Spaceflight liver lipid dysregulation, reproduced from the OKN federation

Framing (non-negotiable). Unit of analysis is the gene, measured as a stored differential-expression result for one vetted Space-Flight-vs-Ground-Control liver assay in spoke-genelab, projected to human orthologs and annotated across six further federation graphs. Coverage is the three of the original paper's four GeneLab liver datasets that survive contrast vetting. This is hypothesis generation from secondary, summary-level data — not a re-analysis of the raw sequencing, and not causal or clinical inference. Mouse-to-human claims are ortholog-inferred; disease and phenotype links are observational associations. Keep both caveats attached to every downstream claim.

Abbreviations. ORA = over-representation analysis · GSEA = gene set enrichment analysis · DEG = differentially expressed gene · FDR = false-discovery rate (Benjamini–Hochberg) · log2FC = log2 fold change · FC = fold change · GO = Gene Ontology · BP = biological process · NAFLD = non-alcoholic fatty liver disease · MASLD/MASH = metabolic dysfunction-associated steatotic liver disease / steatohepatitis · HP = Human Phenotype Ontology · MONDO = Mondo Disease Ontology · DOID = Disease Ontology · EFO = Experimental Factor Ontology · UBERON = Uber-anatomy ontology · OSD/GLDS = NASA Open Science Data Repository / GeneLab Data System accession · RR = Rodent Research · STS = Space Transportation System (Shuttle) · CASIS = Center for the Advancement of Science in Space · IPA = Ingenuity Pathway Analysis · KG = knowledge graph · PPAR = peroxisome proliferator-activated receptor.

1. Executive summary

The original study's central claim — that spaceflight alone, without the confound of live return to Earth, drives lipid dysregulation in mouse liver — reproduces from the OKN federation, but from a narrower and differently-shaped slice of evidence than the paper used. Of the paper's four GeneLab liver datasets, 3 of 4 are present in spoke-genelab as contrast-vetted Space-Flight-vs-Ground-Control assays. The fourth, GLDS-168 (RR-1 NASA), is excluded: all 9 of its liver flight-vs-ground assays fail the within-assay comparability rule because the dataset pools two missions (SpaceX-4 / RR-1 and SpaceX-8 / RR-3) and two library preparations, so its flight and ground arms differ in covariates the contrast is supposed to hold fixed.

Across the 3 vetted assays, 4678 genes carry a stored liver DE result and 130 pass adj-p ≤ 0.05 with |log2FC| ≥ 1 (93 up in flight, 38 down), collapsing to 124 human orthologs. Against an explicit prokn-annotated background of 2160 genes, the single most enriched GO biological process is lipid catabolic process (7.08×, k = 4/20, FDR = 0.027); the only Reactome pathway to survive FDR correction alongside it is "MLL4 and MLL3 complexes regulate expression of PPARG target genes in adipogenesis and hepatic steatosis" (4.48×, FDR = 0.027) — an independent, ontology-level hit on the paper's PPAR and steatosis argument. The flight-responsive gene core is significantly over-represented for curated liver-disease genes (rdkg, 2.0×, p = 0.0074) and, more sharply, for the GWAS-scale NAFLD trait set in digcfdekg (3.54×, k = 10/105, FDR = 0.0074).

The genes carrying that signal are the ones a hepatologist would name: PNPLA3 — the strongest common human genetic risk factor for fatty liver disease and the only gene in the core reaching a hepatic steatosis phenotype (HP:0001397) through the graph — plus PNPLA2/ATGL, CIDEC, FGF21, MLXIPL/ChREBP, PLIN5, APOA4, ACOT1/ACOT2, GPAT3, ELOVL3/ELOVL6 and the peroxisome-biogenesis set PEX3/PEX11A/PEX19. None of these is named in the original paper. Ranking on recurrence, effect size and cross-KG disease support tiers 6 genes A, 22 B and 96 C.

What this adds is two things the original could not show. First, an independent ontology-level confirmation of the lipid and PPAR claims, reached from curated human disease and pathway graphs rather than from the same expression data that generated the hypothesis. Second, a methodological finding about the original: the paper's own selection rule (FC ≥ 1.2) admits 1560 of the 2160 testable genes — 72% of the universe — at which point over-representation is arithmetically incapable of detecting anything (0 terms at FDR ≤ 0.05). The paper's rank-based GSEA was immune to this; a knowledge-graph reproduction is not, and needs the stricter cut. The lipid signal survives it.

