Upload
Through the initial curation process described previously, 700,000+ referenced structure-organism pairs were validated for Wikidata upload. To automate this process, a set of programs were written to automatically process the curated outputs, group references, organisms and compounds, check if they are already present in Wikidata (using SPARQL and direct Wikidata querying) and insert or update the entities as needed (i.e., upserting). These scripts can be used for future batch upload of properly curated and referenced structure-organism pairs to Wikidata. Programs for data addition to Wikidata can be found in the repository lotus-wikidata-importer. The following Xtools page o ers an overview of the latest activity performed by our NPimporterBot, using those programs.

Data Editing
Even if correct at a given time point, scienti c advances can invalidate or update previously uploaded data. Thus, the possibility to continuously edit the data is desirable and guarantees data quality and sustainability. Community-maintained knowledge bases such as Wikidata encourage such a process. Wikidata presents the advantage of allowing both manual and automated correction. Field-speci c robots such as SuccuBot, KrBot, Pi_bot and ProteinBoxBot or our NPimporterBot went through an approval process. The robots are capable of performing thousands of edits without the need for human input. This automation helps reduce the amount of incorrect data that would otherwise require manual editing. However, manual curation by human experts remains irreplaceable as a standard. Users who value this approach and are interested in contributing are invited to follow the manual curation tutorial in SI-4.
The Scholia platform provides a visual interface to display the links among Wikidata objects such as researchers, topics, species or chemicals. It now provides an interesting way to view the chemical compounds found in a given biological organism (see here for the metabolome view of Eurycoma longifolia). If Scholia currently does not o er a direct editing interface for scienti c references, it still allows users to proceed to convenient batch editing via Quick Statements. The adaptation of such a framework to edit the referenced structure-pairs in the LOTUS initiative could thus facilitate the capture of future expert curation, especially manual e orts that cannot be replaced by automated scripts.

Data Interpretation
To illustrate the nature and dimensions of the LOTUS dataset, some selected examples of data interpretation are shown. First, the repartition of chemical structures among four important NP reservoirs: plants, fungi, animals and bacteria (Table 3). Then, the distribution of biological organisms according to the number of related chemical structures and likewise the distribution of chemical structures across biological organisms are illustrated ( Figure 4). Furthermore, the individual electronic NP resources participation in LOTUS data is resumed using the UpSet plot depiction, which allows the visualization of intersections in data sets ( Figure 5). Across these gures we take again the two previous examples, i.e, β-sitosterol as chemical structure and Arabidopsis thaliana as biological organism because of their well-documented statuses. Finally, a biologically-interpreted chemical tree and a chemically-interpreted biological tree are presented ( Figure 6 and Figure 7). The examples illustrate the overall chemical and biological coverage of LOTUS by linking family-speci c classes of chemical structures to their taxonomic position. Table 3, Figures 4, 6 and 7 were generated using the frozen data (2021-05-23 timestamp), which is available for download at https://osf.io/eydjs/. Figure 5 required a dataset containing information from DNP and the complete data used for its generation is therefore not available for public distribution. All scripts used for the generation of the gures (including SI-5) are available in the lotus-processor repository in the src/4_visualizing directory for reproducibility. Table 3 summarizes the repartition of chemical structures and their chemical classes (according to NPClassi er (kim et al., 2020)) across the biological organisms reported in LOTUS. For this, biological organisms were grouped into four arti cial taxonomic levels (plants, fungi, animals and bacteria). These were built by combining the two highest taxonomic levels in the OTL taxonomy, namely Domain and Kingdom levels. When the chemical structure/class was reported only in one taxonomic grouping, it was counted as "speci c".

Distributions of Organisms per Structure and Structures per Organism
Readily achievable outcomes from LOTUS show that the depth of exploration of the world of NP is rather limited: as depicted in Figure 4, on average, three organisms are reported per chemical structure and eleven structures per organism. Notably, half of all structures have been reported from a single organism and half of all studied organisms are reported to contain ve or fewer structures. Metabolomic studies suggest that these numbers are heavily underrated (Noteborn et al., 2000;Wang et al., 2020) and indicate that a better reporting of the metabolites detected in the course of phytochemical investigations should greatly improve coverage. This incomplete coverage may be partially explained by the habit in classical NP journals to accept only new and/or bioactive chemical structures for publication.

Contribution of Individual Electronic NP Resources to LOTUS
The added value of the LOTUS initiative to assemble multiple electronic NP resources is illustrated in Figure 5 : Panel A shows the contributions of the individual electronic NP resources to the ensemble of chemical structures found in one of the most studied vascular plants, Arabidopsis thaliana ("Mouseear cress"; Q147096). Panel B shows the ensemble of taxa reported to contain the planar structure of the widely occuring triterpenoid β-sitosterol (Q121802). Figure 5 A. also shows that according to NPClassi er, the chemical pathway distribution across electronic NP resources is unconserved. Note that NPClassi er and ClassyFire (Djoumbou Feunang et al., 2016) chemical taxonomies are both available as metadata in the frozen LOTUS export and LNPN. Both classi cation tools return a chemical taxonomy for individual structures, thus allowing their grouping at higher hierarchical levels, in the same way as it is done for biological taxonomies. The UpSet plot in Figure 5 indicates the poor overlap of preexisting electronic NP resources and the added value of an aggregated dataset. This is particularly well illustrated in Figure 5 B., where the number of organisms for which the planar structure of β-sitosterol (KZJWDPNRJALLNS) has been reported is shown for each intersection. NAPRALERT has by far the highest number of entries (2,085 in total), while other electronic NP resources complement this well: e.g., UNPD has 532 reported organisms with β-sitosterol that do not overlap with those reported in NAPRALERT. Of note, β-sitosterol is documented in only 3 organisms in the DNP, highlighting the importance of a better systematic reporting of ubiquitous metabolites and the interest of multiple data sources agglomeration.

