ChEMBL is a manually curated database of bioactive molecules and their measured activities, maintained by EMBL-EBI. It is the backbone of public bioactivity data, it is free, and for antibacterial work it will answer a lot of questions and refuse to answer the one you actually asked.
What it holds
- Compounds with structures, identifiers and calculated properties.
- Assays extracted from the primary literature and from deposited datasets, each tied to a document.
- Activities: the measured value, its type, its units and its relation, for example an IC50 in nM or a MIC in µg/mL with a greater than relation.
- Targets, which for antibacterial work are frequently a whole organism rather than a protein.
For antibacterials the important part is that whole organism assays are first class citizens. A MIC against Staphylococcus aureus is stored as an activity against an organism target, not against a purified enzyme, which is exactly the phenotype level data a screening question needs.
The querying pitfalls
Anyone who has assembled a MIC picture from ChEMBL has hit most of these:
| Pitfall | What goes wrong |
|---|---|
| Units | MIC appears in µg/mL, mg/L, µM and nM. Converting needs the molecular weight and care with salts. |
| Relations | A value of 64 with a greater than relation is not the same as a value of 64, and averaging them silently is wrong. |
| Strain granularity | Target records may be a species, a specific strain, or a resistant clinical isolate. These are not interchangeable. |
| Method variation | Broth microdilution, agar dilution and gradient strips give related but not identical numbers. |
| Assay conditions | Media, inoculum and serum content change the result, and are often only in the paper, not the record. |
| Duplication | The same measurement can appear through several documents. |
| Sparse coverage | Most published compounds have been tested against a handful of organisms, not a panel. |
None of this is a criticism of ChEMBL. It is a faithful record of a heterogeneous literature, and a faithful record of a messy field is messy. The conventions behind those numbers are explained under minimum inhibitory concentration.
What it will not tell you
- What your compound will do. If it is not in the database, the database is silent. Similarity searching gives you neighbours, not an answer.
- Which compound to test next. There is no ranking, because ranking is an opinion and a database does not have opinions.
- Why something failed. A high MIC is recorded. Whether it was efflux, porin loss or target mutation is in the discussion section of a paper, if anywhere.
- A strain level roll-up. Assembling one means reconciling assays, units and strain designations by hand.
The other public sources worth knowing
- CO-ADD, the Community for Open Antimicrobial Drug Discovery, which screens submitted compounds against a standard panel and publishes the data.
- SPARK and other open antibacterial data initiatives aimed specifically at the Gram-negative accumulation problem.
- PubChem, broader and less curated, useful for coverage.
- Surveillance reports, which carry MIC50 and MIC90 distributions for circulating isolates rather than for individual compounds. See antimicrobial resistance.
What a screening layer adds on top
Three things, all of which are work you would otherwise do by hand:
- Roll-up per strain rather than per assay record, with the units and relations reconciled.
- A mechanism hypothesis for each organism, drawn from the resistance literature rather than from the activity table.
- A ranking with confidence, so a series becomes a plate list and the rows with no published basis say so instead of being quietly omitted.
The honest caveat: if you screen one compound a month, the manual ChEMBL query is free and perfectly adequate. The case for a screening layer starts when a series arrives and the same afternoon of work has to happen four hundred times.