2. Sources used

All 7 graphs were queried directly; 12 non-exploratory SPARQL queries back every number in this report and appear verbatim in the reproducibility record.

3. Design & rules

What was reproduced, and why only part. The original analysed four GeneLab liver datasets: GLDS-25 (STS-135 Shuttle, 13 days, C57BL/6, microarray), GLDS-47 (RR-1 CASIS, 21 days, C57BL/6, RNA-seq), GLDS-168 (RR-1 NASA, 37 days, C57BL/6, RNA-seq) and GLDS-137 (RR-3, 42 days, BALB/c, RNA-seq). All four exist in spoke-genelab as OSD studies with liver assays, but a spaceflight contrast is only readable when its two arms differ in nothing but the condition. Applying the graph's contrast rules to liver (UBERON:0002107) returns 13 clean and 20 confounded Space-Flight-vs-Ground-Control assays across all studies. Three of the paper's four datasets contribute a clean assay; GLDS-168 contributes none, because every one of its 9 liver flight-vs-ground assays pairs arms across missions or spike-in protocols. That is not a defect in the graph: GLDS-168 is itself a combined RR-1 + RR-3 dataset, which the original paper acknowledges when it reports separate DEG counts "for the RR1 and RR3 data from the GLDS-168".

Selection rule. A gene is called differentially expressed when adj-p ≤ 0.05 and |log2FC| ≥ 1 in a vetted assay, with group 1 = Space Flight and group 2 = Ground Control so a positive log2FC means up in flight. The original used a more permissive FC ≥ 1.2 cut; §6.3 runs both and explains why the permissive cut cannot support over-representation analysis. Human orthologs come from spoke-genelab's own ortholog edge, collapsed one-row-per-human-gene by maximum |log2FC| with a mean-rule sensitivity check (no sign flips; 2 of 124 genes are many-to-one).

Joins. Every cross-graph step runs on a shared identifier, never on a study accession — OSD/GLDS numbers are a federation island. Genes travel on Entrez node IRIs (identical between spoke-genelab, digcfdekg and spoke-okn; identifiers.org form in rdkg), on human gene symbol into prokn, and diseases travel on MONDO and DOID through ubergraph's subclass closure. The exact predicates, backgrounds and scoring formula are in the reproducibility record.

The reproducible slice is dominated by one dataset. GLDS-25 contributes 4617 of the 4689 stored DE rows and 116 of the 130 DEGs; GLDS-47 contributes 68 rows and 13 DEGs; GLDS-137 contributes 4 rows and

1. This inverts the paper's design: its argument rested on the ISS datasets precisely

because the Shuttle animals were returned live, and the ISS arms are the ones the graph carries thinly.

4. Confidence tiers

Genes are scored on six evidence axes — recurrence across vetted assays, effect size, curated MONDO liver-disease membership (rdkg), NAFLD trait-set membership (digcfdekg), lipid GO/Reactome membership (prokn), and reaching an HP phenotype profile — plus a bonus for the number of corroborating graphs. The exact weights are in the reproducibility record.

5. Findings by axis

5.1 Primary signal — direction and magnitude of the flight response

Of the 130 mouse DEGs, 93 are up in flight and 38 down, a 2.4:1 up-skew consistent with the original's report that "the majority of pathways being regulated in the liver are upregulated". The largest-magnitude changes in the ISS arm (GLDS-47) are structural and calcium-handling genes (Tpm3-rs7, Cacna1h, Ttn, Cacna1c) rather than lipid genes; the lipid signal sits almost entirely in the deeper GLDS-25 payload.

The lipid genes cluster tightly on the up-in-flight side at high significance — Pnpla2, Cidec, Pex11a, Plin5, Apoa4, Acot2, Sult1e1 all sit above −log10 p ≈ 4 — while the two strongest NAFLD risk genes, Pnpla3 and Mlxipl, move down. That split (lipid-droplet and fatty-acid-handling machinery up, lipogenic/risk transcription factors down) is the shape of a lipid-overload response rather than a lipogenic drive.