A Biologically-interpreted Chemical Tree
The chemical diversity captured in LOTUS is here displayed using tmap ( Figure 6), a visualization library allowing the structural organization of large chemical datasets as a minimum spanning tree . Using Faerun, an interactive HTML le is generated to display metadata and molecule structures by embedding the SmilesDrawer library Reymond, 2018a, 2018b). Planar structures were used for all compounds to generate the TMAP (chemical space treemap) using MAP4 encoding (Capecchi et al., 2020). As the tree organizes structures according to their molecular ngerprint, an anticipated coherence between the clustering of compounds and the mapped NPClassi er chemical class is observed (Figure 6 A.). For clarity, the eight most represented chemical classes of LOTUS plus the quassinoids and carotenoids (C40, β-β) classes are mapped, with examples of a quassinoid (green star) and a carotenoid (yellow star) and their corresponding location in the TMAP.
To explore relationships between chemistry and biology, it is possible to map taxonomical information such as the most reported biological family per chemical compound (Figure 6 B.) or the biological speci city of chemical classes (Figure 6 C.) on the TMAP. The biological speci city score at a given taxonomic level for a given chemical class is calculated as the number of structure-organism pairs within the taxon where the chemical class occurs the most, divided by the total number of pairs. See Equation 1: This visualization allows to highlight chemical classes speci c to a given taxon, such as the quassinoids in the Simaroubaceae family. In this case it is striking to see how well the compounds of a given chemical class (quassinoids) (Figure 6 A.) and the most reported plant family per compound (Simaroubaceae) (Figure 6 B.) overlap. This is also evidenced on Figure 6 C. with a chemical class speci city of 0.95 at the biological family level for quassinoids. In this plot, it is also possible to identify chemical classes that are widely spread among living organisms, such as the carotenoids (C40, β-β), which exhibit a speci city of 0.12 at the biological family level. This means that among all the carotenoids (C40, β-β) -organism pairs, about one tenth belong to the most common family.

A Chemically-interpreted Biological Tree
An alternative view of the biological and chemical diversity covered by LOTUS is illustrated in Figure 7.
Here chemical compounds are not organized but biological organisms are placed in their taxonomy. To limit bias due to underreporting in the literature and keep a reasonable display size, only families with at least 50 reported compounds were included. Organisms were classi ed according to the OTL taxonomy and structures according to NPClassi er. The tips were labeled according to the biological family and colored according to their biological kingdom. The bars represent structure speci city of the most characteristic chemical class of the given biological family (the higher the more speci c). This speci city score was calculated as in Equation 2: Figure 7 makes it possible to spot highly speci c compound classes such as trinervitane terpenoids in the Termitidae, the rhizoxin macrolides in the Rhizopodaceae, or the quassinoids and limonoids typical, respectively, of Simaroubaceae and Meliaceae. Similarly, tendencies of more generic occurrence of NP can be observed. For example, within the fungal kingdom, Basidiomycotina appear to have a higher biosynthetic speci city toward terpenoids than other fungi, which mostly focus on polyketides production. When observed at a ner scale, down to the structure level, such chemotaxonomic representation can give valuable insights. For example, among all chemical structures, only two were found in all biological kingdoms, namely heptadecanoic acid (KEMQGTRYUADPNZ-UHFFFAOYSA-N) and β-carotene (OENHQHLEOONYIE-JLTXGRSLSA-N). Looking at the repartition of β-sitosterol (KZJWDPNRJALLNS-VJSFXXLFSA-N) within the overall biological tree, SI-5 plots its presence/absence versus those of its superior chemical classi cations, namely the stigmastane, steroid and terpenoid derivatives, over the same tree used in Figure 7. The comparison of these ve chemically-interpreted biological trees clearly highlights the increasing speciation of the β-sitosterol biosynthetic pathway in the Archaeplastida kingdom, while the superior classes are distributed across all kingdoms. Figure 7 is zoomable and vectorized for detailed inspection.
As illustrated, the possibility of data interrogation at multiple precision levels, from fully de ned chemical structures to broader chemical classes, is of great interest, e.g., for taxonomic and evolution studies. This makes LOTUS a unique ressource for the advancement of chemotaxonomy, a discipline pioneered by Augustin Pyramus de Candolle and pursued by other notable researchers (Robert Hegnauer, Otto R. Gottlieb) (de Candolle, 1816; Gottlieb, 1982;Hegnauer, 1986a). Six decades after Hegnauer's publication of "Die Chemotaxonomie der P anzen" (Hegnauer, 1986b) much remains to be done for the advancement of this eld of study and the LOTUS initiative aims to provide a solid basis for researchers willing to pursue these exciting explorations at the interface of chemistry, biology and evolution.
As shown recently in the context of spectral annotation (Dührkop et al., 2020), lowering the precision level of the annotation allows a broader coverage along with greater con dence. Genetic studies investigating the pathways involved and the organisms carrying the responsible biosynthetic genes would be of interest to con rm the previous observations. These forms of data interpretation exemplify the importance of reporting not only new structures, but also novel occurrences of known structures in organisms as comprehensive chemotaxonomic studies are pivotal for a better understanding of the metabolomes of living organisms.
The integration of multiple knowledge sources, e.g. genetics for NP producing gene clusters (Kautsar et al., 2019) combined to taxonomies and occurrences DB, also opens new opportunities to understand if an organism is responsible for the biosynthesis of a NP or merely contains it. This understanding is of utmost importance for the chemotaxonomic eld and will help to understand to which extent microorganisms (endosymbionts) play a role in host development and its NP expression potential (SAIKKONEN, 2004).

