rowansci.com tools
Run Molecular Modelling & Computational Chemistry workflows — pKa, tautomers, bond strengths, Fukui indices, redox potentials, solubility, docking, descriptors and FEP — through your own rowansci.com account.
rowansci.com tools
Rowan is RxCubed's Molecular Modelling & Computational Chemistry provider, which is a different kind of answer from everything else in RxCubed. The local tools apply published rules; the Neurosnap tools run machine-learned models. Rowan solves the electronic structure of the molecule via quantum chemistry, so the numbers it returns are physics rather than pattern-matching — and slower and more expensive for the same reason.
They run on your own rowansci.com account, using your API key, and they spend your own credits.
Before you start
Add your key under Settings → External services. Create one at labs.rowansci.com under Account → API keys. Without a key, Rowan workflows appear in the Tool Selector but cannot be submitted. See API keys.
Two things differ from RxCubed's own tools:
- They cost credits. RxCubed does not mark this up and does not pay it for you. Your remaining balance is shown in Settings → Usage.
- Your inputs leave RxCubed. A submitted molecule or protein is sent to Rowan and is subject to their terms and data handling.
The nine workflows
| Workflow | Answers | Input |
|---|---|---|
| pKa prediction | Which sites gain or lose a proton, and at what pH | A SMILES, or a 3D structure for the two structure-based methods |
| Tautomer search | Which drawn form of the molecule actually exists | A 3D structure |
| Bond dissociation energies | Which bond breaks first — a map of metabolic soft spots | A 3D structure |
| Fukui indices | Which atom is attacked first, and by what (electrophile, nucleophile, radical) | A 3D structure |
| Redox potential | How readily the molecule gains or loses an electron | A 3D structure |
| Solubility | Predicted solubility, in one or more solvents and temperatures | A SMILES |
| Docking (Rowan) | A third redocking engine, alongside the local one and Neurosnap's GNINA | A protein and a 3D ligand |
| Molecular descriptors | A broad panel of computed properties, for comparing a series | A 3D structure |
| Relative binding free energy (FEP) | Binding free energy across a ligand series | A protein and an SDF of ligands |
Eight of them are cheap enough to run casually. FEP is not — see below.
Most of them need real 3D coordinates
Rowan optimises a geometry, so a flat structure has nothing to optimise. A 2D SDF — which is what PubChem returns by default — is refused rather than quietly used, because a tautomer search or a bond strength computed on a flat molecule returns a confident number that is wrong.
Run 3D conformer generation first and pick the SDF it writes. pKa's starling and chemprop_nevolianis2025 methods, and Solubility, are the exceptions: they work from a SMILES string, and refuse a 3D structure.
Estimate before you run
Every Rowan form has an Estimate first button. It sends the exact job that Submit would send and asks Rowan to price it — how long, on what hardware — without queueing anything and without spending a credit. Changing any field clears the estimate, because it priced the previous version of the form.
The credit cap field is Rowan's own spending ceiling. Rowan refuses a submission whose own estimate exceeds the cap, at no charge, so setting one costs nothing and bounds what a job can spend. For FEP it is required.
FEP is the expensive one
Relative binding free energy is the most accurate affinity prediction that exists — typically within about 1 kcal/mol of experiment, where a docking score is lucky to rank correctly. It is also days of GPU time and credits in the thousands for a real ligand series.
Two constraints come from the physics rather than from RxCubed:
- Every ligand must carry the same formal charge. Charge-changing perturbations are not supported. Split a mixed series by charge state and run each separately.
- Every ligand needs a real 3D pose in the site, not a SMILES — so dock or align them first.
It runs as one RxCubed run in two stages: Rowan first plans a perturbation graph (which ligand to mutate into which), then runs molecular dynamics along every edge of it. Pressing Check status advances between the stages, and the run says which stage it is in while it waits. The planning stage's own result is saved too, because it is the only record of how the ligands were connected.
Rowan gates FEP per account. If yours is not enabled for it, the form says so before you fill it in rather than after.
Checking progress and collecting results
Rowan workflows go to the Run Results tab in Tool Selector: Pending Runs while Rowan is working, then Results not collected once finished.
Checking status and collecting files don't spend your own Rowan allowance — only the job's own compute does — so RxCubed checks for you automatically: on opening Run Results or Results Hub, and again whenever you switch back to that tab. The moment a check finds a job finished, its files are fetched automatically. Press ⟳ Check status for all any time you want an answer right now, or if a run somehow finished without ever being checked, it waits under Results not collected for a manual check.
Where results go
Rowan returns its results as data rather than as files, so collecting writes the files rather than downloading them. Every workflow produces at least two evidence rows:
- a JSON file holding the complete result, which is the record;
- a TSV table of the numbers, which is the readable view and opens in the table viewer.
Where a workflow references optimised geometries — pKa microstates, tautomers, the neutral/cation/anion structures behind a redox potential, Docking's own poses — those come back as XYZ files too, capped at eight per run so one collection cannot make dozens of extra calls against your account. Every structure's ID is in the JSON regardless.
All of it lands in Results Hub under Category → Tool → Run → Evidence, with the run settings and selected inputs recorded as provenance.
Reading the results
Energies come back in the units Rowan reports them in, with a readable unit beside — never instead of — the original. Bond dissociation energies, for instance, carry both kcal/mol and the Hartree value Rowan actually returned.
A low C–H bond strength marks a likely site of metabolic oxidation; it says nothing about whether an enzyme can reach it. A predicted pKa near physiological pH means the charge state is genuinely uncertain in the body, not that it is definitely one or the other. See Drug discovery basics.