5.2 Internal replication across datasets

Cross-dataset recurrence is the weakest axis in this reproduction. Exactly 1 gene, DEPP1 (mouse Depp1), passes threshold in two independent vetted assays — down in flight in both GLDS-25 (log2FC −1.73) and GLDS-47 (log2FC −1.02). Given that GLDS-137 stores only 4 genes and GLDS-47 only 68, the intersection available for replication is arithmetically tiny, so the low recurrence measures the graph's payload depth, not the biology. Every other cross-dataset statement in this report therefore rests on shared pathway or disease membership, not on shared genes.

6. Domain analyses

Enrichment families run and skipped. GO biological process — RUN (§6.1). Reactome pathway — RUN (§6.1); the two are separate families and both were executed. Disease / trait gene-set — RUN, in both flavours the method requires: curated (rdkg) and broad GWAS-scale (digcfdekg), §6.2. Phenotype (HP) — RUN, §6.2. Drug / target linkage — SKIPPED by scope decision: this reproduction was scoped to NAFLD/liver-disease evidence, and the original paper makes no therapeutic claim to reproduce. Chemical / adverse-outcome (biobricks-aopwiki, biobricks tox) — SKIPPED: there is no exposure or toxicant in the question. Microbiome / taxon alignment — SKIPPED: spoke-genelab's microbial-abundance layer carries no liver assay in the paper-matched studies. Place-based / geospatial linkage — SKIPPED: the unit of analysis is a mouse liver assay, which has no geography.

Suppliers considered and not used, with reasons: biomarkerkg (joins on only 191 Entrez genes — too sparse to test), pankgraph (pancreatic-cancer disease payload, GO not joinable on Entrez), biobricks-mesh, biohealth, nde, nestkg, gene-expression-atlas-okn and ncipidkg (carry disease/pathway payloads but do not join on Entrez, and add no liver-specific evidence the three used disease suppliers lack).

6.1 Functional enrichment — GO and Reactome

Both families were tested with a hypergeometric test and Benjamini–Hochberg FDR against an explicit background: the genes measured in the three vetted assays that prokn actually annotates (2160 for GO, 1720 for Reactome), never an implicit whole-genome universe.

Lipid catabolic process is the top GO term (7.08×, FDR = 0.027), driven by ADORA1, APOA4, NCEH1, PNPLA2; lipid metabolic process (2.15×, k = 10/165) and fatty acid metabolic process (2.78×) follow at nominal significance. Relaxing the k floor to 2 as a sensitivity check surfaces a coherent set of finer lipid terms — positive regulation of triglyceride catabolic process, peroxisome fission, fatty acid elongation (saturated and monounsaturated), and in Reactome Synthesis of very long-chain fatty acyl-CoAs and Activation of gene expression by SREBF (SREBP). On the Reactome side, only two pathways cleared the floor at all, and one of them names the phenotype directly: MLL4/MLL3 complexes regulate PPARG target genes in adipogenesis and hepatic steatosis (4.48×, FDR = 0.027; PEX11A, PNPLA2, TBL1XR1). Inflammatory response (4.25×, FDR = 0.027; ADORA1, BCL6, FOS, NR4A1, TLR5, VNN1) is the one strong non-lipid theme, consistent with the "second hit" the original invokes for NASH progression.

6.2 Liver-disease, trait and phenotype linkage

The federation supplies three independent disease→gene layers joinable on Entrez, and the method requires testing both a curated and a broad set rather than picking one.

The curated rdkg layer — 1216 genes across 60 MONDO liver diseases — is significantly over-represented: 15 of 124 core genes against 7.51 expected (2.0×, p = 0.0074). Those 15 are ACOT1, ADRA1A, BDH1, E2F8, ERN1, FGF21, FOS, MLXIPL, PDGFRL, PEX11A, PNPLA3, PPTC7, SLC22A10, SULT1E1, TLR5. The broad digcfdekg layer behaves in the way the method warns broad GWAS sets usually do — most of its 19 liver trait sets are null — with one striking exception: the EFO non-alcoholic fatty liver disease set is the single most enriched category tested anywhere in this study (3.54×, k = 10/105, p = 4.6e-04, FDR = 0.0074 across the 16 testable trait sets). The third supplier is the honest counterweight: spoke-okn's DOID:409 association layer is a single undifferentiated bucket of 1289 genes, and against it the core shows no enrichment at all (0.87×, p = 0.72) — exactly the null a permissive, uncurated set produces, and a reminder that the rdkg and digcfdekg results are informative because those sets are specific.

