Skip to content

Biomarkers of aging in dismech: what we have, what to add

2026-08-31. Landscape and gap analysis. Report-only — no KB changes are made by this document.

Summary

dismech has an essentially complete mechanism representation of aging and an essentially empty biomarker representation of it. Eleven of the twelve hallmarks of aging exist as modules with proper trigger→consequence node chains, and 63 disease-entry nodes conform to them. But of those eleven modules, exactly one — cellular_senescence — carries a biochemical: block. It is also the only module of all 166 in kb/modules/ that carries one.

(Status at time of writing. All eleven are now curated — see the status note under Recommendations for what changed and what the tail taught.)

The larger gap, found after this report was first written, is one layer up: dismech has no representation of the clinical endpoints aging biomarkers would be surrogates for — disability-free survival, multimorbidity, frailty index. See The endpoint layer below; it is a bigger hole than the biomarker gap and it is partly a schema question.

The good news on the biomarker layer is that no schema work is needed. BiomarkerReadout already encodes the qualification vocabulary that the NIA/Biomarkers of Aging Consortium literature asks for, and ExperimentalReadout on modeled_mechanisms is its model-system twin. Both carry a biomarker_term and both point at a pathophysiology target. The clinical-versus-model-system reconciliation the field treats as an open problem is, in dismech's data model, already a join on a shared pathograph node — nobody has used it for aging yet.

What we have

Modules: hallmark coverage is complete

Hallmark (López-Otín 2023) Module Nodes biochemical:
Genomic instability genomic_instability_aging 4 —
Telomere attrition telomere_attrition 3 —
Epigenetic alterations epigenetic_alterations 3 —
Loss of proteostasis loss_of_proteostasis 4 —
Disabled macroautophagy disabled_macroautophagy 3 —
Deregulated nutrient sensing deregulated_nutrient_sensing 7 —
Mitochondrial dysfunction mitochondrial_dysfunction 5 —
Cellular senescence cellular_senescence 5 2 markers
Stem cell exhaustion stem_cell_exhaustion 3 —
Altered intercellular communication (folded into inflammaging) — —
Chronic inflammation inflammaging 4 —
Dysbiosis gut_dysbiosis 4 —

Adjacent modules also exist: il11_erk_ampk_mtor_aging, senescence_tumor_suppression, photoaging, cytokine_storm_hyperinflammation.

"Altered intercellular communication" is not a standalone module; it appears as the inflammaging node Systemic Propagation via Altered Intercellular Communication. That is a defensible lump — the hallmark's aging-relevant content is largely the inflammatory secretome — but it means a curator looking for a conformance target by hallmark name will not find one, and it should be stated in the inflammaging description if it is not already.

The one worked example

cellular_senescence carries two markers, and they are curated correctly — worth reading before adding any others:

  • p16INK4a → NCIT:C129948 (CDKN2A Gene Product), presence: INCREASED, with a readouts: link to Senescence-Associated Cell Cycle Arrest (relationship: READOUT_OF, direction: POSITIVE, endpoint_context: PROGNOSTIC), each layer separately evidenced.
  • SA-β-galactosidase → NCIT:C107438 (Beta-Galactosidase).

This is the template. The readouts: link is what makes a marker part of the pathograph rather than a disconnected list entry — the same rule that governs modeled_mechanisms and influences_mechanisms.

Conformance uptake is thin and inverted

Diseases conforming to each aging module:

loss_of_proteostasis        11      gut_dysbiosis                3
mitochondrial_dysfunction   10      deregulated_nutrient_sensing 3
genomic_instability_aging    9      stem_cell_exhaustion         1
disabled_macroautophagy      9      inflammaging                 1
epigenetic_alterations       6
cellular_senescence          6
telomere_attrition           4

inflammaging at 1 conformer is the anomaly. It is the hallmark with the most mature clinical biomarker panel (IL-6, hsCRP, TNF-α) and the widest disease reach, and it is nearly unused.

