Skip to main content
All solutions

Solutions · 4 workflows

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.

Not sure which? Describe your problem
[Predictive modeling]Fig. 04
Workflows
4
Stages
4
Tools
6
Each line is one workflow. Lines share a station where they run the same step; select a workflow below the map to jump to its details.
[Predictive modeling · what goes in, what comes out]

  1. Your target

  2. Its measured inhibitors

    Imatinib
    Nilotinib
    Dasatinib

  3. 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.

Workflows

Pick the one that fits.

Open any of them in the builder: signed out you get a read-only preview, signed in you can add your inputs, see the estimate for your plan, and run. The cost cap is the default; you can change it at launch, and a run pauses instead of spending past it.

  • 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: Train an activity model from ChEMBL data for your targets (opens in a new tab)
  • moderate5 steps

    Train an activity model from ChEMBL data and score your library with it

    ChEMBL activities → curate → train MPNN → score your library

    Pull one activity type for your targets from ChEMBL, curate it, train an MPNN, then predict activity for your own molecules and keep the 20 highest. ChEMBL data is CC BY-SA 3.0, and every table carries its licence notice.

    You provide
    Molecules (SMILES), Model name, Targets
    Runs on
    CPU
    Cost cap
    $15.00 default
    Open in the builder: Train an activity model from ChEMBL data and score your library with it (opens in a new tab)
  • simple2 steps

    Train a reaction yield model from your reactions

    Reactions → atom-map → train a reaction model

    Atom-map a table of reactions with RXNMapper, then train a reaction model on a measured column such as yield. Models work best on consistent in-house HTE data; mixed literature yields train weak models.

    You provide
    Model name, Table (CSV)
    Runs on
    CPU
    Cost cap
    Set at launch
    Open in the builder: Train a reaction yield model from your reactions (opens in a new tab)
  • moderate4 steps

    Enumerate products and keep the reactions your model predicts best

    Reaction enumerate → atom-map → predict yield → keep the best

    Enumerate products from a reaction, atom-map each enumerated reaction, predict its yield (or another reaction property) with one of your saved reaction models and keep the 50 highest. A model trained with condition components (ligand, base, ...) takes them as fixed component_values, one SMILES per component.

    You provide
    Model id, Molecules (SMILES), Reaction smarts
    Runs on
    CPU
    Cost cap
    Set at launch
    Open in the builder: Enumerate products and keep the reactions your model predicts best (opens in a new tab)

Related tools

Run any step on its own.

The tools behind these workflows also run individually.

  • ChEMBL Target Activities

    Pull measured IC50, Ki, Kd, and EC50 values for your targets from ChEMBL into one table

  • Bioactivity Dataset Curator & Splitter

    Normalize assay units, censoring, and replicates into SAR/QSAR datasets with scaffold, random, or time splits

  • MPNN Models

    Fast structure-property inference that highlights key molecule substructures

  • Reaction Atom Mapper

    Atom-map reaction SMILES and check element, charge, and stereochemistry balance

  • Reaction Models

    Predict reaction yield, selectivity and rates from atom-mapped reactions and their conditions

  • Reaction Enumeration

    Enumerate product libraries from reaction SMARTS or RXN files and reactant SMILES lists