Phenotypes route gene → disease → HP, since the federation has no direct gene→HP edge. Inside rdkg, one path completes end-to-end: PNPLA3 → NAFLD1 (MONDO:0021105) → Hepatic steatosis (HP:0001397) — the graph independently reaching the very phenotype the original measured by Oil Red O staining. Broadening to oard-kg's EHR profiles, the MONDO liver-disease category is covered for 36 diseases and 1763 HP terms, in which hepatic steatosis recurs across 21 alongside cirrhosis, elevated hepatic transaminases and hepatomegaly. oard-kg does not itself carry MASH or NAFLD1, so the specific terms our genes reach are absent from its EHR layer — the phenotype breadth comes from the surrounding category, and these are co-occurrence associations, not causal ones.

6.3 Threshold sensitivity — why the original's cut-off cannot be reused

The original selected genes at FC ≥ 1.2 (|log2FC| ≥ 0.263) with adj-p ≤ 0.05, a deliberately low bar it justified as standard for the rank-based GSEA and IPA workflows it ran. Re-running both enrichment families at that rule is the cleanest way to separate "this finding does not reproduce" from "this finding reproduces only under a different statistic".

At FC ≥ 1.2 the signature becomes 3216 human genes, of which 1560 fall inside the 2160-gene annotated background — 72% of the testable universe. Every fold enrichment collapses towards 1 and 0 GO terms and 0 Reactome pathways survive FDR correction; the lipid terms are still the top-ranked ones (fatty acid metabolic process, p = 0.03), but at fold 1.17 they carry no evidential weight. This is not a failure of the original — GSEA ranks the whole gene list and never forms a signature/ background split, so it is unaffected — but it does mean a knowledge-graph reproduction that uses over-representation must impose the stricter cut, and that the paper's headline lipid claim is recoverable only at |log2FC| ≥ 1. Circadian terms behave the same way: they are present and testable in the background (15 GO terms over 41 genes) but reach fold 1.06–1.08 at the permissive cut and only ADORA1 at the strict one.

7. Discussion

Read together, the axes describe a liver that is handling more lipid than it should. The up-in-flight set is dominated by machinery for storing and turning over lipid droplets — CIDEC (the lipid-droplet fusion protein FSP27), PLIN5, PNPLA2/ATGL, GPAT3, ACOT1/ACOT2, ELOVL3/ELOVL6, PEX3/PEX11A/PEX19 — while the down-in-flight set contains the lipogenic and risk-allele transcription machinery, MLXIPL/ChREBP and PNPLA3. Peroxisomal genes and acyl-CoA thioesterases are canonical PPARα targets, so their coordinate induction is the transcriptional shadow of the PPAR signalling the original inferred by IPA — recovered here from an entirely different evidence path, through curated human pathway annotation. That the one Reactome pathway to survive correction names hepatic steatosis in its own label is about as direct an ontology-level corroboration as the federation can give.

The disease layer sharpens this from "lipid genes changed" to "NAFLD genes changed". The core is enriched for curated liver-disease genes and, far more specifically, for the GWAS NAFLD set, and the single completed gene→disease→phenotype path terminates on hepatic steatosis — the phenotype the original quantified histologically. PNPLA3 deserves separate emphasis: it is the top-ranked gene in this analysis on integrated evidence, it is corroborated by five graphs, and it is not mentioned anywhere in the original paper. Its I148M variant is the strongest common genetic determinant of human fatty liver, which makes a flight-associated change in its hepatic expression a specific, testable proposition rather than a generic lipid observation.

