Data extraction and curation

Compound records were mainly extracted from the remarkable annual MNPs reviews published by the late D. John Faulkner in Nat. Prod. Rep. (11), a series continued by the team of John W. Blunt (12), and now Anthony R. Carroll (13). The cited references list was retrieved through Web of Science (https://apps.webofknowledge.com, 2018) and then imported into EndNote (version 9.0.0, Clarivate Analytics Inc. 2018) to obtain full text. After manual curation, more than 20 000 articles covering the period from the1960s to December 2018 were collected and integrated into a main document library. Most of these articles focused on the reports of new compounds and the relevant biological activities, together with previously reported compounds where there had been a structural revision or a newly established stereochemistry. The chemical structure, compound name, source organism and other information of small molecules were manually extracted and curated from these publications. The structure of MNPs was then used as a query to search in generic chemical databases (e.g. Reaxys, PubChem, ChEMBL), and the scientific literature and patents of hit compounds were integrated into a general document library.

In order to improve the efficiency of structure extraction, the optical chemical structure recognition tool CLiDE (version 5.12.1, Keymodule Inc. 2017) was used to convert the graphical representations of chemical structures to machine-readable format and transfer them into the chemical editor ChemDraw (version 19.1, PerkinElmer Inc. 2020) for manual inspection and correction. The conformation of the chemical structure was kept consistent with the molecular image in the literature as far as possible to improve recognizability. The chair conformation, Haworth projection and Fischer projection were converted to wedge-dash diagrams to ensure that the stereochemistry could be recognized correctly by the computer software and the 3D shapes could be accurately depicted on the 2D screen. When substituent group abbreviations were expanded, the angle and length of some bonds were adjusted to prevent stacking. The 2D structures were saved in MOL format and then converted into multiple formats (e.g. SMILES, InChI, InChIKey) via Pipeline Pilot (version 18.1, BIOVIA Inc. 2018). Each structure was classified into its corresponding chemical classes using the ClassyFire web server (14). NMR, IR, Raman, UV-Vis, and mass spectra of the structures were presented if available from Wiley SpectraBase (https://spectrabase.com, 2020).

The 3D conformers were generated using OMEGA (version 3.0.0.1, OpenEye Scientific Software Inc. 2018). When undefined stereocenters existed, each stereoisomer was enumerated and conformers independently generated. A maximum of 100 000 conformers per stereoisomer were allowed, and the lowest-energy conformer of each stereoisomer was retained at last. Since a certain number of complicated MNPs contain many undefined stereocenters and/or rotatable bonds, it makes no sense to compute 3D descriptions for all records. Therefore, according to the criteria of PubChem3D conformer models, CMNPD provides a 3D representation for each compound that satisfies the following conditions: (i) not too large (with no more than 50 heavy atoms), (ii) not too flexible (with no more than 15 rotatable bonds), (iii) has only a single covalent unit (salt, mixture or polymer keep only the largest fragment in the calculation), (iv) consists of only supported elements (H, C, N, O, F, Si, P, S, Cl, Br and I), (v) contains only atom types recognized by the MMFF94s force field, (vi) has fewer than six undefined atom or bond stereocenters (15). As a result, 79.9% of all records have 3D information.

Calculated properties

To estimate the drug-likeness of each compound, some physicochemical and pharmacokinetic properties were calculated using widely accepted algorithms. Physicochemical properties were calculated using RDKit (https://www.rdkit.org, 2020), including molecular weight, molecular mass (i.e. monoisotopic mass), octanol/water partition coefficient ALogP (16), polar surface area (17) and quantitative estimate of drug-likeness QED weighted (18), as well as the number of rotatable bonds, hydrogen bond acceptors, hydrogen bond donors, aromatic rings and heavy atoms. Predictions of pharmacokinetic properties were calculated using Pipeline Pilot ADMET models, including blood brain barrier penetration, human intestinal absorption (19,20), aqueous solubility (21), CYP2D6 binding (22), hepatotoxicity (23) and plasma protein binding (24).