Conclusion & Perspectives Advancing Natural Products Knowledge
At its current development stage, data harmonized and curated throughout the LOTUS initiative remain imperfect and, by the very nature of research, at least partially biased (see Introduction). In the context of bioactive NP research, and due to global editorial practices, it should not be ignored that many publications tend to emphasize new compounds and/or those for which interesting bioactivity has been measured. Near-ubiquitous (primarily plant-based) compounds tend to be overrepresented in the NP literature, yet the implication of their wide distribution in nature and associated patterns of broad, non-speci c activity are often underappreciated (Bisson et al., 2015). Ideally, all characterized compounds independent of structural novelty and/or bioactivity pro le should be documented, and the expansion of veri ed structure-organism pairs is fundamental to the advancement of NP research.
The LOTUS initiative provides a framework for rigorous review and incorporation of new records and already presents a valuable overview of the distribution of NP occurrences studied to date. While current data presents a reasonable approximation of the chemistries of a few well-studied organisms such as Arabidopsis thaliana, they remain patchy for many other organisms represented in the dataset. Community participation is the most e cient means of achieving more comprehensive documentation of NP occurrences, and the comprehensive editing opportunities provided within LOTUS and through the associated Wikidata distribution platform open new opportunities for collaborative engagement. In addition to facilitating the introduction of new data, it also provides a forum for critical review of existing data, as well as harmonization and veri cation of existing NP datasets as they come online.

Fostering FAIRness and TRUSTworthiness
The LOTUS harmonized data and dissemination of referenced structure-organism pairs through Wikidata, enables novel forms of queries and transformational perspectives in NP research. As LOTUS follows the guidelines of FAIRness and TRUSTworthiness, all researchers across disciplines can bene t from this opportunity, whether the interest is in ecology and evolution, chemical ecology, drug discovery, biosynthesis pathway elucidation, chemotaxonomy, or other research elds that connect with NP.
The introduction of LOTUS even provides a new opportunity to advance the FAIR guiding principles for scienti c data management and stewardship originally established in 2016 (Wilkinson et al., 2016). Researchers worldwide uniformly acknowledge the limitations caused by the intrinsic unavailability of essential (raw) data (Bisson et al., 2016). The lack of progress is, at least in part, due to elements in the dissemination channels of the classical print and static PDF publication formats that complicate or sometimes even discourage data sharing, e.g., due to page limitations and economically motivated mechanisms, including those involved in the focus on and calculation of journal impact factors. In particular raw data such as experimental readings, spectroscopic data, instrumental measurements, statistical, and other calculations are valued by all, but disseminated by only very few. The immense value of raw data and the desire to advance the public dissemination has recently been documented in detail for nuclear magnetic resonance (NMR) spectroscopic data by a large consortium of NP researchers (McAlpine et al., 2019). However, to generate the vital ow of contributed data, the e ort associated with preparing and submitting content to open repositories as well as data reuse should be better acknowledged in academia, government, regulatory, and industrial environments (Cousijn et al., 2019(Cousijn et al., , 2018Pierce et al., 2019).

Opening New Perspectives for Spectral Data
The possibilities for expansion and future applications of the Wikidata-based LOTUS initiative are signi cant. For example, properly formatted spectral data, e.g., data obtained by MS or NMR, can be linked to the Wikidata entries for the respective chemical compounds. MassBank (Horai et al., 2010) and SPLASH (Wohlgemuth et al., 2010) identi ers are already reported in Wikidata, and this existing information can be used to report MassBank or SPLASH records for example for Arabidopsis thaliana compounds (https://w.wiki/3PJD). Such possibilities will help to bridge experimental data results obtained during the early stages of NP research with data that has been reported and formatted in di erent contexts. This opens exciting perspectives for structural dereplication, NP annotation, and metabolomic analysis. The authors have previously demonstrated that taxonomically-informed metabolite annotation is critical for the improvement of the NP annotation process (Rutz et al., 2019). Alternative approaches linking structural annotation to biological organisms have also shown substantial improvements (Ho mann et al., 2021). The LOTUS initiative o ers new opportunities for linking chemical objects to both their biological occurrences and spectral information and should signi cantly facilitate such applications.