Three testable predictions follow. First, PNPLA3, PNPLA2 and CIDEC should show concordant flight-associated change in an independent liver dataset with adequate payload depth — the RR-1 NASA and RR-6 liver assays, re-derived as clean single-mission contrasts, would be the natural test and would also settle whether the GLDS-168 pooling is the only obstacle. Second, if the peroxisomal and thioesterase induction is genuinely PPARα-driven, a PPARα-null flight cohort should lose it while retaining the CIDEC/PLIN5 droplet response, which is downstream of substrate load rather than of PPARα transcription. Third, because PPARA and PPARG themselves are significantly changed but sub-threshold here (log2FC +0.67 and +0.84, adj-p ≤ 0.014), the effect is a broad, low-amplitude transcriptional shift rather than a few large-effect genes — which predicts that rank-based methods will keep outperforming threshold-based ones on this tissue, exactly as the original found.

8. Comparison with prior work

Claims were checked against the primary literature retrieved through the PubMed and Paperclip MCP connectors. The complete per-claim record, with citations, is in Liver-Lipid_literature_comparison.md.

Claims 1, 2, 6, 7 and 8 were checked against the full text of the original [1]; claim 2 was additionally checked against the full abstract and results of Jonscher et al. [2], and claims 4 and 5 against a corpus-wide full-text search rather than abstracts alone.

Where the KG evidence diverges from the literature. The divergences are all differences of scope and instrument, not errors in the graphs. Claims 6 and 8 diverge because the federation carries no liver proteomics and no record of the pancreatic hormone transcripts, so the original's proteomic and IPA-derived arguments have no counterpart to test — an evidence gap, not a contradiction. Claim 7 diverges because ORA and GSEA answer different questions on the same data. The one finding that is genuinely about the graphs rather than the biology is claim 9, and it reflects how GeneLab packaged the dataset, which the graph faithfully records, rather than a mis-assignment introduced by ingestion. No entity-resolution collision or mis-assigned annotation was found in any of the seven graphs used.

9. Full ranked results

The complete ranked table — all 124 human genes with their effect sizes, evidence axes, corroborating graphs, scores and tiers — is in Liver-Lipid_results.xlsx (sheet Ranked Results) and as data/ranked_candidates.tsv. The interactive version below is sortable by any column, filterable by free text, and sliceable by tier, direction and dataset; the sources (n) column counts how many federation graphs corroborate each gene, with one pill per graph — spoke-genelab supplies the expression change, prokn the GO/Reactome annotation, rdkg and spoke-okn curated disease associations, digcfdekg the GWAS trait sets, and oard-kg the phenotype profile.

The ranking is steeply top-heavy and its shape is informative: the tier-A genes are exactly those where a large effect coincides with independent disease evidence, and all 6 of them sit in the lipid/NAFLD axis rather than being scattered across the transcriptome. Below roughly rank 25 the score is carried almost entirely by effect size with no cross-KG corroboration, which is why 96 of 124 genes are tier C — a reminder that most of a DEG list, even a small one, has no disease context in the federation.

10. Summary of findings & limitations

Findings. Three of the four GeneLab liver datasets behind Beheshti et al. 2019 are reproducible from spoke-genelab as contrast-vetted spaceflight comparisons; the fourth is excluded because it pools two missions. Across those three, 130 mouse genes pass a strict differential- expression threshold and collapse to 124 human orthologs, whose most enriched biological process is lipid catabolic process (7.08×, FDR 0.027) and whose only FDR-surviving Reactome pathway alongside inflammation names PPARG targets in hepatic steatosis. The core is over-represented for curated liver-disease genes (2.0×, p 0.0074) and markedly so for the GWAS NAFLD set (3.54×, FDR 0.0074), while showing no enrichment against a deliberately coarse DOID liver bucket. The paper's central claim therefore reproduces, independently of the expression data that generated it.

The genes carrying that signal — PNPLA3 (tier A, five corroborating graphs, and the only gene reaching hepatic steatosis HP:0001397 through the graph), PNPLA2, CIDEC, FGF21, MLXIPL, SULT1E1, PEX11A — are not named in the original, and DEPP1 is the only gene to recur across two independent vetted assays. Separately, re-running the analysis at the original's own FC ≥ 1.2 rule shows why: that cut admits 72% of the testable universe, at which point over-representation detects nothing.

Limitations.

1. **spoke-genelab stores a filtered subset of each dataset's differential-expression results, not

the full transcriptome** — 4617, 68 and 4 genes for the three assays. Every count here is bounded by that payload, and the two ISS datasets are effectively too thin to contribute independent evidence.

