Chemical Databases Every ML Scientist Should Know for Drug Discovery

Large-scale bioactivity databases have made data-driven drug discovery practical. A broad ecosystem of cheminformatics tools and web services supports retrieval, curation, and analysis of experimental data, enabling the development of machine learning models for lead identification and optimization.

1. ChEMBL database

ChEMBL is a manually curated database of bioactive molecules with drug-like properties. It brings together chemical and bioactivity data to aid the translation of information into effective new drugs. Bioactivity data is reported in Ki, Kd, IC50, % of inhibition and EC50. Data can be filtered and analyzed to develop compound screening libraries for lead identification during the drug discovery process. The availability of curated bioactivity data on small drug-like molecules opens new opportunities for data-driven drug discovery and makes it possible to apply machine learning methodologies to pharmaceutical research field.

Figure 1: Web interface of ChEMBL database (https://www.ebi.ac.uk/chembl/)

2. BindingDB

BindingDB is a public database of experimental binding affinities of interactions between macromolecules and small molecules. It contains more than 1.2 million binding data points for more than 6,400 protein targets and around 550,000 small molecules. Furthermore, affinities for protein-protein, protein-peptide as well as host-guest interactions are provided. BindingDB provides a wide range of searches, queries, tools, and datasets. A virtual compound screening tool allows users to screen an external dataset for similar bioactivity. Data can be accessed via download (SD-file, TSV, Oracle dump) or programmatically (RESTful API, structured URLs, or KNIME).

2.1. Find My Compound’s Target tool

Find My Compound’s Target is a tool integrated in BindingDB that aims to predict the target of a small molecule of interest or possible off-targets. The query compound is first compared to other compounds in the database and targets of these compounds are selected if the compounds’ similarity is above the chosen cut-off and the affinity is respectively above a certain threshold.

2.2. Find Compounds for My Target

Find Compounds for My Target is another tool that tries to find compounds for a specific target. It features advanced search and extraction of curated information from various data sources such as ChEMBL database, PubChem bioassays, D3R, etc.

Figure 2: Web interface of BindingDB (https://www.bindingdb.org/)

3. Pubchem Bioassays Database

PubChem is an American database of chemical molecules managed by the National Center for Biotechnology Information (NCBI), branch of the National Library of Medicine of the United States under the authority of the National Institutes of Health (NIH). PubChem lists several million compounds by putting a large amount of data of various kinds online for each substance free of charge: chemical, biochemical, pharmacological, production, toxicological, environmental, etc. PubChem BioAssay stores the activity data of small molecules or RNAi and contains curated parts of ChEMBL for which flags for active or inactive compounds are assigned depending on whether the IC50, EC50 or Ki is above 50 μM or not. The data can be accessed and analyzed via a broad range of provided web services and tools (Figure1). Besides using a name, a smiles code can be used or a structure can be drawn to search for an identical molecule, a similar molecule or a substructure. This leads to information about bioassay results or substance descriptions. Variations of assay results can be analyzed using detailed description of the performed experiments. The data is additionally clustered, e.g., according to the protein or gene target, the type of assay (e.g., cell-based, protein-protein interaction), an assay project or more complex kinds of relationships like target similarity or common active compounds.

3.1. PubChemPy

PubChemPy offers a way to use and interact with PubChem database directly with Python. It allows chemical searches by name, substructure and similarity, chemical standardization, conversion between chemical file formats, depiction and retrieval of chemical properties. For more information on installing and using PubChemPy package visit the official website (https://pubchempy.readthedocs.io/en/latest/index.html).

Figure 3: PubChem web interface (https://pubchem.ncbi.nlm.nih.gov/)

4. PDBbind Database

The PDBbind database is a comprehensive collection of experimentally measured binding affinity data (Kd, Ki, and IC50) for the protein-ligand complexes deposited in the Protein Data Bank (PDB). It thus provides a link between energetic and structural information of protein-ligand complexes, which is of great value to various studies on molecular recognition occurred in biological systems. The basic information of each complex in PDBbind is totally open for browsing. Users are however required to register for access under a license agreement in order to utilize the full functions provided on this web site or to download the contents of PDBbind. The registration is free of charge to all academic and industrial users.

Figure 4: PDBbind web interface (http://www.pdbbind.org.cn/index.php)

5. BRENDA Enzyme Database

BRENDA is an enzyme database. It is maintained and developed by the Institute of Biochemistry of the University of Cologne. Data on enzyme functions are taken directly from the primary literature. The database covers 40 entries, with information about enzyme nomenclature, reactions, specificity, structure, method of isolation or preparation, references in scientific literature and cross-references for the sequence or 3D structure. The database is accessible free of charge for academic and non-profit uses, commercial uses need to acquire a license. To use the database, it is necessary to register by email. The database can be searched by EC nomenclature, enzyme name, organism or an advanced search combining these entries.

Figure 5: Brenda Enzymes Database web interface (https://www.brenda-enzymes.org/)

Conclusion

ChEMBL, BindingDB, PubChem, PDBbind, and BRENDA cover the most commonly needed data types for ML-based drug discovery, from bioactivity and binding affinity to enzyme kinetics. All are freely accessible, with Python APIs or bulk download options to support automated pipelines.