Paper deep dive
A Semantic Model of Genetic Evidence: A Step Toward Bridging the Basic-Science-Clinic Gap
Michael Bouzinier, Dmitry Etin
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 9/7/2026, 5:23:07 AM
Summary
The paper introduces a semantic model for genetic evidence to bridge the gap between basic science and clinical practice. It defines three core classes (ScientificEvidence, EvidenceVariable, EvidenceAssertion) specialized for genetics, aligns them with FHIR Evidence, and uses a dimensional vocabulary validated by SHACL schemas. A human-AI annotation pilot on six papers demonstrated the model's feasibility for capturing fine-grained, source-anchored genetic evidence.
Entities (10)
Relation Signals (6)
GeneticEvidence → alignswith → FHIR Evidence
confidence 95% · align it structurally to FHIR Evidence with a SEPIO-anchored credibility decomposition
ScientificEvidence → specializesto → GeneticEvidence
confidence 95% · ScientificEvidence, EvidenceVariable, and EvidenceAssertion specialize to GeneticEvidence, GeneticEvidenceVariable, and GeneticEvidenceAssertion, respectively.
Conditional Activation → validatedby → SHACL
confidence 95% · conditional-activation rules are validated by a SHACL schema
Credibility → anchoredby → SEPIO
confidence 90% · The overall rating is anchored on SEPIO confidence
Human-AI annotation pilot → evaluates → GeneticEvidence
confidence 90% · we evaluate the model through a human-AI annotation pilot over six genetics papers
Mode of Inheritance → requiredwhen → Knowledge Domain contains human genetics AND Target Type is gene
confidence 90% · Mode of Inheritance is required when Knowledge Domain contains human genetics and Target Type is gene
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Scientific and clinical decision-making depends on evidence from the primary literature, but existing standards for representing that evidence (FHIR Evidence, ECO, SEPIO, and the GA4GH Genomic Knowledge Standards) are oriented toward clinical-trial workflows, evidence codes, or single-variant assertions, and do not capture the fine-grained, domain-specific structure of claims in basic and pre-clinical research. We introduce a semantic model for scientific evidence with three core classes, specialize it for genetics, align it structurally to FHIR Evidence with a SEPIO-anchored credibility decomposition, and attach a compact dimensional vocabulary whose conditional-activation rules are validated by a SHACL schema for the implemented constraints. Using clinical variant interpretation as the driving use case, we evaluate the model through a human-AI annotation pilot over six genetics papers, yielding 28 evidence items and 95 source-anchored assertions, with a workflow that keeps curator-authored reference annotations distinct from AI-drafted annotations. Treating the pilot as a feasibility study rather than a benchmark, we argue that the model is a useful increment toward trustworthy, AI-ready infrastructure for variant interpretation: a reference data model and validation schema for representing genetic evidence.
Tags
Links
- Source: https://arxiv.org/abs/2609.04509v1
- Canonical: https://arxiv.org/abs/2609.04509v1
Trouble viewing inline? Open PDF directly →
Full Text
145,315 characters extracted from source content.
Expand or collapse full text
A Semantic Model of Genetic Evidence: A Step Toward Bridging the Basic-Science–Clinic Gap Michael Bouzinier ∗1,2,3 and Dmitry Etin 3,4 1 Harvard University, Cambridge, MA, USA 2 IDEXX Laboratories, Westbrook, ME, USA 3 Forome Association, Newton, MA, USA 4 Deggendorf Institute of Technology, Germany Abstract Scientific and clinical decision-making depends on evidence from the primary literature, but existing standards for representing that evidence (FHIR Evidence, ECO, SEPIO, and the GA4GH Genomic Knowledge Standards) are oriented toward clinical-trial workflows, evidence codes, or single-variant assertions, and do not capture the fine-grained, domain-specific structure of claims in basic and pre-clinical research. We introduce a semantic model for scientific evidence with three core classes, specialize it for genetics, align it structurally to FHIR Evidence with a SEPIO-anchored credibility decomposition, and attach a compact dimensional vocabulary whose conditional-activation rules are validated by a SHACL schema for the implemented constraints. Using clinical variant interpretation as the driving use case, we evaluate the model through a human–AI annotation pilot over six genetics papers, yielding 28 evidence items and 95 source- anchored assertions, with a workflow that keeps curator-authored reference annotations distinct from AI-drafted annotations. Treating the pilot as a feasibility study rather than a benchmark, we argue that the model is a useful increment toward trustworthy, AI-ready infrastructure for variant interpretation: a reference data model and validation schema for representing genetic evidence. Keywords: genetic evidence; variant interpretation; ontology; FHIR; ECO; SEPIO; GA4GH; AI-ready data. 1 Introduction Scientific and clinical decision-making depends on evidence from heterogeneous sources, especially the primary literature. Today, this evidence is largely consumed by humans and encoded indirectly into annotation databases and local rules. Existing resources such as HL7 Fast Healthcare Interoperability Resources (FHIR) Evidence and EvidenceVariable [9], the Evidence and Conclusion Ontology (ECO) [6], the Scientific Evidence and Provenance Information Ontology (SEPIO) [11], and the Global Alliance for Genomics and Health (GA4GH) Genomic Knowledge Standards [7] advance interoperability in clinical research. However, they do not comprehensively represent the evidence types and workflows of the basic sciences. They focus mainly on clinical trials, high-level assertions, and variant representations, with ∗ Correspondence: michael_bouzinier@harvard.edu 1 arXiv:2609.04509v1 [cs.DB] 3 Sep 2026 limited support for the fine-grained, domain-specific evidence from basic and pre-clinical research needed for automated reasoning and AI-ready infrastructure. Each scientific domain demands a tailored approach for structuring and communicating evidence. For example, the genetic evidence discussed in this paper is generated in diverse contexts, from experimental model systems to large-scale population studies, that are not fully supported by current standards. To properly use evidence derived from scientific literature in clinical workflows, two principal challenges require standardization: Relevance and Specificity, that is, how precisely a genetic finding maps to an individual patient, disease context, or biological hypothesis; and Confidence, that is, whether a given piece of evidence can be robustly and safely used, ensuring defensible compliance, interpretability, and reproducibility. A well-designed schema for genetic evidence helps make literature review a semantic-parsing problem: mapping natural-language claims to a shared, source-anchored representation. Large language models are increasingly capable of this kind of parsing, but dependable automation still requires a carefully structured target language. The present work offers one such increment. Building on decades of work by the biomedical-ontology community, it takes a small step toward literature review that is materially assisted by automated parsing. Contributions. This paper makes the following contributions: •A domain-analysis-driven semantic model for genetic evidence, with core classes (ScientificEvidence, EvidenceVariable, EvidenceAssertion), genetic specializations, a compact dimensional vocabulary, and a validation schema in the W3C Shapes Constraint Language (SHACL) [16]. •An application of the model to six genetic-evidence papers spanning human, population, model-organism, and polygenic-score genetics, supported by extraction tooling. 1 •A human–AI collaborative annotation workflow, packaged as executable AI skills for annotation and review that apply the model under curator supervision. • A structural alignment of the core classes with HL7 FHIR Evidence, including a credibility representation that decomposes certainty into SEPIO-anchored facets. The dimensions are designed to align with ECO, SEPIO, IAO, and OBI; Supplementary Note SN5 provides a partial term-level mapping. •A curated term-level UMLS crosswalk of the dimensional vocabulary, in which every value is mapped with a typed relation or recorded as an argued gap. The crosswalk was produced with the open-source, model-agnosticGEM Mapping Studio, a standalone curation tool distributed as a Python package that supports the full workflow, from defining dimension axes in an empty workspace to adjudicating value-level mappings. The workflow follows a reproducible protocol with pre-registered search sweeps, structured rejection rationales, and a public decision log in the repository. The remainder of the paper reviews related standards (Section 2), defines the semantic model (Section 3), and illustrates it with two worked examples (Section 4). It then describes the annotation workflow and pilot statistics (Section 5), discusses findings, candidate extensions, limitations, and future work (Section 6), lists the released artifacts (Section 7), and concludes (Section 8). 1 https://github.com/ForomePlatform/genetic-evidence-model 2 2 Background and Relation to Existing Standards The representation of scientific and clinical evidence has been developed across clinical informatics, biomedical ontology, and genomic knowledge representation. Seven families of standards bear directly on our work: HL7 FHIR Evidence [9]; the Evidence and Conclusion Ontology (ECO) [6] and the Scientific Evidence and Provenance Information Ontology (SEPIO) [11]; the GA4GH Variant Annotation (VA) and Variation Representation (VRS) specifications [7,17]; the ACMG/AMP variant- interpretation guidelines [13]; the upper ontologies IAO and OBI [2,5]; the PROV-O provenance model [15]; and the ClinGen, MONDO, and HPO curation and terminology resources [10,12,14]. Each captures part of what a basic-science genetic publication reports, but none captures the whole: a single conditional-knockout study may report molecular, cellular, and clinical phenotypes at once, and a polygenic-score study may aggregate millions of variants into one score. A per-family review, with the specific coverage gaps that motivate our model, is given in Supplementary Note SN1. Supplementary Table ST1 presents the same analysis as a matrix comparing ten evidence patterns across the standards and our model (GEM). For each pattern, the matrix indicates whether a resource represents it natively, only in part or by extension, or not at all. It also marks patterns that lie outside a resource’s subject matter. 2.1 Relation to FHIR Evidence FHIR is an HL7 information model rather than an OBO Foundry ontology. A genetic-evidence model could, in principle, be built directly on the OBO side, but doing so first would be premature. ECO and SEPIO provide, respectively, an evidence-type vocabulary and a model of assertion provenance and qualifiers. Neither provides the variable-and-certainty content model for an evidence unit that FHIR Evidence offers and that basic-science evidence needs to support translation from bench to bedside. Aligning a genetics-only model directly to OBO before such a general content model exists would yield a narrow resource. It would neither compose cleanly with other basic-science domains nor reuse OBO’s orthogonal vocabularies, creating exactly the redundancy that OBO’s emphasis on orthogonality and reuse is intended to prevent. We therefore take the structural content model from FHIR, generalize it from clinical trials to basic science, and design the dimensions to align with ECO, SEPIO, IAO, and OBI. Supplementary Note SN5 provides a term-level crosswalk to these vocabularies, including verified identifiers and the remaining gaps. Genetics is the first specialization; extending the model to other basic-science domains and completing the term-level alignment are future work. The HL7 FHIR Evidence, EvidenceVariable, and EvidenceReport resources [9] were designed with systematic-review and guideline-development workflows in mind, where evidence concerns an intervention tested in a population. FHIR Evidence therefore has a different native framing and does not by itself represent the evidence produced by basic and pre-clinical research. Our three core classes address this gap. The core classes are general across the basic sciences rather than specific to genetics. They retain FHIR’s separation between the evidence variable and the certainty of a finding, incorporate FHIR’s separatestatisticinto EvidenceVariable as the measured value, and replace the clinical-trial content frame of population, intervention, comparator, and outcome with a genetics-specific dimen- sional frame. Genetics is the specialization developed in this paper (section 3): ScientificEvidence, EvidenceVariable, and EvidenceAssertion specialize to GeneticEvidence, GeneticEvi- denceVariable, and GeneticEvidenceAssertion, respectively. Table 1 gives the class-level correspondence. The field-level alignment, including the certainty subcomponents discussed in 3 Table 1: Class-level correspondence between the core model and FHIR Evidence R5. The alignment is architectural: the genetics-specific dimensional frame replaces FHIR’s clinical-trial frame. Field-level detail is provided in Supplementary Table ST16. This workFHIR Evidence R5Relationship ScientificEvidence EvidenceStructural sibling; GEM replaces the PICO frame with a genetics-specific dimensional frame EvidenceVariable EvidenceVariable + Evidence.statisticBroader; GEM combines the variable definition and measured value in one class EvidenceAssertion no equivalentMeta-predicate over decision rules; part of the reasoning layer described in [4], which is outside the scope of this paper Credibility Evidence.certaintyStructural parallel; both provide an overall rating with separable, genetics-specific subcomponents section 3.2, is given in Supplementary Table ST16. 2 3 A Semantic Model of Genetic Evidence The model contains three parts: a small set of core classes (section 3.1) that define what a unit of evidence is and how evidence items compose; a dimensional vocabulary (section 3.2) that describes each evidence item along roughly twenty axes; and a conditional-activation mechanism (section 3.3) that keeps the vocabulary compact by requiring dimensions only when they are semantically appropriate. 3.1 Core Classes Three abstract classes form the core of the model: ScientificEvidenceA unit of evidence, typically linked to a publication or dataset and described by a set of dimensions (see tables 2 and 3). It carries anassertionproperty containing the object-level, source-anchored claims made by the source, for example that a variant is associated with a phenotype. These are the claims recorded in the pilot (section 5). A single paper may give rise to multiple ScientificEvidence items when it reports distinct evidential claims, such as a case–control association and an in vivo functional assay of the same variant. 2 We distinguish two notions that the literature often conflates. Theassertionproperty of a ScientificEvidence item records the object-level claim made by a paper. It is what the pilot records, and its closest analogues are the GA4GH VAStatementand the SEPIOAssertion, not the FHIREvidence.assertionelement, which is a human-readable summary. The EvidenceAssertion class is retained for the reasoning layer developed in [4], where it is a meta-predicate over decision rules that use such claims. That reasoning layer is not instantiated in this paper and has no direct analogue in FHIR, VA, or SEPIO. In the released RDF serialization,GeneticEvidenceAssertioncurrently types object-level assertions so that assertion-level SHACL validation applies to them. This temporary serialization choice does not assign the meta-predicate role to the pilot assertions. The property and the class share a name; a future schema revision will separate them. 4 EvidenceVariableA named quantity extracted from one or more ScientificEvidence items, such as the gnomAD allele frequency of a variant, the LOD score in a family, or the odds ratio reported by a GWAS. Evidence variables are the terms from which predicates over Evidence are constructed. They are also described by dimensions, because the same quantity, such as an odds ratio, may carry different epistemic weight depending on how it was obtained. EvidenceAssertion A meta-predicate, that is, a predicate about a predicate. It determines whether EvidenceVariables are used appropriately in a predicate over Evidence, namely in a decision rule. This class is the model’s point of contact with the reasoning layer described in [4], where such meta-predicates assert which evidence a rule may use. This paper constructs no decision rules and therefore instantiates no EvidenceAssertion in its meta-predicate role. The class remains part of the model, but the interpretability and AI-readiness of the reasoning layer are the primary focus of [4], not of the present pilot. Theassertionproperty and the EvidenceAssertion class share a name but denote different things. The property records the object-level claims captured in the pilot, whereas the class is a meta-predicate over any decision rule that would use those claims. The property name is retained for compatibility with the released annotations and will be renamed, for example toselector, in a future schema revision. Specializing these classes for the genetic domain yields GeneticEvidence, GeneticEvi- denceVariable, and GeneticEvidenceAssertion. These specializations inherit the provenance requirement and add the dimensional annotations described next. 3.2 Dimensions Each GeneticEvidence instance is described using a dimensional vocabulary summarized in table 2, which lists the always-required dimensions, and table 3, which gives representative conditional dimensions. 