Pith. sign in

REVIEW 3 major objections 6 minor 20 references

Koza and Koza-Hub for born-interoperable knowledge graph generation using KGX

T0 review · 3 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Koza reduces knowledge-graph building to three primitives and a YAML file.

desk verdict A real, working KG ingest pipeline with overclaimed breadth and unsupported generalizability; the core artifact deserves a serious referee. read the letter →

arxiv 2509.09096 v1 pith:BIT6WID2 submitted 2025-09-11 cs.DB

Daniel R Korn , Patrick Golden , Aaron Odell , Katherina Cortes , Shilpa Sundar , Kevin Schaper , Sarah Gehrke , Corey Cox
show 5 more authors
Harry Caufield Justin Reese Evan Morris Christopher J Mungall Melissa Haendel
This is my paper · ORCID
classification cs.DB
keywords knowledgegraphconstructionKGXBiolinkmodelbiomedicaldataintegrationmodularingesttransformationpipelineKozaharmonization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper introduces Koza, a Python package that treats every knowledge-graph ingest as the same three steps—read, transform, write—configured through YAML and a small decorator API, with output forced into the KGX/Biolink standard. The authors argue this makes graph construction modular, testable, and reusable: instead of every project writing bespoke parsers for the same biomedical sources, a shared hub (Koza-Hub) publishes per-source ingests that anyone can feed into their own graph build. If the approach holds, the redundant labor of converting raw data into graph format shrinks to selecting pre-built ingests, and graphs become 'born interoperable' because they share a schema. Sympathetic readers should care because the barrier to assembling a custom biomedical knowledge graph drops from a multi-month engineering effort to a configuration exercise.

What carries the argument

The load-bearing mechanism is the reader–writer–transform primitive trio. Readers parse structured files declaratively (CSV/JSON and filter expressions); writers emit KGX-formatted node and edge files; transforms are Python functions selected by the @transform_record or @transform decorator, receiving a KozaTransform object whose write, state, and log methods carry records through the pipeline. YAML configuration declares readers, writers, and file paths, and the decorator API enforces that every output record conforms to the Biolink model. The same mechanism makes per-source ingests modular: each ingest is a standalone repository with its own tests, and Koza Merge concatenates chosen ingest

What would settle it

Pick a biomedical source with a structure not representable by the shipped readers and decorator API—say, an Excel workbook with multiple interlinked sheets or an XML stream requiring state across records—and write an ingest using only Koza's public interfaces. If the source cannot be ingested without patching Koza internals, the claim that ingests reduce to these primitives fails.

Watch

Extended reading notes

Core claim

The central claim is that KG construction can be decomposed into reader, writer, and transform primitives, with the transform expressed as a decorated Python function that writes Biolink-compliant records through a KozaTransform object. Koza implements this and, together with the Koza-Hub collection of modular ingests, provides conversion processes for thirty gold-standard biomedical data sources into KGX-format artifacts. To the authors' knowledge, it is the first tool that combines modularity, developer simplicity, declarative parameters, a standard biomedical data model (KGX/Biolink), and a standalone Python package. The discovery is that this combination is sufficient to build real graph

Load-bearing premise

The architecture assumes every source can be squeezed into the current reader set plus Python transform hooks; the paper demonstrates 14 repositories (25 artifacts) while the abstract claims 30 sources, so the gap between demonstrated coverage and claimed generality is load-bearing.

Editorial extensions

If this is right

  • New biomedical knowledge graphs can be assembled by selecting existing Koza-Hub ingests and merging them, skipping the usual parsing/cleanup phase for common sources.
  • Curatorial decisions become explicit per-ingest code and configuration, making it possible to audit what was included, transformed, or excluded from a graph.
  • Processing a source once and re-releasing the KGX artifact reduces repeated heavy downloads on source providers; downstream builds stream compact KGX files instead.
  • The 2 GB GitHub artifact limit constrains ingests to refined datasets, but removing that limit would let the modular model extend to terabyte-scale resources.
  • As more groups adopt KGX output, graphs built independently become directly connectable, not just within one project but across projects.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The three-primitive model is general enough that it could outgrow biomedicine: any domain with a standardized node/edge schema could use the same reader/transform/writer pattern, though the paper only demonstrates KGX/Biolink.
  • The 'born-interoperable' value proposition is network-effect driven: the more data providers that publish KGX artifacts directly (the paper suggests petitioning them), the less transformation is needed anywhere, and the hub's reusability compounds.
  • A quantitative comparison—developer hours or defect rates for Koza ingests versus bespoke scripts on the same sources—would test the developer-simplicity claim directly, and the paper does not yet provide that measurement.
  • Koza Merge suggests a path toward customizable, on-demand KGs: instead of one monolithic graph, users could query a registry of ingests and assemble a source subset tailored to a question, which has implications for reproducible biomedical analyses.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper introduces Koza, a Python package that constructs KGX-format knowledge graphs from raw biomedical sources using a declarative reader/writer/transform pipeline configured by YAML files and Python decorators. It also presents Koza-Hub, a collection of modular per-source ingests maintained in public GitHub repositories with monthly automated releases. The authors claim Koza is the first KG construction tool combining modularity, developer simplicity, declarative parameters, KGX-centricity, and standalone Python packaging, and they support the claim with two use cases: the ClinGen ingest and the Model Organism Knowledge Graph (MOKG). The paper reports current coverage of 14 repositories / 25 artifacts, with adoption by Monarch KG, MOKG, the Biomedical Data Translator, and KG-Microbe.