Integrating Chemodiversity, Biodiversity, and Human Health
As shown in SI-5, observing the chemical and biological diversity at various granularities can o er new insights. Regarding the chemical objects involved, it will be important to document the taxonomies of chemical annotations for the Wikidata entries. However, this is a rather complex task, for which stability and coverage issues will have to be addressed rst. Existing chemical taxonomies such as ChEBI, ClassyFire, or NPClassi er are evolving steadily, and it will be important to constantly update the tools used to make further annotations. Repositioning NP within their greater biosynthetic context is another major challenge -and active eld of research. The fact that the LOTUS disseminates data through Wikidata will help facilitate its integration with biological pathway knowledge bases such as WikiPathways and contribute to this complex task (Martens et al., 2021;Slenter et al., 2018).
In the eld of ecology, for example, molecular traits are gaining increased attention (Kessler and Kalske, 2018;Sedio, 2017). Conceptually, LOTUS can help associate classical plant traits (e.g., leaf surface area, photosynthetic capacities, etc.) with Wikidata biological organisms entries, and, thus, allow their integration and comparison with chemicals that are associated with the organisms. Likewise, the association of biogeography data documented in repositories such as GBIF could be further exploited in Wikidata to pursue the exciting but understudied topic of "chemodiversity hotspots" (Defossez et al., 2021).
Further NP-related information of great interest remains poorly formatted. One example of such a shortcoming relates to traditional medicine, including ethnomedicine and ethnobotany, which is the historical and empiric approach of mankind to discover and use bioactive products from Nature, primarily plants. The amount of knowledge generated in human history on the use of medicinal substances represents fascinating yet underutilized information. Notably, the body of literature on the pharmacology and toxicology of NP is compound-centric, increases steadily, and relatively scattered, but still highly relevant (not necessarily: su cient) for exploring the role and potential utility of NP for human health. To this end, the LOTUS initiative represents a resource for new concepts by which such information could be valued and conserved in the digital era, as LOTUS provides a blueprint for appropriate formatting and sharing of such data (Allard et al., 2018;Geo rey A. Cordell, 2017a, 2017b. This underscores the transformative value of the LOTUS initiative for the advancement of Traditional Medicine and its drug discovery potential in health systems worldwide.

Summary & Outlook
The various facets discussed above connect with ongoing and future developments that the tandem of the LOTUS initiative and its Wikidata integration can accommodate through a broader knowledge base. The information of the LOTUS initiative is already readily accessible by third party projects build on top of Wikidata such as the SLING project (https://github.com/ringgaard/sling, see entry for gliotoxin) or the Plant Humanities Lab project (https://lab.plant-humanities.org/, see entry for Ilex guayusa).
Behind the scenes, all underlying resources represent data in a multidimensional space and can be extracted as individual graphs that can be interconnected. The craft of appropriate federated queries allows users to navigate these graphs and fully exploit their potential (Kratochvíl et al., 2018;Waagmeester et al., 2020). The development of interfaces such as RDFFrames (Mohamed et al., 2020) will also facilitate the use of the wide arsenal of existing machine learning approaches to automate reasoning on these knowledge graphs.
Overall, the LOTUS initiative aims to make more and better data available. This project paves the way for the establishment of an open and expandable electronic NP resource. The design and e orts of the LOTUS initiative re ect our conviction that the integration of NP research results is long-needed and requires a truly open and FAIR knowledge base. We believe that the LOTUS initiative has the potential to fuel a virtuous cycle of research habits and, as a result, contribute to a better understanding of Life and its chemistry.

Data Curation
Gathering Before their inclusion, the overall quality of the source was manually assessed to estimate, both, the quality of referenced structure-organism pairs and the lack of ambiguities in the links between data and references. This led to the identi cation of thirty-six electronic NP resources as valuable LOTUS input. Data from the proprietary Dictionary of Natural Products (DNP v 29.2) was also used for comparison purposes only and is not publicly disseminated. FooDB was also curated but not publicly disseminated since its license proscribed sharing in Wikidata. SI-1 gives all necessary details regarding electronic NP resources access and characteristics.
Manual inspection of each electronic NP resource revealed that the structure, organism, and reference elds were widely variable in format and contents, thus requiring standardization to be comparable. The initial stage consisted of writing tailored scripts that are capable of harmonizing and categorizing knowledge from each source (Figure 1). This transformative process led to three categories: elds relevant to the chemical structure described, to the producing biological organism, and the reference describing the occurrence of the chemical structure in the producing biological organism. This process resulted in categorized columns for each source, providing an initial harmonized format for each table.
For all thirty-eight sources, if a single le or multiple les were accessible via a download option including FTP, data was gathered that way. For some sources, data was scraped (cf. SI-1). All scraping scripts can be found in the lotus-processor repository in the src/1_gathering directory (under each respective subdirectory). Data extraction scripts for the DNP are available and should allow users with a DNP license only to further exploit the data (src/1_gathering/db/dnp). The chemical structure elds, organism elds, and reference elds were manually categorized into three, two, and ten subcategories, respectively. For chemical structures, "InChI", "SMILES", and "chemical name" (not necessarily IUPAC). For organisms, "clean" and "dirty", meaning lot text not referred to the canonical name was present or the organism was not described by its canonical name (e.g. "Compound isolated from the fresh leaves of Citrus spp."). For the references, the original reference was kept in the "original" eld. When the format allowed it, references were divided into: "authors", "doi", "external", "isbn", "journal", "original", "publishing details", "pubmed", "title", "split". The generic "external" eld was used for all external cross-references to other websites or electronic NP resources (e.g. "also in knapsack"). The last subcategory, "split", corresponds to a still non-atomic eld after the removal of parts of the original reference. Other eld titles are self-explanatory. The producing organism eld was kept as a single eld.

