Abstract

Most knowledge of how a molecular lesion or exposure produces a human disease phenotype remains distributed across narrative reviews, primary studies, clinical resources, and model-system reports. These sources are individually valuable but do not provide a shared, computable account of the causal steps connecting aetiology, pathophysiology, clinical outcomes, and treatment. DisMech is an open, mechanism-centred knowledge resource that represents these accounts as pathographs: typed, directed graphs whose nodes and edges are grounded in community ontologies and linked to provenance-bearing evidence. Pathographs preserve competing mechanistic hypotheses, identify explicit knowledge gaps and human/model mismatches, and record experiments proposed to resolve them. DisMech is maintained through a hybrid curation workflow in which human curators direct language-model agents while schema, ontology, and reference validators provide reproducible quality-control gates. Here we describe the resource, quantify its content and evidentiary structure, and evaluate whether access to pathographs improves pathophysiological reasoning. We further show how a disease mechanism graph can connect new approach methodology readouts to candidate surrogate endpoints and unresolved organ-specific outcomes. DisMech provides an inspectable substrate for mechanistic diagnosis, evidence-gap discovery, translational model selection, and auditable biomedical AI.

Introduction

Most of what is known about human disease mechanism is written as prose. Textbooks, review articles, and primary studies describe how molecular events propagate through cells and tissues to produce signs, symptoms, and clinical course. This literature is authoritative, but it is not directly computable. A researcher who wants to compare a mechanism across diseases, determine which phenotype is downstream of a treatment target, or identify where a model system fails to explain a human outcome must repeatedly reconstruct causal models that domain experts already hold implicitly.

Existing biomedical resources solve important adjacent problems. OMIM and the Human Phenotype Ontology capture disease–phenotype associations; ClinVar and ClinGen represent variant and gene–disease evidence; Reactome and KEGG represent biological pathways; and drug resources connect compounds to targets and indications [CITE]. What remains uncommon is a disease-centred representation of the complete explanatory chain:

aetiology → molecular dysfunction → cellular process → tissue pathology → clinical phenotype → treatment response

The missing intermediates matter. Diseases that resemble one another phenotypically may arise through different causal processes and require different interventions. Conversely, diseases assigned to different clinical categories may share a pathological process that creates a common therapeutic vulnerability. A mechanistic representation must also preserve uncertainty: several causal accounts may coexist, evidence may support different branches, and the most useful statement may be that a particular edge remains unknown.

This need is becoming more acute as new approach methodologies (NAMs), including patient-derived cells, organoids, organs-on-chips, and high-throughput perturbation systems, are used to study disease and support drug development. A NAM readout is not self-interpreting. Its translational meaning depends on which part of a human disease mechanism it recapitulates, which clinical outcome lies downstream, and which organ-specific branches it does not test. The same problem arises for surrogate endpoints: association between a biomarker and an outcome is not equivalent to a mechanistic account of why treatment-induced change in the biomarker should predict clinical benefit.

Large language models and agentic systems make it possible to assemble structured biomedical content at a scale that was previously impractical, but they also introduce fabricated citations, invented identifiers, semantic misgrounding, and confidently stated unsupported claims. The relevant question is therefore not whether an agent can generate a disease summary, but whether agent-assisted curation can produce a resource whose claims, uncertainties, and provenance remain inspectable.

We address these linked problems with DisMech, the Disorder Mechanisms Knowledge Base. Each DisMech entry is a structured disease model containing a pathograph, phenotypes, genetic and environmental factors, biochemical findings, treatments, experimental models, datasets, and evidence. Reusable mechanism modules capture conserved pathological motifs across diseases. Mechanistic hypotheses label alternative or superimposed causal subgraphs, while discussion objects record knowledge gaps, controversies, and human/model mismatches attached to particular nodes or edges. The resource is openly versioned and rendered for both human browsing and computational use.

Here we present DisMech primarily as a biomedical data resource. We describe its data model and curation governance, characterize its current content and evidence structure, examine its representation of mechanistic uncertainty, and define an evaluation of pathophysiological reasoning with and without pathograph context. We use the relationship between Fabry disease mechanisms, surrogate endpoints, and patient-derived cellular models as a translational case study. A companion manuscript describes the agentic curation and validation architecture in depth.

Results

Pathographs make disease explanations computable

A pathograph is a typed, directed graph whose nodes represent states or processes on a disease-causation chain and whose edges represent established or hypothesized causal relationships. DisMech stores the graph alongside the other parts of a disease record rather than treating it as an isolated pathway diagram. This allows a mechanism node to connect directly to a phenotype, treatment, biomarker, experimental model, or explicit knowledge gap.

