Bulk work repeats a calculation across a defined input library. The crucial first decision is how inputs pair, not which button submits them.
Choose the pairing rule
| Intended work | Input organization | Expected calculations |
|---|---|---|
| One receptor against a ligand library | One receptor broadcasts across N ligands | N |
| Matched construct and partner pairs | Two N-entry lists in corresponding order | N |
| Every receptor against every ligand | Explicitly enumerate R × L pairs or orchestrate them | R × L |
| Multiple stages per candidate | Preserve candidate identity and filter deliberately | Depends on surviving candidates and stage settings |
The pipeline builder accepts size-one broadcasting and matching-size parent lists. It does not infer an all-combinations experiment. A file containing several records is also not universally equivalent to several jobs: field and node semantics determine the expansion.
Prepare a manifest
Give every input a stable identifier. Keep a table mapping identifiers to exact sequences, structures, compounds, pairing, intended settings, and resulting job IDs. Validate sequence-structure correspondence and avoid duplicate filenames that obscure identity.
Do not apply a generic CSV format to every service. A molecule list, fine-tuning dataset, and grid input have different contracts. Use the current form or service schema for required columns and uploads.
Pilot before scaling
Select a small representative subset that includes difficult formats, longer sequences, modified chemistry, or missing-data cases you expect in the full library. Check:
- Actual number of created jobs and candidates.
- Input pairing and retained identifiers.
- Expected file names and downstream selectors.
- Runtime, credits, failures, and useful scientific outputs.
Only expand after that handoff works. Increasing sample counts within a service can multiply the amount of downstream work even when the number of input entries is unchanged.
Execute and reconcile
Use Pipelines for supported visual workflows or the API for custom orchestration. Preserve each returned ID as soon as submission succeeds. Limit concurrency and request rate rather than launching an uncontrolled number of requests.
Track failed, pending, completed, and missing items explicitly. Do not silently remove failed cases from aggregate performance statistics or compare only the attractive survivors. On interruption, reconcile the manifest with job history before retrying uncertain submissions.
Summarize the campaign
Export a candidate table with provenance, native scores and units, status, and selection decisions. Rank with a declared method and retain diversity where the research objective requires it. A computational shortlist should identify experimental hypotheses, not claim validated hits.