Harmonization
To perform the harmonization of all previously gathered sources, sixteen columns were chosen as described above. Upon electronic NP resources harmonization, resulting subcategories were divided and subject to further cleaning. The "chemical structure" elds were divided into les according to their subcategories ("InChI", "names" and "SMILES"). A le containing all initial structures from all three subcategories was also generated. The same procedure was followed for organisms and references.

Cleaning
To obtain an unambiguously referenced structure-organism pair for Wikidata dissemination, the initial sixteen columns were translated and cleaned into three elds: the reported structure, the organism canonical name, and the reference. The structure was reported as InChI, together with its SMILES and InChIKey translation. The biological organism eld was reported as three minimal necessary and su cient elds, namely its canonical name and the taxonID and taxonomic DB corresponding to the latter. The reference was reported as four minimal elds, namely reference title, DOI, PMCID, and PMID, one being su cient. For the forthcoming translation processes, automated solutions were used when available. However, for speci c cases (common or vernacular names of the biological organisms, Traditional Chinese Medicine (TCM) names, and conversion between digital reference identi ers), no solution existed, thus requiring the use of tailored dictionaries. The initial entries (containing one or multiple producing organisms per structure, with one or multiple accepted names per organism) were cleaned into 2M+ referenced structure-organism pairs.

Chemical Structures
To retrieve as much information as possible from the original structure eld(s) of each of the sources, the following procedure was followed. Allowed structural elds for the sources were divided into two types: structural (InChI, SMILES) or nominal (chemical name, not necessarily IUPAC). If multiple elds were present, structural identi ers were preferred over structure names. Among structural identi ers, when both identi ers led to di erent structures, InChI was preferred over SMILES. SMILES were translated to InChI using the RDKit (2021.03.1) implementation in Python 3.8 (src/2_curating/2_editing/structure/1_translating/smiles.py). They were rst converted to ROMol objects which were then converted to InChI. When no structural identi er was available, the nominal identi er was translated to InChI rst thanks to OPSIN (Lowe et al., 2011), a fast Java-based translation open-source solution. If no translation was obtained, chemical names were then submitted to the CTS (Wohlgemuth et al., 2010), once in lower case only, once with the rst letter capitalized. If again no translation was obtained, candidates were then submitted to the Chemical Identi er Resolver via the cts_convert function from the webchem package (Szöcs et al., 2020). Before the translation process, some typical chemical structure-related greek characters (such as α, ß) were replaced by their textual equivalents (alpha, beta) to obtain better results. All pre-translation steps are included in the preparing_name function and are available in src/r/preparing_name.R.
The chemical sanitization step sought to standardize the representation of chemical structures coming from di erent sources. It consisted of three main stages (standardizing, fragment removal, and uncharging) achieved via the MolVS package. The initial standardizer function consists of six stages (RDKit Sanitization, RDKit Hs removal, Metals Disconnection, Normalization, Acids Reionization, and Stereochemistry recalculation) detailed in the molvs documentation. In a second step, the FragmentRemover functionality was applied using a list of SMARTS to detect and remove common counterions and crystallization reagents sometimes occurring in the input DB. Finally, the Uncharger function was employed to neutralize molecules when appropriate.
Molconvert function of the MarvinSuite was used for traditional and IUPAC names translation, Marvin 20.19, ChemAxon. When stereochemistry was not fully de ned, (+) and (-) symbols were removed from names. All details are available in the following script: src/2_curating/2_editing/structure/4_enriching/naming.R. Chemical classi cation of all resulting structures was done using classy reR (Djoumbou Feunang et al., 2016) and NPClassi er API.
After manual evaluation, structures remaining as dimers were discarded (all structures containing a "." in their SMILES were removed).

Biological Organisms
The cleaning process at the biological organism's level had three objectives: convert the original organism string to (a) taxon name(s), atomize elds containing multiple taxon names, and deduplicate synonyms. The original organism strings were treated with Global Names Finder (GNF) and Global Names Veri er (GNV), both tools coming from the Global Names Architecture (GNA) a system of web services that helps people to register, nd, index, check and organize biological scienti c names and interconnect on-line information about species. GNF allows scienti c name recognition within raw text blocks and searches for found scienti c names among public taxonomic DB. GNV takes names or lists of names and veri es them against various biodiversity data sources. Canonical names, their taxonID, and the taxonomic DB they were found in were retrieved. When a single entry led to multiple canonical names (accepted synonyms), all of them were kept. Because both GNF and GNV recognize scienti c names and not common ones, common names were translated before a second resubmission.

