PubChem and ChEMBL

PubChem and ChEMBL are the two primary open-access relational repositories of small-molecule chemical structures, calculated and experimental physicochemical properties, and target-linked quantitative bioactivity measurements that support modern chemoinformatics, virtual screening, quantitative structure-activity relationship (QSAR) modeling, and translational drug discovery. Prior to the mid-2000s, large-scale structure–activity relationship (SAR) tables and high-throughput screening (HTS) outcomes were restricted to proprietary pharmaceutical corporate vaults or commercial subscription databases (such as CAS Registry, Beilstein, and MDDR). The launch of PubChem in 2004 by the U.S. National Center for Biotechnology Information (NCBI) and ChEMBL in 2009 by the European Molecular Biology Laboratory's European Bioinformatics Institute (EMBL-EBI) opened chemical biology data, providing >118 million standardized chemical structures and >320 million bioactivity data points under public licenses.

Although the two databases cross-reference one another and share standardized SMILES, International Chemical Identifier (InChI), and InChIKey representations, they fulfill complementary scientific roles. PubChem operates as an archive-scale aggregator of depositor-contributed chemical substances and automated robotized high-throughput screening bioassays, whereas ChEMBL focuses on manual and semi-automated expert curation of medicinal chemistry literature, dose–response binding affinities, ADMET pharmacokinetic parameters, and clinical development milestones.

PubChem Substance (SID), Compound (CID), and BioAssay (AID)

Established in 2004 at the National Library of Medicine (NLM) within the National Institutes of Health (NIH) as the informatics core of the NIH Molecular Libraries Roadmap Initiative (guided by Stephen H. Bryant, Evan E. Bolton, Yanli Wang, and Sunghwan Kim), PubChem is organized into three interlinked relational databases:

1. PubChem Substance (SID): Archives raw, unmodified chemical records as deposited by >950 external data contributors (including chemical vendors, patent offices, academic screening centers, ChEMBL, and DrugBank). Exceeding 318 million SID records, PubChem Substance preserves provenance and allows multiple depositors to submit separate records for the same molecule.
2. PubChem Compound (CID): Contains >118 million non-redundant, standardized chemical structures extracted and deduplicated from PubChem Substance via an automated chemical standardization pipeline. During standardization, PubChem verifies atomic valences, normalizes functional group representations (such as nitro groups and hypervalent sulfur/phosphorus centers), standardizes tautomeric states, computes canonical SMILES and InChI/InChIKey identifiers, and links mixture components or salt forms to their parent covalently bonded connectivity record.
3. PubChem BioAssay (AID): Stores experimental screening protocols, readouts, and biological target annotations for >1.6 million bioassays encompassing >300 million bioactivity outcomes across >11,000 protein and gene targets. PubChem BioAssay integrates primary single-concentration high-throughput screens (reporting percent inhibition at 10 μM) alongside confirmatory dose–response titration curves.

PubChem 881-Bit Substructure Fingerprint and PUG-REST API

To support interactive similarity and substructure searching across >10⁸ compounds, every PubChem CID is indexed with a custom 881-bit structural molecular fingerprint divided into seven hierarchical sections: element counts (bits 0–114), ring systems up to 10 atoms (bits 115–262), simple atom pairs (bits 263–326), atom neighborhoods (bits 327–415), detailed atom types (bits 416–459), simple SMARTS patterns (bits 460–712), and multi-group SMARTS patterns (bits 713–880). Users and automated workflows query PubChem programmatically via the Power User Gateway (PUG-REST) and PUG-View HTTP APIs or download bulk RDF/SDF archives.

