Computed & derived values

Every cited value carries a hover card naming its source. The underline tells you how far that source was checked, and the three states are deliberately different claims:

• Solid underline - the source was checked directly: either a document read at a recorded location, which the hover card then shows, or a structure or registry record re-derived and matched. This is the strongest claim the Atlas makes, and it applies to about 4% of cited values.
• Dotted underline - a source is named, but no location within it was recorded, or the value is computed, inherited from a shared payload/linker reference, or a registry lookup. About 90% of cited values. The hover card says which. These are not claims of verification.
• Dashed amber, with a marker - the cited source was checked and could not be shown to carry the value (unver.), or the value is an absence claim inferred from the construct class rather than measured (inferred). About 6%. The value is retained because it is usually right; the provenance is not established.

A value with no underline carries no hover card. Usually that means no citation row exists, but a field shown as — (no value recorded) is also rendered without one even where a citation row exists. Aggregate pages - payload, linker, target, chemical space and the query explorer - carry no hover cards at all; use the ADC profile pages for provenance.

The corneal, limbal, conjunctival, and RPE detection rates (hca_*) are the fraction of cells in which the target gene is detected in single-cell RNA-seq atlases of human ocular tissue, from the Human Cell Atlas. The retina value (hpa_retina_ntpm) is bulk retina expression in normalized transcripts per million (nTPM) from the Human Protein Atlas. Both are properties of the target gene, so they are shared across every ADC against the same target.

A value of 0% means the target was assessed and not detected in that tissue; a blank means the tissue was not assessed.

severity_ratio = grade-3+ rate ÷ any-grade rate (the fraction of ocular events that are severe). severity_weighted = 0.3 × any-grade + 0.7 × grade-3+ (a single composite weighting severe events more heavily). Both are computed from the reported ocular-AE rates, which carry their own primary sources.

A three-way provenance flag separating a true documented zero from missing data: reported (a measured ocular-AE rate > 0 at the recommended dose), documented-absent (the eye was assessed and explicitly clear - a real 0%), and unknown (never assessed; excluded from the Atlas).

Mechanistic subtype, Fcγ-receptor and C1q binding status, effector-silencing annotation, dose basis, linker symmetry, and the released-catabolite identity are inferred deterministically from the construct's composition (isotype, Fc mutations, linker/payload chemistry) rather than measured. The composition fields they derive from each link to their own public source.

The Atlas is a curation of the public record, assembled largely by automated literature passes and then audited. A full traceability audit was run on 2026-08-25; these are its findings, stated here rather than left for a reader to discover.

• Most values have a source named but not re-read. Across the ADC profile pages, roughly 4% of cited values have been read against the document and pinned to a location in it. The rest name a source without establishing where in it the value appears. Historically the label “Verified” in the underlying registry was assigned from the confidence tier rather than from a reading; that is no longer displayed as a verification claim.
• Per-trial adverse-event rows carry a weaker check than the profile fields. Of 55,176 trial-level AE rows: 42,391 were confirmed by a script that re-derives the reported percentage from the source record’s own affected/at-risk counts - mostly ClinicalTrials.gov, with about 6% from FDA label tables instead. That check does not distinguish grade, trial arm, or denominator, so it establishes that a matching rate exists in the record, not that it is the rate for the grade and cohort shown; these are marked rate reproduced. A further 7,964 rows are extraction-stage and marked to verify. The remaining 4,821 are the original hand-curated pivotal-trial rows and carry no marker.
• The comparability layer and the OAE data-status flag are internally derived. Ascertainment, native and comparable grading scale, denominator basis and the comparable flag carry no external citation, and the three-state OAE status is computed from our own reading of each trial’s reporting rather than quoted. Treat them as this project’s judgement, not as sourced fact.
• 313 annotation cells are absence claims inferred from construct class. Fc modification, glycoengineering, C1q binding, effector silencing and hydrophobicity masking are frequently recorded as “None”, “Standard” or “Retained” because most ADCs use an unmodified IgG1 Fc - not because a source measured it. These are marked inferred.
• Some citations point at the wrong document. The audit resolved every identifier in the registry against PubMed, PMC, Crossref, ClinicalTrials.gov and DailyMed, and found three separate problems. Fifteen cited documents, backing 26 values, have no bearing on ADCs - mostly a plainly wrong identifier, and in two cases a link built from the drug’s own development code or CAS number; the affected values are marked, though four rows whose identifier field bundles several references are not yet caught. Fifty-three further rows carried a link assembled from a patent number, a compound id or a PMC id stripped of its prefix, which resolved to a real but unrelated record; those have been re-pointed to the correct registry or removed. Separately, about 30 documents behind roughly 740 values resolve to a genuine paper that is not the one the citation names - sometimes a different study of the same drug, sometimes a different drug entirely. Where a cited document looks irrelevant to the value, it may be.
• We hold about 58% of the cited literature in full text. 440 distinct external documents are cited; 254 are held as full text, 59 as abstract only, and the remainder not at all. A value cited to a paper we hold only as an abstract has not been checked against its methods section.
• Ocular adverse-event rates are not comparable across trials without care. Trials that mandated ophthalmic examination report far higher ocular AE rates than trials that recorded only patient-reported symptoms - a difference larger than most of the biological effects in this dataset. The ascertainment field on each ADC page records which applies.
• A blank is not a zero. An empty cell means the value is not in the public literature. A documented 0% means the tissue or endpoint was assessed and nothing was found. The Atlas renders these differently everywhere, and never converts a reporting bound such as “<1%” into a zero.

Corrections are welcome. Where a value here disagrees with the source you hold, the source wins - please tell us.