Dictionaries
To perform the translations from common biological organism name to latin scienti c name, specialized dictionaries included in DrDuke, FooDB, PhenolExplorer were aggregated together with the translation dictionary of GBIF Backbone Taxonomy. The script used for this was src/1_gathering/translation/common.R. When the canonical translation of a common name contained a speci c epithet that was not initially present, the translation pair was discarded (for example, "Aloe" translated in "Aloe vera" was discarded). Common names corresponding to a generic name were also discarded (for example "Kiwi" corresponding to the synonym of an Apteryx spp. (https://www.gbif.org/species/4849989)). When multiple translations were given for a single common name, the following procedure was followed: the canonical name was split into species name, genus name, and possible subnames. For each common name, genus names and species names were counted. If both the species and genus names were consistent at more than 50%, they were considered consistent overall and, therefore, kept (for example, "Aberrant Bush Warbler" had "Horornis avolivaceus" and "Horornis avolivaceus intricatus" as translation; as both the generic ("Horornis") and the speci c (" avolivaceus") epithets were consistent at 100%, both ("Horornis avolivaceus") were kept). When only the generic epithet had more than 50% consistency, it was kept (for example, "Angelshark" had "Squatina australis" and "Squatina squatina" as translation, so only "Squatina" was kept). Some unspeci c common names were removed (see https://osf.io/gqhcn/) and only common names with more than three characters were kept. This resulted in 181,891 translation pairs further used for the conversion from common names to scienti c names. For TCM names, translation dictionaries from TCMID, TMMC, and coming from the Chinese Medicine Board of Australia were aggregated. The script used for this was src/1_gathering/translation/tcm.R. Some unspeci c common names were removed (see https://osf.io/zs7ky/). Careful attention was given to the Latin genitive translations and custom dictionaries were written (see https://osf.io/c3ja4/, https://osf.io/u75e9/). Organ names of the producing organism were removed to avoid wrong translation (see https://osf.io/94fa2/). This resulted in 7,070 translation pairs. Both common and TCM translation pairs were then ordered by decreasing string length, rst translating the longer names to avoid part of them being translated incorrectly.

Translation
To ensure compatibility between obtained taxonID with Wikidata, the taxonomic DB 3 (ITIS), 4 (NCBI), 5 (Index Fungorum), 6 (GRIN Taxonomy for  All other available taxonomic DB are listed at http://index.globalnames.org/datasource. To retrieve as much information as possible from the original organism eld of each of the sources, the following procedure was followed: First, a scienti c name recognition step, allowing us to retrieve canonical names was carried (src/2_curating/2_editing/organisms/subscripts/1_cleaningOriginal.R). Then, a subtraction step of the obtained canonical names from the original eld was applied, to avoid unwanted translation of parts of canonical names. For example, Bromus mango contains "mango" as a speci c epithet, which is also the common name for Mangifera indica. After this subtraction step, the remaining names were translated from vernacular (common) and TCM names to scienti c names, with help of the dictionaries. For performance reasons, this cleaning step was written in Kotlin and used coroutines to allow e cient parallelization of that process (src/2_curating/2_editing/organisms/2_translating_organism_kotlin/).

References
The Rcrossref package (Chamberlain et al., 2020) interfacing with the Crossref API was used to translate references from their original subcategory ("original", "publishingDetails", "split", "title") to a DOI, the title of its corresponding article, the journal it was published in, its date of publication and the name of the rst author. The rst twenty candidates were kept and ranked according to the score returned by Crossref, which is a tf-idf score. For DOI and PMID, only a single candidate was kept. All parameters are available in src/functions/reference.R. All DOIs were also translated with this method, to eventually discard any DOI not leading to an object. PMIDs were translated, thanks to the entrez_summary function of the rentrez package (Winter, 2017). Scripts used for all subcategories of references are available in the directory src/2_curating/2_editing/reference/1_translating/. Once all translations were made, results coming from each subcategory were integrated, (src/2_curating/2_editing/reference/2_integrating.R) and the producing organism related to the reference was added for further treatment. Because the crossref score was not informative enough, at least one other metric was chosen to complement it. The rst metric was related to the presence of the producing organism's generic name in the title of the returned article. If the title contained the generic name of the organism, a score of 1 was given, else 0. Regarding the subcategories "doi", "pubmed" and "title", for which the same subcategory was retrieved via crossref or rentrez, distances between the input's string and the candidates' one were calculated. Optimal string alignment (restricted Damerau-Levenshtein distance) was used as a method. Among "publishing details", "original" and "split" categories, three additional metrics were used: If the journal name was present in the original eld, a score of 1 was given, else 0. If the name of the rst author was present in the original eld, a score of 1 was given, else 0. Those three scores were then summed together. All candidates were rst ordered according to their crossref score, then by the complement score for related subcategories, then again according to their title-producing organism score, and nally according to their translation distance score. After this re-ranking step, only the rst candidate was kept. Finally, the Pubmed PMCID dictionary (PMC-ids.csv.gz) was used to perform the translations between DOI, PMID, and PMCID (src/2_curating/2_editing/reference/3_cleaning.R).
After the curation of all three objects, all of them were put together again. Therefore, the original aligned table containing the original pairs was joined with each curation result. Only entries containing a structure, an organism, and a reference after curation were kept. Each curated object was divided into minimal data (for Wikidata upload) and metadata. A dictionary containing original and curated object translations was written for each object to avoid those translations being made again during the next curation step (src/2_curating/3_integrating.R).