Origin, Literature Extraction, and Assay Ontology

ChEMBL was established at the EMBL-EBI on the Wellcome Genome Campus in Hinxton, UK, in 2008–2009 under team leader John P. Overington (and subsequently curated under Bissan Al-Lazikani, Anna Gaulton, and Andrew R. Leach), following a Wellcome Trust award that enabled the acquisition and open-data conversion of the commercial StARLITe database originally built by Inpharmatica Ltd. Unlike aggregator repositories, ChEMBL centers on manual scientific curation of primary medicinal chemistry journals—including the Journal of Medicinal Chemistry, Bioorganic & Medicinal Chemistry Letters, European Journal of Medicinal Chemistry, and ACS Medicinal Chemistry Letters—supplemented by deposited datasets from neglected tropical disease screens (DNDi, MMV), chemical probes, and regulatory drug approvals.

In ChEMBL release 34, the database encompasses >2.4 million individual compounds, >20.7 million bioactivity measurements extracted from >89,000 scientific publications, and >15,500 annotated biological targets mapped to UniProt accessions. Every ChEMBL assay is classified by Assay Type: B (Binding, measuring direct biomolecular interaction such as radioligand displacement or surface plasmon resonance), F (Functional, measuring cellular or whole-tissue phenotypic response), A (ADMET, measuring microsomal clearance, Caco-2 permeability, or CYP inhibition), and T (Toxicity). Furthermore, ChEMBL assigns a Target Confidence Score ranging from 0 (uncurated target) to 9 (direct binding to a single, unambiguously assigned homologous protein target, as opposed to a multi-subunit protein assembly [score 6–8] or whole-cell line [score 1]).

The pChEMBL Standardized Potency Metric

Medicinal chemistry papers report biological potencies across heterogeneous units (M, mM, μM, nM, pM, or μg/mL) and diverse endpoint types: half-maximal inhibitory concentration (IC₅₀), half-maximal effective concentration (EC₅₀), equilibrium inhibition parameter (K_i), dissociation parameter (K_d), or 50% growth inhibition (GI₅₀). To allow direct statistical comparison and training of graph neural networks, ChEMBL computes a unified logarithmic metric termed the pChEMBL value:

pChEMBL = −log₁₀(molar IC₅₀, EC₅₀, XC₅₀, AC₅₀, K_i, K_d, or Potency)

Under this logarithmic scale, an IC₅₀ of 1 μM (10⁻⁶ M) equals pChEMBL = 6.00, while a 10 nM (10⁻⁸ M) inhibitor equals pChEMBL = 8.00. Only equality numerical measurements (standard_relation = '=') with verified dose–response data receive a pChEMBL score, filtering out censored inequalities (> 100 μM).

BindingDB, ZINC, DrugBank, and SureChEMBL

PubChem and ChEMBL anchor a broader ecosystem of specialized chemical bioinformatics resources:

• BindingDB (created by Michael Gilson's group) focuses on quantitative protein–ligand binding thermodynamics (Δ*G*°, Δ*H*°, −*T*Δ*S*° from isothermal titration calorimetry, alongside K_i and K_d) and curates SARS-CoV-2 and kinase patent bioactivities.
• ZINC (ZINC22) (developed by John J. Irwin and Brian K. Shoichet at UCSF) provides 3D protonated, tautomer-enumerated, partial-charge-assigned structures of >37 billion commercially available and make-on-demand compounds (such as Enamine REAL space) formatted directly for high-throughput structure-based molecular docking.
• SureChEMBL mines chemical structures daily from full-text world patent corpora (USPTO, EPO, WIPO, JPO) using optical structure recognition and chemical named-entity recognition.
• PDBbind links 3D crystal structures in the Protein Data Bank with experimentally measured binding affinities (K_d/K_i) to train docking scoring functions.

