Core Examples

The worked scripts live in the repository's examples/ folder. Each worked guide below explains its inputs, pipeline stages, prerequisites, selection criteria, and validation limits. They demonstrate pipeline composition; inclusion here does not certify scientific performance or successful execution on every supported cluster.

Example	What it demonstrates or measures
Nanoparticle scaffold	C3 RFdiffusion, public I3 RPXdock, tied sequence design, and AF3 trimer refolding; full-cage checks remain required.
Peptide binder	Sequence-conditioned flexible peptide design and two rounds of AF3 selection.
Multi-MPNN comparison	Three model families on a shared backbone, with optional AF3 refolding.
Biotin binder	Ligand-conditioned generation, polar sequence design, AF3 refolding and Rosetta interface analysis.
DNA binder	Explicit DNA coordinates and donor/acceptor conditioning with public RFD3.
Antibody design	Antibody framework conditioning and sequence/refold validation; see the guide for required inputs.
Glycosidase	Catalytic motif scaffolding and apo, covalent-intermediate, and ligand-bound refolding branches.
Protease	Substrate-conditioned design, sequence/refinement cycles, and separate enzyme-only/complex predictions.
Hammerhead ribozyme	RNA motif rescaffolding and crystal-reference catalytic geometry after AF3 refolding.
FGFR2 binder campaign	Backbone generation, sequence design, AF3 refolding, and interface confidence ranking with ipSAE for FGFR2.
PPI benchmark	Two inverse-folding backends in an RFD3 → MPNN → RF3 workflow for the insulin receptor; compare refolding RMSD and interface confidence under matched inputs and sampling.
DISCO benchmark	Heme B ligand-conditioned co-design followed by AF3; measures backbone and ligand pose recovery after refolding.
Active-site motif benchmark	Enzyme design around catalytic motifs; measures recovery of motif sidechain geometry after sequence-only AF3 folding.
Prediction validation cohort	Structure predictors evaluated against reference PDB assemblies with lDDT and PEPPR metrics, with explicit target and chain coverage.
CatCIF storage example	Archiving, reading, and materializing structures; an engineering demonstration, not a scientific benchmark.

CatCIF storage preserves design IDs, parent lineage, metadata, and metrics. Staged and materialized filenames include an archive/entry digest so matching entry names from different archives cannot overwrite each other. Use handler design IDs for identity rather than inferring identity from extracted filenames.

What “benchmark” means here

A benchmark script specifies a task, an input set, and metrics to compare a workflow or model variant. It is not itself a results report. Design benchmarks measure computational proxies such as refolding consistency and motif recovery; they do not measure experimental binding or catalytic activity. The prediction cohort measures agreement with reference structures. Record model/checkpoint versions, seeds, sample counts, failures, and compute cost alongside scores when comparing runs.

The PPI scripts make a comparison between FoundryMPNN and the mpnn block (public LigandMPNN) possible, but their presence does not establish matching sequence distributions or equivalent scientific performance. A controlled comparison must reuse the same generated backbones and scoring protocol and include enough samples to estimate uncertainty.

Validation status and first runs

These GPU campaigns are not run in routine CI. No release-wide, fresh-cluster end-to-end validation is asserted for this list. Unit tests and small integration checks cover different questions from full scientific benchmarking. Where an example has run evidence, consult its README and retained results for the exact hardware, inputs, and versions; otherwise treat it as requiring validation.

Install the relevant containers and weights, follow the example's preparation steps, and start with --test or a small bounded input. Inspect outputs and missing metrics before a pilot or full campaign. A failed or zero-pass pilot needs a diagnosis, not a larger sample count. See the simple example and Slurm site settings.

Why contribute an example?

An example records a reusable composition of existing blocks, including its inputs, settings, and acceptance criteria. It helps both people and coding agents find a relevant starting point. Agents can read examples, but adding one is not required for an agent to discover blocks or create a new workflow; the skills index and block reference serve those purposes.

Contribute a focused script when it adds a distinct task or useful composition. Include a README with prerequisites, preparation commands, a small test command, expected output fields, scientific metrics, and an honest record of what was actually tested. Keep large generated outputs and site credentials out of the repository, and put generally reusable operations into blocks rather than copying orchestration into each example.

Agent benchmark isolation

These worked examples and workflow documentation are public learning material. The Harbor benchmark's agent snapshot excludes examples/ and these pages; its only included documentation page is the generic block reference. The benchmark therefore does not hand these task-specific recipes to agents as workspace context. This filesystem boundary does not prevent an agent with external network access from discovering publicly available material.