Validation
The pairs obtained after curation were of di erent quality. Globally, structure and organism translation was satisfactory whereas reference translation was not. Therefore, to assess the validity of the obtained results, a randomized set of 420 referenced structure-organism pairs was sampled in each reference subcategory and validated or rejected manually. Entries were sampled with at least 55 of each reference subcategory present (to get a representative idea of each subcategory) (src/3_analysing/1_sampling.R). An entry was only validated if: i) the structure (as any structural descriptor that could be linked to the nal sanitized InChIKey) was described in the reference ii) the producing organism (as any organism descriptor that could be linked to the accepted canonical name) was described in the reference and iii) the reference was describing the occurrence of the chemical structure in the biological organism. Results obtained on the manually analyzed set were categorized according to the initial reference subcategory and are detailed in SI-2. To improve these results, further cleaning of the references was needed. This was done by accepting entries whose reference was coming from a DOI, a PMID, or from a title which restricted Damerau-Levenshtein distance between original and translated was lower than ten or if it was coming from one of the three main journals where NP occurrences are commonly expected to be published (i.e., Journal of Natural Products, Phytochemistry, or Journal of Agricultural and Food Chemistry). For "split", "publishingDetails" and "original" subcategories, the year of publication of the obtained reference, its journal, and the name of the rst author were searched in the original entry and if at least two of them were present, the entry was kept. Entries were then further ltered to keep the ones where the reference title contained the rst element of the detected canonical name. Except for COCONUT, exceptions to this lter were made for all DOI-based references. To validate those ltering criteria, an additional set of 100 structure-organism pairs were manually analyzed. F0.5 score was used as a metric. F0.5 score is a modi ed F1 score where precision has twice more weight than recall.
The F-score was calculated with ß = 0.5, as in Equation 3: Based on this rst manually validated dataset, ltering criteria (src/r/ lter_dirty.R) were established to maximize precision and recall. Another 100 entries were sampled, this time respecting the whole set ratios. After manual validation, 97% of true positives were reached on the second set. A summary of the validation results is given in SI-2. Once validated, the ltering criteria were established to the whole curated set to lter entries chosen for dissemination (src/3_analysing/2_validating.R).

Unit Testing
To provide robustness of the whole process and code, unit tests and partial data full-tests were written. They can run on the developer machine but also on the CI/CD system (GitLab) upon each commit to the codebase.
Those tests assess that the functions are providing results coherent with what is expected especially for edge cases detected during the development. The Kotlin code has tests based on JUnit and code quality control checks based on Ktlint, Detekt and Ben Mane's version plugin.

Data Dissemination Wikidata
All the data produced for this work has been made available on Wikidata under a Creative Commons 0 license according to Wikidata:Licensing. This license is a "No-right-reserved" license that allows most reuses.

Lotus.NaturalProducts.Net (LNPN)
The web interface is implemented following the same protocol as described in the COCONUT publication (Sorokina et al., 2021) i.e. the data are stored in a MongoDB repository, the backend runs with Kotlin and Java, using the Spring framework, and the frontend is written in React.js, and completely Dockerized. In addition to the diverse search functions available through this web interface, an API is also implemented, allowing programmatic LNPN querying. The complete API usage is described on the "Documentation" page of the website. LNPN is part of the NaturalProducts.net portal, an initiative aimed at gathering diverse open NP resources in one place.

Data Retrieval
Bulk retrieval of a frozen (2021-05-23) version of LOTUS data is also available at https://osf.io/eydjs/. lotus-wikidata-exporter allows the download of all chemical compounds with a "found in taxon" property. That way, it does not only get the data produced by this work, but any that would have existed beforehand or that would have been added directly on Wikidata by our users. It makes a copy of all the entities (compounds, taxa, references) into a local triplestore that can be queried with SPARQL as is or converted to a TSV le for inclusion in other projects. It is currently adapted to export directly into the SSOT thus allowing direct reuse by the processing/curation pipeline.

Data Addition Wikidata
Data is loaded by the Kotlin importer available in the lotus-wikidata-importer repository under a GPL V3 license and imported into Wikidata. The importer processes the curated outputs grouping references, organisms, and compounds together. It then checks if they already exist in Wikidata (using SPARQL or a direct connection to Wikidata depending on the kind of data). It then uses update or insert, also called upsert, the entities as needed. The script currently takes the tabular le of the referenced structure-organism pairs resulting from the LOTUS curation process as input. It is currently being adapted to use directly the SSOT and avoid an unnecessary conversion step. To import references, it rst double checks for the presence of duplicated DOIs and utilizes the Crossref REST API to retrieve metadata associated with the DOI, the support for other citation sources such as Europe PMC is in progress. The structure-related elds are only subject to limited processing: basic formatting of the molecular formula by subscripting of the numbers. Due to limitations in Wikidata, the molecule names are dropped if they are longer than 250 characters and likewise the InChI strings cannot be stored if they are longer than 1500 characters.
Uploaded taxonomical DB identi ers are currently restricted to ITIS, GBIF, NCBI Taxon, Index Fungorum, IRMNG, WORMS, VASCAN, and iNaturalist, and newly OTL. The taxa levels are currently limited to family, subfamily, tribe, subtribe, genus, species, variety. The importer checks for the existence of each item based on their InChIKey and upserts the compound with the found in taxon statement and the associated organisms and references.

LNPN
From the onset, LNPN has been importing data directly from the frozen tabular data of the LOTUS dataset (https://osf.io/6kc2m/). In future versions, LNPN will directly feed on the SSOT.

Data Edition
The bot framework lotus-wikidata-importer was adapted such that, in addition to batch upload capabilities, it can also edit erroneously created entries on Wikidata. As massive edits have a large potential to disrupt otherwise good data, progressive deployment of this script is used, starting by editing progressively 1, 10, then 100 entries that are manually checked. Upon validation of 100 entries, the full script is run and check its behavior checked at regular intervals. An example of a corrected entry is as follows: https://www.wikidata.org/w/index.php? title=Q105349871&type=revision&di =1365519277&oldid=1356145998

Curation interface
A web-based (Kotlin, Spring Boot for the back-end, and TypeScript with Vue for the front-end) curation interface is currently in construction. It will allow mass-editing of entries and navigate quick navigation in the SSOT for the curation of new and existing entries. This new interface is intended to become open to the public to foster the curation of entries by further means, driven by the users. In line with the overall LOTUS approach, any modi cation made in this curation interface will be mirrored after validation on Wikidata and LNPN.