The progeroid entries are the obvious pilot

Werner_Syndrome conforms to four aging hallmarks (cellular_senescence, genomic_instability_aging, mitochondrial_dysfunction, telomere_attrition) and Hutchinson-Gilford_Progeria_Syndrome to six. Both carry zero biochemical markers. These are the entries where a segmental-progeroid biomarker panel would be least speculative and most informative, and they already have the pathograph nodes to hang readouts on. Nestor-Guillermo_progeria_syndrome has no conforms_to at all and should get one.

What the NIA/BAC landscape says

The Biomarkers of Aging–NIA Joint Symposium 2024 (PMID:40525821, held 12 Sep 2024, published 2025) and the Consortium's translation papers converge on a few points that bear directly on what dismech should curate.

There is no gold standard, and that is the stated problem. The symposium report is explicit that there is "no gold-standard measurement of biological aging … nor consensus about what one should be," and that "systematic validation of biomarkers of aging for clinical use has remained elusive." A knowledge base should therefore record aging biomarkers with their context of use attached, not as facts about biological age. dismech's endpoint_context enum already forces this, and its CANDIDATE_SURROGATE value is the honest setting for nearly every aging biomarker today.

Biomarker classes in play. Epigenetic clocks (pan-mammalian, DunedinPACE, PRC2-AgeIndex, and foundation-model clocks like CpGPT/MethylGPT); plasma-proteomic organ-specific clocks; metabolomic markers of mitochondrial function (glycerophospholipids) predicting cognitive and mobility decline; senescent-cell burden from blood; and functional proxies (gait speed, grip strength, healthspan as time to first chronic condition).

Organ-specific and individual-level heterogeneity is the frontier. Gladyshev's point that "aging within individuals may likewise not occur at uniform rates," and Barzilai's proteome findings varying by genetic heritage and sex, argue against a single whole-organism marker and in favour of markers attached to specific mechanisms — which is what a pathograph node is.

Evidence triangulation. Belsky's framing — predictive modelling, in vitro mechanistic experiments, and intervention-response — maps onto dismech's existing evidence_source axis (HUMAN_CLINICAL / IN_VITRO / MODEL_ORGANISM), and the existing check-snippet-grading gate already prevents the same quote being re-graded across those categories.

Also worth tracking: the FAST initiative (mining biospecimens from completed trials of metformin, SGLT-2 inhibitors, GLP-1 agonists, bisphosphonates) and the Hevolution Alliance for Aging Biomarkers, both of which will produce citable intervention-response biomarker data over the next cycle; and NIA's Fifth Geroscience Summit, Revisiting the Geroscience Hypothesis — Focus on Health, 2–3 December 2026.

Clinical markers vs model-system markers, and how to reconcile them

This is the part where dismech has an actual structural answer, so it is worth being precise about the problem first.

The problem

AFAR's criteria for a "true" biomarker of aging include applicability in both humans and model organisms. That criterion systematically favours molecular and subcellular markers and disfavours organism-specific ones. The clean illustration is the FRIGHT clock, built on mouse frailty indices: it is a good mouse biomarker and it does not translate, because the measurement itself has no human counterpart. The inverse holds too — gait speed and grip strength are excellent human markers with no faithful mouse equivalent. Meanwhile epigenetic clocks do cross species (pan-mammalian clocks, EnsembleAge HumanMouse), which is exactly why they dominate the field.

A useful worked case sits in the inflammaging block added alongside this report: the IgG N-glycome is reported in healthy old people, centenarians and their offspring and in calorie-restricted mice (PMID:22353383). That is a marker satisfying AFAR's cross-species criterion on its face — and it is still not a clean case, because IgG-G0 is itself pro-inflammatory, so a mouse-to-human concordance in the marker is partly a concordance in the mechanism. Cross-species agreement in a marker that is also a mechanism is weaker evidence of a shared clock than it first appears.

So "reconciliation" is not one problem but three:

  1. Same marker, both species (DNAm age, p16, IL-6) — needs a concordance claim.
  2. Species-specific marker of a shared mechanism (mouse frailty index vs human SPPB gait speed) — needs a mechanism-level bridge, not a marker-level one.
  3. Marker measurable in only one system (SA-β-gal in tissue/NAMs; organ-specific plasma proteomic clocks in humans) — needs to be recorded as such, not silently generalised.

How the existing schema handles all three

The join is the pathophysiology node, not the marker.