Figure 1 | DisMech and the pathograph data model. The central Disease object connects genetic and environmental aetiology to pathophysiology, biochemical states, phenotypes, treatments, surrogate endpoints, datasets, and experimental models. Typed causal edges form the pathograph. Evidence items, mechanistic hypotheses, and discussions attach to the claims they qualify. [FIGURE TO GENERATE FROM THE CURRENT SCHEMA]

Table 1 | Core pathograph objects and ontology grounding.

Object	Example	Principal standards
Disease	Fabry disease	MONDO, Orphanet
Genetic factor	GLA loss of function	HGNC, ClinGen, GENO
Environmental factor	ultraviolet exposure	ECTO or other mapped vocabularies
Pathophysiology	lysosomal glycosphingolipid accumulation	GO, CL, UBERON
Biochemical state	elevated phenylalanine	ChEBI, LOINC
Phenotype	renal insufficiency	HPO
Treatment	enzyme replacement therapy	MAXO, NCIT, ChEBI
Experimental model	patient-derived iPSC cardiomyocyte	NAMO and source dataset
Surrogate endpoint	renal peritubular-capillary GL-3 clearance	FDA source table and assay metadata
Knowledge gap	uncertain cardiac surrogacy	Discussion attached to a graph branch

Causal edges can carry evidence specific to the source–target relationship, while nodes carry evidence for the state or process itself. This distinction is important: evidence that a process occurs in a disease does not automatically establish that it causes a particular downstream phenotype. The current resource has substantially higher node-level than edge-specific evidence coverage, making edge-level provenance a measurable curation target rather than an assumed property of every pathograph.

DisMech spans diseases, reusable mechanisms, and evidence types

The publication analysis will be run against a frozen DisMech release. A preliminary inventory of the working tree on 29 July 2026 found the following content; these numbers are included here to define the analysis, not as final publication values.

Table 2 | Preliminary working-snapshot inventory.

Object	Count
Disorder files	1,635
Mechanism modules	118
Groupings	48
Pathophysiology nodes, disorders and modules	9,114
Causal edges, disorders and modules	17,912
Phenotypes, disorders and modules	17,314
Treatments, disorders and modules	6,372
Evidence items, disorders and modules	79,862
Mechanistic hypotheses, disorders and modules	360
Discussions, disorders and modules	656
Reference-cache records	31,623

Nearly all disorder files contain pathophysiology and phenotype content, but the resource is intentionally heterogeneous in depth. The frozen-release analysis will report distributions rather than totals alone: nodes and edges per disease, connected components, cross-scale transitions, ontology coverage, evidence-source composition, treatment-to-mechanism links, and coverage by disease class. It will also report explicit denominators such as the relevant MONDO or Orphanet disease space.

Figure 2 | Scope and depth of DisMech. Proposed panels: disorders by clinical/mechanistic class; distributions of pathograph nodes and edges per disease; evidence items by source type; edge-specific evidence coverage; and mechanism-module conformance. [ANALYSIS AND FIGURE REQUIRED]

Mechanism modules expose shared pathological motifs

Mechanism modules represent conserved motifs such as fibrosis, amyloid formation, thrombogenesis, lysosomal substrate accumulation, immune checkpoint blockade, and cardiac-ion-channel dysfunction. A module has the same structural form as a disease entry. Disease-specific nodes declare conforms_to relationships to module nodes while retaining organ-, cell-, gene-, and disease-specific detail.

Conformance is not inheritance. A hepatic stellate-cell activation node and a pulmonary fibroblast activation node may both conform to a generic mesenchymal-cell-activation step in the fibrotic-response module, but the disease records duplicate and specialize the relevant content. This design allows module consistency to be checked without erasing clinically meaningful context.

The frozen-release analysis will quantify how often modules are reused, whether conforming disease subgraphs preserve expected processes and edges, and which high-frequency motifs remain unrepresented. This converts mechanism modules from illustrative examples into an evaluated cross-disease layer.

DisMech represents uncertainty rather than collapsing it

Pathophysiological literature rarely supports a single uncontested graph. DisMech represents four complementary forms of epistemic qualification.

First, mechanistic hypotheses organize edges into canonical, alternative, emerging, or deprecated explanatory models. Several hypotheses may share upstream nodes and diverge only at a contested causal step. Second, evidence items classify whether a source supports, partially supports, refutes, or fails to provide evidence for a claim. Third, evidence-source typing distinguishes human clinical, model-organism, in vitro, computational, and other evidence. Fourth, discussion objects record open questions, controversies, interpretations, knowledge gaps, and human/model mismatches.