Code Availability General Repository
All programs written for this work can be found in the following group: https://gitlab.com/lotus7.

Processing
The source data curation system is available at https://gitlab.com/lotus7/lotus-processor. This program takes the source data as input and outputs curated data, ready for dissemination. The rst step involves checking if the source data has already been processed. If not, all three elements (biological organism, chemical structures, and references) are submitted to various steps of translation and curation, before validation for dissemination.

Wikidata Import
The Wikidata importer is available at https://gitlab.com/lotus7/lotus-wikidata-importer. This program takes the processed data resulting from the lotusProcessor subprocess as input and uploads it to Wikidata. It performs a SPARQL query to check which objects already exist. If needed, it creates the missing objects. It then updates the content of each object. Finally, it updates the chemical compound page with a "found in taxon" statement complemented with a "stated in" reference.

Export
The Wikidata exporter is available at https://gitlab.com/lotus7/lotus-wikidata-exporter. This program takes the structured data in Wikidata corresponding to chemical compounds found in taxa with a reference associated as input and exports it in both RDF and tabular formats for further use. Two subsequent options are (a) that the end-user can directly use the exported data.; or (b) that the exported data, which can be new or modi ed since the last iteration, is used as new source data in lotusProcessor.

LNPN
The LNPN website and processing system is available at https://github.com/mSorok/LOTUSweb. This system takes the processed data resulting from the lotusProcessor as input and uploads it on https://lotus.naturalproducts.net. The repository is not part of the main GitLab group as it bene ts from already established pipelines developed by CS and MS. The website allows searches from di erent points of view, complemented with taxonomies for both the chemical and biological sides. Many chemical molecular properties and molecular descriptors that otherwise are unavailable in Wikidata are also provided.

Code Freezing
All repository hyperlinks in the manuscript point to the preprint branches by default. The links contain all programs and code before submission (2021-02-23) and will eventually be updated to a publication branch using modi cations resulting from the peer-reviewing process. As the code evolves, readers are invited to refer to the main branch of each repository for the most up-to-date code. A frozen version (2021-02-23) of all programs and code is also available in the LOTUS OSF repository (https://osf.io/pmgux/).

SI 1 Data Sources List
If your target taxon is not yet present on Wikidata and you are sure you have a valid taxon name that is spelled correctly, then you can go to https://www.wikidata.org/wiki/Special:NewItem, as described in the Add your data manually to Wikidata section. For items about taxa, the instance of statement should have a value taxon (i.e. Q16521). As for chemical compounds, the user interface will then suggest to you further statements to add. For taxa, these include taxon name, parent taxon and taxon rank.

Add the reference documenting the structure-organism pair
Finally, since we report referenced structure-organisms pairs, we need to add the reference for this newly created compound found in taxon relationship. This happens on the item about the compound, just below the found in taxon statement. Click on the 0 references link and then on add reference : Here, we use the stated in property (P248): Now, type in the rst letters or word of the scienti c publication documenting the natural product occurence, autocompletion happens again. Note that multiple publications might have the same title, and that there could be minor di erences in punctuation or special characters between the information you and Wikidata have about the same reference. If you are not sure whether your target reference is already in Wikidata, you can use its DOI to check, as outlined in the Check whether your target reference is already on Wikidata section.
Click publish to save your changes and make them public.

Check whether your target reference is already on Wikidata
If you are not sure whether your target reference is already in Wikidata, you can use its DOI to check. For our DOI 10.1021/ol2030907 , the URL https://scholia.toolforge.org/doi/10.1021/ol2030907 will lead you to a Scholia page about this publication: https://scholia.toolforge.org/work/Q83059010. Scholia visualizes information from Wikidata, so if it has an entry for your target reference, then so does Wikidata, and both of them will use the same identi er (in this case Q83059010). If you prefer to resolve your DOI to Wikidata directly, you can do so by using the uppercase-normalized DOI in the following URL pattern: https://hub.toolforge.org/P356:10.1021/OL2030907, which will lead you to the respective Wikidata page, in this case Q83059010.
If you think that no Wikidata entry exists for your target reference, you can use the DOI in the URL pattern https://sourcemd.toolforge.org/index_old.php?id=10.1021/ol2030907&doit=Check+source, which will trigger a check with both Crossref and Wikidata, and if no Wikidata entry can be found, the metadata from Crossref will be fetched and presented to you for creating the respective Wikidata item semi-automatically. Using such semi-automated work ows does require and account that is a minimum number of days old and has made a minimum number of edits on Wikidata.
If you are interested the annotation of article with topics in Scholia here is a tutorial https://laurendupuis.github.io/Scholia_tutorial/

Voilà !
You have added your rst referenced structure-organism relationship to Wikidata and made a valuable contribution to the community. You can add further statements, e.g. molecular formula , or SPLASH code for linking to spectral data.
You can run a SPARQL query and check that everything went smoothly by modifying the InChiKey line in the following SPARQL query: