FDA Surrogate Endpoints Project

Planned REGULATORYBIOMARKERSPATHOGRAPHCROSS_CUTTING
Home / Projects / FDA Surrogate Endpoints Project

FDA Surrogate Endpoints Project

Overview

FDA publishes a table of the surrogate endpoints that have been the basis of a drug or biologic approval. dismech holds a row-level transcription of it in kb/surrogate_endpoints/fda_surrogate_endpoints.yaml (225 rows, FDA content current as of 2026-04-29), imported by scripts/import_fda_surrogate_endpoints.py and tracked in issue #2495.

A row says that FDA accepted a measurement for a disease, in a population, for a drug mechanism, at a stated validation level. It does not say what the measurement is a measurement of. The dismech pathograph can answer that: a Biochemical marker's readouts[] names the pathophysiology node it reports on, and readouts[].regulatory_endpoint_refs points that readout at the FDA row. Once that link exists, the graph shows which downstream outcomes the surrogate sits upstream of, and which sibling branches it never sees.

The aim is to make that link for every disease-specific row that has a dismech entry, and to make the coverage gaps readable.

Worked example: Fabry_Disease

FDA lists one Fabry marker across three rows: clearance of GL-3 (Gb3) inclusions in biopsied renal peritubular capillaries.

Row Population Approval Drug mechanism Validation level
FDA-SE-adult-noncancer-024 Adults Traditional Enzyme replacement therapy Validated
FDA-SE-pediatric-noncancer-019 Age 2 and older Traditional Enzyme replacement therapy Validated
FDA-SE-adult-noncancer-025 Amenable GLA variant Accelerated Pharmacological chaperone Reasonably likely

All three are attached to the Renal Globotriaosylceramide Inclusions marker, whose readout targets Renal glycosphingolipid storage and podocyte injury. That node leads to proteinuria, nephrotic syndrome and chronic kidney disease. The other five branches of Tissue-specific glycosphingolipid storage (cardiac, vascular endothelial, small-fiber/autonomic, ocular, cutaneous) carry no FDA-recognized marker. The entry already records this for the heart: its cardiomyocyte autophagy node is described as "not measured by the renal GL-3 surrogate". Plasma lyso-Gb3 is curated as a monitoring readout and is not on the FDA table.

The same reading should be possible for every disease below.

Current state

Counts from the committed table on 2026-10-10.

mapping_status Rows Meaning
EXACT_DISMECH_MATCH 32 Disease label matches a dismech entry
CURATED_DISMECH_MAPPING 12 Matched by a hand-written rule in the import script
CANDIDATE_DISMECH_MAPPING 1 Proposed, not confirmed
NEEDS_CURATION 133 No mapping yet
NOT_DISEASE_SPECIFIC 47 Antiseptics, vaccines across pathogens, and similar

The 45 mapped rows point at 27 disorder entries. Of the FDA row IDs, 24 are cited by a regulatory_endpoint_refs somewhere in kb/, across 15 files.

Mapped but not wired. These entries have FDA rows mapped to them and no regulatory_endpoint_refs citing those rows:

Unmapped rows whose disease has a likely entry. Spot-checked by file name; each needs a curator to confirm the entity is the same before mapping: Type 1 Gaucher disease (Gaucher_Disease has-subtype), alkaptonuria, lysosomal acid lipase deficiency (Wolman_Disease covers the infantile form only), cystinuria, primary hyperoxaluria type 1, arginase 1 deficiency, cerebrotendinous xanthomatosis, familial chylomicronemia syndrome, hypercholesterolemia, hepatitis B, C and D, influenza, pertussis, Alzheimer's disease, Cushing's disease.

The largest unmapped groups (17 hematological-malignancy rows, 9 solid-tumor rows, 8 nonmalignant-hematology rows) are FDA umbrella labels and will usually map to several entries or to none. Treat them last.

Work plan

Tranche 1: wire the mapped-but-unwired entries

For each entry in the list above:

  1. Find or add the Biochemical marker the FDA endpoint measures.
  2. Give it a readouts[] item targeting the pathophysiology node the marker reports on, with regulatory_endpoint_refs naming the FDA row IDs.
  3. Cite evidence for the readout link itself. The FDA row is a regulatory record, not evidence that the marker reads out that node.
  4. Where no marker can honestly be placed on a node (a composite clinical score, a functional test), say so in the entry's notes rather than forcing a readout.

Tranche 2: confirm the mapping candidates

Work through the unmapped-with-likely-entry list. A confirmed match goes into MANUAL_MAPPINGS in scripts/import_fda_surrogate_endpoints.py and the table is regenerated. Never hand-edit fda_surrogate_endpoints.yaml. A row whose population is narrower than the entry (an amenable-variant population, a subtype) should map to the subtype where one exists.

Tranche 3: branch-coverage report

Write a report-only script (in the style of just list-disconnected-phenotypes) that, for each wired FDA row, walks the pathograph from the readout's target node and lists:

Fabry is the reference output. The report must not be a gate: a surrogate that covers one branch is a property of the regulatory record, not a curation defect.

Tranche 4: surrogacy evidence

For rows marked REASONABLY_LIKELY_SURROGATE_ENDPOINT (29 rows), the open question is whether the surrogate-to-outcome link has since been confirmed. just research-surrogacy <provider> <disease> "<surrogate>" "<outcome>" exists for this. Results are leads for the entry's evidence, not KB content.

Open questions

Commands

just validate-surrogate-endpoints        # schema-validate the table
just research-surrogacy openscientist Fabry_Disease \
  "renal peritubular capillary GL-3 clearance" "chronic kidney disease"
just validate-disorders kb/disorders/<Entry>.yaml
just gen-project-pages                   # render this page