3 Six dimensions are always required: Knowledge Domain, Method, Target Type, Resolution, Credibility, and Phenotype Scale. No well-formed genetic-evidence item is fully specified without them. Together, they tell a downstream reader what kind of evidence the item is, how it was obtained, what it concerns, and how credible it is in context. The dimensions differ in internal structure, and each is modeled according to its own semantics rather than forced into a single hierarchy. Knowledge Domain values are sibling framings rather than a subsumption hierarchy. A single item may carry several values at once, as when a functional study is interpreted in a clinical-genetics context; the multi-valued cardinality captures this directly. Phenotype Scale is an ordinal scale, from molecular to clinical, rather than a class hierarchy. Modeling it through subsumption would therefore misrepresent the dimension. Method, by contrast, is genuinely hierarchical, and itsis_arelations are encoded in the SHACL distribution. The chosen structure is therefore part of the definition of each dimension. Phenotype Scale and Variant Ascertainment were added during the annotation work (section 5). Phenotype Scale distinguishes molecular phenotypes, such as binding or activity, from cellular, histological, organismal, and clinical phenotypes. It is independent of the scale of the genetic target. Variant Ascertainment records how a variant entered a study. It is meaningful only for variant- level targets and is therefore a conditional dimension, required when Target Type is variant. Both dimensions are revisited in section 6. 3 The full enumeration of conditional dimensions, including those omitted from table 3 for space, is maintained in the project repository athttps://github.com/ForomePlatform/genetic-evidence-model/blob/master/schema/ dimensions.md. 5 Table 2: Always-required dimensions of the GeneticEvidence model. These apply to every evidence item regardless of context. Phenotype Scale was added to this set during the annotation work (section 5). KD abbreviates Knowledge Domain; “obj.” and “subj.” denote objective and subjective, respectively. DimensionValue typeCardinality Objectivity Example values Knowledge Domain categoricalmultipleobj.human genetics, population genet- ics, model organism, gene function Methodcategoricalmultipleobj.statistical genetics, in vivo, in vitro, clinical evidence Target Typecategoricalsingleobj.gene, variant, interval, transcript Resolutioncategoricalsingleobj.window, gene, functional element, position, variant Credibilityrating+ facets multiplesubj.overall rating+ sample_size, replication,multiple_testing, stratification_control, . . . Phenotype Scalecategoricalsingleobj. molecular, cellular, histological, organismal, clinical Credibility records the methodological soundness of an asserted finding, and nothing else. It is a structured decomposition rather than an open key-value field: one overall ordinal rating together with separately rated facets for statistical power or sample size, independent replication, multiple-testing control, ancestry or population-stratification control, and ascertainment. The facets remain separate rather than being collapsed into a single score because clinical and research users may weigh them differently. This separation is inspired by GRADE’s general practice of assessing the factors that affect certainty separately [8]. GEM does not adopt the GRADE framework or its rating rules, which were developed for bodies of clinical evidence, particularly randomized and observational studies. The GEM facets instead support credibility judgments about individual evidence items from basic and pre-clinical research. The overall rating is anchored on SEPIO confidence, defined as the degree of belief that an asserted proposition is true [11], and is implemented as a genetics-specific value set bound to that attribute, as prescribed by SEPIO profiles (Supplementary Table ST15). Two constructs are excluded from credibility by design. First, the direction or hedging of a finding belongs to the assertion itself and is represented by the width of the value range it constrains. An honestly reported null result from a rigorous study may therefore have high credibility. Second, the extent to which evidence bears on a particular question is applicability. It is captured by Relevance and Specificity and computed for each question. Combining methodological soundness, assertion content, and applicability in one field would conflate distinct constructs. The remaining dimensions are conditional: they are required, or even meaningful, only when other dimensions take particular values. For example, mode of inheritance is not meaningful for a polygenic score, and knockout type is not meaningful for a human case–control study. Rather than marking such dimensions as “optional” and relying on convention, the model assigns explicit activation conditions to each, as described in section 3.3. These conditions specify when a dimension must be present; otherwise, it may be omitted. This makes the requirement status transparent to both human curators and automated validators. 6 Table 3: Conditional dimensions of the GeneticEvidence model. Each is required only when its activation condition holds. KD, MT, and T abbreviate Knowledge Domain, Method, and Target Type, respectively;⊇denotes set membership for multi-valued dimensions. Dimensions marked ‡ have their activation condition enforced by an executable SHACL shape; for the others the condition is documented in the schema and checked at curation time, with executable shapes scoped as future work. DimensionValue type Cardinality Activation condition Mode of Inheritance ‡ categorical singleKD⊇ human genetics AND T = gene Mendelian Segregation booleansingleKD⊇ human genetics AND T = gene Penetrancecategorical singleKD⊇ human genetics AND T = variant Variant Ascertainment ‡ categorical multipleTT = variant Measurement Targetcategorical multipleKD⊇ gene function Gene Relationcategorical singleKD⊇ gene function Organism ‡ categorical multipleMT⊇ in vivo OR KD⊇ model organism Knockout Typecategorical singleKD⊇ model organism 3.3 Conditional Activation Conditional activation is the least conventional part of the model, but it keeps the vocabulary compact without sacrificing expressive power. Rather than using a fixed schema in which every GeneticEvidence item has every dimension, with many dimensions marked as optional and left empty, the model requires particular dimensions only when specified conditions over other dimensions hold. This mirrors expert reasoning. A curator reading a paper about an X-linked recessive pedigree asks about mode of inheritance because the evidence makes that question meaningful, not because a global schema requires it. Three activation conditions are currently encoded as SHACL shapes with SPARQL-based constraints: Variant Ascertainment, Mode of Inheritance, and Organism (marked in table 3). SPARQL-based constraints are the standard SHACL mechanism for expressing requirements that cannot be captured by simple cardinality constraints on a single property. Two representative shapes illustrate the pattern. For brevity, we show only thesh:sparqlbody of each and omit the outer NodeShape wrapper. The full shapes, together with the baseGeneticEvidenceShapeand the simplervariant_ascertainmentrule, are available in the project repository with an annotated walkthrough. The remaining conditional dimensions are documented in the schema and assessed during curation; executable shapes for them are future work. 4 Checking activation conditions is a closed-world question: whether an annotation contains every dimension required by the evidence item. SHACL can answer this question directly. OWL’s open-world semantics cannot express this form of completeness, which is one reason the constraints are encoded in SHACL rather than as ontology axioms. The first example encodes a conjunction across two classification axes. Mode of Inheritance is required when Knowledge Domain contains human genetics and Target Type is gene (listing 1). sh:sparql [ sh:select """ SELECT $this WHERE $this gem:knowledgeDomain gem:HUMAN_GENETICS . $this gem:targetType gem:GENE . FILTER NOT EXISTS $this gem:modeOfInheritance ?m """ ] . 4 https://github.com/ForomePlatform/genetic-evidence-model/blob/master/schema/examples.md 7 Listing 1: Compact form of the conditional-activation shape for Mode of Inheritance. The outer NodeShape wrapper is omitted. The two triple patterns inWHEREform the conjunction;FILTER NOT EXISTS identifies instances that satisfy the condition but lack the required dimension. The second example encodes a disjunction across two classification axes. Organism is required when Method contains in vivo or Knowledge Domain contains model organism (listing 2). The two disjuncts refer to different dimensions: one is a Method value and the other is a Knowledge Domain value. This illustrates that activation conditions can span the dimensional vocabulary rather than being confined to a single axis. sh:sparql [ sh:select """ SELECT $this WHERE $this gem:method gem:IN_VIVO . UNION $this gem:knowledgeDomain gem:MODEL_ORGANISM . FILTER NOT EXISTS $this gem:organism ?o """ ] . Listing 2: Compact form of the conditional-activation shape for Organism. The disjunction is expressed by a SPARQL UNION whose two branches refer to different dimensions. In each case, theSELECTreturns every GeneticEvidence instance that satisfies the activation condition but lacks the required dimension. The validator reports these instances as constraint violations with a localized diagnostic. Writing each conditional rule explicitly adds some complexity, but it improves downstream clarity: a reader of an annotation file can see why a particular dimension is required or absent, and a validator can check conformance automatically. Section 5 returns to conditional activation in the context of the annotation workflow. 4 Illustrative Examples from the Literature To demonstrate the model’s coverage and identify where it reaches its current limits, we applied it to six published genetic-evidence papers that span the major epistemic shapes encountered in literature- based variant interpretation (Supplementary Table ST14). The set covers classical Mendelian human genetics (Gupta), population-scale association (Duerr), case–control resequencing (Davis), model-organism and in vitro functional genetics (Jossin, Nelson, and the functional components of Davis), and a genome-wide polygenic score (Inouye). Four of the six papers were annotated manually from curator PDF highlights and callouts and form the curator-authored reference set for this feasibility corpus. The remaining two were drafted by an AI assistant and reviewed by a curator. All annotations and per-paper case reports are available in the project repository. 5 Duerr et al. 2006 (section 4.1) is the reference example of a classical variant-level association study. Inouye et al. 2018 (section 4.2) is a polygenic risk score study that exposes limitations of the current variant-level model and motivates a cluster of candidate extensions (section 6). 4.1 A Genome-Wide Association Study Example Duerr et al. 2006 identified IL23R as an inflammatory bowel disease gene through a genome-wide association study of non-Jewish patients with ileal Crohn’s disease, followed by replication in a 5 https://github.com/ForomePlatform/genetic-evidence-model/tree/master/case-reports 8 Jewish cohort and family-based transmission testing. Under our decomposition, the paper yields seven GeneticEvidence items (Supplementary Table ST14 and Case Report CR-DU): variant-level discovery and replication, a family-based transmission analysis, conditional fine-mapping, negative results at related pathway genes, and cited functional and mouse-model support. Six of the seven items fit the current schema without strain. The seventh, a cited anti-p40 antibody trial, concerns drug-response evidence for which the model has no fitting genetic Knowledge Domain, revealing a translational-scope gap. Review of this annotation also identified three candidate extensions that remain provisional under the two-papers rule: an explicit direction-of-effect dimension for protective and risk-conferring alleles, a construct for cross-item therapeutic synthesis, and an extension of conditional activation from gene-level to variant-level Mendelian context. Further detail is given in Supplementary Note SN6.1. 4.2 A Polygenic Risk Score Example Inouye et al. 2018 developed and externally validated a meta-genomic risk score (metaGRS) for coronary artery disease by combining three component scores and evaluating performance in UK Biobank participants. Unlike Duerr’s variant-level association study, the evidential target is a derived, whole-genome score and the primary measurements are epidemiological. Our annotation decomposes the paper into four GeneticEvidence items: predictive and strati- fying performance; independence from conventional risk factors; persistence of stratification among medicated individuals; and the headline finding that combining three component scores outperforms each component (Case Report CR-IN). All four require deliberate placeholder assignments in the current model. In particular, the score is assigned the placeholder Target Type variant, and Variant Ascertainment is marked not applicable. These forced fits are recorded as six candidate extensions: a score target type, whole-genome-aggregate resolution, a target-composition axis, epidemiological measurement targets, a structured cohort descriptor, and a flag for derived-artifact targets. The model’s limitations are thus recorded explicitly rather than left implicit. Further detail is given in Supplementary Note SN6.2. 5 Human–AI Annotation Workflow and Evaluation We designed and piloted an annotation workflow in which every annotation is auditable. Each assertion is anchored to the specific text that supports it; each dimension value is defensible under the schema’s activation conditions; and each judgment that a downstream user might contest carries an explicit reviewer or AI flag. The same commitment applies to AI assistance. When an AI system drafts an annotation, its uncertainties and assumptions are recorded for curator review rather than silently absorbed into the annotation. Section 5.2 reports the practical effect of that review. The manual and AI-drafted tracks have different epistemic roles, primary and secondary knowledge work, respectively, and use distinct flag vocabularies to record provenance. We treat the pilot as a feasibility study, not as a benchmark. Six papers and a single annotator do not support claims about inter-annotator agreement or generalizability. The pilot shows that the workflow’s epistemic commitments can be implemented in practice and that the schema exposes, rather than hides, its points of strain. 5.1 Protocol and Coverage Annotations are created through two tracks (fig. 1). 9 Source publi- cation (PDF) Curator high- lights & callouts Structured JSON record YAML an- notation (curator ref- erence set) AI draft (annotation skill) Curator review (review skill) Reviewed YAML annotation SHACL val- idation (conditional activation) Figure 1: The two annotation tracks. Curator-authored annotations (left) form the feasibility corpus’s reference set. AI-drafted annotations (right) are produced under an annotation skill from the source PDF, the schema, and the manual annotations as exemplars, then reviewed by a curator under a review skill. Both tracks are validated against the same SHACL shapes. In every case, the curator, not the model, is the authority for the recorded annotation. Manual track. Four of the six publications, Jossin, Davis, Nelson, and Gupta, were annotated manually by Vladimir Seplyarskiy, a researcher in human and population genetics. The curator read each paper, marked relevant passages in the PDF with highlights and callouts, exported those annotations with a helper script to a structured JSON record, and mapped the result, together with a per-paper summary row, to a YAML annotation conforming to the schema. When this conversion required normalization, the change was recorded inline as anormalization_note withtype = ai_normalization. For example, the free-text “localisation and protein expression” was mapped to the enumeration values localization and existence. This record allows later audits to compare the original and normalized values. The manual-track protocol is described in more detail in Supplementary Note SN2. AI-drafted track. For two publications, Duerr and Inouye, an AI assistant drafted the annotations directly from the PDFs under an executable annotation skill [1]. The assistant received the schema (dimensions.md), the curator-authored reference annotations for the four manual papers as exemplars, and the source PDFs. It produced YAML annotations in the same format as the manual annotations. When the assistant judged a mapping to be uncertain, or concluded that a judgment call required human review, it recorded the issue inline as eitherai_uncertaintyorai_assumption. One of the authors then reviewed the AI-drafted annotations assertion by assertion. The reviewer has expertise in genomic data curation and the evidence model, rather than in genetics domain science. The review assessed source fidelity and schema conformance, not the underlying science. AI-generated drafts and AI-assisted review suggestions were subject to final adjudication by the human curator; the reviewed annotations are human-approved records, and no AI system was treated as the authority for any scientific or curation decision. The four manual annotations are treated as the curator-authored reference annotations for their own papers and as exemplars for AI drafting. The executable skills 10 and review protocol are described in Supplementary Notes SN2 and SN3. Source anchoring. In both tracks, every assertion, that is, theassertionproperty of a Genet- icEvidence item, carries asource_span. The span identifies a specific page and includes a short key phrase, typically fewer than fifteen words, from the source text. The phrase is long enough for a reader to locate the passage by search while remaining short enough to respect fair-use norms for copyrighted material. Source anchoring is mandatory: an annotation with any assertion lacking a source span is rejected at validation time. Across the six-paper pilot, all 95 assertions are source-anchored. This is a property of the workflow’s validation requirement, not an observed outcome. Coverage. Dimension coverage across the 28 GeneticEvidence items is summarized in Supple- mentary Note SN7 (Table ST9) and in the repository’s coverage report. 