Structure activity prediction for antibacterials

QSAR software and QSAR modeling software for antibacterial activity, without a training set of your own

Every QSAR package expects you to arrive with a curated dataset. Most AMR teams do not have one. Paste a structure or a sequence here and the predicted MIC band per strain comes back in about ten seconds, with the mechanism named.

From $149 a month

No training set required
A band per strain, not one number
Confidence stated on every row

A prediction, before you decide whether to build one

of 3 left

Compound, by name or SMILES

Strain panel

S. aureus (MRSA) E. coli K. pneumoniae (CRE) A. baumannii P. aeruginosa E. faecium (VRE)

Read-out

Include closest published analogs

No account needed. 3 screens per session.

What comes back

Rows: One per strain in the panel
Per row: MIC band, S I R call, mechanism
Support: Confidence 1 to 4 and analogs
Typical run: About ten seconds

Sample output Ciprofloxacin Fluoroquinolone

Worked example, replaced when you run

Strain	MIC (µg/mL)	Call	Resistance risk
S. aureus (MRSA)	0.5->32	R Resistant	target mutation, grlA and gyrA Fluoroquinolone resistance is widespread in methicillin resistant lineages.
E. coli	<=0.015-0.06	S Susceptible	efflux, AcrAB-TolC in resistant isolates Potent against wild type; qnr carriage and gyrA changes shift the band sharply.
K. pneumoniae (CRE)	0.5->64	R Resistant	gyrA mutation with plasmid qnr Carbapenemase producing isolates almost always carry quinolone resistance too.
A. baumannii	8->64	R Resistant	efflux, AdeABC Constitutive efflux plus target changes leave little room at achievable exposure.
P. aeruginosa	0.25-2	I Intermediate	efflux, MexAB-OprM Borderline: active on many isolates, lost quickly once efflux is derepressed.
E. faecium (VRE)	4->32	R Resistant	target mutation, parC Enterococci are intrinsically poor fluoroquinolone targets.

Why

Ciprofloxacin is a well characterised fluoroquinolone, so the Gram-negative bands are strongly supported by published activity data. The deciding factor across this panel is not target affinity but exposure: efflux in P. aeruginosa and A. baumannii, and acquired target mutation everywhere resistance is already common. Against a modern ESKAPE panel it reads as a Gram-negative agent with two reliable losses.

Closest published analogs

Levofloxacin Broadly similar Gram-negative bands, better Gram-positive coverage
Delafloxacin Retains activity against many ciprofloxacin resistant staphylococci

Strain	MIC (µg/mL)	Call	Resistance risk	Conf.

Why

Closest published analogs

Want the rest of your series? Batch runs are on the paid tiers. screen left in this session. This session is used up.

Computational prediction from published literature. Research use only, not a lab measurement and not clinical guidance.

The short answer

QSAR software builds a model. This gives you the prediction.

QSAR software fits a relationship between chemical structure and a measured property, then applies it to compounds you have not made yet. Every commercial package in the category, Schrodinger AutoQSAR, the Auto-Modeller module in StarDrop, QSAR in Cresset Flare, ADMET Modeler in ADMET Predictor and QSAR inside MOE, assumes the same starting point: you already hold a consistent dataset of measured values to fit on.

For antibacterial potency, most teams do not. That is the gap this page is about. Antibacterial reads the published antibacterial neighbourhood of your compound and returns a predicted MIC band for each organism in a 6 strain panel, with an S, I or R call, the resistance mechanism expected to limit it and a confidence level per row. No training set, no descriptor selection, no applicability domain to defend in a meeting.

The two approaches are not rivals so much as different purchases. If you run one assay at volume and employ someone to own the model, a QSAR package is the better instrument and this page says so plainly further down. If the recurring question is which compounds are worth plate time against which pathogens, a built model is a long route to an answer you can have in about ten seconds.

At a glance

Training set: Not needed, the model is already built
Read-out: Predicted MIC band per strain, with mechanism
Inputs: SMILES, InChI, compound name, peptide sequence
Priced here: $149 to $1,490 a month, monthly not annual

The options, side by side

QSAR modeling software compared, including what each one needs from you

The column that decides most purchases is not the feature list, it is what the tool expects you to bring. Every row below was read at the vendor's own page. Where a vendor publishes no price, the row says so rather than repeating a figure that circulates second hand.

