A PD-L1 minibinder campaign, end to end

One campaign page. goal_statement, binder_len and order_budget are structured fields, so the filter and ordering screens can read them later.

Read the A–O progress table in the body next to the structured fields in the infobox on the right.

One target page plus a filedb/external cache. Domain bounds, glycan sites, disulfides and the transmembrane span are filled in from UniProt annotations.

Look at the automatic check section — it reports the transmembrane span warning, and the gap between the N-X-[ST] sequons computed from the sequence and the annotated glycan sites. This data is real.

target.hotspots (list) plus one decision note. The note is tied to the target and campaign by wiki links.

Four separate reasons are written down for choosing that face. Being able to revisit them after a failure is the whole point of this note.

The engine registry (in code) plus a job document. Engines that are not installed stay in the list with their reason attached inline.

Read the reason on a greyed-out engine. They are not hidden because this list is the roadmap — and results can be imported before anything is installed.

A run page. The raw parameter JSON, the number generated, the number passing and the output path all stay, which is what makes the run reproducible.

The mode reads as an import — the output was produced elsewhere and brought in. That is the honest state of this run.

The designs collection. No body or revisions, but per-metric indexes, so sorting in bulk stays fast. It shares the title namespace with the wiki, so a design can be linked before it is ever promoted.

A dashed design name means it is not a wiki page yet. Promote it and it becomes a real page. Note the red DEMO chips too.

Immutable versioned filter_profiles, a filters.<key> verdict on every design, and a filter_runs audit log.

Apply it and save. The result screen shows which cut is doing the killing, criterion by criterion — not just how many passed. It also spells out why i_pAE (normalised) and PAE (raw) are separate fields.

designs.cluster, designs.props.liabilities and the shortlisted flag.

The warning column in the design table is a developability score from seqprops — deamidation, free cysteines, furin sites, hydrophobic runs. It is computed on CPU, with no GPU involved.

Assay pages. They point at designs by multiref, so they show up automatically in the backlinks of each design page.

Every assay instrument reads as not measured. Synthetic data is never left looking like a real measurement.

A note page plus the next campaign page (round=2). Revisions accumulate, so the history of the judgement stays.

Finally, open Recent changes and the History tab of any page — every edit stacks a revision, and structured fields and body text can be rolled back separately.