6 5.2 Flags and Extension Promotion The two tracks use small flag vocabularies to make their epistemic provenance explicit. Manual track. The manual track uses three reviewer flags. Areviewer_disagreementrecords a clear factual correction, such as the organism-set correction in Davis 2011. Areviewer_suggestion records an optional addition whose inclusion depends on curator preference, such as HPO terms for patient phenotypes in Davis, Nelson, and Gupta. Areviewer_queryrecords a case in which the mapping to the schema is unclear and should not be silently corrected. For example, the Nelson summary assignedspecificity_of_phenotype = SNV, which does not fit the dimension’s semantics. AI-drafted track. Because there is no prior reference annotation with which to disagree, the AI-drafted track uses two flags. Anai_uncertaintyrecords a mapping that the assistant judged unreliable and referred for a human decision, such as a borderline assignment of Mode of Inheritance. Anai_assumptionrecords a judgment made without strong textual support, such as whether cited mouse-model evidence should be represented as a separate GeneticEvidence item or treated as background context. Across the six-paper pilot, manual-track annotations carried five queries, six suggestions, and one disagreement, whereas AI-drafted annotations carried six uncertainties and six assumptions. Of the nine AI-drafted items from Duerr and Inouye, none was rejected: one was approved as drafted, three were approved with non-blocking flags, and five were edited. The later Inouye draft required fewer edits than the earlier Duerr draft, but the papers differ in genre and protocol version; this comparison is therefore suggestive rather than controlled. The edits mainly affected dimension assignment and credibility calibration rather than the factual content of assertions. The reviewer also added a small number of higher-level items, such as a paper’s overall thesis or a cited translational study. Per-paper counts and detailed rationales are given in Supplementary Note SN4. Extension promotion. The workflow adopts a conservative rule for evolving the schema. A representational gap observed in one paper is recorded only in that paper’scandidate_extensions block. A gap observed independently in a second paper is promoted to a first-class dimension or enumeration value. 6 https://github.com/ForomePlatform/genetic-evidence-model/blob/master/annotations/coverage.md 11 In the six-paper pilot, this rule promoted two extensions. Phenotype Scale was first proposed in Jossin and then reused in Davis, Nelson, and later papers. Variant Ascertainment was first proposed in Davis and then reused in Nelson and Duerr. Both now appear in the vocabulary (tables 2 and 3): Phenotype Scale among the always-required dimensions and Variant Ascertainment among the conditional dimensions, where it is required when Target Type is variant. Fifteen other candidates remain unpromoted, pending a second independent use or an explicit curator decision. The rule also prevents premature additions. For example, a proposed polarity dimension for negative assertions in Duerr was withdrawn when review confirmed that the existing assertion predicate already represents both absence and presence. 6 Discussion Two features of the proposal merit brief emphasis. First, the always-required vocabulary is small, and conditional activation keeps the extended vocabulary compact without sacrificing expressive power. Second, every assertion is anchored to a specific source span in its source publication, making each annotation independently auditable. 6.1 Positioning and Model Extensions The released schema formalizes the implemented conformance requirements as SHACL shapes, makes the implemented activation conditions machine-checkable, and anchors each assertion to its source text rather than only to a style guide. The flag vocabularies of section 5.2 make provenance explicit: a reader can tell whether a value was drawn from the author’s summary, proposed by a reviewer, added as an AI assumption, or left deliberately unpopulated. More broadly, the schema is intended to make literature review tractable as a semantic-parsing problem: the mapping of natural-language claims to a shared, source-anchored representation. This is not a task a language model can perform in isolation. The target of each parse is a structured, source-anchored representation, so the quality of the parse depends on the quality of that target language. We see the present contribution as one increment along a longer path, building on decades of biomedical-ontology work (ECO, SEPIO, HPO, IAO, OBI, VRS, and others), not as a replacement for any of it. A complementary downstream layer, in which an epistemological type system and machine-checkable meta-predicates constrain the decision logic that consumes such evidence, is addressed in [4]. The pilot identified two candidate dimensions that were exercised by multiple papers and promoted to first-class status, Phenotype Scale and Variant Ascertainment. It also identified fifteen further candidate extensions that remain unpromoted under the two-papers rule (section 5.2). The pattern of candidates is summarized in table 4; per-candidate rationale, affected dimensions, and proposed enumeration values are documented in a companionschema/EXTENSIONS.mdin the repository. 7 Supplementary Table ST1 places these candidates in a broader context. Of the ten evidence patterns considered, five common patterns in basic-science genetic publications are represented natively by the model. Three patterns were exposed by the pilot as forced fits and define documented areas for extension. Two further patterns, therapeutic-response evidence and cross-publication synthesis, are outside the model’s intended scope and are left to standards designed to represent them. The Gupta annotation, which identified no candidate extensions, is also a positive finding. It reports a family with an X-linked splice-site mutation and careful pedigree segregation, a canonical evidence shape for which the model was designed. The schema accommodates this evidence without strain. 7 https://github.com/ForomePlatform/genetic-evidence-model/blob/master/schema/EXTENSIONS.md 12 Table 4: Unpromoted candidate schema extensions surfaced by the six-paper pilot, grouped by concern. Per-candidate detail is in the companionschema/EXTENSIONS.md; paper keys are J = Jossin, D = Davis, N = Nelson, DU = Duerr, IN = Inouye. Gupta surfaced none. GroupPapers Count and examples Finer structural resolutionJ, N4 (subdomain, interaction, gene-relation, cross-level variant) Polygenic-score machineryIN6 (score target, whole-genome aggregate, target composition, epidemiological measurement, cohort sub-block, derived-artifact flag) Effect-size and curationDU, D3 (direction of effect, knowledge-domain priority, curator critique) Cross-item and activation scope DU2 (cross-item/translational synthesis, conditional-activation scope) 6.2 Limitations and Future Work The pilot has significant limitations. Six papers and a single annotator do not support claims about inter-annotator agreement or generalizability; we treat the work as a feasibility study, not a benchmark. The evaluation is not a clinical deployment: although the model is motivated by variant interpretation in clinical-genomics programs, the pilot stops short of applying the model to a live workflow. Three near-term directions follow from these limitations. First, annotate a second corpus selected to exercise the strongest candidate extensions. Second, generate OWL, SHACL, JSON Schema, and LinkML renderings from a single source of truth. This would open the schema to the OBO and Bridge2AI ecosystems and support term-level alignment of the dimension vocabularies with relevant ontologies, for example by aligning the molecular and cellular phenotype scales with Gene Ontology biological processes. Third, conduct a clinical-deployment study to test whether the feasibility claim translates into utility. A further limitation revealed by the UMLS crosswalk is representational rather than empirical. Several GEM tokens are intersective compounds: statistical_genetics denotes statistics×genet- ics (C0038215× C0017398), and bioinformatics_inference denotes inference×bioinformatics. Relational tokens have the same structure. For example, related_gene combines a relation operator with Genes (C0017337). Because the current crosswalk maps each GEM token to one UMLS concept, it must sometimes use the nearest available anchor rather than a concept that expresses the full compound meaning. For statistical_genetics, that anchor is Biostatistics, recorded with the relation narrower. Concept-level post-coordination would address this limitation by mapping a token to a small set of concepts whose intersection represents its intended meaning. The UMLS Metathesaurus does not provide such a mechanism at the aggregate level. As an aggregator, it records the concepts asserted by its source vocabularies and the links among them, but it does not define a compositional grammar of its own. Composition remains a feature of individual source vocabularies, with the compositional grammar of SNOMED CT being the best-known example. It cannot be transferred directly across the aggregate: the same intended intersection may be pre-coordinated in one source, absent from another, and decomposed differently in a third [3]. The problem is particularly clear for related_gene. UMLS often contains pre-coordinated specific instances while lacking a generic concept: gene-relatedness appears in terms such as modifier, homologous genes, and linked genes, whereas the generic sense of “related in some way” is represented only by the bare operator Associated with (C0332281). Thus related_gene can be represented asC0332281× C0017337, but not by a single concept. Extending the crosswalk with an explicit 13 composite relation, represented as a structured set of concepts, is future work. At present, the intended composition is preserved in the mapping’s curator note; where the relation is the semantic core, the operator concept itself is recorded as the anchor with relation related. The crosswalk review also highlighted a representational limitation of the Resolution axis. Its current enumeration presupposes genomic coordinates, but genetic evidence may be localized in several coordinate systems. A variant may be described at the genomic level (HGVSg.), the transcript level (c.), or the protein level (p.). For example, an assertion of the form “variantVhas effectEin transcriptT” is resolved in transcript coordinates. Such assertions are often tissue-restricted and are characteristic of bioinformatics inference. The domain mapping underlying the protein_subdomain candidate is instead resolved in protein coordinates. In addition, variant itself has two components: a position and a discrete nucleotide change. We therefore record variant_in_transcript as a curator-surfaced candidate value for Resolu- tion. It is separate from the fifteen pilot-surfaced extensions listed in table 4 and awaits independent use in a corpus paper under the promotion rule. 7 Availability of Code, Data, and Artifacts All artifacts referenced in this paper are open-source and available in the project repository at https://github.com/ForomePlatform/genetic-evidence-model. This includes the SHACL schema, the human-readable dimension reference (dimensions.md), per-paper annotations for all six pilot publications, case reports, the companionschema/EXTENSIONS.md, the SHACL walk- through (schema/examples.md), the annotation and review protocol suite (protocols/) and the two executable annotation and review skills (skills/) described in Supplementary Note SN3, and the extraction tooling used to convert PDF highlights and callouts into annotation scaf- folds. The namespacehttps://w3id.org/genetic-evidence-model/resolves to the same repos- itory. This paper describes releasev0.2.5, archived on Zenodo under the concept DOIhttps: //doi.org/10.5281/zenodo.22260686, which resolves to the latest archived release; the GEM Mapping Studio is distributed on PyPI as gem-mapping-studio. Reproducibility and conformance All counts reported in this paper were generated from the accompanying tagged release. From a clean checkout, corpus validation and regeneration of the coverage table require two commands: gem-validate and gem-coverage. Continuous integration runs both checks on every change. The paper distinguishes SHACL-enforced activation conditions, marked in table 3, from docu- mented rules that are not yet executable. The enforced activation shapes are unit-tested in both directions: a conforming instance passes, whereas an instance that satisfies an activation condition but lacks the required dimension is reported as a violation. 14 Table 5: Artifact-to-claim traceability. Paths are relative to the repository root of the tagged release. ClaimWhere to verify Canonical schema schema/genetic_evidence.shacl.ttl(three enforced activation shapes: Vari- ant Ascertainment, Mode of Inheritance, and Organism) YAML-to-RDF transform src/python/main/forome/gem/extraction/yaml_to_rdf.py(normalizes the v0/v1 field renameskey_phrase/phraseandresolution/target_resolution) Corpus validation gem-validateoverannotations/;CIworkflow .github/workflows/validate.yml Coverage counts gem-coverage regenerates annotations/coverage.md Activation-shape tests src/python/test/forome/gem/validation/ UMLS crosswalk data/umls/ (harnessgem-umls-crosswalk;decisionlog data/umls/DECISIONS.md; SN8 regenerated by gem-umls-render) Archived releaseZenodo deposit of the tagged release; metadata in .zenodo.json Acknowledgements The idea of a semantic model for the analysis of genetic literature grew out of 2018 conversations with Prof. Shamil Sunyaev; without his early support this work could not have been completed. The authors are also grateful to colleagues at the Brigham Genomics Medicine (BGM) program, where much of the thinking about clinical-genomics evidence curation took shape, and to Vladimir Seplyarskiy for his meticulous curation of the four manually annotated papers in the pilot corpus. This work received no external funding. 8 Conclusion We have introduced a semantic model for genetic evidence from the biomedical literature: a compact hierarchy of core classes, a dimensional vocabulary with conditional-activation rules, three of which are presently machine-enforced in SHACL, and a SHACL encoding for automated conformance validation. The model is grounded in an open corpus of six annotated papers and accompanied by a collaborative workflow that keeps curator-authored reference annotations distinct from AI-drafted annotations and promotes candidate schema extensions only when a second independent paper confirms them. The model is also released with executable annotation and review skills that let an AI assistant apply it under curator supervision. Taken together, these elements are offered as a small but concrete contribution toward trustworthy, AI-ready infrastructure for variant interpretation: a reference data model and validation schema for representing genetic evidence from the biomedical literature. References [1]Anthropic. Agent Skills.https://platform.claude.com/docs/en/agents-and-tools/agent- skills/overview, 2025. Accessed 2026. [2]Anita Bandrowski, Ryan Brinkman, Mathias Brochhausen, et al. The ontology for biomedical investigations. PLoS ONE, 11(4):e0154556, 2016. doi: 10.1371/journal.pone.0154556. [3]Olivier Bodenreider. The unified medical language system (UMLS): integrating biomedical terminology. Nucleic Acids Research, 32(Database issue):D267–D270, 2004. doi: 10.1093/nar/ gkh061. 15 [4]Michael Bouzinier, Sergey Trifonov, Michael Chumack, Eugenia Lvova, and Dmitry Etin. Trustworthy clinical decision support using meta-predicates and domain-specific languages, 2026. URL https://arxiv.org/abs/2604.21263. [5]Werner Ceusters and Barry Smith. Aboutness: Towards foundations for the Information Artifact Ontology. In Proceedings of the Sixth International Conference on Biomedical Ontology (ICBO 2015), volume 1515 of CEUR Workshop Proceedings, pages 1–5, 2015. URLhttps: //ceur-ws.org/Vol-1515/regular10.pdf. [6]Michelle Giglio, Rebecca Tauber, Suvarna Nadendla, et al. ECO, the Evidence and Conclusion Ontology: Community standard for evidence information. Nucleic Acids Research, 47(D1): D1186–D1194, 2019. doi: 10.1093/nar/gky1036. [7] Global Alliance for Genomics and Health. GA4GH Variant Annotation (VA) Specification. https://va-spec.ga4gh.org/en/1.0/, 2024. VA-Spec v1.0. [8]Gordon H. Guyatt, Andrew D. Oxman, Gunn E. Vist, Regina Kunz, Yngve Falck-Ytter, Pablo Alonso-Coello, and Holger J. Schünemann. GRADE: an emerging consensus on rating quality of evidence and strength of recommendations. BMJ, 336(7650):924–926, 2008. doi: 10.1136/bmj.39489.470347.AD. [9] HL7 International. FHIR Evidence, EvidenceVariable, and EvidenceReport Resources.https: //hl7.org/fhir/R5/evidence.html , 2023. FHIR Release 5 (v5.0.0), published 26 March 2023. [10]Sebastian Köhler, Michael Gargano, Nicolas Matentzoglu, et al. The Human Phenotype Ontology in 2021. Nucleic Acids Research, 49(D1):D1207–D1217, 2021. doi: 10.1093/nar/gkaa1043. [11]Monarch Initiative. SEPIO: Scientific Evidence and Provenance Information Ontology.https: //github.com/monarch-initiative/SEPIO-ontology, 2026. Accessed April 26, 2026. [12]Heidi L. Rehm, Jonathan S. Berg, Lisa D. Brooks, et al. ClinGen—The Clinical Genome Resource. New England Journal of Medicine, 372:2235–2242, 2015. doi: 10.1056/NEJMsr1406261. [13] Sue Richards, Nazneen Aziz, Sherri Bale, et al. Standards and guidelines for the interpretation of sequence variants: A joint consensus recommendation of the American College of Medical Genetics and Genomics and the Association for Molecular Pathology. Genetics in Medicine, 17 (5):405–424, 2015. doi: 10.1038/gim.2015.30. [14]Nicole A. Vasilevsky, Sabrina Toro, Nicolas Matentzoglu, et al. Mondo: Integrating disease terminology across communities. Genetics, 232(4):iyaf215, 2026. doi: 10.1093/genetics/iyaf215. [15] W3C. PROV-O: The PROV ontology. https://w.w3.org/TR/prov-o/, 2013. [16]W3C. Shapes Constraint Language (SHACL).https://w.w3.org/TR/shacl/, 2017. W3C Recommendation, 20 July 2017. [17] Alex H. Wagner, Lawrence Babb, Gil Alterovitz, et al. The GA4GH variation representation spec- ification: A computational framework for variation representation and federated identification. Cell Genomics, 1(2):100027, 2021. doi: 10.1016/j.xgen.2021.100027. 