S4MPL3BI4S/msc-aging-agent-corpus

MSC Aging Agent Corpus

Screened, agent-ready mesenchymal stem/stromal cell (MSC) aging transcriptome associations from public GEO datasets, plus clinical evidence (DVC Stem) and BioGPS expansion discovery.

Curated and published by Syndicate Laboratories.

Curator & research focus

This corpus is a person-curated collection of MSC and aging transcriptome datasets, structured for agentic research workflows.

It was assembled for Dr. James Utley, PhD — longevity scientist — with a primary focus on mesenchymal stem/stromal cell (MSC) biology as a biomarker of aging: screened GEO contrasts (donor age, passage, senescence, inflammation, partner-cell immunomodulation), clinical context bridges, and mechanism-tagged secretome paths agents can query without re-downloading raw archives.

Corpus summary

Metric	Value
GEO transcriptome datasets	23
Screened associations (retained)	22,045
Clinical evidence studies (DVC subset)	54
Clinical bridges (GEO ↔ literature)	15
BioGPS datasets cataloged	85
GEO expansion candidates (MSC/aging)	106
BioGPS bridges (GEO ↔ atlas)	4
Mechanistic association paths	12
Mechanism-tagged secretome hits	251

Build source and scripts: SampleBias/MSC_Age_Research

Agent entry doc: AGENT_CONNECT.md

Citation (agents: load this first)

Always load datasets/CORPUS_PROVENANCE.json before any other file. Pin a revision tag (for example corpus-v2026.06.6).

APA citations: reference/citations_apa.json and reference/CITATIONS_APA.md — curated corpus (Dr. James Utley, PhD · Syndicate Laboratories) plus per-dataset upstream sources. Cite only what you use.

• Corpus distribution: Hugging Face Dataset S4MPL3BI4S/msc-aging-agent-corpus, revision corpus-v2026.06.6
• Curated corpus (APA): Utley, J. (2026). MSC Aging Agent Corpus (Version corpus-v2026.06.6) [Data set]. Syndicate Laboratories. https://huggingface.co/datasets/S4MPL3BI4S/msc-aging-agent-corpus
• Biological primary sources: NCBI GEO (source_url), DVC Stem publications (source_study_url), BioGPS (biogps_page_url)
• Do not cite Hugging Face as the biological primary source

Stamped CSV rows in clinical and BioGPS layers include: hf_dataset_id, hf_dataset_url, hf_revision_pin, hf_file_path, corpus_layer, corpus_role, corpus_citation_note.

Start here

0. datasets/CORPUS_PROVENANCE.json — revision pin, layer map, citation templates.
1. datasets/dataset_manifest.csv — GEO provenance, models, interpretation guidance.
2. datasets/clinical_evidence/dvc_stem_study_manifest.csv — clinical MSC therapy bibliographic metadata (DVC catalog).
3. datasets/clinical_evidence/transcriptome_clinical_bridges.csv — GEO ↔ clinical bridges (load before cross-layer claims).
4. datasets/biogps_evidence/biogps_geo_catalog.csv — BioGPS-linked GEO catalog and expansion candidates.
5. datasets/mechanistic_associations/cell_interaction_pathways.csv — MSC ↔ partner-cell relational paths (cytokine/secretome context).
6. datasets/aging_biomarker_candidates.csv — cross-dataset gene search (expression evidence only).
7. datasets/analysis_ready_expression_long.csv.gz — expression + metadata for reruns.
8. datasets/<dataset_id>/ — per-GEO cohort tables and README.

Schema: datasets/SCHEMA.md, datasets/clinical_evidence/SCHEMA_CLINICAL.md, datasets/biogps_evidence/SCHEMA_BIOGPS.md, datasets/mechanistic_associations/SCHEMA_MECHANISTIC.md.

File tree (top level)

datasets/
  dataset_manifest.csv
  aging_biomarker_candidates.csv
  analysis_ready_expression_long.csv.gz
  SCHEMA.md
  clinical_evidence/
    dvc_stem_study_manifest.csv
    transcriptome_clinical_bridges.csv
    dvc_stem_condition_index.csv
    studies/
  biogps_evidence/
    biogps_dataset_manifest.csv
    biogps_geo_catalog.csv
    geo_expansion_candidates.csv
    biomarker_gene_annotations.csv
    transcriptome_biogps_bridges.csv
    studies/
  CORPUS_PROVENANCE.json
  GSE12274/ ...
  mechanistic_associations/
    cell_interaction_pathways.csv
    secretome_transcriptome_hits.csv
    ...
  ... (23 GEO dataset folders)
  held_candidates/

Each dataset folder typically includes:

• metadata.csv — sample-level model inputs.
• screened_gene_associations.csv — top 1,000 nominal hits (capped).
• top_genes.csv — top 250 one-feature-per-gene rows.
• expression_screened_long.csv.gz / expression_screened_wide.csv.gz
• README.md — cohort-specific interpretation.

