Overview
Appendix B. Chemical Data Repositories
This appendix introduces the major public chemical and biomedical data repositories that support AI-driven drug discovery, while emphasizing a central warning: more data does not automatically mean better models. Drug-discovery datasets are often incomplete, imbalanced, noisy, or context-limited, and many resources combine primary experimental records with secondary curation. Because databases change frequently, users are urged to verify counts, record version or access dates, and pay close attention to provenance, licensing, and the intended use of each source.
The text then surveys key resources by function. For bioactivity and binding, it highlights curated sources such as ChEMBL, BindingDB, and the Guide to Pharmacology; for chemical identity and properties, PubChem, DrugBank, DrugCentral, ChemSpider, HMDB, and SureChEMBL; for structure, the Protein Data Bank and AlphaFoldDB, along with related structural repositories. It also covers ultra-large virtual libraries like ZINC22 and Enamine REAL, benchmark suites such as MoleculeNet, TDC, Polaris Hub, PoseBusters, PLINDER, and CrossDocked2020, and specialized resources for toxicity, reactions, RNA targets, target discovery, systems biology, and clinical evidence.
A major theme throughout is careful data handling. The appendix explains how to distinguish primary from secondary sources, standardize units and molecular representations, resolve duplicates, and track how integrated databases may propagate errors. It stresses that random train/test splits are often misleading in drug discovery, recommending scaffold, time-based, and similarity-aware splits instead, and it warns about benchmark leakage, assay noise, and activity cliffs. The closing message is that model quality is bounded by data quality and context: good curation, realistic evaluation, and awareness of clinical relevance are essential if machine learning is to produce useful drug-discovery tools.
B.13 Where to Look Next
The resources catalogued here represent a snapshot. New databases appear, older ones change license terms, scales grow by orders of magnitude, and curation practices evolve. Community review articles (e.g., Nucleic Acids Research publishes an annual database issue each January) can help you keep current. Whatever resource you use, cite the specific version or release (the ChEMBL release number, the UniProt release date, the Open Targets quarterly tag) and archive the data snapshot you trained on. Models are only reproducible if the data they were trained on is.
FAQ
What is the main caution when using public chemical databases for AI drug discovery?
Public databases are extremely valuable, but their data can be noisy, incomplete, biased, or inconsistent across sources. You should check provenance, version, update frequency, licensing, and the quality of labels and annotations before using them in model training.Which repositories are recommended as a good starting point for different drug-discovery tasks?
PubChem is recommended for broad chemical lookup and identifier resolution; ChEMBL and BindingDB for bioactivity and binding affinity; PDB and AlphaFold DB for protein structure; DrugBank for approved-drug annotations; Tox21/ToxCast for toxicity assays; and ZINC22 or Enamine REAL for screening libraries.Why is ChEMBL especially useful for machine learning in drug discovery?
ChEMBL is manually curated, literature-derived, and built around FAIR data principles. Its standardized identifiers, structured access, and detailed metadata make preprocessing, feature engineering, validation, and benchmarking more reproducible for ML workflows.How is BindingDB different from ChEMBL?
BindingDB focuses on measured binding affinities such as Ki, Kd, IC50, EC50, and thermodynamic quantities. It also includes many patent-derived measurements and protein–small-molecule crystal structures, which helps extend coverage beyond literature-curated bioactivity data.What is the difference between experimental protein structures in PDB and predicted structures in AlphaFold DB?
PDB contains experimentally resolved structures from methods like X-ray crystallography, cryo-EM, and NMR, while AlphaFold DB contains computational predictions. Experimental structures are preferred when available, but AlphaFold DB is useful when no experimental structure exists and should be checked using confidence scores like pLDDT.Why are ZINC22 and Enamine REAL important for virtual screening?
They provide ultra-large purchasable or make-on-demand chemical libraries for large-scale virtual screening. ZINC22 is widely used for docking campaigns, and Enamine REAL Space offers an even larger enumerated set of synthesizable molecules based on validated reactions and building blocks.What problems do MoleculeNet and TDC have as benchmarks?
Both are widely used, but some datasets contain leakage, duplicated records, near-duplicates, and overly optimistic splits. The book recommends using more robust benchmark platforms like Polaris Hub and checking split strategy carefully before trusting reported performance.Why are Tox21, ToxCast, and CompTox important but limited for toxicity modeling?
They are the dominant sources of high-throughput in vitro toxicity data and are very useful for ML, but they are assay-dependent, noisy, and not direct clinical ground truth. Their labels should not be interpreted as the full real-world toxicity outcome for a drug.What should researchers verify before integrating data from multiple repositories?
Researchers should standardize units, normalize chemical identifiers, detect and resolve duplicates, decide how to handle conflicting measurements, document provenance, and apply quality filters. This helps prevent error propagation across linked databases and improves reproducibility.Why does the appendix emphasize split strategy and benchmark leakage?
Because random splits often inflate performance in drug-discovery tasks, and even scaffold splits can still leak near-identical compounds or related structures. Time-based and similarity-aware splits are more realistic, especially for docking and structure-based ML, and benchmarks like PLINDER and PoseBusters are designed to reduce leakage.
Build AI Drug Discovery Pipelines ebook for free