Significance. If the central claims hold, Koza addresses a real pain point in biomedical KG construction: repeated, bespoke data-wrangling pipelines for each source. The manuscript's strengths are concrete and verifiable: 14 public GitHub repositories, monthly release automation, schema enforcement toward KGX/Biolink, and demonstrated adoption in multiple real KGs. The existence claim is well supported by the public artifacts. However, the significance depends on the breadth and expressiveness of the pipeline; those are currently under-supported because the demonstrated ingests are all file-based and the advertised source count is inconsistent with the evidence. The paper would be strengthened by correcting the count, adding a comparison with existing tools, and either broadening the demonstrated input types or explicitly scoping the claims.

major comments (3)
  1. [Abstract vs. Conclusion and Table 1] The manuscript presents inconsistent counts for the same contribution. The Abstract states 'conversion processes for thirty gold standard biomedical data sources,' while the Conclusion states 'Koza-Hub currently integrates data from 25 data artifacts across 14 data repositories.' Table 1 lists 14 ingest repositories, and even that overcounts distinct biomedical providers: three entries are Alliance repos and two are ZFIN repos. The breadth claim is load-bearing for the stated reduction of redundant labor, yet the evidence supports roughly half the advertised coverage. The Abstract should be corrected to match Table 1, or the missing 16 sources must be documented.
  2. [Methodology (Koza Primitives) and Table 1] The central expressiveness premise is not established. All fourteen ingests in Table 1 consume flat-file formats (CSV/JSON-like); no API, SQL database, or very large (>2 GB) source is demonstrated. The paper itself notes that GitHub Actions imposes a 2 GB artifact limit and that larger artifacts are future work. The three-primitive model may well be extensible, but as written the claim that Koza 'streamlines ingesting raw biomedical information' in general rests on 14 file-based examples. The authors should either add one or two non-file-based/large-source ingests, or explicitly scope the claims to the demonstrated source types.
  3. [Introduction (novelty claim)] The paragraph beginning 'To our knowledge, the Koza pipeline is the first KG construction tool...' makes a strong novelty claim but provides no comparison with existing frameworks (e.g., KGTK, R2RML-based mappers, ROBOT, or other KGX-compatible pipelines). Since this claim is part of the paper's framing, the absence of a related-work comparison makes it impossible to assess whether the specific combination is new. Add a brief related-work section and cite concrete alternatives; if the claim is retained, explain why each existing tool lacks at least one of the five listed properties.
minor comments (6)
  1. [Abstract / Conclusion] Align the source/artifact counts: 'thirty gold standard biomedical data sources' in the Abstract should be reconciled with '25 data artifacts across 14 data repositories' in the Conclusion.
  2. [Table 1] Add a column or note distinguishing artifact count from provider count; the current 'Number of Artifacts Processed' can be confused with the number of distinct data sources.
  3. [Methodology] Minor typos and style: 'Github' should be 'GitHub'; 'e.g. membership in a list' needs a comma; 'Renaissance Computing Institute' could be expanded on first use.
  4. [Use case 2] The phrase 'producing one additional modular ingest resource' is vague; specify which new ingest was created for MOKG rather than reused from Monarch.
  5. [Reproducibility] Consider adding version/commit hashes or Zenodo DOIs for Koza and the ingest repositories so that downstream users can cite exact artifacts.
  6. [Methodology (Motivation)] The text mentions 'complex data loads, such as SQL databases' as a motivation, but no SQL reader is listed among the current readers. Clarify whether such a reader exists, is planned, or is out of scope.