16 Supplementary Material: A Semantic Model of Genetic Evidence Michael BouzinierDmitry Etin Contents Supplementary Notes3 SN1 Background and Related Work3 SN1.1 Requirements and coverage matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 SN2 The annotation and review protocol suite10 SN2.1 Manual-track protocol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 SN2.2 AI-drafted-track protocol . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 SN2.3 Source anchoring and validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 SN2.4 Provenance and protocol versions . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 SN3 Executable skills13 SN4 AI-track review outcomes13 SN5 Standards crosswalk14 SN6 Case reports: Duerr 2006 and Inouye 201817 SN6.1 Case Report CR-DU: Duerr et al. 2006 (genome-wide association) . . . . . . . . . 17 SN6.2 Case Report CR-IN: Inouye et al. 2018 (polygenic risk score) . . . . . . . . . . . . 19 SN7 Dimension coverage over the annotated corpus21 SN8 UMLS crosswalk and illustrative OMOP CDM positioning22 SN8.1 UMLS concept crosswalk (generated) . . . . . . . . . . . . . . . . . . . . . . . . . 22 SN8.2 OMOP CDM positioning (illustrative) . . . . . . . . . . . . . . . . . . . . . . . . 25 SN9 Credibility tiers: definitions, the UMLS finding, and external alignment26 SN9.1 What the tiers mean . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 SN9.2 Why the tiers have no UMLS proxy . . . . . . . . . . . . . . . . . . . . . . . . . . 27 SN9.3 External alignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 Supplementary Tables30 1 List of Tables ST1 Requirements and coverage matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 ST2 Per-paper AI-track review verdicts . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 ST3 Crosswalk: structural classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 ST4 Crosswalk: assertion property and class . . . . . . . . . . . . . . . . . . . . . . . . . 15 ST5 Crosswalk: dimensions and values . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 ST6 Crosswalk: coverage gaps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 ST7 Duerr 2006: evidence-item decomposition . . . . . . . . . . . . . . . . . . . . . . . . 18 ST8 Inouye 2018: evidence-item decomposition . . . . . . . . . . . . . . . . . . . . . . . 19 ST9 Dimension population across 28 GeneticEvidence items . . . . . . . . . . . . . . 21 ST10 GEM dimensional vocabulary mapped to UMLS concepts (generated by the map- ping harness and curator adjudication). . . . . . . . . . . . . . . . . . . . . . . . . . 22 ST11 Illustrative positioning of GEM dimensions against the OMOP CDM (hand-authored, not exhaustive). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 ST12 Credibility sweep: rejected concept families . . . . . . . . . . . . . . . . . . . . . . . 28 ST13 Credibility: external alignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 ST14 The six annotated publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 ST15 Credibility facets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 ST16 FHIR field-level alignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 2 Supplementary Notes SN1 Background and Related Work Scientific and clinical evidence has been the subject of sustained standards work in clinical informat- ics, biomedical ontology, and genomic knowledge representation. Our schema lies at the intersection of these communities. It is oriented toward downstream decision-making, grounded in formal con- straints and explicit provenance, and treats target resolution, including variant, gene, region, and whole-genome aggregate, as a first-class concern while relying on GA4GH standards for sequence identifiers. This note positions the schema relative to seven families of standards and identifies the coverage gap that motivated the work. A first term-level alignment is provided in Section SN5. Completing the coverage and formalizing the mappings in OWL remain future work. FHIR Evidence Resources. The HL7 FHIR Evidence, EvidenceVariable, and EvidenceReport resources [9] provide structured representations of clinical-trial-style evidence, with explicit fields for population, intervention, comparator, outcome, and statistical measures. They were designed with systematic-review and guideline-development workflows in mind, where evidence concerns an intervention tested in a population. This framing does not fit basic-science and pre-clinical genetic evidence cleanly. A mouse knockout study, for example, does not have an “intervention” in the clinical-trial sense, and a coding-variant case–control study has a target, the variant, that is not the same kind of object as a drug or procedure. Our schema shares FHIR Evidence’s commitment to structured, dimensional description but replaces its clinical-trial frame with a vocabulary suited to genetic claims. Evidence Codes: ECO and SEPIO. The Evidence and Conclusion Ontology (ECO) [6] pro- vides a large hierarchical vocabulary of evidence types, including inference methods, assay types, and source-document categories. It is organized within the Open Biological and Biomedical Ontolo- gies (OBO) framework and is widely used in Gene Ontology annotation. The Scientific Evidence and Provenance Information Ontology (SEPIO) [11] complements ECO with machine-readable rep- resentations of statements, qualifiers, and provenance relationships. Our Method and Measurement Target dimensions overlap substantially with areas already covered by ECO and SEPIO. Our long- term goal is to ground those dimensions in ECO and SEPIO identifiers rather than maintain a separate vocabulary. GA4GH Genomic Knowledge Standards. The most relevant comparison within the GA4GH Genomic Knowledge Standards effort is the Variant Annotation (VA) Specification [7]. VA is a broad information model for structured statements about variants, organized around a Statement class whose instances carry a proposition, direction of support, strength, classification, evidence lines, and contribution metadata. VA is information-centric: its top-level abstraction, InformationEntity, represents what is known about a variant rather than the process by which primary-literature claims become evidence. Our schema is narrower and complementary. It focuses on evidence drawn from the primary literature, with the Relevance and Specificity and Confidence challenges identified in the main-text introduction as concrete targets. The Variation Representation Specification (VRS) [18], a separate product of the same effort, addresses canonical variant identity across coordinate systems and is orthogonal to the present work. 3 ACMG/AMP variant interpretation. The American College of Medical Genetics and Ge- nomics and Association for Molecular Pathology (ACMG/AMP) guidelines [13] specify a rule-based framework for variant classification from weighted evidence codes and criteria. Our schema is in- tended to supply inputs to such frameworks, rather than replace them: its per-publication evidence items are the kind of input an ACMG/AMP-compatible classifier would consume. A term-level alignment of the Method and Credibility dimensions with ACMG/AMP evidence codes remains future work. Upper ontologies: IAO and OBI. The Information Artefact Ontology (IAO) [5] and the Ontol- ogy for Biomedical Investigations (OBI) [2] provide upper-level scaffolding for many OBO-compliant biomedical ontologies. IAO supplies classes such as information-content-entity and textual-entity; OBI supplies assay and protocol hierarchies relevant to the Method dimension. Formal alignment of the model’s top-level classes with IAO and OBI, together with an OWL/LinkML rendering of the schema, remains future work. Provenance: PROV-O. The W3C PROV Ontology (PROV-O) [17] provides standard prove- nance relationships. Aligning the model with PROV-O remains future work. Related annotation efforts. The Clinical Genome Resource (ClinGen) evidence-curation frame- work [12] provides a structured workflow for gene–disease validity and variant-pathogenicity assess- ments with explicit evidence rubrics. Our schema shares its commitment to source-anchored claims and explicit categorization, but operates at a lower level by representing the individual claims that such frameworks aggregate. The Monarch Disease Ontology (MONDO) [16] and the Human Phe- notype Ontology (HPO) [10] provide established vocabularies for disease and phenotype terms. We expect curator-authored HPO terms to populate the schema’s phenotype-mapping fields, as illus- trated by the HPO-term suggestions in the Davis, Nelson, and Gupta case studies (Section 5 of the main paper). SN1.1 Requirements and coverage matrix The per-family paragraphs above describe each standard on its own terms. Table ST1 takes the complementary perspective: it begins with ten evidence patterns that a genetic publication may report and asks whether each pattern can be represented by each standard and by GEM. The six- paper corpus (Table ST14) exercises GEM’s representation of the first eight patterns, either natively or through a forced fit. The final two are not represented in the current model. Cross-publication synthesis is outside GEM’s scope by design, whereas therapeutic-response evidence remains an open scope question under CE-DU3 (row 9). The corpus exposed therapeutic response directly through the cited anti-p40 antibody trial in Duerr (GE-new-1). It also exposed a related within-paper synthesis problem: a higher-order claim over several items in one paper. Both issues are recorded under candidate extension CE-DU3. The matrix is a coverage inventory, not a ranking. The standards were developed for different purposes, and a pattern marked as not represented is a limitation only if that pattern lies within the standard’s intended scope. The five standards columns represent the resources in this comparison that carry evidence con- tent. ECO and SEPIO are shown separately, whereas the ClinGen and ACMG/AMP frameworks share one column. IAO, OBI, and PROV-O are omitted because they are orthogonal to all ten patterns. They provide upper-level classes, assay hierarchies, and provenance relations that the represented models may reuse, but do not themselves represent these patterns. 4 Cell values are interpreted relative to each standard’s role. For an information model, including FHIR Evidence, GA4GH VA, and SEPIO, indicates that a class or published profile represents the pattern. G# indicates that the pattern is representable only through the standard’s extension mechanism, such as a custom VA profile, a FHIR extension, or an external predicate vocabulary; only by reference to a resource outside the module named in the column; or only in one aspect. # indicates that the model has no representation for the pattern. For ECO, indicates that a term for the pattern exists, G# that only a generic term applies, and # that no suitable term was identified. The ClinGen gene–disease validity framework [15] and the ACMG/AMP variant-interpretation guidelines [13] share one column because they score evidence at complementary levels: ClinGen at the gene level (gene–disease validity) and ACMG/AMP at the variant level (variant pathogenicity). The column is read as their union. indicates that the pattern is a scored evidence category in the framework that operates at the pattern’s level, or in both frameworks when it is scored at both levels (rows 1, 4, and 10); in row 10 the scored category is the classification the criteria combine into. The row note names the scoring framework or frameworks. The absence of a criterion in one framework is not counted as a gap when the pattern lies outside its level, as a gene-level claim does for ACMG/AMP (rows 2, 5, and 6). G# indicates partial or indirect treatment, including a required input that is not itself scored (row 7) or treatment by a ClinGen product other than the validity framework (rows 3, 8, and 9). # indicates that neither framework provides a corresponding scored category. For GEM, indicates that the pattern is represented natively and was annotated in the corpus without a forced fit. G# indicates that the pattern was annotated only through a forced fit recorded as a candidate extension, identified by a CE code in schema/EXTENSIONS.md. # indicates that the pattern is not represented. In rows 9 and 10 the pattern belongs to what a genetic publication may report, so GEM’s open circle marks a scope boundary rather than a limitation: settled by design in row 10, left open pending CE-DU3 in row 9. In any column, – indicates that the pattern is outside the standard’s subject matter altogether, so its absence is not a coverage gap. Every identifier cited in the row notes was verified against the relevant source vocabulary or specification at the time of writing. 5 Table ST1: Ten evidence patterns across five standards and GEM. = represented natively; for ECO, a suitable term exists; for the combined ClinGen/ACMG column, the framework at the pat- tern’s level (gene or variant) provides a scored category or resulting classification. G# = represented only through an extension mechanism, only by reference to a resource outside the column’s module, or only in part; for ECO, only a generic term applies; for GEM, a forced fit is recorded as a candidate extension. # = not represented. – = outside the standard’s subject matter. The remaining gap column is relative to GEM and names a candidate extension where one exists. Row notes follow the table. FHIR = HL7 FHIR Evidence R5; VA = GA4GH Variant Annotation 1.0 with VRS 2.0 and Cat-VRS 1.0; ClinGen/ACMG = the ClinGen gene–disease validity framework, with other ClinGen products where noted, and the ACMG/AMP variant-interpretation guidelines. Evidence patternFHIR VA SEPIO ECO ClinGen/ ACMG GEM Remaining gap (GEM) 1. Mendelian pedigree: variant co-segregates with disease in a family G# G# G# no structured segregation strength (meioses, LOD); activation scope for Variant targets (CE-DU4) 2. Case–control burden: gene-level aggregate of rare variants G# G# G# G# cohort composition (CE-IN5) 3. Common-variant association: GWAS discovery, replication, fine mapping G# G# G# G#G# direction of effect (CE-DU1); activation scope (CE-DU4); no discovery–replication link 4. Functional assay: variant or gene effect measured in vitro G# G# G# protein-subdomain resolution (CE-J1); assay type recorded as free text 5. Animal model: knockout or conditional-knockout phenotype G# G# G# HPO-only phenotype mapping; rescue not distinguished 6. Gene–gene relation: X inhibits, binds, or regulates Y # G# G# G# relation vocabulary (CE-J3); interaction as target (CE-J2) 7. Variant identity across coordinate systems (protein, cDNA, genomic) G# –G#G# cross-level description (CE-N1); transcript coordinates (CE-C1); VRS/HGVS binding 8. Polygenic score: construction and validation of a genome-wide aggregate G# G# G# G#G#G# score machinery (CE-IN1 to CE-IN6) 9. Therapeutic response: intervention tested in patients G# G## no intervention construct; translational scope open (CE-DU3) 10. Cross-publication synthesis: evidence lines combined into a classification # out of scope by design Row notes. Paper keys follow Table ST14; item identifiers (GE-n) follow the released annotations. 1. Pedigree segregation. FHIR can represent the family as a Group population and the segregation statistic as an ad hoc statistic, but it has no dedicated segregation construct. VA covers this pattern through the ACMG 2015 Variant Pathogenicity Evidence Line profile, in which co- segregation (P1) is a criterion that an evidence line may carry. SEPIO provides an assertion and evidence-line frame, but the segregation content must come from a domain model, as in the SEPIO-based ClinGen interpretation model. ECO has no term specific to co-segregation, pedigree, or linkage evidence; the only applicable term is the general ECO:0005613, inference by association of genotype from phenotype, whose definition covers alleles linked to Mendelian disease and whose usage note names segregation of genetic markers (the same term credited 6 in rows 2, 3, and 8). ClinGen scores segregation as case-level evidence [15], and ACMG/AMP codes it as P1 [13]. GEM represents the pattern natively through Mode of Inheritance and Mendelian Segregation, as exercised by Gupta GE-1 (X-linked recessive) and Davis GE-5 (au- tosomal recessive). Both are Variant-target items, so they populate the Mendelian dimensions outside the documented human genetics+ gene activation condition; this activation-scope gap is logged as CE-DU4 (row 3). Because Mendelian Segregation is Boolean, segregation strength (meioses, LOD) has no structured field; pedigree size is recorded only as free text (Gupta GE-1). 2. Case–control burden. FHIR represents the study design and statistic natively (studyDesign, statistic), but the gene-level aggregate exposure must be coded ad hoc. VA 1.0 has no gene- level proposition among its base profiles. Its nearest study result, Cohort Allele Frequency, is variant-level, so a custom profile is required. ECO offers only the generic ECO:0005613. ClinGen scores case–control evidence, including aggregate variant analysis [15]. ACMG/AMP classifies single variants and has no gene-level burden criterion; its case–control criterion, PS4, concerns the prevalence of one variant. The cell therefore reflects ClinGen alone. GEM represents the pattern natively in Davis GE-1, with Population Genetics and Human Genetics, Statistical Genetics, and a Gene target. Case and control counts are recorded only as free text (Davis GE-1, special_considerations), not as structured fields; a structured cohort sub-block is candidate extension CE-IN5. 