This representation makes uncertainty queryable. A user can ask which edges rest only on model-system evidence, where two hypotheses diverge, or which clinical outcome lacks a mechanistic bridge from an experimentally measured biomarker.

Knowledge gaps become research objects

A knowledge gap in DisMech is not free-floating prose. It has a stable identifier, a prompt, a lifecycle status, a rationale, and an attaches_to pointer identifying the disease object, node, or edge to which the gap applies. It may also include proposed experiments and supporting evidence. Human/model mismatch is represented separately from general absence of evidence: some evidence exists, but its fidelity to human biology is the unresolved question.

The preliminary working snapshot contains 447 KNOWLEDGE_GAP discussions, 112 HUMAN_MODEL_MISMATCH discussions, and 375 proposed experiments. The publication analysis will classify these gaps by biological scale, evidence source, disease class, graph position, and proposed experimental modality. It will also distinguish unresolved gaps from resolved or archived discussions and test whether automated literature and knowledge-gap scans preferentially identify particular gap classes.

Figure 3 | The anatomy of a mechanistic knowledge gap. Proposed panels: gap typology; graph positions at which gaps occur; evidence-source deficits; human/model mismatches by tissue or cell type; and proposed experiments by modality. [ANALYSIS AND FIGURE REQUIRED]

Fabry disease connects mechanism, surrogate endpoints, and NAMs

Fabry disease illustrates why surrogate-endpoint reasoning requires an explicit disease mechanism. Pathogenic GLA variants reduce alpha-galactosidase A activity, producing lysosomal accumulation of globotriaosylceramide (GL-3) and related glycosphingolipids. That upstream storage process branches into renal, cardiac, vascular, and neurological pathology.

The FDA surrogate-endpoint table records clearance of GL-3 from renal peritubular capillaries as an endpoint used in support of drug approval [CITE FDA SOURCE]. The pathograph places this measurement on one renal branch of a multi-organ mechanism. It therefore makes two propositions inspectable: the endpoint is mechanistically proximal to enzyme replacement and substrate clearance, but it is not a direct measurement of podocyte, cardiomyocyte, or dorsal-root-ganglion outcomes. A response in renal capillary endothelium cannot simply be assumed to establish benefit in each parallel tissue branch.

This localization identifies a concrete knowledge gap: which treatment- responsive cellular readout best bridges substrate clearance to cardiac outcomes? Patient-derived iPSCs differentiated into cardiomyocytes and other relevant lineages could test enzyme uptake, GL-3 clearance, electrophysiological effects, hypertrophic response, and drug response in cells carrying the same patient genotype. The experiment would not by itself validate a clinical surrogate, but it would supply evidence on the presently weak mechanistic bridge.

Figure 4 | Fabry disease mechanism-to-endpoint map. GLA deficiency and GL-3 accumulation branch to renal, cardiac, and neurological outcomes. The FDA renal peritubular-capillary endpoint and candidate iPSC cardiomyocyte readouts are localized to the branches they test. Unsupported extrapolations are shown as explicit gaps. [FIGURE AND PRIMARY EVIDENCE REQUIRED]

Pathographs provide a substrate for pathophysiological reasoning

The central evaluation asks whether explicit pathograph context improves reasoning rather than merely supplying more text. The evaluation set will sample diseases and mechanism modules across genetic, complex, infectious, neoplastic, and environmental categories. Each item will be derived from curated graph structure but withheld from the model context used for the test.

Tasks will include:

Explain why a phenotype follows from a molecular lesion.
Distinguish two diseases with similar phenotype profiles but different causal mechanisms.
Identify a patient feature not explained by a candidate pathograph.
Predict which outcomes should respond to perturbation of a treatment target.
Distinguish canonical, alternative, and emerging causal accounts.
Identify where human evidence ends and model-system extrapolation begins.
Select an experiment that addresses an attached knowledge gap.
Judge whether a biomarker is upstream of, parallel to, or disconnected from a clinical outcome.

The same reasoner will be tested with no resource context, with conventional disease-summary context, and with a structured DisMech pathograph and its evidence. Blinded domain experts will score biological correctness, faithfulness to the supplied path, unsupported-claim rate, recognition of uncertainty, and usefulness of the explanation. Deterministic graph-traversal questions will provide a separate machine-scored subset.