Side Where it lives Class Key slots
Clinical biochemical[].readouts[] BiomarkerReadout target, relationship, direction, endpoint_context, regulatory_endpoint_refs
Animal animal_models[].modeled_mechanisms[].readouts[] ExperimentalReadout target, biomarker_term, direction, assays
NAM experimental_models[].modeled_mechanisms[].readouts[] ExperimentalReadout same

Both readout classes carry biomarker_term and both carry a target that must name a node in the same entry. That gives the three cases clean, distinct representations:

  • Case 1 — the same biomarker_term appears on a clinical BiomarkerReadout and a model-side ExperimentalReadout pointing at the same target node. Concordance is then a query, not an assertion, and the directions can be compared (BiomarkerReadoutDirectionEnum POSITIVE vs ModelReadoutDirectionEnum INCREASED).
  • Case 2 — different biomarker_term values, same target node. The mechanism bridges what the markers cannot. This is precisely what the FRIGHT-vs-SPPB case needs, and it is why attaching markers to hallmark nodes beats maintaining a flat cross-species marker table.
  • Case 3 — a marker that exists on only one side simply has no counterpart readout, and ModelMechanismLink.fidelity plus limitations carry the translational caveat. FAILS_TO_RECAPITULATE is available for the genuinely negative result and requires both limitations and evidence.

One thing to be careful about. A model-side readout tempts a curator to grade its evidence HUMAN_CLINICAL because the claim is about human aging. It is not: evidence_source classifies the cited publication. A mouse epigenetic-clock result is MODEL_ORGANISM however translatable the clock is. just check-snippet-grading enforces this per quoted sentence and will catch the drift.

The endpoint layer, and why it is the bigger gap

Everything above treats aging biomarkers as the object. But a surrogate is only a surrogate for something, and dismech currently represents the "something" not at all. The CANDIDATE_SURROGATE values curated in the two finished modules point nowhere: kb/surrogate_endpoints/fda_surrogate_endpoints.yaml holds 225 FDA rows and not one of them is an aging endpoint.

The field has converged on three endpoint families, each with a flagship trial:

Endpoint family Definition as operationalized Flagship
Disability-free survival Composite of death, dementia, and persistent physical disability ASPREE (PMID:30221596)
Multimorbidity Accumulation of new age-related diseases TAME (PMID:30151729)
Deficit-accumulation frailty 36-item frailty index, fit / less fit / frail SPRINT (PMID:26755682), Look AHEAD

The TAME Biomarkers Workgroup report is the document that ties this layer to the biomarker layer, and it should be read before any further biomarker curation. It states the purpose of a geroscience biomarker exactly: they exist to show an intervention is engaging aging biology before enough clinical events accrue to power the trial. Its panel — IL-6, TNFα-receptor I or II, CRP, GDF15, insulin, IGF1, cystatin C, NT-proBNP, HbA1c — is a third independent list to set against the NIA symposium's and the 2025 Delphi consensus'. Where three such lists agree, that is about as close to a validated core as this field currently gets: IL-6, CRP and GDF-15 appear on all three.

Two things follow that changed work already done:

  • TAME selected the TNF receptors, not TNF-α. The soluble receptors are the more stable analytes. The inflammaging TNF entry is curated against the ligand and now records that sTNFR-I/II should supersede or accompany it.
  • The unqualified step is the one that matters. None of these markers has been shown to satisfy the condition that makes a surrogate legitimate — that a treatment-induced change in the marker predicts a change in the endpoint. Association with the endpoint, which is what the curated evidence supports, is a weaker claim and is not sufficient.

TAME's own summary of the gap is the sharpest statement of it in the literature, and it is phrased in exactly dismech's terms: the work "revealed the scarcity of well-vetted biomarkers for human studies that reflect underlying biologic aging hallmarks." A hallmark module's biochemical: block is precisely that object.

