Tasks and Workflows🔗
Tasks: reusable scientific operations🔗
A task performs one defined operation. Examples include preparing structure data, fetching a multiple sequence alignment (MSA), predicting a structure or calculating evaluation metrics. Each task has an input contract, parameters and a defined output format, so it can be used independently or connected to other tasks to form a workflow.
Tasks are packaged in container images called modules. The task is the scientific operation; the container supplies the software needed to run it.
Workflows: connected steps and useful results🔗
A workflow connects tasks to carry out a scientific process. It passes data between steps, applies the selected parameters and collects outputs. The following examples describe the scientific work covered by our current catalogue.
| Scientific work | What you provide | What you get |
|---|---|---|
| Structure prediction | Molecular queries, such as protein sequences and ligand descriptions; optional alignments or supported structure templates. | Predicted structures and model-specific confidence scores. For OpenFold3, outputs include CIF structures, JSON confidence scores and a manifest identifying each prediction. |
| Structure benchmarking | Reference structures, such as experimentally determined structures supplied for the study. | Predicted structures, per-sample structural quality metrics and aggregate summaries for comparison across models or checkpoints on the same samples. |
| Structure fine-tuning | Training and validation structure files (CIF or PDB) plus selected training parameters. | A prepared dataset, training and validation metrics by epoch, and retained model checkpoints. |
| Binding-affinity prediction | Same as structure prediction. | Predicted 3D structures and binding-affinity estimates. |
| Binding-affinity fine-tuning | Protein–ligand query sets and their alignments, measured-affinity labels, separated into training and validation sets. | Molecular embeddings, retained affinity-head checkpoints and validation metrics. |
| Structure-data preparation | Structure files and preprocessing settings. Depending on the file and the supported repair steps, additional inputs can include SDF files for ligand information missing from CIF or PDB files, or protein sequences to help repair incomplete or inconsistent structure data. | Prepared structures in CIF format (.cif) and processing outputs for use in downstream tasks and workflows. |
| ADMET prediction | Compound SMILES in a CSV file. | CSV of molecular property estimates for the available ADMET endpoints. |
| ADMET fine-tuning | Compound SMILES in a CSV file with ADMET measurements, separated into training and validation data, with training settings. | New model checkpoint and its configuration. |
Example: fine-tune and evaluate a structure model🔗
The OpenFold3 structure fine-tuning workflow connects several steps:
- Provide separate datasets. Provide CIF or PDB files, as training and validation sets, so evaluation measures performance on the intended validation examples.
- Fine-tune on the dataset. The workflow fine-tunes OpenFold3 on the provided data.
- Review the outputs. The dataset outputs describe the processed structures and preprocessing outcomes. Fine-tuning outputs include losses, validation metrics, CSV summaries and retained checkpoints.
These outputs let you compare checkpoints on the same validation set. When the computation finishes, the validation results help assess whether the adapted model is useful for your scientific question.
Example: benchmark against reference structures🔗
For benchmarking, provide reference structures—for example, experimentally determined structures of the complexes being studied. The workflow generates predicted structures and evaluates them against the references, providing per-sample structural quality metrics and aggregate summaries. You can use these to compare models or checkpoints on a consistent dataset.
Running workflows through Foundry Hub🔗
In Foundry Hub, you select a workflow, supply its inputs and parameters, and start a run. Foundry Hub manages execution and exposes run status, logs and outputs. A workflow's instructions describe the expected files and how to interpret them.
Currently, structure prediction, benchmarking and structure fine-tuning offer guided browser interactions, while all workflows offer command-line and skills for agentic interaction.
All tasks can also be executed on their own, so you can build your own workflows yourself.