Tool	What it is	Builds your own model	What you have to bring	Published price	Best for
Schrodinger AutoQSAR and DeepAutoQSAR	Automated model building inside the Schrodinger suite	Yes, and it is the point of the product	Your own assay data, one endpoint at a time	Not published, quote only	A team with a computational chemist and a clean internal dataset
Optibrium StarDrop Auto-Modeller	A module of StarDrop that builds and validates QSAR models on your chemistry	Yes	Your own assay data	StarDrop from about $10,000 a year, rising into six figures, published by the vendor	Design teams already running multi parameter optimization
Cresset Flare QSAR models	Field based 3D QSAR plus RDKit fingerprints and custom descriptors	Yes	Your own assay data, with aligned structures for the 3D methods	No commercial price published, academic licences offered	Ligand based design where 3D field shape carries the signal
Simulations Plus ADMET Predictor	175 or more shipped models for ADME, toxicity and physchem, plus ADMET Modeler to retrain them	Yes, by retraining the shipped models	Optional, the shipped models run without any	Not published, evaluation on request	ADME and tox triage rather than antibacterial potency
Chemical Computing Group MOE	QSAR and QSPR model building inside a full molecular modelling environment	Yes	Your own assay data	Not published, seat or site licence on quote	Groups that want modelling, docking and QSAR in one environment
OECD QSAR Toolbox	Free regulatory read-across and category formation, with about 63 databases behind it	No, it fills data gaps from analogues	None, it uses its own databases	Free of charge, published by the OECD	REACH style hazard assessment, not potency against a pathogen
Antibacterial, this site	A pre-trained antibacterial read-out: predicted MIC band per strain, with the mechanism named	No, the model is already built	None, it reads the published antibacterial neighbourhood of your compound	$149, $499 or $1,490 a month, on the pricing page	AMR programs with no in-house MIC dataset to train on

Vendor pricing and module names move. Treat each row as the position at the time of writing and confirm with the vendor before a budget depends on it. What the wider category charges, and which suites disclose anything at all, is worked through on drug discovery software cost.

Before you budget for a modelling seat

Four reasons an antibacterial QSAR project costs more than the licence

None of these are arguments against QSAR. They are the things that turn a three week modelling exercise into a six month one, and they are specific to whole cell antibacterial data rather than to the software.

You need one assay, run one way, at scale

A QSAR model is only as consistent as the data under it. Broth microdilution reads in doubling dilutions, so the same isolate measured in two labs can legitimately differ by a full dilution step, and a public MIC set stitched together from many papers carries that spread as noise. Modellers who have tried this on antibacterial data usually end up spending the first month on curation rather than on modelling.

Antibacterial activity is really two problems stacked

Whole cell potency is target engagement times getting into the cell and staying there. In Gram-negatives the second term usually decides the outcome, which is why the physicochemical property rules that predict accumulation in E. coli look nothing like a classical potency QSAR. A single regression on MIC is quietly modelling both at once, and it generalizes badly outside the series it was fitted on.

One model per organism, not one model

The unit of work in an antibacterial program is compound times strain. A model that predicts activity against S. aureus tells you very little about A. baumannii, so a QSAR approach here means building and maintaining a family of models, each with its own training set, its own applicability domain and its own revalidation when the panel changes.

Somebody has to own it

A built model is an asset with running costs: retraining as data arrives, watching for drift, defending the applicability domain when a program meeting asks why a compound scored well. That is a job, and in a small AMR team it is usually the same person who is meant to be running the chemistry.

What happens when you paste a structure

Read-across, not a regression you have to maintain

Signal 1

Measured activity first

Where the compound itself has published antibacterial activity, that is the strongest signal and it is used directly.

Signal 2

Then the close analogs

The shape of the series around it: what the nearest published neighbours do against each organism, and how steeply activity moves with structure.

Signal 3

Then the species

What the resistance and surveillance literature says about each organism, because a compound class that is dead against KPC producers is dead regardless of its potency on paper.

The read-out is a band rather than a point value, because broth microdilution resolves in doubling dilutions and a prediction with more precision than the reference method would be dishonest. Each row carries its own confidence level, so a compound sitting in a well published class and a compound with almost no neighbourhood do not look the same on the screen.

What it does not do is replace a plate. It decides what goes on the plate. The full method, including the limits, is written out on how it works, and the reasoning behind the band itself is on minimum inhibitory concentration.

Where teams use it

Six jobs it does without a model of your own

Triage before a plate run

Rank a series against the panel and spend plate time on the top of the list instead of the whole set.

A library you were quoted for

Score a vendor deck before you buy it, so the purchase decision has a predicted hit picture behind it.

Peptides as well as small molecules

A sequence goes in the same way a SMILES string does, which classical descriptor based QSAR handles poorly.

A defensible exclusion

Every compound you drop carries a written reason and a confidence level, which is what a program meeting actually argues about.

Grant and partner reporting

Export the matrix as CSV, SDF or PDF and attach the reasoning rather than a bare score.

Inside an existing pipeline

The same engine answers over REST, so a design loop can call it between enumeration and synthesis.

