Discovery of new antibiotics: the questions that block the decision

A prediction. Antibacterial estimates how a compound is likely to behave against a strain, based on published antibacterial activity data and the structural neighbourhood of the compound. It is not a measurement, it does not replace a plate, and every row of every result carries a confidence level and the standing label "computational prediction, research use only". A screen tells you what is worth measuring. The plate still tells you what is true.

A ChEMBL query returns records. A screen returns a decision. Antibacterial rolls activity up per strain rather than per assay, names the resistance mechanism it expects and why, ranks the compounds in your series against each other, and does it in about ten seconds instead of an afternoon. ChEMBL is one of the sources underneath, and the honest answer is that if you only ever screen one compound a month, the manual query is free.

No. Antibacterial is a research tool for discovery teams. It is not a diagnostic, not an antibiogram, and not clinical guidance. Nothing it outputs should reach a patient decision, directly or indirectly. This is stated in the product, in the exports and in the terms.

No. Submitted structures are never used to train or fine-tune any model, they are not shared with other customers, and they are deletable on request. This is a written commitment in the terms, not a preference setting.

The ESKAPE panel (S. aureus including MRSA, E. coli, K. pneumoniae including CRE, A. baumannii, P. aeruginosa, E. faecium including VRE), plus standard Gram-negative and Gram-positive clinical isolate sets. Custom panels are available on Discovery and above, and Enterprise can load your own internal isolate library.

It says so. A scaffold with no published antibacterial activity nearby returns a low confidence level and, where there is genuinely nothing to reason from, no MIC band at all rather than an invented number. A screen that refuses to guess is more useful than one that always answers.

Yes. CSV on every plan, SDF and a PDF report from Discovery, scheduled exports on Program and above. There is a REST API from Discovery upwards so results can go straight into your own registry or notebook. Nothing you put in is locked in.

SSO through SAML 2.0 and OIDC is on Program and Enterprise, along with the audit log and a 99.9% SLA. Invoicing, purchase orders and a signed DPA are available from Program. Private or VPC deployment and data residency options are Enterprise.

It depends entirely on what the model has to stand on, which is why every row carries a confidence level rather than a single headline accuracy figure. A marketed agent against a well surveyed species is a strong prior. A novel scaffold with no antibacterial neighbourhood is a level one row and often no band at all. Publishing one accuracy number across those two cases would be misleading, so we do not.

Because a broth microdilution plate is a doubling dilution series. A measured MIC is already a band in disguise, and two labs testing the same compound and strain routinely land one dilution apart. Reporting 0.73 µg/mL would be false precision dressed up as science.

No. This is a whole cell, phenotype level question. Uptake and efflux usually decide the outcome before target affinity does, which is exactly why target based tools struggle with antibacterials.

Yes, with the same caveat that applies everywhere: the confidence level tells you how much published neighbourhood exists. Antimicrobial peptides often come back with useful mechanism reasoning and lower numeric confidence, which is an honest reflection of the literature.

They are used to answer your screen and nothing else. They are never used to train or fine tune a model, never shared with other customers, and deletable on request. That commitment is in the terms, not in a settings toggle.

No. The screen on the homepage runs without an account and is the free taste. Every plan is paid, starting at $149 a month.