Table 3 | Pathophysiological-reasoning evaluation. [REPORT TASK COUNTS, DISEASE STRATA, HUMAN RATERS, AGREEMENT, MODEL CONDITIONS, EFFECT SIZES, AND CONFIDENCE INTERVALS]

Figure 5 | Effect of pathograph context on mechanistic reasoning. [RESULTS REQUIRED; DO NOT SUBSTITUTE COMPLIANCE SCORES FOR THIS EVALUATION]

Additional resource applications

Mechanism-aware phenotype matching can use a pathograph to distinguish phenotypically similar diseases and to explain secondary manifestations. Any diagnostic probability emitted by such a system must be independently calibrated and is not treated here as a validated clinical score.

Treatments linked to the mechanisms they target enable transparent queries for diseases sharing a therapeutic vulnerability. Such matches are hypothesis-generating repurposing candidates, not treatment recommendations. Evaluation should measure whether known cross-indication relationships are recovered before novel candidates are emphasized.

Pathographs can also evaluate gene→program→trait relationships proposed by GWAS and perturbation studies. A proposed triple may be confirmed, partially represented, absent and therefore a curation candidate, or in conflict with the curated direction of effect. The existing autoimmune/T-cell analysis will be regenerated against the frozen release before its numerical results are included [CITE].

Methods

Resource scope and governance

DisMech models disorders and pathological processes rather than attempting to reproduce the MONDO disease hierarchy. Entry scope, disease-versus-subtype decisions, disease groupings, module conformance, ontology reuse, evidence policy, and BioLink/KGX export policy follow the project decision register. The publication will identify the frozen release and the decision-register version used for analysis.

The source of truth is a collection of YAML records validated against a LinkML schema. Changes are proposed through version-controlled branches and pull requests. Human maintainers define schema and editorial policy, review scientifically consequential changes, and adjudicate disagreements. Language-model agents may draft or revise records, but agent provenance is recorded separately from the type of scientific evidence cited.

Data model

The central Disease class contains identifiers, definitions, classifications, inheritance, pathophysiology, phenotypes, biochemical findings, genetic and environmental factors, treatments, trials, datasets, models, hypotheses, and discussions. Pathophysiology nodes connect through CausalEdge objects. Treatments connect to pathophysiology or symptomatic phenotype targets through mechanism-target objects. Experimental models connect to the graph nodes they recapitulate, perturb, or read out.

Ontology-grounded descriptors preserve both a canonical ontology label and a curator-facing preferred term. Major sources include HPO, MONDO, GO, CL, UBERON, MAXO, NCIT, ChEBI, HGNC, GENO, and LOINC. BioLink reuse is primarily an export-layer concern rather than the internal modelling principle.

Curation workflow

Candidate content may originate from primary literature, structured databases, clinical resources, or retained deep-research reports. Deep research is treated as a source of leads rather than ground truth. Before research content is curated, the disease identity is checked against authoritative identifiers and causal genes to reduce named-entity confusion. References, quotations, and ontology identifiers are independently verified.

Each curation event may record actors, tools, models, affected sections, links, outcome, and review detail in an append-only history record. The companion agentic-framework paper describes the interactive and scheduled agent workflows.

Evidence and validation

Evidence items may include an authoritative reference, title, support direction, evidence-source classification, exact source snippet, explanation, and images. Reference fidelity means that the quoted text exists in the cited source; it does not establish by itself that the quotation entails the curated claim. Scientific relevance and causal interpretation therefore remain separate review dimensions.

Validation includes schema conformance, ontology identifier and canonical- label validation, reference-snippet validation against tool-generated cache records, foreign-key and data-integrity tests, and compliance reporting. Compliance is a coverage measure and is not used as a proxy for biological correctness.

Mechanistic hypotheses and discussions

Mechanistic hypotheses carry stable group identifiers and maturity states. Causal edges opt into one or more hypothesis groups. Discussions record open questions, knowledge gaps, controversies, emerging hypotheses, interpretations, and human/model mismatches. Proposed experiments are stored on the discussion they are intended to resolve.

Mechanism modules and groupings

Mechanism modules use the same schema as disease entries and define recurring pathological motifs. Disease nodes may conform to module nodes while retaining disease-specific content. Groupings are curated unions of distinct entries with explicit membership rationales and criteria; they are not inferred ontology classes.

Frozen-release resource analysis

[TO IMPLEMENT] A committed analysis script will compute all resource counts, coverage estimates, evidence distributions, graph-topology measures, gap typologies, and module-conformance summaries from a tagged release. The script will emit machine-readable tables used directly to generate Tables 2 and the corresponding figures. Every reported percentage will include its numerator, denominator, and treatment of missing values.