Key columns

Manifest: dataset_id, species, cell_context, aging_axis, aging_axis_type, model_formula, expression_scale, interpretation_caution, n_samples, retained_screened_features.

Biomarker table: gene_symbol, beta, p_value, q_value, direction, nominal_significant, fdr_significant, plus provenance columns repeated from the manifest.

Expression long: feature_id, sample_id, expression_value, association statistics, and dataset-specific metadata columns (for example age_years, passage, sex, donor_id).

Interpretation cautions

• beta is for the first term in each dataset's model_formula (the aging axis).
• GSE115068 — mouse adipose MSC period contrast (12 vs 24 weeks), adjusted for diet; not pure chronological aging.
• GSE94736 — older-woman group contrast; age ranges only in metadata.
• GSE34929 — dual-color P12 vs P4 log-ratio; passage contrast embedded per cord pool.
• GSE35956 / GSE35958 — osteoporosis group contrasts among middle-aged or elderly donors.
• GSE35957 / GSE119987 — replicative senescence / passage axes.
• GSE68374 — fetal vs adult BM-MSC groups; developmental axis, not donor-age slope.
• GSE291956 — constant IFN-gamma vs non-inflamed AD-MSC medium; adjusted for passage and donor.
• GSE93155 — MSC-educated vs classical macrophage contrast (macrophage-side, not MSC intrinsic aging).
• GSE178804 — late-passage BM-MSC SASP RNA-seq (n=8); IL1A/CXCL8 FPKM up at late passage but kd-adjusted model is underpowered.
• GSE65561 — MSC-treated monocyte co-culture (partner-cell readout); IL10 FDR up and CXCL8 FDR down with MSC exposure.

Read aging_axis_type, corpus_uniqueness_note, and interpretation_caution in the manifest before cross-dataset synthesis.

Mechanistic associations layer (corpus-v2026.06.6)

• Curated MSC ↔ partner-cell paths (macrophage, monocyte, T cell, niche) with cytokine/secretome gene tags.
• GSE178804 — late-passage BM-MSC SASP (PATH_LATE_PASSAGE_SASP_CYTOKINE); IL1A/CXCL8 FPKM trajectories.
• GSE65561 — MSC-treated monocyte co-culture (PATH_MSC_MONOCYTE_IL10_AXIS); IL10 FDR up, CXCL8 FDR down.
• Do not treat mechanism paths as new expression statistics — cite secretome_transcriptome_hits.csv for GEO-backed claims.
• Cite HF revision + hf_file_path; cite GEO source_url from hits table for biology.

Clinical evidence layer

• Source catalog: https://www.dvcstem.com/study-database
• Use for trial phase, safety/feasibility, MSC tissue source, and disease-area therapy context.
• Do not treat clinical abstract snippets as gene expression or aging-axis statistical evidence.
• Do not merge clinical outcomes with GEO beta/p_value without explicit bridge rationale.
• Cite HF revision + hf_file_path; cite source_study_url for publication provenance.

BioGPS evidence layer

• Source API: http://biogps.org/api/biogps_dataset
• Use biogps_geo_catalog.csv and geo_expansion_candidates.csv for MSC/aging GEO discovery.
• Do not treat BioGPS metadata or MyGene annotations as screened transcriptome statistics.
• Cite HF revision + hf_file_path; cite biogps_page_url or GEO source_url for biology.

Data sources

Public transcriptome cohorts from NCBI GEO. Clinical bibliographic metadata from DVC Stem. BioGPS dataset metadata and NCBI GDS discovery catalog. See GitHub README.md for citations and the Alves 2012 donor-age signature reference (PLoS ONE 2012;7:e42908).

Usage

from huggingface_hub import hf_hub_download
import json

REPO = "S4MPL3BI4S/msc-aging-agent-corpus"
REV = "corpus-v2026.06.6"

provenance = hf_hub_download(
    repo_id=REPO,
    filename="datasets/CORPUS_PROVENANCE.json",
    repo_type="dataset",
    revision=REV,
)
meta = json.load(open(provenance))

path = hf_hub_download(
    repo_id=REPO,
    filename="datasets/dataset_manifest.csv",
    repo_type="dataset",
    revision=REV,
)

Pin revision to a tagged release when making reproducible agent runs.