3. Common-variant association. FHIR and VA have the same limitations described in row 2. Neither distinguishes a genome-wide discovery stage from a replication stage, and VA 1.0 has no population-association proposition among its base profiles. ECO has no genome-wide- association or replication term beyond ECO:0005613. The ClinGen validity framework scores Mendelian gene–disease evidence and does not score complex-trait association. ACMG/AMP, the framework at the variant level, does not score it either. The guidelines state that they are not intended for variants in genes associated with multigenic non-Mendelian complex disorders [13]; their one case–control criterion, PS4, is written for rare variants in Mendelian genes (a relative risk or odds ratio above 5.0, with a fallback for very rare variants absent from controls), an effect size that common-variant associations rarely approach; and a common allele itself is treated as benign evidence by the population-frequency criteria BA1 and BS1. Within the two frameworks the pattern is therefore unscored. ClinGen’s Low Penetrance/Risk Allele Working Group has, however, issued separate recommendations for curating, classifying, and reporting common risk alleles [14], a ClinGen product outside the validity framework; by the convention applied to PRS-RS in row 8, this earns the half circle. GEM represents the pattern natively through the Method hierarchy: GWAS, Association Study, and Fine Mapping under Statistical Genetics. This is exercised by Duerr GE-1, GE-2, and GE-4. The same paper exposed the missing direction-of-effect dimension for a protective allele (CE-DU1) and, through the family-based transmission test in GE-3, an activation gap for Human Genetics with a Variant target (CE-DU4). 4. Functional assay. FHIR has no population–intervention frame for a cell-based assay; the assay system would have to be cast as a population. VA covers the variant-level case through the Experimental Variant Functional Impact proposition; its companion study-result profile is scoped to multiplexed assay scores, and gene-level assays such as the Jossin binding and phosphorylation experiments require a custom profile. ECO:0000181, in vitro assay evidence, names the evidence type generically; ECO’s assay-specific terms sit in sibling branches such as direct assay evidence (ECO:0000002) rather than beneath it. ClinGen scores experimental 7 evidence, including biochemical function and functional alteration [15]; ACMG/AMP codes it as PS3 or BS3. GEM represents this pattern natively, as illustrated by Nelson GE-1 to GE-3 and the in vitro Jossin items. Jossin exposed the missing protein-subdomain resolution (CE-J1). The assay itself is described only at the Method level, In Vitro. The crosswalk maps that value to ECO:0000181 at the same granularity (section SN5); grounding assay type in ECO’s assay-specific terms or in OBI assay classes remains future work. 5. Animal model. FHIR can cast the genotype as an exposure and the animal cohort as a Group, but it has no model-organism frame. VA has no animal-model proposition; the study can enter only as an evidence item behind a custom profile. ECO provides detailed evidence-type coverage through ECO:0000179, animal model system study evidence; ECO:0001091, knockout phenotypic evidence; ECO:0001030, conditional knockout evidence; and ECO:0000013, trans- genic rescue experiment evidence. ClinGen scores non-human model organisms and rescue as experimental evidence [15]. A knockout phenotype is a gene-level claim outside ACMG/AMP’s variant-level scope; PS3 and BS3 admit in vivo functional studies only as evidence of the ef- fect of the variant under interpretation on its gene or gene product. The cell therefore reflects ClinGen alone. GEM represents the pattern natively with the conditional Organism and Knockout Type dimensions, exercised in full by Jossin GE-1, a conditional mouse knockout with organismal-scale evidence and the corpus’s only knockout item. The Davis zebrafish (GE-2, a morpholino knockdown with rescue) and rat (GE-4) items exercise Organism but not Knockout Type. The phenotype-mapping fields currently expect HPO terms, so they do not bind model-organism phenotype vocabularies, and rescue is not distinguished from other in vivo evidence. 6. Gene–gene relation. FHIR Evidence has no frame for a relation between two molecular enti- ties. VA includes a Gene domain entity but no interaction proposition. A SEPIO assertion can carry the relation through a predicate from the Relation Ontology. ECO:0000021, physical interaction evidence, covers the evidence type. ClinGen scores protein interaction as experi- mental evidence in its Function category, alongside the biochemical function evidence of row 4 [15], although the relation serves a gene–disease claim rather than standing as a claim of its own. ACMG/AMP has no corresponding criterion: a gene–gene relation is a gene-level claim outside its variant-level scope, and a variant’s effect on an interaction would enter only as generic functional evidence under PS3 or BS3. The cell therefore reflects ClinGen alone. GEM has a Gene Relation dimension, of whose three current values the corpus exercised one: X inhibits Y in Jossin GE-5. The physical-binding relation in GE-4 matched no value and was left unpopulated (CE-J3, which also proposes a trafficking-regulation value), and the LLGL1–N-cadherin interaction targeted by GE-6 was a forced fit as a Gene target (CE-J2, which proposes Interaction and Complex values). 7. Variant identity across coordinate systems. This pattern is outside the FHIR Evidence mod- ule. FHIR handles it through genomics reporting resources that an Evidence resource can reference. VA delegates variant representation to VRS [18] and Cat-VRS: a VRS Allele sits on a genomic, transcript, or protein sequence reference and can carry HGVS g., c., or p. expressions, and the Cat-VRS CanonicalAllele and ProteinSequenceConsequence classes group congruent alleles across assemblies, transcripts, and proteins, a linkage that VRS it- self confines to a single sequence context. SEPIO and ECO do not address variant identity. Neither curation framework scores identity, but both require it as input: ACMG/AMP man- dates HGVS descriptions against a versioned reference sequence [13], and ClinGen’s curation interfaces identify a variant by ClinVar or ClinGen Allele Registry identifier, whose canonical 8 allele links the genomic, transcript, and protein forms. GEM relies on GA4GH identifiers by design, but its Resolution values currently presuppose genomic coordinates. The Nelson insertion allele, a 12-bp insertion in cDNA coordinates but a single-nucleotide substitution in genomic coordinates, was a forced fit under Resolution variant and was recorded as CE-N1, resolved by reference to HGVS and VRS rather than by a new value. A transcript-coordinate Resolution value (CE-C1) was surfaced by curator review during the UMLS crosswalk and has no corpus instance yet. Binding target identifiers to VRS or HGVS is planned rather than implemented. 8. Polygenic score. FHIR can represent a risk-prediction study with the score as a coded variable, but it has no dedicated score construct. VA 1.0 and VRS provide no entity for a genome-wide aggregate, so a custom profile would be required. ECO offers only the generic computational term ECO:0007672 and association term ECO:0005613. ClinGen’s relevant contribution is the PRS-RS reporting checklist [19], not a curation rubric. In GEM, all four Inouye items are forced fits: Target Type and Resolution both carry the placeholder value variant, and Variant Ascertainment is set to the documented sentinel. These forced values motivated CE- IN1 to CE-IN3; the same annotation surfaced CE-IN4 to CE-IN6 from its epidemiological outcome measures, free-text cohort description, and the score’s status as a derived artifact. All six await a second polygenic-score paper under the two-papers rule. 9. Therapeutic response. This pattern fits FHIR Evidence’s native population–intervention– comparator–outcome frame. VA covers it through the Variant Therapeutic Response proposi- tion. ECO:0000180, clinical study evidence, covers the evidence type. ClinGen addresses thera- peutic response through actionability curation rather than the gene–disease validity framework. GEM has no intervention construct. The cited anti-p40 antibody trial in Duerr, GE-new-1, was recorded with Knowledge Domain left unresolved by curator decision rather than forced into a genetic domain. It is the second CE-DU3 instance in the same paper; the first is the au- thors’ therapeutic-strategy recommendation, a higher-order claim over several of the paper’s items. Two instances in one paper do not satisfy the two-papers rule, so CE-DU3 remains a candidate whose proposal leaves open whether to add a higher-order claim construct or to declare translational claims out of scope. 10. Cross-publication synthesis. FHIR Evidence carries a synthesized result natively (a rated certainty; synthesisType, the method used to combine studies; and numberOfStudies un- der statistic.sampleSize), with EvidenceReport as the container that composes the con- tributing resources. Combining evidence lines into a conclusion is likewise the purpose of VA Statements with evidence lines, including ACMG classification profiles; of SEPIO evidence lines; and of the ClinGen and ACMG/AMP rubrics. ECO covers the evidence type through ECO:0000212, combinatorial evidence, and its descendant ECO:0007012, combinatorial exper- imental and curator inference evidence; as an evidence-type vocabulary it names the combined evidence, not the resulting classification. GEM does not synthesize evidence by design. The within-paper form of this pattern, a higher-order claim over several of a paper’s own items, is the open candidate CE-DU3 (row 9). GEM represents the per-publication evidence items consumed by such frameworks, while the reasoning layer that combines them is developed separately [4]. Read by column, the matrix clarifies GEM’s contribution. Rows 1 through 5 cover common patterns in basic-science genetic publications. Under the operational definitions used in this coverage inventory, GEM is the only resource in this comparison that represents all five patterns natively 9 within one common, genetics-specific dimensional profile. Other resources require an extension mechanism because they were designed for a different primary purpose, provide an evidence-type vocabulary that names the evidence without structuring the claim, or, in the case of the ClinGen and ACMG/AMP frameworks, score Mendelian gene- and variant-level evidence as inputs to a classification while leaving common-variant association to a separate risk-allele product (row 3). Rows 6 through 8 identify patterns that the pilot exposed as forced fits and therefore as docu- mented areas for extension. Row 10 is outside GEM’s scope by design. Row 9, therapeutic response, is not represented in the current model, but CE-DU3 leaves open whether translational claims should become part of the model’s scope or should be left to standards designed for intervention evidence. The coverage gap. None of these standards was designed to treat the heterogeneous evidence patterns of basic-science and pre-clinical genetic publications as a first-class concern. A Llgl1 conditional-knockout study, for example, may report molecular, cellular, and clinical phenotypes in the same paper. A polygenic-score study may aggregate millions of variants into one score. Each existing standard captures part of what such papers report, but none captures the whole. The schema introduced in the main paper aims to represent this heterogeneity in a form that is compact enough for manual annotation, machine-checkable enough for automated validation, and flexible enough to accommodate the evidence patterns produced by the literature. SN2 The annotation and review protocol suite The annotation workflow described in the main paper has been consolidated into a versioned, openly published protocol suite available in the project repository. 1 The suite distinguishes what a well- formed annotation or review must contain from how it is performed, either manually or with AI assistance. It governs both pilot tracks: the curator-authored manual track and the AI-drafted track. A mode-agnostic annotation protocol (PROTOCOL.md) defines the requirements for a valid annota- tion, including classes, dimensions, source anchoring, and flag semantics. Two mode-specific special- izations adapt this protocol for autonomous drafting (PROTOCOL_AUTONOMOUS.md) and interactive, curator-in-the-loop work. A separate review protocol (REVIEW_PROTOCOL.md, with an interactive variant) governs assertion-by-assertion review of drafted annotations. A labeling-examples docu- ment records worked cases for recurring judgment calls. These written protocols are also packaged as executable skills (section SN3), one for annotation and one for review, so that the same rules can be applied by a human curator or by an AI assistant under curator supervision. SN2.1 Manual-track protocol In the manual track, four of the six publications, Jossin 2017, Davis 2011, Nelson 1992, and Gupta 2015, were annotated by Vladimir Seplyarskiy, a researcher in human and population genetics specializing in human germline mutation and mutational processes. He is an Assistant Professor at UT Southwestern and was formerly affiliated with Harvard Medical School and Brigham and Women’s Hospital. This expertise covers the human-genetics and population-genetics knowledge domains represented in the corpus. For each paper, the curator: 1. read the PDF in full and marked relevant passages with highlights and short callouts; 1 https://github.com/ForomePlatform/genetic-evidence-model/tree/master/protocols 10 2. exported those annotations with a helper script to a structured JSON record containing page, offset, highlighted text, and free-text callout fields; 3. constructed a one-row summary table containing high-level judgments about knowledge do- main, method, credibility, resolution, mode of inheritance, and related dimensions; and 4. mapped the highlights and summary row to the schema, producing a YAML file conforming to the GeneticEvidence model. When this conversion required normalization, the change was recorded inline as a normalization_note with type: ai_normalization, while preserving the original text. For example, the free-text value “localisation and protein expression” was mapped to the enumeration values LOCALIZATION and EXISTENCE. These notes allow later audits to compare the original and normalized values and support subsequent ontology alignment. All manual-track annotations were validated against the SHACL schema (schema/genetic_ evidence.shacl.ttl). Violations, such as a missing always-required dimension or a conditional dimension whose activation condition held but whose value was absent, had to be resolved before an annotation was admitted to the corpus. The four manual annotations are treated as the curator-authored reference annotations for this work. They are not a gold standard in the statistical sense, because the corpus has only one annotator. They are, however, the authoritative reference against which the AI-drafted track is evaluated. SN2.2 AI-drafted-track protocol In the AI-drafted track, an AI assistant annotated two publications, Duerr 2006 and Inouye 2018, directly from their PDFs under the same schema. The four manual annotations served as worked examples. For each paper, the assistant received: • the schema (schema/dimensions.md and the SHACL file); • the four manual annotations; • the source PDF; and • the autonomous-annotation skill (section SN3). The assistant produced draft YAML annotations in the same format as the manual annotations. When it judged a mapping uncertain, or concluded that a judgment required human review, it recorded the issue inline using one of two flag types: • ai_uncertainty, for a mapping judged unreliable and referred for a human decision; and • ai_assumption, for a judgment made without explicit textual support and requiring recon- sideration at review rather than silent acceptance. AI-drafted annotations are not accepted without review. One of the authors, a curator with expertise in genomic data curation and the evidence model rather than in human or population genetics, reviewed each annotation assertion by assertion under the review skill. The review assessed: 11 • faithful representation of the source text; • conformance to the schema, including conditional-activation rules; and • the appropriateness of any ai_uncertainty or ai_assumption flags. Corrections were applied directly to the YAML, and every change and ruling was logged under annotations/reviews/. Table ST2 summarizes the resulting verdicts for each GeneticEvidence item, and section SN4 analyzes them qualitatively. In all cases, the curator, not the model, is the authority for the recorded annotations. SN2.3 Source anchoring and validation A central epistemic commitment of the workflow is that every assertion must be independently auditable against its source. • Each assertion in a GeneticEvidence item carries a source_span containing a page number and a concise key phrase, typically fewer than fifteen words, sufficient to locate the passage by search in the PDF. • The phrase-length limit is intended to respect fair-use norms for copyrighted publisher material while keeping verification practical. • The SHACL schema treats source_span as mandatory. An annotation with any assertion lacking a source anchor is rejected at validation time. Across the six-paper pilot, all 95 assertions carry source spans. This is a property of the valida- tion requirement, not an observed outcome. When a verbatim key phrase could not be reconstructed from a public PDF, for example because a value appeared in a publisher PDF table not included in the repository, the annotation records this explicitly with a source_span_is_paraphrase flag and an explanatory note. Future curators can then refine the anchor against the original file. SN2.4 Provenance and protocol versions Two provenance points are important for interpreting the corpus. First, the pilot corpus described in the main paper was initially annotated under an earlier procedure, “protocol v0.” The suite described here is its subsequent, generalized formalization. The pilot annotations remain valid under the current protocol. Some artifacts, such as slightly different span-selection conventions in the earliest AI-drafted annotations, are recorded as forward- looking notes rather than retrofitted edits. Two such artifacts are field-name changes. Protocol v0 used key_phrase for the source-span phrase and resolution for the target resolution, whereas v1 uses phrase and target_resolution. The corpus retains both spellings. Before RDF validation, the YAML-to-RDF transformation (src/python/main/forome/gem/extraction/yaml_to_rdf.py) normalizes them to the RDF properties gem:phrase and gem:resolution, so gem-validate treats the two spellings identically; the legacy names are retained solely for backward compatibility. Second, the AI-drafted annotations were produced with Claude Opus 4.7 for Duerr and Claude Opus 4.8 for Inouye, and the assertion-by-assertion review was conducted by the human curator interacting with Claude Opus 4.8 under the executable review skill. These implementation details are reported for reproducibility and are not central to the scientific contribution. AI-generated drafts and AI-assisted review suggestions were subject to final adjudication by the human curator; the reviewed annotations are human-approved records, and no AI system was treated as the authority for any scientific or curation decision. 