Chemical Standardization and Assay Noise in Machine Learning

Combining bioactivity records across PubChem and ChEMBL for machine learning requires rigorous chemical and biological data hygiene, formalized by Denis Fourches, Eugene Muratov, and Alexander Tropsha (2010). Raw records must undergo salt stripping (removing counterions such as Na⁺, Cl⁻, or trifluoroacetate to isolate the parent active pharmacophore), neutralization of ionizable groups, canonical tautomer generation, and stereochemical verification. On the biological side, Christian Kramer et al. (2012) analyzed replicate K_i and IC₅₀ measurements across independent laboratories in ChEMBL, demonstrating that inter-laboratory experimental noise exhibits a standard deviation of σ ≈ 0.45 to 0.55 log units (roughly a 3-fold uncertainty in molar potency), setting a theoretical upper bound (R²_max ≈ 0.80–0.85) on the predictive accuracy of any QSAR model trained on heterogeneous literature compilations.

What is the primary difference between a PubChem Substance (SID) and a PubChem Compound (CID)?

A PubChem Substance (SID) is an archival record of a specific chemical sample submitted by an individual depositor (such as a university lab, chemical vendor, or ChEMBL); if 50 different vendors deposit aspirin, PubChem creates 50 separate SID records. A PubChem Compound (CID) is the standardized chemical structure generated by PubChem's automated chemical normalization pipeline. All 50 aspirin SID records are mapped and linked to a single canonical aspirin record (CID 2244), eliminating structural duplication for similarity searching and property calculation.

Why should machine-learning pipelines filter ChEMBL records by Target Confidence Score and Assay Type?

A raw ChEMBL target query returns measurements from heterogeneous experimental designs: cell-based cytotoxicity assays, multi-protein homogenates, ortholog proteins from rats or mice, and purified human recombinant enzymes. Mixing functional whole-cell assays (Assay Type F) with direct biochemical binding assays (Assay Type B) introduces label noise. Filtering for assay_type = 'B', target_confidence_score = 9 (confirming direct assignment to a single human protein target), and valid pchembl_value fields ensures that all training points measure direct ligand–protein interaction on a uniform −log₁₀(M) scale.

How much experimental noise exists in public ChEMBL IC₅₀ and K_i data, and how does it affect QSAR models?

When the same small molecule is tested against the same protein target by different laboratories in ChEMBL, variations in substrate concentration (via the Cheng–Prusoff relationship relating IC₅₀ to K_i), buffer ionic strength, pH, protein construct length, and incubation time produce a root-mean-square error of ~0.40 log units for K_i and ~0.55 to 0.68 log units for IC₅₀. Consequently, a QSAR or graph neural network model achieving a test RMSE below 0.45 pChEMBL units is likely overfitting to experimental noise rather than learning physical binding energetics.

1. Kim, S., Chen, J., Cheng, T., Gindulyte, A., He, J., He, S., Li, Q., Shoemaker, B. A., Thiessen, P. A., Yu, B., Zaslavsky, L., Zhang, J., & Bolton, E. E. (2023). "PubChem 2023 Update." Nucleic Acids Research, 51(D1), D1373–D1380. https://doi.org/10.1093/nar/gkac956
2. Gaulton, A., Bellis, L. J., Bento, A. P., Chambers, J., Davies, M., Hersey, A., Light, Y., McGlinchey, S., Michalovich, D., Al-Lazikani, B., & Overington, J. P. (2012). "ChEMBL: A Large-Scale Bioactivity Database for Drug Discovery." Nucleic Acids Research, 40(D1), D1100–D1107. https://doi.org/10.1093/nar/gkr777
3. Zdrazil, B., Felix, E., Hunter, F., Manners, E. J., Blackshaw, J., Corbett, S., de Veij, M., Ioannidis, H., Lopez, D. M., Mosquera, J. F., et al. (2024). "The ChEMBL Database in 2023: A Drug Discovery Platform Spanning Multiple Bioactivity Data Types and Time Periods." Nucleic Acids Research, 52(D1), D1180–D1192. https://doi.org/10.1093/nar/gkad1004
4. Kramer, C., Kalliokoski, T., Gedeck, P., & Vulpetti, A. (2012). "The Experimental Uncertainty of Heterogeneous Public K_i Data." Journal of Medicinal Chemistry, 55(11), 5165–5173. https://doi.org/10.1021/jm300131x