How it works

High throughput screening drug discovery, one layer above the plate

Antibacterial sits between a library and a plate. It reads the published antibacterial neighbourhood of your compound, rolls the evidence up per strain, and returns a ranked read-out you can defend in a program meeting. Here is exactly what happens, and what does not.

Run a screen

The method

Four steps, about ten seconds

01

Submit the compound

A SMILES string, an InChI, or the name of a known agent. Batch runs take an SDF or a CSV with one structure per row, up to the limit on your plan.

Accepted: SMILES, InChI, compound name
Batch: SDF or CSV upload
Not needed: A target, a protein structure or a docking setup

02

Choose the strain panel

The unit of work is compound times strain. Pick one of the standard panels or assemble a custom set of clinical isolates. Enterprise can load an internal isolate library so the panel matches your own collection.

ESKAPE: S. aureus, E. coli, K. pneumoniae, A. baumannii, P. aeruginosa, E. faecium
Gram-negative: Adds E. cloacae and S. Typhimurium
Gram-positive: Adds MSSA, E. faecalis, S. pneumoniae, S. epidermidis

03

The model reads the published neighbourhood

Reported antibacterial activity for the compound itself where it exists, then for its closest structural analogs, then what the surveillance and mechanism literature says about resistance in each species. The scaffold, the charge, the size and the polarity all matter, because in antibacterials uptake usually decides the outcome before affinity does.

Signal 1: Measured activity for the compound, when published
Signal 2: Activity of close analogs and the shape of the series
Signal 3: Species level resistance mechanisms and how common they are

04

Read the matrix and rank the series

One row per strain: a predicted MIC band in µg/mL, an S, I or R call, the mechanism the model expects, a one sentence reason, and a confidence level from one to four. Batch runs add a ranking across the whole series.

Export: CSV on every plan, SDF and PDF from Discovery
API: REST, from Discovery upwards
Honesty: No published basis returns no number, not a guess

What it does

Triage, ranking and a reason

Predicts a doubling dilution MIC band per strain, not a single false precision number.
Names the resistance mechanism it expects for that species, such as efflux through AcrAB-TolC or a KPC carbapenemase.
Returns the closest published analogs it can actually name, with their reported activity.
Ranks a series so the plate list is a decision, not a shortlist of everything.
Says when it has no basis, with a confidence level of one and no invented number.

What it does not do

The limits, stated up front

It does not measure anything. Every number is a computational prediction from published data.
It is not a diagnostic and not an antibiogram. Nothing it returns should reach a patient decision.
It does not model your specific in-house isolate unless you are on Enterprise and have loaded it.
It does not predict toxicity, PK or formulation behaviour. Those are different questions.
It does not invent citations, accession numbers or p-values, so it will sometimes say less than you want.

Confidence

Every row says how much it is standing on

A prediction without a confidence level is a guess with good manners. Each row carries a four segment meter, and the meaning of each level is fixed.

Level	Meaning	Typical case	How to use it
4	Extensively characterised	A marketed agent against a species with decades of surveillance data	Treat the band as a strong prior
3	Well supported	A known scaffold with several published MIC datasets nearby	Good enough to order a series
2	Inferred from close analogs	A new analog in a well described family	Use for ranking, confirm the top calls
1	Little or no published data	A novel scaffold with no antibacterial neighbourhood	Treat as unknown, the model will often decline to give a band

The same discipline applies to the numbers themselves. A band such as 0.5-2 is a doubling dilution range, which is the resolution a broth microdilution plate actually has. If you want the detail behind that, the minimum inhibitory concentration guide walks through how the number is read, and high throughput screening covers where in silico triage sits relative to a real plate run.

Reference pages behind the method

The science the model reasons over, written out in full.

Screen the compound you are arguing about this week

Run a screen on the homepage without an account. When the read-out is useful, create one and bring the rest of the series.

Run a screen first

What you are agreeing to

Your compounds stay yours
Never used for model training
Deletable on request
No card required to run a screen