12 SN3 Executable skills Because the term may be unfamiliar outside agent-based AI tooling, we briefly define an executable skill. A skill is a version-controlled package of natural-language instructions and supporting re- sources, such as reference files, templates, and worked examples. An AI assistant loads the package on demand to perform a defined task in a repeatable way [1]. A skill is the executable counterpart of a written protocol. A protocol tells a human reader what a task requires; a skill expresses those requirements in a form that an AI assistant can apply directly. In practice, a skill is a folder containing a top-level instruction file, conventionally SKILL.md, and the resources referenced by that file. The assistant reviews the skills available to it, loads the one relevant to the task, and follows its instructions. Its behavior is therefore governed by the versioned package rather than improvised anew for each session. The general mechanism is not specific to a particular model or vendor, and a cross-vendor specification is emerging. 2 The two skills accompanying this work are implemented as Claude Agent Skills. The annotation skill packages the annotation protocol, and the review skill packages the review protocol (section SN2). They allow an AI assistant to apply the same rules that guide a human curator, under curator supervision. Both skills are published with the protocols in the project repository. 3 Their purpose in this work is reproducibility and reuse, not autonomy. The curator remains the authority for every recorded annotation, as described in section SN2. SN4 AI-track review outcomes This note reports the review outcomes for the two AI-drafted annotations: Duerr et al. 2006, a conventional genome-wide association example, and Inouye et al. 2018, a polygenic-score case that exposes several model limitations. Each draft was reviewed assertion by assertion against REVIEW_PROTOCOL v1.0 by the human curator, who interacted with an AI assistant operating under the review skill. AI-generated drafts and AI-assisted review suggestions were subject to final adju- dication by the human curator; the reviewed annotations are human-approved records, and no AI system was treated as the authority for any scientific or curation decision. The complete review logs, including every flag and ruling, are available in the repository under annotations/reviews/. Table ST2 summarizes the verdicts for the drafted items. PaperDrafted Approved With flags Edited Rejected Added on review Duerr 2006602401 (approved) Inouye 2018311101 (edited) Total913502 Table ST2: Per-paper review verdicts for AI-drafted items. “With flags” denotes approval with non-blocking flags and no edits. “Added on review” denotes evidence omitted from the draft and entered by the reviewer as a new item. The Duerr results reflect the v0 annotation. The Inouye results, and all related analyses in this note, use the v1 re-annotation produced under the protocol refined during the Duerr review. The original five-item v0 Inouye draft is retained in the repository for comparison. 2 https://agentskills.io 3 https://github.com/ForomePlatform/genetic-evidence-model/tree/master/skills 13 The drafts were produced under different protocol versions. The later Inouye draft required fewer edits: two of its three items were accepted without edits and one was edited, whereas four of Duerr’s six items were edited. This comparison is only suggestive, not controlled, because the papers differ in both genre and difficulty. It is nevertheless consistent with the provenance record: all thirty-seven cited Inouye spans were verbatim and within the length limit, whereas the earlier Duerr draft contained notation drift and one incorrect page citation. These issues are retained as forward-looking notes because they predate the current rules. A second pattern concerns omitted evidence. The reviewer-added items were not granular as- sociation findings, which the drafts captured well. Instead, they were higher-level or interpretive items: in Inouye, the paper’s headline claim that combining component scores outperforms each component; and in Duerr, a cited anti-p40 antibody trial supporting the therapeutic-target argu- ment. The drafts were therefore most reliable for specific, directly stated claims and less reliable for synthetic conclusions and cited translational evidence that required cross-item interpretation. Edits to existing items primarily concerned schema-level judgment rather than the factual con- tent of assertions. They included Knowledge Domain assignment, distinguishing family-based from allele-frequency evidence; Method placement, distinguishing targeted replication from a new genome- wide scan; Credibility calibration, lowering the score-combination item from high to medium because the paper gave no stated mechanism; and the applicability of Variant Ascertainment to a polygenic- score target. In the last case, the reviewer marked the dimension as not applicable and recorded the resulting constraint as support for a candidate extension. No item was rejected, and no item was found to misrepresent its source. The revision also affects the promotion record. Because Variant Ascertainment is no longer exercised by Inouye, its promotion rests on its initial proposal in Davis and its independent reuse in Nelson and Duerr, which still satisfies the two-papers rule. Updating the promotion record is a separate curation step and is not undertaken here. Both reviews recommend a second pass over the changed items because each paper gained a new item and several existing items were edited. SN5 Standards crosswalk This section extends the class-level FHIR alignment in the main paper with term-level mappings for GEM classes, properties, dimensions, and selected enumerated values. The mappings cover the Sequence Ontology (SO), the Human Phenotype Ontology (HPO), NCBITaxon, the Evidence and Conclusion Ontology (ECO), SEPIO, and the GA4GH Variant Annotation (VA) model. Each mapping was verified against its source vocabulary. This standards crosswalk is distinct from the UMLS crosswalk in Section SN8.1. It aligns GEM constructs with source vocabularies, whereas that note records concept-level mappings and curation decisions within the UMLS Metathesaurus. The relation column describes the relation between a GEM construct and an external term. Exact indicates the same scope; close a near match with a material scope difference; narrower that GEM is more specific; broader that GEM is more general; related that the external term captures a constituent or associated aspect of the GEM construct; and none that no equivalent has been identified. For structural correspondences that are not term-level equivalences, the tables instead use structural labels (structural sibling, structural parallel). The crosswalk prioritizes semantic fidelity over a fixed preference for one vocabulary. When the most faithful concept is available only in a specialized vocabulary, the crosswalk retains that exact match rather than substituting a broader term from a preferred vocabulary. For example, animal_genetics maps exactly to NCI Thesaurus C1510895 (Animal Genetics), for which MeSH 14 has no corresponding descriptor. The relation field records the nature of the mapping; it does not express a preference among vocabularies. A second source-level issue is sense merging within a single UMLS concept. For variant, UMLS concept C0042333 (Genetic Variation) combines two senses. The MeSH descriptor defines a population-level phenomenon (“genotypic differences observed among individuals in a population”) and determines the concept’s semantic type, Natural Phenomenon or Process. The NCI Thesaurus instead defines an entity, a deviation in the nucleotide sequence of an individual, and places the ACMG classification variants (pathogenic, likely pathogenic, and benign) directly under that concept. Its synonym list also includes Sequence Variant, the exact SO term shown in table ST5. The UMLS mapping is therefore recorded as close, rather than exact, because the concept includes the population-level MeSH sense. We nevertheless accept its semantic type even though it lies outside the expected axis. The entity-level NCI sense is the relevant genomics sense, and UMLS provides no generic variant concept that is typed as an entity: the classified child concepts are typed Nucleotide Sequence, but their generic parent is not. Table ST3: Structural correspondences between GEM and FHIR R5. The table extends the class- level alignment in the main paper with term-level detail. GEM constructExternal termVocab.Relation ScientificEvidence (evidence item) EvidenceFHIR R5 structural sibling; GEM replaces the PICO frame with a genetics-specific dimensional frame EvidenceVariable (definition and measured value) EvidenceVariable + Evidence.statistic FHIR R5 broader; GEM combines the variable definition and measured value in one class Credibility (overall rating slot) Evidence.certaintyFHIR R5 structural parallel; both provide an overall rating with separable subcomponents Table ST4: Crosswalk for the assertion property and the EvidenceAssertion class. The prop- erty records an object-level, source-anchored claim; the class is a meta-predicate over a decision rule. GEM constructExternal termVocab.Relation assertion property (source-anchored claim) StatementGA4GH VA narrower; a close, genetics-specific match assertion property Assertion (SEPIO:0000001) SEPIOnarrower assertion property (false friend) Evidence.assertion (human-readable summary) FHIR R5none; do not map EvidenceAssertion class (meta-predicate over a decision rule) no equivalent—none The assertion property and the EvidenceAssertion class are distinct despite their similar names. The property records the object-level, source-anchored claim made by a paper. Its clos- est structural analogues are the GA4GH VA Statement and the SEPIO Assertion, not FHIR’s Evidence.assertion, which is a human-readable summary. 15 The EvidenceAssertion class is a meta-predicate over a decision rule and is the model’s point of contact with the reasoning layer of [4]. We found no close analogue for this decision-rule meta-predicate in FHIR, GA4GH VA, ECO, or SEPIO. Its lack of an external analogue is therefore expected rather than a coverage failure. A future schema revision will rename the property to remove the naming collision; the pilot annotations will remain unchanged. Table ST5: GEM dimensions and selected values mapped to external vocabularies. Identifiers were verified against their source vocabularies. SO = Sequence Ontology; VA = GA4GH Variant Annotation. GEM dimension / valueExternal term + IDVocab.Relation Target Type: genegene SO:0000704SOexact Target Type: related genegene SO:0000704 (relation recorded in Gene Relation) SOexact Target Type: variantsequence_variant SO:0001060SOexact Target Type: transcripttranscript SO:0000673SOexact Target Type: interval / segment region SO:0000001SOnarrower Mode of Inheritance (axis)Mode of inheritance HP:0000005 HPOexact autosomal dominant HP:0000006HPOexact autosomal recessive HP:0000007HPOexact Organism: humanHomo sapiens NCBITaxon:9606 NCBITaxon exact Organism: mouseMus musculus NCBITaxon:10090 NCBITaxon exact Organism: ratRattus norvegicus NCBITaxon:10116 NCBITaxon exact Organism: zebrafishDanio rerio NCBITaxon:7955NCBITaxon exact Credibility (rating and facets) confidence level (SEPIO:0000187), via has_confidence_level (SEPIO:0000167), on a rated Assertion (SEPIO:0000001) SEPIOexact (rating) / narrower (facets) Method: in vivo (animal model) animal model system study evidence ECO:0000179 ECOexact Method: in vitroin vitro assay evidence ECO:0000181 ECOexact Method: clinicalclinical study evidence ECO:0000180 ECOexact Method: statistical geneticsno clean ECO leaf; broadest parent is computational evidence (ECO:0007672); OBI is a candidate vocabulary for study designs (unconfirmed) ECO / OBI none (ECO) / broader (OBI) The following table lists dimensions for which the selective review reported here identified no single external equivalent. These entries are documented coverage gaps in the present crosswalk, not claims that no relevant term exists in any vocabulary. 16 Table ST6: Coverage gaps in the present crosswalk. The relation is none in every row. GEM dimensionWhy no external term Knowledge Domainno single external vocabulary represents this axis Gene Relation (conditional)GEM-specific gene–gene relations, such as X_regulates_Y and X_has_same_function_as_Y; RO may cover some relations (unconfirmed) Phenotype Scale (ordinal)no external vocabulary provides this ordinal axis; GO biological_process is a possible future alignment for the molecular and cellular end of the scale Variant Ascertainmentno single external term captures this dimension Resolutionrepresented implicitly by the selected SO target-type term, rather than as an independent external axis EvidenceAssertion (meta-predicate) belongs to a reasoning layer beyond the scope of ECO, SEPIO, VA, and FHIR Taken together, the mappings and gaps show two broad patterns. First, the dimensions that describe what the evidence concerns and which organism it concerns map directly to established ontologies. The three experimental Method values also map cleanly. Statistical genetics is the main Method gap: ECO represents evidence types rather than study designs such as genome- wide association studies and linkage analyses, and OBI is a candidate vocabulary for that level of description. Second, this crosswalk identifies no single external equivalent for Knowledge Domain, Phenotype Scale, Variant Ascertainment, Resolution, or the Gene Relation enumeration. These are GEM- specific epistemological axes, rather than failed attempts to duplicate existing vocabularies. Their partial or absent coverage in the reviewed standards constitutes a coverage gap in the present crosswalk and helps motivate a dedicated reference data model. SN6 Case reports: Duerr 2006 and Inouye 2018 The two AI-drafted annotations illustrate contrasting parts of the model’s operating range. The main paper (Section 4) summarizes them; this note provides the item-level detail omitted there for space. Duerr is the more conventional case: six of its seven GeneticEvidence items fit the current schema without strain, and the remaining item raises an open question about translational scope. Inouye tests the limits of the variant-level model: all four items require forced-fit assignments, which motivate six candidate extensions. Assertion-by-assertion review verdicts are given in section SN4; the term-level standards crosswalk is in section SN5; and the complete annotations are available as the repository files annotations/v0/duerr2006.yaml and annotations/v1/inouye2018.yaml. SN6.1 Case Report CR-DU: Duerr et al. 2006 (genome-wide association) Duerr and colleagues identified IL23R as an inflammatory bowel disease gene through a genome- wide association study of non-Jewish patients with ileal Crohn’s disease, followed by replication in a Jewish cohort and family-based transmission testing in 883 nuclear families. Our annotation decomposes the paper into seven GeneticEvidence items (table ST7). 17 Table ST7: Duerr 2006: the seven GeneticEvidence items, with target type, leaf-level Method, and Credibility. IDClaimTargetMethodCred. GE-1rs11209026 protective in non-Jewish ileal CD (discovery)variant gwasHigh GE-2rs11209026 protective in Jewish ileal CD (replication)variant association study High GE-3Gln undertransmitted to affected offspring (FBAT)variant transmission disequilibrium test High GE-4Multiple independent IL23R signals (conditional)variant fine mappingHigh GE-5No association at IL12RB1/IL23A/IL12BgenegwasHigh GE-6Cited mouse-model and functional supportgenein vivo, in vitro High GE-new-1 Cited anti-p40 (IL12B) antibody trialgeneclinical evidence High Six of the seven items fit the current schema without strain. The seventh, GE-new-1, is a cited anti-p40 antibody trial added during the missing-evidence check. It concerns drug-response evidence for which the model has no fitting genetic Knowledge Domain, revealing a translational- scope gap. The discovery item, GE-1, is shown in abridged form below. It illustrates how dimensions, a credibility comment, and source-anchored assertions are recorded. id: GE-1 label: "rs11209026 (Arg381Gln) in IL23R confers protection against ileal CD in non-Jewish discovery cohort" knowledge_domain: [POPULATION_GENETICS] method: [GWAS] target_type: VARIANT target: "rs11209026 (c.1142G>A, p.Arg381Gln) in IL23R" resolution: VARIANT variant_ascertainment: [OBSERVED_IN_CASES, OBSERVED_IN_CONTROLS] phenotype_scale: CLINICAL credibility: HIGH credibility_comment: > Genome-wide significance after Bonferroni correction (corrected P = 1.56e-3); large effect (OR = 0.26); independent replication follows in GE-2. candidate_extension_exercised: CE-DU1 # direction of effect assertions: - id: GE-1.A1 statement: "rs11209026 is genome-wide significantly associated with ileal Crohn’s disease" source_span: page: 1, key_phrase: "rs11209026 (P = 5.05e-9, corrected P = 1.56e-3)" - id: GE-1.A2 statement: "The Gln allele is protective (OR = 0.26)" source_span: page: 2, key_phrase: "glutamine allele appears to protect against development of CD" 18 Candidate extensions. Curator review identified three candidate extensions, all provisional un- der the two-papers promotion rule: • CE-DU1, direction of effect: a first-class dimension distinguishing protective from risk- conferring alleles. In this case, the protective direction is implicit in an odds ratio of 0.26. The distinction may matter for downstream ACMG/AMP-style use. • CE-DU3, cross-item synthesis: a construct for higher-order claims that jointly reference several evidence items. Here, the relevant claim is the paper’s recommendation that blocking the IL-23 pathway is a rational therapeutic strategy. The issue arises twice, in that recommen- dation and in GE-new-1, and therefore also raises a question about the model’s translational scope. • CE-DU4, conditional-activation scope: widen the activation condition for mode_of_ inheritance, mendelian_segregation, exact_variant, subdomain beyond human genet- ics+ gene to also cover human genetics+ variant. GE-3, a family-based transmission test of a specific variant, illustrates this need. The extension would also require a complex mode-of-inheritance value and a family-based subdomain value. A fourth candidate, CE-DU2, proposed a separate dimension for negated assertions. It was retracted on review because an assertion is already a predicate and can express either absence or presence. This retraction is the corpus’s clearest example of the conservative promotion rule preventing a premature addition. SN6.2 Case Report CR-IN: Inouye et al. 2018 (polygenic risk score) Inouye and colleagues developed and externally validated a meta-genomic risk score (metaGRS) for coronary artery disease from 1,745,180 variants in 482,629 UK Biobank participants. The evidential target is a derived, whole-genome score, and the primary measurements are epidemiological, such as hazard ratios and C-indices. The paper therefore exposes limitations of a model whose target types are variant-level. The annotation reported here is the v1 re-annotation produced under the current protocol, which was refined during the Duerr review. The original five-item v0 draft remains in the repository for comparison. The v1 annotation decomposes the paper into four claim-coherent items (table ST8). Each can be represented only through deliberate placeholder assignments, recorded as forced fits and candidate extensions rather than left implicit. Table ST8: Inouye 2018 (v1): four GeneticEvidence items. Each uses the same placeholder Target Type and Resolution value, variant, and is flagged against CE-IN1 and CE-IN2. The actual target is a whole-genome polygenic score. IDClaimTarget † Cred. GE-1Prediction and quintile stratification (HR 1.71/SD; 4.17-fold)variantHigh GE-2Independence from six conventional risk factorsvariantHigh GE-3Stratification persists among medicated individualsvariantHigh GE-new-1 Combining three component scores outperforms eachvariantMedium † Placeholder assignment. The actual target is a polygenic score (CE-IN1) at whole-genome-aggregate resolution (CE-IN2). Method is recorded at the intermediate association study/meta-analysis level because no leaf represents polygenic-score association (flag F-METH). Variant Ascertainment is marked not applicable because a score has no single ascertainment mode; this supports the CE-IN1 SHACL argument. 