What this implies for dismech. Three of the endpoint families are clinical outcome assessments, not analytes, so none of them fits Biochemical. Frailty index and disability-free survival are composites over deficits and events; multimorbidity is a count of incident diseases. This is the same structural problem as composite biomarkers and epigenetic clocks, arriving from the outcome side rather than the measurement side, and it is now recorded as a knowledge gap in inflammaging. Whether dismech should carry a clinical-outcome-assessment class, extend SurrogateEndpoint beyond the FDA table, or represent these as Groupings is an open design question and should go to the decision register rather than being settled by a curator mid-tranche.

Ontology gaps

BiomarkerTerm validates term existence against NCIT with no hierarchy constraint, so binding is permissive. Availability, checked against NCIT via OLS:

Already available and mostly already cached — NCIT:C129765 Telomere Length; NCIT:C181406 GDF-15 Measurement; NCIT:C74834 Interleukin 6 Measurement; NCIT:C157114 High Sensitivity C-Reactive Protein Measurement; NCIT:C127624 Klotho Protein Measurement; NCIT:C88043 Neurofilament Light Polypeptide; NCIT:C20535 Tumor Necrosis Factor; NCIT:C17783 Cyclin-Dependent Kinase Inhibitor 1 (p21); NCIT:C107438 Beta-Galactosidase; NCIT:C129948 CDKN2A Gene Product; NCIT:C165222 DNA Methylation Array; NCIT:C63328 DNA Methylation Analysis; NCIT:C129903 Global DNA Methylation Profile. Functional markers are covered too: NCIT:C139210 Grip Strength, NCIT:C181968 SPPB Gait Speed Test, NCIT:C185373 Clinical Frailty Scale.

The real gap: there is no NCIT term for an epigenetic clock or for biological age. Searching NCIT for "Epigenetic Clock" returns nothing, and "Biological Age" returns only Biological Agent and its descendants. The single most important class of aging biomarker is unbindable.

Per the dismech-terms rule that no term beats a bad one, the answer is not to bind a clock to NCIT:C17961 (DNA Methylation) or to NCIT:C16269 (Aging) — neither is the measurement. Two defensible options:

  1. Bind the assay (NCIT:C165222 DNA Methylation Array) and carry the clock identity in a free-text preferred_term — e.g. preferred_term: DunedinPACE pace-of-aging estimate. This is the documented pattern for a preferred_term more specific than the best available term, and it validates today.
  2. Leave term: off, record in notes that NCIT was searched and what was missing, and submit an NCIT term request. Given how central these measures are, a request is probably worth making regardless of which option is used in the interim.

Note that a composite clock is not really a biomarker in the Biochemical sense at all — it is a model output over many features. If clocks become a substantial part of the KB, whether they belong in biochemical: or in computational_models: is a genuine open design question, not a curation detail.

Recommendations

Tier 1 — close the module biomarker gap. Add biochemical: blocks with readouts: links to the ten bare hallmark modules, following the cellular_senescence pattern. Highest value first, because these are the markers with real clinical evidence behind them:

Module Candidate markers Node to read out
inflammaging IL-6, hsCRP, TNF-α Chronic Low-Grade Sterile Inflammation
telomere_attrition Leukocyte telomere length Progressive Telomere Attrition
epigenetic_alterations DNAm age / pace-of-aging (see ontology gap) Age-Associated Epigenetic Drift
mitochondrial_dysfunction GDF-15, lactate, glycerophospholipids Bioenergetic Decline and Oxidative Stress
deregulated_nutrient_sensing IGF-1, FGF21 mTORC1 Hyperactivation; Attenuated FGF21 Response

inflammaging first: best-evidenced markers, worst conformance uptake (1 disease), and adding markers gives curators a reason to conform to it.

Status update. inflammaging is done — a five-marker block (IL-6, hsCRP, TNF, cf-mtDNA, IgG-G0 N-glycans) with BiomarkerReadout links was added on the same branch as this report, grounded in Franceschi 2018 (PMID:30046148), Harris 1999 (PMID:10335721), Pinti 2014 (PMID:24470107) and Dall'Olio 2013 (PMID:22353383). It is now the second module in the KB with a biochemical: block and the worked example for the rest of Tier 1. Two patterns established there are worth copying: IL-6 carries two readouts against different nodes to show one marker serving distinct contexts of use (MONITORING on the mechanism, PROGNOSTIC on the outcome), and cf-mtDNA/IgG-G0 are annotated as both marker and mechanism, which is common in aging biology and which the READOUT_OF relationship deliberately does not assert away.

