Product · Workflows

Built around the questions chemists actually ask.

24 ready-made workflows in 4 problem families. Each one is a tested chain of tools with its file formats and hand-offs already connected: add your inputs, check the estimate, and run.

Browse workflows Or describe your problem

Family 01 · 6 stages

Virtual screening

Find the compounds in a library most likely to bind your target, then rescore the best hits with a stronger method.

Explore virtual screening

[Virtual screening]Fig. 01

[Selected workflow]Pick another in the legend

simple3 steps

Screen a compound library against a target

SMILES → 3D embed → dock

Convert SMILES to 3D structures, detect pockets, then dock with AutoDock Vina.

You provide: Molecules (SMILES), Protein structure (PDB)
Runs on: CPU
Cost cap: $10.00 default

Open in the builder

[Virtual screening · what goes in, what comes out]

Your library
Imatinib
Nilotinib
Dasatinib
▸
Your target
▸
A pose in the pocket

Target: ABL1 kinase domain
Structure: PDB 2HYY · 2.4 Å
Polar contacts: 6

Shown: imatinib, nilotinib, and dasatinib as a sample library, and the ABL1 kinase domain with imatinib bound, from its crystal structure. Docking returns a pose like this for every compound, ranked by score. Structure: Cowan-Jacob et al. (2007), Acta Crystallogr D Biol Crystallogr.

Family 02 · 6 stages

Structure prediction to docking

Dock against a protein with no solved structure: predict it from sequence, find its pockets, then dock.

Explore structure prediction to docking

[Structure prediction to docking]Fig. 02

[Selected workflow]Pick another in the legend

simple3 steps

Dock into a predicted structure without choosing a box

Sequence → fold → DiffDock

Fold a protein sequence with ESMFold2, dock a SMILES library blind with DiffDock, and keep the top-confidence poses.

You provide: Molecules (SMILES), Protein sequence
Runs on: A10 GPU · T4 GPU
Cost cap: $110.00 default

Open in the builder

[Structure prediction to docking · what goes in, what comes out]

Your sequence
▸
Its structure
▸
A pose in the pocket

Sequence: 273 residues
Structure: PDB 2HYY · 2.4 Å
Ligand: Imatinib

Shown: the ABL1 kinase domain sequence from PDB 2HYY and that entry's crystal structure. The workflow predicts the structure from the sequence with ESMFold2, finds its pockets, then docks your compounds into them. Structure: Cowan-Jacob et al. (2007), Acta Crystallogr D Biol Crystallogr.

Family 03 · 6 stages

Library design

Generate or enumerate new compounds for a target or an SAR series, then dock them.

Explore library design

[Library design]Fig. 03

[Selected workflow]Pick another in the legend

moderate4 steps

Dock analogs enumerated from your SAR data

Free Wilson → 3D embed → dock

Enumerate analogs with Free Wilson, embed to 3D, detect pockets, then dock with AutoDock Vina.

You provide: Molecules (SMILES), Protein structure (PDB)
Runs on: CPU
Cost cap: $15.00 default

Open in the builder

[Library design · what goes in, what comes out]

Your series
Imatinib
Nilotinib
Radotinib
▸
A combination not in it
Radotinib head · Imatinib tail
▸
A pose in the pocket

Series: 3 ABL inhibitors
Shared core: Phenylaminopyrimidine
New combination: C28H30N8O

Shown: imatinib, nilotinib, and radotinib, which share a core (highlighted). Free Wilson analysis splits a series into that core and its substituents and enumerates the combinations not yet made, like radotinib's head on imatinib's tail; the workflow then docks them. The pose is imatinib's crystal pose. Structure: Cowan-Jacob et al. (2007), Acta Crystallogr D Biol Crystallogr.

Family 04 · 4 stages

Predictive modeling

Pull measured activities for your targets, curate them into a clean dataset, train a model that predicts activity, and score your own compounds with it.

Explore predictive modeling

[Predictive modeling]Fig. 04

[Selected workflow]Pick another in the legend

simple3 steps

Train an activity model from ChEMBL data for your targets

ChEMBL activities → curate → train MPNN

Pull one activity type for your targets from ChEMBL, curate it into a scaffold-split dataset, then train an MPNN on it. ChEMBL data is CC BY-SA 3.0, and every table carries its licence notice.

You provide: Model name, Targets
Runs on: CPU
Cost cap: Set at launch

Open in the builder

[Predictive modeling · what goes in, what comes out]

Your target
▸
Its measured inhibitors
Imatinib
Nilotinib
Dasatinib
▸
A model that reads them

Target: ABL1 kinase domain
Measured: 3 ABL inhibitors
Model: MPNN

Shown: the ABL1 kinase domain from its crystal structure and imatinib, nilotinib, and dasatinib, inhibitors whose measured potencies are in ChEMBL. The workflow pulls measurements like these for your targets, curates them into a dataset, and trains a message-passing neural network (MPNN) that predicts activity for your own compounds. The MPNN reads a molecule as a graph: over 3 steps, each atom gathers messages from its bonded neighbours (drawn for one of imatinib's nitrogens), then every atom is pooled into one prediction. The drawing shows the method, not a prediction. Structure: Cowan-Jacob et al. (2007), Acta Crystallogr D Biol Crystallogr.

Under the hood

Change any step, or build your own.

Every template opens in the visual builder. The same building blocks are available for your own workflows, and the builder checks each connection as you make it.

Bindings

Each step input is bound to one source:

Binding	Connects to
`$param`	A workflow input you fill in at launch: a file, a list, or a setting.
`$from`	Another step's output port, e.g. the poses from a docking step.
`$artifact`	A file already in your library, such as a result from an earlier run.
`$const`	A fixed value written into the workflow.

Typed ports

Every port carries a data kind, so a step only accepts what it can use. Known format conversions, such as SMILES to 3D SDF or CIF to PDB, run between steps automatically.

Archive
Docked poses
File
Force-field system
Image
Kinetic mechanism
Molecules
Protein sequences
Protein structure
Sequences
Structures
Table
Trained model

Filters and fan-out

Filter steps reshape results between jobs: deduplicate, filter rows, limit, merge, partition, sample, select top k, split, zip join. A step can also run once per item (scatter): one job per ligand, for example.

Up to 64 steps

Workflows can grow to 64 steps. Each run is estimated before it starts and pauses at its cost cap rather than overspend; you can resume it.

Versions

Saved workflows keep their version history, so you can look back at an earlier version or restore it.

Share and fork

Share a workflow with a link; colleagues fork their own copy to change and run. Export any workflow as JSON.