Circularity Check

0 steps flagged · score 2.0 of 10

No definitional or fitted-input circularity; only minor self-referential validation evidence.

full rationale

This is a software/tool description rather than a derivation with fitted parameters, equations, or predictions, so the high-burden circularity patterns do not apply. The central claims—three primitives (reader/writer/transform), YAML configuration, decorator API, KGX-compliant output—are directly implemented and demonstrated via Table 1's ingests and the ClinGen and MOKG use cases. No quantity is fitted to a subset of data and then renamed a prediction; no uniqueness theorem is imported; no ansatz is smuggled in via citation. The main self-referential element is that the strongest evidence of value comes from the authors' own ecosystem: the Koza-Hub repositories are all under monarch-initiative, and the Monarch KG and MOKG use cases are built by the same team. This weakens the independence of the adoption evidence but is not circular, because the ingests are externally checkable artifacts and the paper also cites an outside user (the Biomedical Data Translator, ref [19]). The abstract/body mismatch—"thirty gold standard biomedical data sources" versus the Conclusion's "25 data artifacts across 14 data repositories"—is an internal consistency and coverage-support problem, not a circularity problem. Under the hard rules, no circular step can be quoted as reducing an output to an input by construction, so the score is 2 for minor self-referential validation rather than a higher circularity score.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

This is a systems paper, so the ledger holds design assumptions rather than fitted quantities. There are no free parameters. The load-bearing assumptions are the sufficiency of the three-primitive architecture, the choice of KGX/Biolink as the target standard, and the unmeasured claim that modular ingests save labor. The abstract/body discrepancy in source counts is a factual error, not a parameter.

assumptions (3)
  • domain assumption KGX with the Biolink Model is an appropriate and sufficient target standard for interoperable biomedical knowledge graphs
    The entire value proposition of 'born-interoperable' ingest rests on this; the paper adopts KGX rather than arguing for it (Abstract, Introduction).
  • domain assumption The three-primitive architecture (reader, writer, transform) with YAML configuration and a decorator API is expressive enough for the claimed range of biomedical sources
    Methodology 'Koza Primitives'. Only 14 repositories are demonstrated (Table 1) while the abstract claims 30 sources; generality is assumed.
  • domain assumption Modular per-source ingests reduce redundant labor and reduce load on data providers
    Modularity section asserts these benefits (e.g., avoiding over-use of GWAS as a source) without measurement.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Koza and Koza-Hub for born-interoperable knowledge graph generation using KGX." pith.science (2026). https://pith.science/paper/BIT6WID2

@misc{pith2026250909096,
  author       = {Pith},
  title        = {Pith review of: Koza and Koza-Hub for born-interoperable knowledge graph generation using KGX},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BIT6WID2}},
  note         = {Machine review of arXiv:2509.09096}
}
read the original abstract

Knowledge graph construction has become an essential domain for the future of biomedical research. But current approaches demand a high amount of redundant labor. These redundancies are the result of the lack of data standards and "knowledge-graph ready" data from sources. Using the KGX standard, we aim to solve these issues. Herein we introduce Koza and the Koza-Hub, a Python software package which streamlines ingesting raw biomedical information into the KGX format, and an associated set of conversion processes for thirty gold standard biomedical data sources. Our approach is to turn knowledge graph ingests into a set of primitive operations, provide configuration through YAML files, and enforce compliance with the chosen data schema.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 4 canonical work pages

  1. [1]

    Fish oil, raynaud’s syndrome, and undiscovered public knowledge

    Swanson DR. Fish oil, raynaud’s syndrome, and undiscovered public knowledge. Perspect Biol Med. 1986;30: 7–18

    1986

  2. [2]

    Deconstructing the translational tower of babel

    Austin CP, Colvis CM, Southall NT. Deconstructing the translational tower of babel. Clin Transl Sci. 2019;12: 85

    2019

  3. [3]

    A comprehensive survey on automatic knowledge graph construction

    Zhong L, Wu J, Li Q, Peng H, Wu X. A comprehensive survey on automatic knowledge graph construction. ACM Comput Surv. 2024;56: 1–62

    2024

  4. [4]

    Biolink Model: A universal schema for knowledge graphs in clinical, biomedical, and translational science

    Unni DR, Moxon SAT, Bada M, Brush M, Bruskiewich R, Caufield JH, et al. Biolink Model: A universal schema for knowledge graphs in clinical, biomedical, and translational science. Clin Transl Sci. 2022;15: 1848–1855

    2022

  5. [5]

    [cited 10 Sep 2025]

    Biomedical Data Translator. [cited 10 Sep 2025]. Available: https://ncats.nih.gov/research/research-activities/translator

    2025

  6. [6]

    Electronic address: splon@bcm.edu, ClinGen Consortium

    ClinGen Consortium. Electronic address: splon@bcm.edu, ClinGen Consortium. The Clinical Genome Resource (ClinGen): Advancing genomic knowledge through global curation. Genet Med. 2025;27: 101228