19 The independence item, GE-2, makes these limitations concrete. Both its target_type and target_resolution are assigned the placeholder value variant and flagged against CE-IN1 and CE-IN2. Variant Ascertainment is marked not applicable because it does not apply cleanly to a score target. Epidemiological measures are recorded in special_considerations and motivate CE-IN4. The item is shown in abridged form below. Protocol v0, used for Duerr, named the fields resolution and key_phrase; protocol v1, used for Inouye, renamed them target_resolution and phrase (section SN2). The corpus retains both forms, which the YAML-to-RDF transformation normalizes to gem:resolution and gem:phrase; gem-validate therefore treats them identically. id: GE-2 label: "metaGRS predicts incident CAD largely independently of six conventional risk factors" knowledge_domain: [POPULATION_GENETICS] method: [ASSOCIATION_STUDY] # intermediate; no leaf fits PGS assoc. (F-METH) target_type: VARIANT # forced; CE-IN1 proposes SCORE target_resolution: VARIANT # forced; CE-IN2 -> WHOLE_GENOME_AGGREGATE variant_ascertainment: not_applicable_or_omitted # CE-IN1 (SHACL); F-VA phenotype_scale: CLINICAL credibility: HIGH # epidemiological measures (C-index, HR) in special_considerations; CE-IN4 assertions: - id: GE-2.A1 statement: "metaGRS has a higher C-index than any single factor" source_span: page: 6, phrase: "higher C-index (C = 0.623; 95% CI: 0.615 to 0.630)" - id: GE-2.A3 statement: "Adding metaGRS to all six factors raised C-index to 0.696" source_span: page: 6, phrase: "C-index of 0.696 (95% CI: 0.688 to 0.703)" Candidate extensions. The Inouye annotation identified six candidate extensions, the most of any paper in the corpus: • CE-IN1, score target type: a metaGRS is neither a gene nor a variant, but a derived composite predictor. A score value, possibly characterized as polygenic, composite, or both, would represent the target directly. Under the current variant placeholder, all four items mark Variant Ascertainment as not applicable. The schema accepts this as an explicit escape, but a first-class score target would remove the forced fit. • CE-IN2, whole genome aggregate resolution: the score aggregates variants across the genome non-contiguously, which the current contiguous or pointwise Resolution values do not capture. • CE-IN3, target-composition dimension: an axis orthogonal to Resolution that distin- guishes single, aggregate, and composite targets. This provides a complementary fram- ing of the same underlying gap as CE-IN2. • CE-IN4, epidemiological measurement targets: an extension of measurement_target, which is currently oriented toward molecular assays, with values such as hazard ratio, 20 odds ratio, c-index, and auroc. These values would be activated for population-genetics evidence. • CE-IN5, structured cohort descriptor: a replacement for free-text cohort properties with named fields for size, ancestry composition, follow-up, and ascertainment. Such a structure could be reused across Duerr, Davis, and future cohort studies. • CE-IN6, natural versus derived-artifact target: a flag, or a natural/derived di- mension, indicating that the metaGRS is a human-designed construct rather than a natural kind. SN7 Dimension coverage over the annotated corpus This note reports how often each dimension is populated across the 28 GeneticEvidence items in the pilot corpus and how often each conditional dimension is applicable under its activation rules. The counts are computed from the released YAML annotations by scripts/compute_coverage.py. A per-paper report is available in the repository at annotations/coverage.md. 4 Table ST9: Dimension population across 28 GeneticEvidence items from six papers (Duerr v0, Inouye v1), computed by compute_coverage.py. The left block lists always-required dimensions; the right block lists conditional dimensions and the number of items in which their activation conditions apply. Always-required dimensionsConditional dimensions DimensionPopulated Applicable DimensionPopulated Applicable Knowledge Domain2728Variant Ascertainment1418 Method2828Mode of Inheritance01 Target Type2828Mendelian Segregation01 Resolution2828Penetrance24 Credibility2828Measurement Target1011 Phenotype Scale2828Gene Relation111 Organism88 Knockout Type16 The always-required dimensions are populated in all 28 items by construction, except Knowledge Domain, which is populated in 27. The exception is a cited anti-p40 antibody trial in Duerr, which concerns drug-response evidence with no fitting genetic Knowledge Domain. Among the conditional dimensions, Variant Ascertainment is populated in 14 of the 18 items for which its activation condition holds. The four Inouye v1 items use variant as a placeholder Target Type but mark Variant Ascertainment as not applicable, because a polygenic score has no single ascertainment mode. This is the CE-IN1 gap. Well-calibrated conditions such as Organism, populated in 8 of 8 applicable items, and Measurement Target, populated in 10 of 11, are generally populated whenever their conditions hold. In contrast, Gene Relation, populated in 1 of 11 applica- ble items, and Knockout Type, populated in 1 of 6, expose enumeration and over-broad-activation gaps that motivate candidate extensions. Mode of Inheritance and Mendelian Segregation are each populated in 0 of 1 applicable items. The single human genetics+ gene item, Davis’s case–control burden test, addresses both as 4 https://github.com/ForomePlatform/genetic-evidence-model/blob/master/annotations/coverage.md 21 not applicable. This indicates that the current activation condition is broader than the context in which those dimensions are meaningful. SN8 UMLS crosswalk and illustrative OMOP CDM positioning SN8.1 UMLS concept crosswalk (generated) This crosswalk binds Genetic Evidence Model dimensions to Unified Medical Language System (UMLS) Semantic Types and maps enumerated values to UMLS concepts where an adequate mapping is available. It is produced by the mapping harness (gem-umls-crosswalk), which re- solves each term against the UMLS Metathesaurus through the UTS REST API or against a locally indexed licensed copy, and by curator adjudication in the accompanying curation studio (gem-mapping-studio). The machine-readable sources are data/umls/umls_crosswalk.yaml (the harness output) and data/umls/adjudications.yaml (the curator’s decisions). The narrative be- hind the decisions is kept in the repository’s data/umls/DECISIONS.md. This UMLS crosswalk is distinct from the structural standards crosswalk in Section SN5 and does not replace it. A concept is reported as mapped only when the harness retrieved it and it was confirmed against UMLS. No concept identifier is asserted unless it was returned by a query and confirmed against UMLS. Each dimension axis is bound to a UMLS Semantic Type. The type scopes the search for that dimension’s values; a curator may deliberately accept a concept outside the axis and records why. Each dimension value is mapped to a Metathesaurus concept. The relation column records how the accepted concept relates to the GEM token: exact, close, narrower (the GEM token is more specific than the concept), broader (the GEM token is more general), or related (a constituent or operator concept anchors a composite token). An unmapped value is an adjudicated conclusion, not an absence. Relation none means that no faithful UMLS mapping was accepted. The verdict is supported by rejected near-miss concepts, each assigned the adequacy criterion it failed, and by the documented search protocol that found no proxy. Where a concept identifier appears in connection with an argued gap, it identifies a retrieved or rejected near miss, not an accepted mapping. The credibility tiers are argued in full in Section SN9. Coverage. The model has 17 dimension axes; 14 are bound to a UMLS Semantic Type, and 3 boolean or free-text dimensions take no type by design. Of the 71 enumerated values, 62 carry an accepted mapping and 9 are argued gaps with a structured rationale (method: FAMILY_BASED, TRANSMISSION_DISEQUILIBRIUM_TEST, SEGREGATION_ANALYSIS; credibility: VERY_HIGH, HIGH, MEDIUM, LOW; measurement_target: EXISTENCE; gene_relation: X_has_same_function_as_Y). Table ST10: GEM dimensional vocabulary mapped to UMLS concepts (generated by the mapping harness and curator adjudication). GEM valueCUI / TUI Relation UMLS concept or semantic type knowledge_domain (axis)T090—Occupation or Discipline (semantic type, tree A2.6) HUMAN_GENETICSC0020124 exactHuman Genetics (Biomedical Occupation or Disci- pline; MSH) ANIMAL_GENETICSC1510895 exactAnimal Genetics (Biomedical Occupation or Disci- pline; NCI) 22 GEM valueCUI / TUI Relation UMLS concept or semantic type POPULATION_GENETICSC0017404 exactGenetics, Population (Occupation or Discipline; MSH) COMPARATIVE_GENOMICSC4704937 exactComparative Genomics (Biomedical Occupation or Discipline; MSH) EPIGENETICSC1655731 exactstudy of epigenetics (Biomedical Occupation or Dis- cipline; MTH) GENE_FUNCTIONC0872316 exactFunctional Genomics (Biomedical Occupation or Discipline; MTH) MODEL_ORGANISMC0599779 exactAnimal Model (Animal; MTH) method (axis)T062—Research Activity (semantic type, tree B1.3.2) STATISTICAL_GENETICSC2717898 narrower Biostatistics (Research Activity; MTH) ASSOCIATION_STUDYC2717878 exactGenetic Association Studies (Molecular Biology Re- search Technique; MSH) GWASC2350277 exactGenome-Wide Association Study (Molecular Biol- ogy Research Technique; MSH) CANDIDATE_GENE_STUDYC2717881 exactCandidate Gene Identification (Molecular Biology Research Technique; MSH) FINE_MAPPINGC1517888 closeLinkage Disequilibrium Mapping (Molecular Biol- ogy Research Technique; MTH) FAMILY_BASED—noneno faithful concept; argued gap (search protocol recorded) LINKAGE_ANALYSISC0796345 exactgenetic linkage analysis (Laboratory Procedure; MTH) TRANSMISSION_DISEQUILIBRIUM_TEST —noneno faithful concept; argued gap (2 rejected near- misses) SEGREGATION_ANALYSIS—noneno faithful concept; argued gap (3 rejected near- misses) META_ANALYSISC0920317 exactMeta-Analysis (statistical procedure) (Research Ac- tivity; MTH) EXPERIMENTC0678323 exactexperimentation (Research Activity; MTH) IN_VIVOC0681829 exactin vivo study (Research Activity; MTH) IN_VITROC0681828 exactin vitro study (Research Activity; MTH) IN_SILICOC0009609 exactComputer simulation (Machine Activity; MTH) BIOINFORMATICS_INFERENCEC0376528 narrower Computational Biology (Biomedical Occupation or Discipline; MTH) CLINICAL_EVIDENCEC0008972 exactClinical Research (Research Activity; MTH) target_type (axis)T028—Gene or Genome (semantic type, tree A1.2.3.5) GENEC0017337 exactGenes (Gene or Genome; MTH) RELATED_GENEC0332281 relatedAssociated with (Qualitative Concept; MTH) VARIANTC0042333 closeGenetic Variation (Natural Phenomenon or Pro- cess; MTH) SEGMENTC0678933 closeGenetic Loci (Gene or Genome; MTH) INTERVALC1707511 relatedCoordinate on an axis or a grid (Spatial Concept; MTH) TRANSCRIPTC1519595 closeRNA Transcript (Nucleic Acid, Nucleoside, or Nu- cleotide; MTH) resolution (axis)T082—Spatial Concept (semantic type, tree A2.1.5) WINDOWC0678933 closeGenetic Loci (Gene or Genome; MTH) GENEC1708726 exactGene Locus (Spatial Concept; MTH) FUNCTIONAL_ELEMENTC1517495 closeGene Feature (Gene or Genome; MTH) POSITIONC1707511 exactCoordinate on an axis or a grid (Spatial Concept; MTH) 23 GEM valueCUI / TUI Relation UMLS concept or semantic type VARIANTC0042333 closeGenetic Variation (Natural Phenomenon or Pro- cess; MTH) credibility (axis)T080—Qualitative Concept (semantic type, tree A2.1.2) VERY_HIGHC0205250 noneno faithful concept; argued gap (11 rejected near- misses) HIGHC0205250 noneno faithful concept; argued gap (11 rejected near- misses) MEDIUMC1561547 noneno faithful concept; argued gap (11 rejected near- misses) LOWC0205251 noneno faithful concept; argued gap (11 rejected near- misses) phenotype_scale (axis)T071—Entity (semantic type, tree A) MOLECULARC0567416 relatedMolecule (Substance; MTH) CELLULARC0007634 relatedCells (Cell; MTH) HISTOLOGICALC0040300 relatedBody tissue (Tissue; MTH) ORGANISMALC0029235 relatedOrganism (Organism; MTH) CLINICALC0037088 relatedSigns and Symptoms (Sign or Symptom; MTH) variant_ascertainment (axis)T062—Research Activity (semantic type, tree B1.3.2) OBSERVED_IN_CASESC0007328 relatedCase-Control Studies (Research Activity; MTH) OBSERVED_IN_CONTROLSC0007328 relatedCase-Control Studies (Research Activity; MTH) FROM_DATABASEC0993637 exactPublished Database (Intellectual Product; MTH) SYNTHETICC0598279 exactsynthetic construct (Nucleic Acid, Nucleoside, or Nucleotide; MTH) mode_of_inheritance (axis)T045—Genetic Function (semantic type, tree B2.2.1.1.4.1) autosomal_dominantC0443147 exactAutosomal dominant inheritance (Genetic Func- tion; MTH) autosomal_recessiveC0441748 exactAutosomal recessive inheritance (Genetic Function; SNOMEDCT_US) x_linked_dominantC1847879 exactX-linked dominant inheritance (Finding; MTH) x_linked_recessiveC1845977 exactX-linked recessive inheritance (Finding; MTH) mitochondrialC0887941 exactMitochondrial Inheritance (Genetic Function; MTH) penetrance (axis)T045—Genetic Function (semantic type, tree B2.2.1.1.4.1) completeC5826795 exactTypified by complete penetrance (Genetic Func- tion; HPO) incompleteC1836598 exactTypified by incomplete penetrance (Finding; MTH) unknownC0439673 closeUnknown (Qualitative Concept; MTH) measurement_target (axis)T044—Molecular Function (semantic type, tree B2.2.1.1.4) EXISTENCEC1159832 noneno faithful concept; argued gap (3 rejected near- misses) EXPRESSIONC0017262 exactGene Expression (Genetic Function; MTH) STABILITYC2350440 closeProtein Stability (Qualitative Concept; MTH) BINDINGC0033618 exactProtein Binding (Molecular Function; MSH) LOCALIZATIONC0597704 closeprotein localization activity (Cell Function; MTH) ACTIVITYC1148560 narrower molecular_function (Molecular Function; GO) CATALYSISC0243102 exactenzyme activity (Molecular Function; MTH) 24 GEM valueCUI / TUI Relation UMLS concept or semantic type gene_relation (axis)T045—Genetic Function (semantic type, tree B2.2.1.1.4.1) X_has_same_function_as_Y—noneno faithful concept; argued gap (search protocol recorded) X_regulates_YC0017263 exactGene Expression Regulation (Genetic Function; MTH) X_inhibits_YC2611924 relatednegative regulation of gene expression (Genetic Function; GO) organism (axis)T001—Organism (semantic type, tree A1.1) knockout_type (axis)T063—Molecular Biology Research Technique (semantic type, tree B1.3.2.1) CONDITIONALC0814041 exactconditional gene knockout technology (Molecular Biology Research Technique; AOD) UNCONDITIONALC0599772 narrower Gene Knockout Techniques (Molecular Biology Re- search Technique; MTH) subdomain (axis)T063—Molecular Biology Research Technique (semantic type, tree B1.3.2.1) GWASC2350277 exactGenome-Wide Association Study (Molecular Biol- ogy Research Technique; MSH) Linkage StudyC0796345 exactgenetic linkage analysis (Laboratory Procedure; MTH) WGS-WES StudyC3640076 broader Whole Genome Sequencing (Molecular Biology Re- search Technique; MSH) Candidate Gene StudyC2717881 exactCandidate Gene Identification (Molecular Biology Research Technique; MSH) mendelian_segregation (axis)—no semantic type (boolean / free-text dimension) genetic_background_considered (axis)—no semantic type (boolean / free-text dimension) environmental_factors (axis)—no semantic type (boolean / free-text dimension) SN8.2 OMOP CDM positioning (illustrative) The following worked examples position GEM dimensions relative to the OHDSI OMOP Com- mon Data Model (CDM) and its standardized vocabularies. This is an illustrative, hand-authored positioning exercise, not a full crosswalk. OMOP CDM is built around observational, patient- level clinical data, including conditions, drugs, measurements, and