Mickey Pentecost, PhD — Built with Claude · Life Sciences Hackathon (Researcher Track)
📹 Watch the 3-minute explainer: https://www.youtube.com/watch?v=i6Fc2_6tdlU
A cross-modality, cross-species test of whether the same microglial inflammatory program that traumatic brain injury (TBI) installs is the one that Alzheimer's disease (AD) risk alleles predispose to — computed entirely from public data during the event.
One-sentence thesis. TBI installs environmentally what AD risk alleles predispose to genetically, and both converge on the same microglial accelerator axis (OPN/SPP1 · TREM2 · APOE · complement), while the pro-resolving brake arm (TSG-6/TNFAIP6 · CD44) stays disengaged — and both arms are read out at a single receptor, CD44.
This repository holds the complete analysis behind the submission: a coordinated microglial accelerator module tested across two species and four assay types (single-nucleus and bulk RNA, spatial transcriptomics, proteomics), anchored by common-variant and regulatory genetics, and integrated at CD44. Every result is computed here from public data — no prior analysis is reused — under an OSI-approved (MIT) license. The genetic-anchoring layer (variant enrichment, allelic chromatin effect, partitioned heritability) is documented in full below.
The full plain-language write-up is SUBMISSION.md, and
every panel is described in FIGURE_CAPTIONS.md. The
five figures tell the story in order — each reads orient → data → interpret,
with a shared color grammar (red = inflammatory accelerator, blue =
brake/resolution, purple = astrocytes, slate = genetics, amber = conclusion):
| Figure | Question it answers |
|---|---|
| Fig 1 — the hook | Is there one shared accelerator program, and does it hold up? (module recovery · up in both diseases · reproduces across species/methods · 31% of inherited AD risk) |
| Fig 2 — where & when | Does the accelerator concentrate at the damage? (real tissue maps: at amyloid plaques in AD, at the lesion in injury; peaks ~7 days) |
| Fig 3 — trigger vs threshold | Are inherited risk and injury-installed genes the same? (only APOE/TREM2 are inherited-risk; the effectors are not — trigger ≠ threshold) |
| Fig 4 — the regulatory switch | Which master-control proteins turn the accelerator on and off? (NFκB activator · MEF2C repressor · SPI1/CEBPB identity proteins) |
| Fig 5 — the CD44 hub | Where do the accelerator and brake meet, and how does the brake fail? (both signal through CD44; CD44 rises across four modalities while TSG-6 falls) |
The design system these figures share is documented in
figures/DESIGN_SYSTEM.md. Three Supplementary
figures — the genetic anchor (variant enrichment, ChromBPNet allelic effect,
S-LDSC heritability, colocalization), the resolution brake (HA machinery,
oxidative-fragmentation routes, imbalance, reproducibility), and regulatory grammar
plus cross-condition validation (motif ablation, trajectory, CTE cohort) — are
captioned in FIGURE_CAPTIONS.md.
Accelerator genes sit high on both the environmental axis (installed by TBI and AD in single-nucleus data) and the genetic axis (carrying AD risk variants in microglial enhancers). The upstream microglial switches — TREM2, APOE, TNF — carry the genetic load (≥3 AD risk variants each); the effectors — SPP1/OPN, the C1q complement genes, C3, TLR2 — are installed environmentally (up in AD and/or TBI microglia, few common risk variants).
The accelerator/resolution axis is hypothesis-driven — defined a priori from the DAM/neuroinflammation literature and the pro-resolving TSG-6→CD44 mechanism, not derived from the datasets it is tested on. Two independent robustness checks confirm it holds up, and one honest limitation is stated plainly:
The accelerator genes co-vary as a genuine module (beating 100% of
detection-matched random sets) and re-emerge in unsupervised NMF; the module is
organized around CD44, the receptor where the accelerator ligand SPP1/OPN and
the brake ligand TSG-6 compete. The resolution arm is near-undetectable in
single-nucleus RNA — its quantification requires bulk RNA-seq
(results/refined_accelerator_geneset.csv).
| Layer | Question | Result | Figure |
|---|---|---|---|
| 1 · Variant enrichment | Do AD-associated variants concentrate in microglial accelerator-gene enhancers? | OR = 1.56, MAF-matched permutation P = 0.0022 (1.59× over null); APOE-excluded P = 9.5×10⁻³; resolution arm OR = 0 | figures/enrichment_results.png |
| 2 · Allelic chromatin effect | Do those variants change predicted microglial chromatin accessibility? | AD-associated accelerator variants disrupt chromatin more than non-associated (ChromBPNet, Mann-Whitney P = 0.005); top hit rs3800342 in the TREM2 enhancer (AD P = 9.3×10⁻¹²) | figures/chrombpnet_results.png |
| 3 · Partitioned heritability | What fraction of AD SNP-heritability sits in microglial regulatory DNA? | Microglial peaks (1.5% of genome) carry 31% of AD h² — 21× enrichment, P = 1.1×10⁻⁵; conditional coefficient z = 3.62. Within axis enhancers the signal is accelerator-specific (z = +0.99) vs resolution (z = −1.04) | figures/sldsc_heritability.png |
Supporting GWAS locus map: figures/gwas_axis_loci.png.