mitochondrial_dysfunction is also done — GDF-15, FGF-21 and blood mtDNA copy number, plus a knowledge gap on the disease-vs-aging scope limit.

telomere_attrition, epigenetic_alterations and deregulated_nutrient_sensing followed, which completes this KB's coverage of the TAME panel and most of the 2025 Delphi consensus list. Notable cases from that tranche:

  • IGF-1 has two directions, and they point opposite ways. Against the node it reports, it is ordinary: higher IGF-1 means more anabolic signalling, so both readouts are POSITIVE. Against outcome, it inverts — lower IGF-1 predicts longer survival, in females only (PMID:24618355). The first draft encoded the outcome inversion as direction: NEGATIVE, which was wrong: direction describes the linked event, not the outcome, and it contradicted the sibling readout to mTORC1 Hyperactivation, the direct downstream of that node. Caught in review and corrected; the longevity inversion now lives in prose where it belongs. Worth knowing about because the biology genuinely is inverted, which is what makes the wrong slot so tempting.
  • HbA1c and hsCRP carry REFUTE items recording that they failed consensus agreement on predicting biological age. The limitation travels with the marker.
  • Telomere length's gap is about measurement, not biology. The population association is settled over 121,749 individuals; whether one person's measurement means anything is not. That is an assay-standardization problem.
  • The epigenetic clock entry documents the NCIT gap in place — assay bound, clock identity in preferred_term, with NCIT:C17961 and NCIT:C16269 considered and rejected as not-the-measurement.

Eight KNOWLEDGE_GAP discussions now exist across five modules, and the recurring one is now recorded in the decision register (§12, Computed indices and composite endpoints in aging biology).

A correction to how this report first framed it. Earlier drafts called composite biomarkers, clinical endpoints, epigenetic clocks and the frailty index "one absence." Working it through for the register, they are two, and merging them would produce the wrong schema:

  1. Computed indices over measurements — epigenetic clock, composite biomarker panel, frailty index. Fitted estimators whose output is a number. Not analytes, so not Biochemical.
  2. Composite clinical outcome endpoints — disability-free survival (a time-to-event composite), multimorbidity (a count of incident diseases). Outcomes, not measurements, and the thing a candidate surrogate is surrogate for.

The frailty index straddles both, being a computed index used as an outcome. That is what makes it look like one problem from a distance.

One loop closed: PhenoAge's transcriptional analysis ties epigenetic age acceleration to pro-inflammatory and interferon programmes, which is the DNAm-age/inflammaging link Franceschi called uninvestigated in 2018 and which this report declined to curate as a gap. Declining it was right.

The sweep is complete. All eleven hallmark modules are curated — ten with biomarker blocks, one (disabled_macroautophagy) deliberately without. The tail behaved as predicted, and the negative results are the most useful part of it:

  • disabled_macroautophagy gets no biomarkers, and that is the finding. It is the only hallmark where the search returns nothing usable, because autophagy is a flux, not a level: LC3-II and p62 are steady-state abundances of proteins the process consumes, so a raised value is equally consistent with more autophagosome formation and with blocked degradation. Telling them apart needs a lysosomal clamp or a tandem reporter, and the cited methods review scopes those to cell culture and animal models. Recorded with an empty-section anchor and an explicit instruction not to import a single LC3 or p62 reading, which would have no defensible direction.
  • CHIP is curated twice, deliberately. In genomic_instability_aging it reports accumulated somatic mutation; in stem_cell_exhaustion it reports hematopoietic stem-cell clonal dominance. Different claims about one measurement, so the readouts and interpretations differ and each entry names the other. It is also the KB's cleanest cross-species case: human association plus mouse bone-marrow-transplant evidence that Tet2-deficient marrow enlarges atherosclerotic lesions.
  • Alpha diversity does not survive the data. The gut_dysbiosis gap carries a REFUTE item: in centenarians followed to death, alpha diversity did not differ before death while ten individual species did. NCIT:C68564 (Microbiome) exists and would validate, but binding it would assert that the microbiome is the biomarker without saying what is measured — so it is deliberately left unbound.
  • Three gaps that look alike are three different kinds. loss_of_proteostasis — marker exists, scoped to one protein in one disease. stem_cell_exhaustion — marker exists, scoped to one compartment for a structural reason (blood is the only tissue where a clone is sequenceable without biopsy). disabled_macroautophagy — the quantity is not measurable by any accessible means, so no further curation closes it.