Organism resources

Isolation from the source marine organism is the only way to obtain a promising MNP if it cannot be synthesized. Therefore, it is necessary to record what organisms this compound was isolated from and where these organisms were collected, which is also critical to guarantee the reproducibility of relevant research and maximize the bioprospecting efficiency of MNPs (25). Taxonomic information and sampling location of source organisms were extracted from the articles in the main document library, some of which referred to the authors’ previous papers. All organisms were classified into seven hierarchies (i.e. kingdom, phylum, class, order, family, genus and species) based on the authoritative taxonomic databases, such as the Catalogue of Life (CoL, version 2019 annual checklist, http://www.catalogueoflife.org, 2020), the World Register of Marine Species (WoRMS, http://www.marinespecies.org, 2020), the Integrated Taxonomic Information System (ITIS, https://www.itis.gov, 2020) and the Index Fungorum (http://www.indexfungorum.org, 2020). The taxonomic names of identified species were normalized to accepted scientific names, and the unaccepted names (e.g. original combination, new combination, replacement name, incorrect spelling) were recorded as synonymised names. Some of the (newly) identified species that had not been published with an adequate taxonomic description were not included in the above taxonomic databases. Hence, their names were marked as nomina nuda (naked names) temporarily. Sampling locations were converted to coordinates using Google Maps (https://www.google.com/maps, 2020). Some articles and their cross references did not describe where the organisms were collected. As a stopgap measure, the address of the author’s affiliation was regarded as the resource location.

Biological activity data

A certain amount of biological activity data based on the initial bioassays of the discoverers was extracted from the main document library, but most of it was brief description of the pharmacological effects, such as ‘cytotoxic’, ‘antibacterial’, ‘antifungal’ and ‘anti-inflammatory’. As one can imagine, substantial bioactivity data is deposited in the generic chemical databases, especially those that store medicinal chemistry data. To capture the greatest quantity of high-quality bioactivity data efficiently, assay and bioactivity information from the ChEMBL database (release 27, 2020), which exchanges data with dozens of datasets such as PubChem BioAssay (26) and BindingDB (27), were incorporated into the CMNPD standardized experimental dataset. This detailed dataset includes target name, target type (e.g. nucleic acid, protein, cell line, tissue, organism), target organism, activity type (e.g. IC50, Ki, ED50, EC50, mortality), activity value, assay type (e.g. binding, functional, ADME, toxicity, physicochemical) and assay description. To provide more authoritative information about the targets, proteins were mapped into the Universal Protein Resource (UniProt) (28), the Protein Data Bank in Europe (PDBe) (29), the Gene Ontology Annotation (GOA) resource (30), the Therapeutic Target Database (TTD) (31) and the Open Targets Platform (32), while cell lines were mapped into the the Cell Line Ontology (CLO) (33), the Cell Ontology (CL) (34), the Experimental Factor Ontology (EFO) (35), the Cellosaurus (36) and the Library of Integrated Network-based Cellular Signatures (LINCS) NIH program (37).

Current content

The content of CMNPD is demonstrated in Figure 1 and the statistical data is summarized in Table 1. In release 1.0, CMNPD contains 31 561 distinct chemical entities of MNPs from over 13 000 sampling organisms. These organisms are distributed in 7 kingdoms, 38 phyla, 93 classes, 289 orders, 682 families, 1480 genera and 3354 species. There are 15 774 active compounds mapped to 2652 targets with 72 343 bioactivities. These targets include 1122 single proteins, 923 cell lines, 459 organisms and several other types. The document library includes 128 488 scientific literature and patents, of which ∼11 000 articles describe the discovery of new compounds and structure revisions.

Figure 1.

Schematic overview of the CMNPD content. (A) Phylogenetic tree of marine organisms. Core nodes include domains, kingdoms, phyla, classes, orders and families. Internal nodes include their corresponding superior and subordinate taxonomic units, which are based on the NCBI Taxonomy database (38). The bar chart shows the number of compounds in each family. This figure was produced using the Interactive Tree Of Life (iTOL) web server (39). (B) Knowledge graph of the chemical entity.

Table 1.

Statistics of various data collections in CMNPD (release 1.0)

Data collection	Count	Description
Compounds		Marine natural products
Chemical entities	31 561a	Unique chemical structures
Compounds with 3D conformers	25 224	Filter according to the criteria of PubChem3D conformer models
Biologically active compounds	15 774	Compounds with biological activity data
Organisms		Source marine organisms
Kingdoms	7	Taxonomic hierarchy
Phyla	38	Taxonomic hierarchy
Classes	93	Taxonomic hierarchy
Orders	289	Taxonomic hierarchy
Families	682	Taxonomic hierarchy
Genera	1480	Taxonomic hierarchy
Species	3354b	Taxonomic hierarchy
Targets	2652	Targets standardized according to ChEMBL target list
Single proteins	1122	Target type
Cell lines	923	Target type
Organisms	459	Target type
Others	148	Target type
Bioactivities	72 343	Biological activity data
Data in brief	15 980	Manual collection from literature
Standardized experimental data	56 363	Incorporation from ChEMBL
Documents	128 488	Scientific literature and patents
Literature	119 543	Literature abstracts/citations
Patents	8945	Patent abstracts/citations

^aChemical entities include different forms of certain compounds (e.g. original structure, revised structure, stereochemically improved structure, controversial structure).