1. The reproducible slice is dominated by the Shuttle dataset (GLDS-25), which is the one arm

the original set out to control for because those animals were returned live. The lipid signal recovered here therefore cannot, on its own, separate space stressors from return stress — the paper's specific contribution.

1. All disease, trait and phenotype links are observational associations, not causal claims.

oard-kg's are EHR co-occurrences; digcfdekg's are statistically inferred gene–trait weights.

1. Mouse-to-human claims are ortholog-inferred. 8 of 130 mouse DEGs have no human

ortholog in the graph and were dropped; 2 human genes derive from more than one mouse gene.

1. The prokn join is an exact gene-symbol label match, which is more fragile than an

identifier join: a symbol synonym or a deprecated name silently fails to match, so GO/Reactome coverage (2160 of 4604 background genes) is a lower bound.

1. No proteomic layer exists in the federation, so the original's quantitative proteomics — the

basis for its apolipoprotein, Cyp7a1, Cyp1a2 and Fgl1 claims — has no counterpart here. Nor do histology, Oil Red O quantification, t-SNE/PCA sample separation, or IPA upstream-regulator activation scores, all of which require sample-level or proprietary data.

1. Over-representation is descriptive, not causal, and with only 61 signature

genes inside the GO background the test is underpowered: several coherent lipid terms sit at k = 2 and are reported only as a sensitivity check.

1. The drug/target axis was not run (scope decision, §6), so no therapeutic or countermeasure

claim is made or implied.

1. Single-graph-version snapshot. Every result is pinned to the KG versions in §2; spoke-genelab

v0.0.2 dates from 2026-03-13 and a later release could change payload depth and therefore counts.

11. Reproducibility

Everything needed to replicate this analysis — the originating prompt, the full replicator specification (contrast rules, thresholds, join recipes, backgrounds, the scoring formula and verified quantities), all 12 supporting SPARQL queries verbatim with their row counts, the pinned KG versions and the run timing — is in Liver-Lipid_reproducibility.md, with the analysis scripts in scripts/ and every intermediate extract in data/.

12. References

Retrieved via the PubMed MCP connector. Full-text verification via the Paperclip MCP connector.

1. Beheshti A, et al. Multi-omics analysis of multiple missions to space reveal a theme of lipid dysregulation in mouse liver. Scientific Reports. 2019. PMID:31844325 · doi:10.1038/s41598-019-55869-2 — full-text-verified (PMC6915713)
2. Jonscher KR, et al. Spaceflight Activates Lipotoxic Pathways in Mouse Liver. PLoS One. 2016. PMID:27097220 · doi:10.1371/journal.pone.0152877
3. Blaber EA, Pecaut MJ, Jonscher KR. Spaceflight Activates Autophagy Programs and the Proteasome in Mouse Liver. International Journal of Molecular Sciences. 2017. PMID:28953266 · doi:10.3390/ijms18102062
4. Mathyk BA, et al. Spaceflight induces changes in gene expression profiles linked to insulin and estrogen. Communications Biology. 2024. PMID:38862620 · doi:10.1038/s42003-023-05213-2
5. Chakraborty N, et al. Gene-Metabolite Network Linked to Inhibited Bioenergetics in Association With Spaceflight-Induced Loss of Male Mouse Quadriceps Muscle. Journal of Bone and Mineral Research. 2020. PMID:32511780 · doi:10.1002/jbmr.4102 — full-text-verified (PMC7689867)
6. Salcher S, et al. C10ORF10/DEPP-mediated ROS accumulation is a critical modulator of FOXO3-induced autophagy. Molecular Cancer. 2017. PMID:28545464 · doi:10.1186/s12943-017-0661-4
7. Casaletto JA, et al. Analyzing the relationship between gene expression and phenotype in space-flown mice using a causal inference machine learning ensemble. Scientific Reports. 2025. PMID:39824847 · doi:10.1038/s41598-024-81394-y
8. Vitry G, et al. Muscle atrophy phenotype gene expression during spaceflight is linked to a metabolic crosstalk in both the liver and the muscle in mice. iScience. 2022. PMID:36267920 · doi:10.1016/j.isci.2022.105213