Final tally, counted from the tree rather than from memory: eleven hallmark modules, 20 curated biomarkers, and 13 KNOWLEDGE_GAP discussions across 10 of the 11 (two of those gaps pre-date this work, in cellular_senescence and deregulated_nutrient_sensing). genomic_instability_aging is the only module with a marker and no open gap, because CHIP is the one aging biomarker here with human, prognostic, and interventional cross-species evidence all present.

A correction to this report's framing

Curating the first two modules surfaced a 2025 Delphi expert consensus statement (PMID:39708300, J Gerontol A) that postdates the NIA symposium summarized above and is more directly useful than anything else cited here. It reached 70–98% agreement on 14 biomarkers: IGF-1, GDF-15, hsCRP, IL-6, muscle mass, muscle strength, grip strength, Timed-Up-and-Go, gait speed, standing balance, frailty index, cognitive health, blood pressure, and DNA methylation/epigenetic clocks. Three findings in it bear directly on the recommendations above:

  • hsCRP, TNF-α, HbA1c and blood pressure did not reach agreement that they predict biological age better than chronological age. They are accepted as measures of state, and unresolved as measures of rate. That is a sharper claim than "unvalidated" and it is now recorded as a gap on the markers themselves.
  • Physiological markers dominate the consensus list — grip strength had the highest agreement of all 14, at 98%, against IGF-1's lowest at 70%. The panel notes this may reflect its own composition, but it cuts against this report's molecular emphasis, and it is a problem for cross-species work, since grip strength and TUG are exactly the human-only measures that do not translate to mice.
  • Composites are preferred and no consensus composite exists — the paper calls this a research priority in as many words. This is the same structural question flagged under ontology gaps: a composite is a model over features, not an analyte.

A further process note. The gap this section originally intended to record was Franceschi 2018's statement that the DNAm-age/inflammaging relationship "has not been investigated." A PubMed check found 51 papers on epigenetic age acceleration and inflammaging, so that claim is eight years stale and was deliberately not curated. Recording a knowledge gap from a review's own framing without checking whether the field has since closed it is a failure mode worth naming.

Tier 2 — the progeroid pilot. Give Werner_Syndrome and Hutchinson-Gilford_Progeria_Syndrome biochemical blocks with readouts to the hallmark nodes they already conform to, and add conforms_to to Nestor-Guillermo_progeria_syndrome. This exercises the whole pattern on entries where the mechanism scaffolding is already in place.

Tier 3 — the cross-species demonstration. Pick one node — Senescent Cell Accumulation is the natural choice — and curate a clinical BiomarkerReadout, an animal-model ExperimentalReadout, and a NAM ExperimentalReadout all pointing at it, with honest fidelity and limitations. That makes the reconciliation pattern concrete and reviewable rather than theoretical, and it is the thing to point at when the question comes up again.

Tier 4 — the design question, now filed. The representation question is recorded in the decision register as Computed indices and composite endpoints in aging biology (docs/explanation/design-decisions.md §12), with the two gaps separated, the candidate shapes laid out, and the NCIT sub-gap noted. It is undecided, not proposed — three candidate shapes are named and none is worked through, because picking one is a maintainer call, not a curator's.

Still outstanding and not filed: an NCIT term request for epigenetic clock / biological age, and a Biomarkers_of_Aging grouping over the progeroid and age-related entries. Neither blocks further module curation.

Caveats

Module and conformance counts are from the working tree at the date above and will drift. The NIA/BAC summary is drawn from the published symposium report and Consortium translation papers, not from primary attendance. No marker in the recommendation tables has had its evidence curated or its snippet verified — they are leads with plausible NCIT bindings, and each still needs the normal just fetch-reference / just count-verified-snippets loop before it goes into an entry.

Sources