^bThis is the minimum estimate of the species count, because many unidentified species have been classified into the same category (e.g. Sinularia sp., unidentified species of Family Spongiidae).

Web interface

CMNPD could be accessed at https://www.cmnpd.org with a user-friendly interface. This interface is modelled on the new web interface of ChEMBL (9) to enhance user experience and reduce learning costs. It allows users to browse, search, and explore MNPs-related information in a variety of ways.

Data query

Quick search is available in the middle of the homepage and on the toolbar in the upper right corner of each page. The free-text search allows users to enter any term of CMNPD identifier, compound name, organism name or target name without specifying a search entity. The search bar will provide suitable suggestions as the term is typed (Figure 2A).

Figure 2.

Data query in the web interface. (A) Free-text quick search. Users can retrieve a list of entities matching a selected keyword (e.g. all organisms matching ‘spongia’) or go directly to a report card page for a selected entry (e.g. organism Demospongiae: CMNPD330032). (B) Advanced search. Query conditions are combined with Boolean operators. The inner Boolean operation of the grouped conditions is executed first, and then the outer Boolean operations are performed in order from top to bottom. (C) Search result of the keyword ‘spongia’. Users can further select the entries through the interactive filters on the left.

In addition, a powerful advanced search capability is provided on the query builder page. This allows users to specify any number of query conditions. Available query conditions, which could be combined with the Boolean operator ‘AND’, ‘OR’ or ‘NOT’, include structure (drawing structure, structural classification), compound representations (e.g. compound name, molecular formula), physicochemical properties (e.g. molecular weight, ALogP), ADMET prediction (e.g. blood brain barrier penetration level, human intestinal absorption level), resources (organism name, collection site), bioactivities (e.g. target name, assay type) and bibliography (e.g. authors, DOI). Multiple query conditions could be easily grouped together by just dragging and dropping them. The inner Boolean operations of the grouped conditions will be executed first (Figure 2B). The structure search module, entered with a drawing or a SMILES or InChI string using Marvin JS sketcher (version 19.9.0, ChemAxon Ltd. 2019), allows users to select among exact match, substructure and similarity. Chemical similarity is computed with the Morgan fingerprints and the Tanimoto metric. These fingerprints are generated using RDKit. Some query conditions (e.g. organism name, target name) have both text entry and browse functions, so users have the choice of typing in a search term or selecting from the index.

Data visualizations

To help users better understand and explore the data inside CMNPD, several interactive data visualizations have been created with the JavaScript graphing libraries Plotly (https://plotly.com/javascript, 2020) and ECharts (version 4.0, https://echarts.apache.org, 2020). The sunburst charts demonstrate different hierarchies of organism taxonomy and structural classification. Any section can be expanded by clicking and the corresponding entries can be retrieved using the button below the chart (Figure 3A). The dot distribution map shows the organism collection or storage sites of each compound. Clicking the dot on the map can retrieve the compounds discovered in the corresponding area. The distribution map is also available on each organism report card page and compound report card page (Figure 3B).

Figure 3.

Data visualizations. (A) Sunburst chart showing different hierarchies of organism taxonomy. The classification is arranged in concentric circles, and from inside to outside are kingdoms, phyla, classes, orders, families, genera and species. Users can click a section to expand the hierarchy and the corresponding entries can be retrieved using the button below the chart. (B) Dot distribution map showing the collection sites of organisms (e.g. sponges). Users can click the dot on the map to retrieve the compounds discovered in the corresponding area. (C) Network showing the relationship between organisms, compounds and targets. Users can click any node to enter the report card page of the corresponding entry.

To intuitively illustrate the relationship between organisms, compounds and targets, a systematic analysis function is provided using vis-network (version 7.7.0, https://visjs.org, 2020), which allows users to build network visualization by selecting a master node from these entries. Clicking on the node can retrieve the detailed information of the corresponding entry (Figure 3C).

Downloads

Some users would like to download the database for data mining or virtual screening besides querying it via the web interface. Bulk downloads of several complete datasets are available at https://docs.cmnpd.org/downloads. In addition, users can customize the compound download list through the advanced search and manual selection.