observations. GEM describes basic-science genetic evidence about gene–phenotype relationships, much of which has no native patient-level representation in the CDM. 25 Table ST11: Illustrative positioning of GEM dimensions against the OMOP CDM (hand-authored, not exhaustive). GEM dimensionOMOP domain / table Standard vocab. Alignment and gap phenotype_scale (CLINICAL) Condition condition_occurrence SNOMED CT A clinical-scale phenotype aligns with an OMOP Condition concept (SNOMED CT), the CDM’s native representation of a clinical finding. Gap: Molecular, cellu- lar, histological and organismal phenotype scales have no Condition-domain repre- sentation; OMOP has no phenotype-scale axis. measurement_target (EXPRESSION, ACTIVITY, ...) Measurement measurement LOINCA quantitative molecular readout is con- ceptually a Measurement; LOINC covers some assay readouts (e.g. enzyme activ- ity panels). Gap: Basic-science functional readouts (binding, localization, stability of a specific gene product in an experimental system) are rarely represented as standard LOINC Measurement concepts. target_type (GENE) / reso- lution (GENE) Observation/ (Genomic CDM extension) observation (or OMOP Genomic CDM) HGNC (via the OMOP Genomic vocabulary effort) A gene identity aligns with HGNC; the OHDSI Genomic CDM extension is the locus where gene- and variant-level data are being standardized. Gap: The base OMOP CDM has no first-class gene or vari- ant entity; this lives only in the evolving Genomic CDM extension, which is not yet a stable standard. organism(none) (none) (none in CDM; NCBITaxon outside it) No alignment. OMOP CDM is patient- centric (Homo sapiens implied) and has no concept of a model organism. Gap: Model- organism evidence is entirely out of scope for the CDM; GEM’s organism dimension maps to NCBITaxon (Section SN5), not OMOP. method(GWAS, ASSOCIA- TION_STUDY, ...) (metadata/ study-level, not CDMclinical domains) (none; study de- sign is not a CDM patient-data domain) (none; partial coverage in MeSH/OBI outside OMOP) No direct alignment.Study design / evidence-generation method is not a patient-level CDM domain. Gap: The CDM records observations, not how up- stream genetic evidence was generated; GEM’s method hierarchy has no OMOP counterpart. SN9 Credibility tiers: definitions, the UMLS finding, and external alignment SN9.1 What the tiers mean The overall credibility rating is the model’s only subjective core dimension. Its four levels are defined by defeasibility: what would lead a curator to stop accepting an evidence item. They are not defined by study design or by the certainty expressed by the authors: 26 very_high accepted by default even in the face of direct contradiction; reconsidered only in light of a stronger opposing argument; high accepted unless directly contradicted by evidence of comparable strength; such a contradic- tion calls the item into question; medium not accepted on its own; accepted when corroborated by an independent source or an orthogonal type of evidence; low not relied upon; recorded pending independent replication. The distinction between high and very_high is deliberate. Evidence of comparable strength is sufficient to call a high item into question, whereas a very_high item remains the default position unless the opposing case is stronger. This definition is a belief-revision policy. It therefore corresponds to SEPIO’s confidence at- tribute: the degree of belief that an asserted proposition is true [11]. The credibility facets, including cohort size, replication, multiple-testing control, ancestry control, and ascertainment, inform that judgment but remain separate. This separation is inspired by GRADE’s general practice of assessing the factors that affect certainty separately [8]. GEM does not adopt the GRADE framework or its rating rules, which were developed for bodies of clinical evidence, particularly evidence from randomized and observational studies. The GEM facets instead support credibility judgments about individual evidence items from basic and pre-clinical research. The scale intensifies at the top, as ACMG evidence strength (supporting to very strong) [13] and ClinGen gene–disease validity (limited to definitive) [12] do. GRADE certainty instead extends downward to very low, because it rates a body of evidence by downgrading from a starting level determined by study design. The GEM and GRADE scales therefore address related but non- equivalent constructs; their alignment in table ST13 is deliberately non-positional. SN9.2 Why the tiers have no UMLS proxy The four credibility tiers have no adequate UMLS proxy. This is a curation conclusion, not merely a missing mapping: the tiers were tested in a pre-registered sweep (data/umls/sweeps/credibility.yaml). The adequacy criteria were fixed before the search. A candidate concept had to (A) denote a degree of epistemic warrant rather than the magnitude of a measured property; (B) represent a scale point rather than the scale dimension itself; (C) belong to a set that is disjoint and ordered in the same way as the GEM scale; and (D) carry that meaning in the relevant domain rather than as a homonym. The sweep used a local index of UMLS 2026A and comprised 174 queries that returned 497 distinct concepts. The results were replicated through the UTS REST API, which returned 505 concepts. The search had three passes: whole-string searches for intensified and degree qualifiers; searches for the names of evidence-grading vocabularies; and searches for the same terms within the UMLS semantic types most likely to contain scale values: Qualitative Concept, Quantitative Concept, Idea or Concept, Intellectual Product, and Finding. Each retrieved concept was assigned the criterion it failed (table ST12); the full generated table is available in the results file. 27 Table ST12: Concepts retrieved by the credibility sweep (UMLS 2026A), grouped by the adequacy criterion they failed. Criterion codes are defined in the text. Concept familyn Criterion failed Examples “grade” (tumour, school, or product)202DAstrocytoma, low grade (C1314694) LOINC/PROMIS self-efficacy items191DCurrent level of confidence I can drive a car (C5142440) SNOMED degree qualifier values21AVery high (C0442804); Extremely high (C5787630) clinical “evidence of” expressions (presence of a sign) 20DBlood alcohol level of less than 20 mg/100 ml (C0481372); Blood alcohol level of 20–39 mg/100 ml (C0481373) qualifiers denoting the degree of a measured property 12AObservation Value - High (C1561957); Message Waiting Priority - High (C1561958) scores, probabilities, risk categories8AIPSS-R Risk Category Very High (C5202916); IPSS Risk Category High (C4522209) patient-perceived credibility of information 7DEstablish teacher credibility, as appropriate (C0511083); credibility (C0870373) NCI/PDQ Level of Evidence I–IV6ALevel of Evidence (C0393009); Level of Evidence I (C2986612) metabolite-identification confidence levels 6DLevel 1 Metabolite Identification Confidence (C5551417); Level 3 Metabolite Identification Confidence (C5551419) SNOMED diagnostic-certainty modality5CDefinite (C0439544); Probable diagnosis (C0332148) NCI confidence answer set4DVery High Confidence (C4331500); High Confidence (C4330261) psychological self-confidence2DEuphoric mood (C0235146); Confidence level (self-esteem) (C0518578) lexical collisions (“media”, “lower”)2DCommunications Media (C0009458); Culture Media (C0010454) degree of a measured property2A Medium Amount of Time (C4522282); A Medium Amount (C4522283) NCI adverse-event causality ladder2DDefinitely Related to Intervention (C1704787); Possibly Related to Intervention (C1705910) LOINC molecular-pathology “level of evidence” 2DLevel of evidence:Find:Pt:Bld/Tiss:Nom (C4760312) diagnostic-certainty rating items2DUHDRS 1999 Version - Diagnosis Confidence Level (C4744064); Level of Diagnostic Certainty (C5909371) TNM certainty factor2DDegree of certainty of TNM classification (C0456878); Qualifier for TNM degree of certainty (C1302640) statistical confidence measures1ALevel of Confidence (statistical) (C5702556) Four families merit explicit discussion because they are the closest constructs retrieved from UMLS. First, the SNOMED qualifier values High (C0205250, SNOMED 75540009), Very high (C0442804), 28 and Low (C0205251), together with related terms, occur under Increased and Decreased. The source terminology therefore treats them as degrees of a measured property, not as degrees of epistemic warrant (A). Second, NCI Thesaurus Level of Evidence I–IV is a study-design hierarchy. Mapping GEM credibility to it would collapse the model’s separate method and credibility dimensions (A, and C for the scale). Third, the NCI Very High / High / Moderate / Low Confidence set has the same four-level, top-intensified form as the GEM scale, but it is a Clinical or Research Assessment Answer: a questionnaire response that records a respondent’s confidence in performing an activity (D). The SNOMED diagnostic-certainty terms definite, probable, and possible are the closest epistemic candidates. They express the likelihood that a proposition is true, but they form a three-level scale with no intensified top category, apply in their source terminologies to diagnostic assertions, and assess likelihood rather than defeasibility (C, D). The same limitation applies to the TNM certainty factor, the LOINC molecular-pathology level of evidence, which is an AMP/ASCO/CAP actionability tier for variant reports, and metabolite-identification confidence levels. These are ordered epistemic scales, but each is defined for a different object. Lexical post-coordination does not supply the missing identifiers. Combining a qualifier concept with an intensifier from the SPECIALIST Lexicon (LRMOD) would add only a lexical modifier. The lexicon’s intensifier class is a syntactic category that includes very, slightly, about, and ago; it encodes neither magnitude nor direction. Such an identifier therefore adds no semantic content beyond the corresponding text string. This is a conceptual argument about the role of identifiers, not an empirical finding. SN9.3 External alignment The following table gives a conceptual, non-positional alignment between the GEM tiers and selected external scales. It does not define equivalences or a common numerical conversion. In particular, the FHIR values are export codes: very_high must be represented as high, and that export is therefore lossy. Table ST13: Conceptual, non-positional alignment of the GEM credibility tiers with selected exter- nal scales. “—” indicates that no counterpart is proposed. The FHIR column gives the lossy export used when a certainty-rating code is required [9]. CIO = Confidence Information Ontology [3]. GEMGRADE certainty [8] FHIR certainty-rating CIO confidence level [3] ACMG evidence strength [13] ClinGen validity [12] very_high — (GRADE caps at high) high (lossy export) high (partial)very strongdefinitive highhigh highhighstrongstrong mediummoderate moderatemediummoderatemoderate lowlow lowlow (“absence of trust”, closer to none) supportinglimited —very low very-low— These alignments are part of the GEM model, not of the UMLS crosswalk. The crosswalk records UMLS mappings, or a considered conclusion that no adequate mapping exists. 29 Supplementary Tables Table ST14: The six annotated publications and the epistemic shape each exercises. Manual annota- tions form the curator-authored reference set; AI-drafted annotations are evaluated against curator review. GE abbreviates GeneticEvidence; the “GE items” column counts how many evidence items each paper yields under our claim-level decomposition. PaperSourceGE items Credibility RoleNotes Jossin 2017 (Llgl1)manual6Very high molecular mechanismmulti-scale phenotype; motivates pheno- type_scale Davis 2011 (TTC21B)manual6Highbreadth exemplarmotivates vari- ant_ascertainment Nelson 1992 (CD18)manual3Mediumclassical molecular genetics single-gene functional; one can- didate (CE-N1) resolved by external reference Gupta 2015 (ATP6AP2) manual2Lowlow-credibility edge caseschema fits cleanly Duerr 2006 (IL23R)AI-drafted 7Highclean GWAS exemplarfeatured in the main text (v0) Inouye 2018 (metaGRS) AI-drafted 4High/Med model-extension stress test featured in the main text (v1) Table ST15: Credibility facets in the GeneticEvidence model. Credibility records methodological soundness only; each facet is rated separately rather than collapsed to one score, and the facets are registered as genetics-specific subcomponents anchored on SEPIO confidence. The overall rating’s levels are analogous in spirit to ClinGen gene-disease validity classifications [12], which however aggregate per gene-disease relationship rather than per assertion. Effect magnitude is not a credi- bility facet: it is a measured value carried on the GeneticEvidenceVariable. Phenotype fit to a clinical question is applicability, captured by Relevance and Specificity. FacetAspect of methodological soundness recorded sample_sizeSize of cohort, pedigree, or sample (statistical power) ascertainmentCase/control makeup and selection of the studied sample multiple_testingWhether and how multiplicity was controlled stratification_control Ancestry restriction or matching, and residual-confounding control replicationIndependent replication or internal resampling 30 Table ST16: Field-level alignment of the core model to FHIR Evidence R5. The certainty rows expand the Credibility entry of the class-level alignment table in the main paper. This workFHIR Evidence R5Note ScientificEvidence item EvidenceOne evidence conceptper claim dimensions Evidence.variableDefinition (roles)PICO roles re- placed by genetic dimensions EvidenceVariable definition EvidenceVariable.characteristicWhat the vari- able is and how measured EvidenceVariable value Evidence.statisticThe measured quantity itself Credibility overall rating Evidence.certainty.ratingCoded ordinal rating anchored on SEPIO confi- dence Credibility facet Evidence.certainty.certaintySubcomponent.type Extension to the subcomponent- type value set facet rating (where graded) ...certaintySubcomponent.ratingPer-domain rat- ing source_span Evidence.note / CitationPage plus key phrase for audit References [1] Anthropic. Agent Skills. https://platform.claude.com/docs/en/agents-and-tools/ agent-skills/overview, 2025. Accessed 2026. [2] Anita Bandrowski, Ryan Brinkman, Mathias Brochhausen, et al. The ontology for biomedical investigations. PLoS ONE, 11(4):e0154556, 2016. doi: 10.1371/journal.pone.0154556. [3] Frederic B. Bastian, Marcus C. Chibucos, Pascale Gaudet, Michelle Giglio, Gemma L. Holliday, Hong Huang, Suzanna E. Lewis, Anne Niknejad, Sandra Orchard, Sylvain Poux, Nives Skunca, and Marc Robinson-Rechavi. The confidence information ontology: a step towards a standard for asserting confidence in annotations. Database, 2015:bav043, 2015. doi: 10.1093/database/ bav043. [4] Michael Bouzinier, Sergey Trifonov, Michael Chumack, Eugenia Lvova, and Dmitry Etin. Trust- worthy clinical decision support using meta-predicates and domain-specific languages, 2026. URL https://arxiv.org/abs/2604.21263. [5] Werner Ceusters and Barry Smith. Aboutness: Towards foundations for the Information Artifact Ontology. In Proceedings of the Sixth International Conference on Biomedical On- tology (ICBO 2015), volume 1515 of CEUR Workshop Proceedings, pages 1–5, 2015. URL https://ceur-ws.org/Vol-1515/regular10.pdf. [6] Michelle Giglio, Rebecca Tauber, Suvarna Nadendla, et al. ECO, the Evidence and Conclusion 31 Ontology: Community standard for evidence information. Nucleic Acids Research, 47(D1): D1186–D1194, 2019. doi: 10.1093/nar/gky1036. [7] Global Alliance for Genomics and Health. GA4GH Variant Annotation (VA) Specification. https://va-spec.ga4gh.org/en/1.0/, 2024. VA-Spec v1.0. [8] Gordon H. Guyatt, Andrew D. Oxman, Gunn E. Vist, Regina Kunz, Yngve Falck-Ytter, Pablo Alonso-Coello, and Holger J. Schünemann. GRADE: an emerging consensus on rating quality of evidence and strength of recommendations. BMJ, 336(7650):924–926, 2008. doi: 10.1136/ bmj.39489.470347.AD. [9] HL7 International. FHIR Evidence, EvidenceVariable, and EvidenceReport Resources. https: //hl7.org/fhir/R5/evidence.html, 2023. FHIR Release 5 (v5.0.0), published 26 March 2023. [10] Sebastian Köhler, Michael Gargano, Nicolas Matentzoglu, et al. The Human Phenotype Ontol- ogy in 2021. Nucleic Acids Research, 49(D1):D1207–D1217, 2021. doi: 10.1093/nar/gkaa1043. [11] Monarch Initiative. SEPIO: Scientific Evidence and Provenance Information Ontology. https: //github.com/monarch-initiative/SEPIO-ontology, 2026. Accessed April 26, 2026. [12] Heidi L. Rehm, Jonathan S. Berg, Lisa D. Brooks, et al. ClinGen—The Clinical Genome Re- source. New England Journal of Medicine, 372:2235–2242, 2015. doi: 10.1056/NEJMsr1406261. [13] Sue Richards, Nazneen Aziz, Sherri Bale, et al. Standards and guidelines for the interpretation of sequence variants: A joint consensus recommendation of the American College of Medical Genetics and Genomics and the Association for Molecular Pathology. Genetics in Medicine, 17(5):405–424, 2015. doi: 10.1038/gim.2015.30. [14] Ryan J. Schmidt, Marcie Steeves, Pinar Bayrak-Toydemir, et al. Recommendations for risk allele evidence curation, classification, and reporting from the ClinGen Low Penetrance/Risk Allele Working Group. Genetics in Medicine, 26(3):101036, 2024. doi: 10.1016/j.gim.2023. 101036. [15] Natasha T. Strande, Erin Rooney Riggs, Adam H. Buchanan, et al. Evaluating the clinical validity of gene-disease associations: An evidence-based framework developed by the Clinical Genome Resource. American Journal of Human Genetics, 100(6):895–906, 2017. doi: 10.1016/ j.ajhg.2017.04.015. [16] Nicole A. Vasilevsky, Sabrina Toro, Nicolas Matentzoglu, et al. Mondo: Integrating disease terminology across communities. Genetics, 232(4):iyaf215, 2026. doi: 10.1093/genetics/iyaf215. [17] W3C. PROV-O: The PROV ontology. https://w.w3.org/TR/prov-o/, 2013. [18] Alex H. Wagner, Lawrence Babb, Gil Alterovitz, et al. The GA4GH variation representation specification: A computational framework for variation representation and federated identifica- tion. Cell Genomics, 1(2):100027, 2021. doi: 10.1016/j.xgen.2021.100027. [19] Hannah Wand, Samuel A. Lambert, Cecelia Tamburro, et al. Improving reporting standards for polygenic scores in risk prediction studies. Nature, 591(7849):211–219, 2021. doi: 10.1038/ s41586-021-03243-6. 32