    2025

  7. [7]

    Genenames.Org: The HGNC resources in 2023

    Seal RL, Braschi B, Gray K, Jones TEM, Tweedie S, Haim-Vilmovsky L, et al. Genenames.Org: The HGNC resources in 2023. Nucleic Acids Res. 2023;51: D1003–D1009

    2023

  8. [8]

    The Unified Phenotype Ontology : a framework for cross-species integrative phenomics

    Matentzoglu N, Bello SM, Stefancsik R, Alghamdi SM, Anagnostopoulos AV, Balhoff JP, et al. The Unified Phenotype Ontology : a framework for cross-species integrative phenomics. Genetics. 2025;229. doi:10.1093/genetics/iyaf027

    work page doi:10.1093/genetics/iyaf027 2025

Show all 20 references
  1. [9]

    The Monarch Initiative in 2024: an analytic platform integrating phenotypes, genes and diseases across species

    Putman TE, Schaper K, Matentzoglu N, Rubinetti VP, Alquaddoomi FS, Cox C, et al. The Monarch Initiative in 2024: an analytic platform integrating phenotypes, genes and diseases across species. Nucleic Acids Res. 2024;52: D938–D949. `

    2024

  2. [10]

    Github; Available: https://github.com/monarch-initiative/phenio

    phenio: An integrated ontology for Phenomics. Github; Available: https://github.com/monarch-initiative/phenio

  3. [11]

    PANTHER: Making genome-scale phylogenetics accessible to all

    Thomas PD, Ebert D, Muruganujan A, Mushayahama T, Albou L-P, Mi H. PANTHER: Making genome-scale phylogenetics accessible to all. Protein Sci. 2022;31: 8–22

    2022

  4. [12]

    PomBase: a Global Core Biodata Resource-growth, collaboration, and sustainability

    Rutherford KM, Lera-Ramírez M, Wood V. PomBase: a Global Core Biodata Resource-growth, collaboration, and sustainability. Genetics. 2024;227. doi:10.1093/genetics/iyae007

    2024 doi

  5. [13]

    dictyBase and the Dicty Stock Center (version 2.0) - a progress report

    Fey P, Dodson RJ, Basu S, Hartline EC, Chisholm RL. dictyBase and the Dicty Stock Center (version 2.0) - a progress report. Int J Dev Biol. 2019;63: 563–572

    2019

  6. [14]

    Updates to the Alliance of Genome Resources central infrastructure

    Alliance of Genome Resources Consortium. Updates to the Alliance of Genome Resources central infrastructure. Genetics. 2024;227. doi:10.1093/genetics/iyae049

    2024 doi

  7. [15]

    Gene: a gene-centered information resource at NCBI

    Brown GR, Hem V, Katz KS, Ovetsky M, Wallin C, Ermolaeva O, et al. Gene: a gene-centered information resource at NCBI. Nucleic Acids Res. 2015;43: D36–42

    2015

  8. [16]

    Improving biomedical knowledge graph quality: A community approach

    Cortes KG, Sundar S, Gehrke S, Manpearl K, Lin J, Korn DR, et al. Improving biomedical knowledge graph quality: A community approach. arXiv [q-bio.OT]. 2025. Available: http://arxiv.org/abs/2508.21774

    2025 arXiv

  9. [17]

    A genomic mutational constraint map using variation in 76,156 human genomes

    Chen S, Francioli LC, Goodrich JK, Collins RL, Kanai M, Wang Q, et al. A genomic mutational constraint map using variation in 76,156 human genomes. Nature. 2024;625: 92–100

    2024

  10. [18]

    [cited 4 Sep 2025]

    gnomAD v4.1. [cited 4 Sep 2025]. Available: https://gnomad.broadinstitute.org/news/2024-04-gnomad-v4-1/

    2025

  11. [19]

    Progress toward a universal biomedical data translator

    Fecho K, Thessen AE, Baranzini SE, Bizon C, Hadlock JJ, Huang S, et al. Progress toward a universal biomedical data translator. Clin Transl Sci. 2022;15: 1838–1847

    2022

  12. [20]

    KG-microbe - building modular and scalable knowledge graphs for microbiome and microbial sciences

    Santangelo BE, Hegde H, Caufield JH, Reese J, Kliegr T, Hunter LE, et al. KG-microbe - building modular and scalable knowledge graphs for microbiome and microbial sciences. bioRxiv. 2025. p. 2025.02.24.639989. doi:10.1101/2025.02.24.639989

    2025 doi

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.