The honest split

When a QSAR package is the right purchase, and when it is not

Buy QSAR modeling software when

You run a consistent in-house assay and the data lands in one place, in one format.
A named person owns modelling as part of their job, not as a favour between experiments.
The endpoint you care about is ADME, solubility or tox, where shipped models are mature.
You need the model itself as an asset, for a partner, a filing or an internal platform.

Use a pre-trained antibacterial read-out when

You have no curated MIC dataset, which is the normal position for a young AMR program.
The question is per organism, across a panel, rather than one endpoint on one series.
Peptides and small molecules both have to go through the same triage.
The budget is a grant line and a five figure annual licence is not survivable this year.

Plenty of groups end up doing both, and the order usually matters: triage the library first, get the assay running on what survives, and only then is there a dataset worth fitting a model to. If that is the path you are on, the volume side is covered by compound library screening software, the single compound path by MIC prediction software, and the same engine is callable from a design loop through the antibacterial screening API. Peptide programs have their own page at antimicrobial peptide prediction tool.

Questions people ask before buying

QSAR software, answered plainly

What software is used for QSAR modeling?

The commercial options are Schrodinger AutoQSAR and DeepAutoQSAR, the Auto-Modeller module in Optibrium StarDrop, QSAR model building in Cresset Flare, ADMET Modeler in Simulations Plus ADMET Predictor, and QSAR inside Chemical Computing Group MOE. The OECD QSAR Toolbox is free and aimed at regulatory read-across. All of the commercial packages except StarDrop quote on request.

How much does QSAR software cost?

Only one vendor in this group publishes figures. Optibrium states that StarDrop annual subscriptions typically start at around $10,000 and can rise into the six figure range with more modules and users. Schrodinger, MOE, Flare and ADMET Predictor publish nothing and price per quote from seats, modules and term, so budget on a quote cycle of several weeks.

Can QSAR predict MIC?

Yes, within a series and with enough consistent data. The practical limits are that MIC is read in doubling dilutions, public values carry roughly a one dilution spread between labs, and whole cell activity in Gram-negatives depends on accumulation as much as on target affinity. A model fitted on one scaffold against one organism rarely transfers to a new scaffold or a new species.

Do I need training data to use QSAR software?

For the model building packages, yes: you supply the measured endpoint and the software fits the model. Two kinds of tool do not need it. Shipped model libraries such as ADMET Predictor come pre-trained on their own data, and read-across tools infer from analogues. The screening on this site sits in the second group, so it runs on a structure with no dataset of your own.

What is the difference between QSAR and read-across?

QSAR fits a mathematical relationship between structural descriptors and a measured property across a training set, then applies it to new compounds. Read-across infers the property of one compound from measured values for structurally and mechanistically similar ones, without fitting a global equation. Regulators accept both, and read-across is the usual answer when the data is too sparse or too noisy to fit a model.

Is QSAR machine learning?

Modern QSAR is machine learning applied to chemical descriptors. AutoQSAR generates descriptors and fingerprints, fits models by several statistical methods, then ranks them by predictive accuracy on held out data, which is a standard supervised learning pipeline. The older literature used linear regression on a handful of descriptors, and the underlying logic has not changed.

How accurate is a QSAR model for antibacterial activity?

Inside the chemical series it was trained on and against the organism it was trained against, a well built model is useful for ranking. Outside its applicability domain the error grows quickly and the model rarely says so. That is why the read-out here carries an explicit confidence level per row and names the resistance mechanism it expects, rather than returning a single number with no error bar.

Should I build a QSAR model or buy a prediction?

Build when you have a consistent in-house assay, a chemist who will own the model, and a series you intend to work for years. Buy a prediction when the recurring question is which compounds deserve plate time against which organisms and nobody has a curated MIC dataset to fit on. Most small AMR teams are in the second case, and discover the first case is a project rather than a purchase.

Pricing

Published, and monthly rather than annual

Full plan comparison

Lab is $149 a month for 250 screens with batches up to 50 compounds. Discovery is $499 for 2,000 screens, batches up to 1,000, and adds the API. Program is $1,490 for 10,000 screens with batches up to 10,000, which is the deck scale plan. A modelling seat in this category is an annual commitment starting in five figures, so the comparison is a different shape as well as a different number.

Yearly billing is two months cheaper, charged once a year.

Lab

$124 /mo

$1,488 charged once a year

250 screens a month against the standard panels, with CSV export.

Get started

Academic AMR lab, 1 to 2 chemists

Discovery

Screen a compound before you price a modelling seat

Paste a structure you know the answer for, read the panel, then decide whether you need a model of your own.

Run a screen first

What you are agreeing to

Your compounds stay yours
Never used for model training
Deletable on request
No card needed to create your account