Pathophysiological-reasoning evaluation

[TO IMPLEMENT] The evaluation protocol will be preregistered within the repository before results are inspected. Disease sampling, question generation, withheld information, comparison conditions, model versions, human-rater instructions, adjudication, statistical tests, and exclusion criteria will be fixed in that protocol. Human evaluation will use at least two independent raters per item and report agreement.

Interface and exports

The public site provides a faceted disorder browser, individual disease and module pages, interactive pathographs, evidence displays, embedding exploration, project views, and quality-control reporting. DisMech exports tabular data and an export-layer mapping to BioLink/KGX-compatible nodes and edges. The publication artifact will archive the release, schema, analysis outputs, and evaluation set under a persistent identifier.

Discussion

DisMech treats a disease mechanism as an inspectable scientific object rather than a narrative summary. The contribution is not simply that causal chains can be drawn. It is that a chain can be queried together with the evidence, alternative hypotheses, model-system limitations, and explicit points at which the explanation fails.

Knowledge gaps are therefore outputs of the resource rather than defects to hide. A missing human bridge, a contested causal direction, or a model that reproduces only one tissue branch can be attached to the exact place where additional evidence is needed. This creates a route from knowledge representation to experiment design, illustrated by the Fabry mechanism–surrogate–cardiac-model example.

The reasoning evaluation is essential to establish whether this additional structure changes performance. If pathograph context does not improve biological correctness, uncertainty recognition, or evidence faithfulness over a well-chosen narrative baseline, the representation may still be useful for curation and integration but its stronger clinical-AI claims would not be supported.

DisMech also provides a production setting for studying agent-assisted biocuration. That aspect is important but conceptually separable: this paper evaluates the biomedical representation and its uses, while the companion paper evaluates the curation and validation architecture.

Limitations

DisMech is incomplete and uneven in depth. Coverage of pathophysiology does not imply complete edge-level evidence, and ontology grounding does not imply that a term is sufficiently specific. Exact-snippet validation establishes source fidelity but not claim–evidence entailment. Evidence available only in full text may be absent when the reference cache contains an abstract alone.

Mechanistic graphs simplify continuous, cyclic, temporal, and context- dependent biology into a tractable representation. Edge direction and granularity reflect editorial judgement. Mechanism modules can expose inconsistency but may also impose a shared abstraction on diseases whose details differ materially.

Knowledge gaps are curated and are consequently subject to ascertainment bias. Diseases investigated by automated scans or domain projects will appear to have more explicit gaps than less-curated diseases. Proposed experiments are hypotheses, not validated study protocols.

The resource is not a clinical decision-support system. Mechanistic matching, surrogate-endpoint interpretation, and treatment-repurposing queries require independent validation in their intended contexts of use.

Data and code availability

DisMech is publicly available at https://dismech.monarchinitiative.org/. Source code and data are available at https://github.com/monarch-initiative/dismech. The final manuscript will cite a frozen release and persistent archive containing the data, schema, analysis code, evaluation set, and generated figures.

Author contributions

[TO COMPLETE USING CRediT ROLES]

Acknowledgements

We thank the LinkML, OAK, Monarch Initiative, OBO Foundry, HPO, MONDO, MAXO, Cell Ontology, Gene Ontology, UBERON, ClinGen, Orphanet, ClinicalTrials.gov, MorPhiC, NAMO, and deep-research-provider communities. Funding and contributor acknowledgements will be reconciled with the consortium author list.

Competing interests

[TO COMPLETE]

References

Moxon, S.A.T. et al. LinkML: an open data modeling framework. GigaScience 15, giaf152 (2026).
[CITE Human Phenotype Ontology]
[CITE Mondo Disease Ontology]
[CITE Gene Ontology]
[CITE Cell Ontology and UBERON]
[CITE ClinGen gene–disease validity]
[CITE Monarch Initiative]
[CITE GA4GH Phenopackets]
[CITE FDA surrogate-endpoint table and regulatory framework]
[CITE Fabry disease mechanism and organ-specific pathology]
[CITE Fabry renal GL-3 surrogate evidence]
[CITE Fabry patient-derived iPSC cardiomyocyte studies]
[CITE MorPhiC]
[CITE NAMO and relevant reporting standards]
[CITE SPIRES and related schema-guided extraction work]
[CITE pathophysiological-reasoning and clinical-AI evaluation literature]