| Layer | Question | Result | Figure |
|---|---|---|---|
| Spatial — AD plaque niche | Does the accelerator concentrate around amyloid plaques? | Subcellular Stereo-seq (18-mo App-NL-G-F, 142k bins at 50 µm): accelerator concentrates at plaques and survives a tissue-geometry control (partial ρ = −0.24 vs distance-to-centroid; near>far within every concentric band, Δ +0.16→+0.23); the resolution/HA brake is excluded (ρ = +0.38); astrocytes co-concentrate peri-plaque (the CD44 shell). WT_F5 negative control shows a general center-high tissue gradient (ρ=−0.44) accounting for ~half the marginal signal | figures/spatial_stereoseq_AD.png |
| Spatial — TBI lesion niche | Does the same axis concentrate at the TBI lesion? | Visium (impact TBI 7 d, 6 TBI / 6 Sham): accelerator up TBI-vs-Sham (P = 4×10⁻²⁶) and tracks lesion proximity (ρ = +0.47, non-circular proxy). Same axis, two injury landmarks | figures/spatial_convergence.png |
| Temporal — TBI time-course | Is the resolution brake a level or a phase? | CEREBRI (24 h → 7 d → 6 mo): the brake is highest acutely (24 h) and collapses by 7 d as the accelerator peaks (accel trend ρ = +0.44, brake ρ = −0.25). The chronic AD plaque niche resembles the 6-mo TBI state — accelerator without the acute brake | figures/cerebri_timecourse.png |
Supporting Visium genotype panel: figures/spatial_visium.png.
Epidemiology links moderate-to-severe TBI to elevated dementia risk, but the mechanism has been a black box. Here we show, from public data computed during the event, that TBI and AD engage the same microglial accelerator program: single-nucleus microglia from human AD (SEA-AD) and mouse TBI (CEREBRI) both up-regulate the accelerator genes, while the pro-resolving arm stays flat — and the genetic side supplies the causal direction: the human genetic architecture of AD is concentrated in exactly the microglial regulatory elements that control those accelerator genes. Two independent kinds of evidence — an environmental perturbation and inherited risk — point at the same molecular axis, which is what a shared, targetable mechanism should look like.
.
├── README.md # this file
├── LICENSE # MIT
├── DATA_PROVENANCE.md # every external dataset, accession, URL, license
├── thesis.md # full causal argument + axis gene modules
├── corces_model_provenance.md # ChromBPNet model source, cluster→cell-type map
├── requirements.txt # Python environments (analysis + ChromBPNet + LDSC)
├── figures/ # publication figures (PNG, 300 dpi)
├── results/ # all result tables (CSV)
├── data/ # axis targets + enhancer/annotation BEDs (GRCh38 + hg19)
├── code/ # scoring + LD-score + munge scripts
└── notebooks/ # reproducible analysis notebooks (see below)
axis_targets.csv— 38 human axis genes (accelerator 21 / resolution 9 / DAM-only 8), GRCh38 coordinates, ±100 kb cis-windows, arm assignment.gwas_axis_gene_summary.csv— per-gene AD-GWAS signal (Bellenguez GCST90027158) in cis-windows.ad_variants_in_microglia_enhancers.csv— 3,522 AD variants intersecting Corces C24 microglial enhancers, annotated by axis gene + arm.enrichment_results.csv— Fisher + MAF-matched permutation enrichment, per arm.chrombpnet_allelic_scores.csv— allelic effect (lfc, JSD) for 3,356 SNVs;chrombpnet_top_variants.csv— ranked chromatin-disruptors.sldsc_results.csv— partitioned-heritability enrichment for 4 annotations (all-microglia, accelerator, DAM, resolution).crossarm_convergence.csv— per-gene environmental (log₂FC) vs genetic (risk-variant load) scores.
novelty_synthesis.md— the full causal narrative: convergence → causation/direction → perturbation/mechanism, plus the drug-target rationale.caqtl_formal_coloc.csv/caqtl_enhancer_coloc.csv— primary-microglia (Kosoy/Raj, n=150) caQTL vs AD-GWAS; enhancer-level dissection (effectors vs inherited-risk loci) and formal shared-SNP coloc.moloc_threeway.csv— pairwise colocalization across three molecular layers (caQTL, eQTL, AD GWAS): three independent shared-SNP Wakefield-ABF tests per gene (not a joint 3-trait model), both molecular QTLs from primary microglia. Columns: PP4_caQTL_eQTL, PP4_eQTL_GWAS, PP4_caQTL_GWAS.caqtl_ep_ad_loops.csv— AD-locus variants in microglial enhancers physically looping (ABC/Hi-C) to axis genes.coloc_mr_results.csv— cell-type-matched myeloid (macrophage/monocyte) eQTL vs AD-GWAS coloc.full_circuit.csv— integrated per-gene circuit (AD-risk / caQTL / E-P loop / expression layers).insilico_perturbation.csv/cebpb_perturbation.csv— ChromBPNet motif-ablation Δaccessibility for NFκB, MEF2, SPI1, CEBPB on the 754 axis enhancers.pseudotime_drivers.csv— diffusion-pseudotime gene/TF trends across the homeostatic→accelerator transition.adni_dod_sidequest.md— standalone prompt for the two-hit (TBI × APOE) clinical test, to run under separate data governance.
The pipeline runs on public data only. Three environments are used
(requirements.txt documents exact versions):
- Analysis (Python 3.11 + pandas/scipy/pyliftover) — GWAS streaming, variant intersection, enrichment, figures.
- ChromBPNet inference (
tf_infer: TensorFlow 2.19 + tf-keras) — allelic scoring. Seecode/chrombpnet_scoring.py; the Corces C24 microglial model is loaded via a bias-free inner-model reconstruction (documented in the script) to bypass a Python-3 Lambda-layer incompatibility. Validated against the authors' own scores (|LFC| Pearson r = 0.986). - S-LDSC (
ldsc: our Python-3 port ofbulik/ldsc) — partitioned heritability.code/run_arm_ldscores.shbuilds per-chromosome annotations and LD scores;code/_munge_stream.pystreams and munges the Bellenguez sumstats.
| Data | Source | Size |
|---|---|---|
| AD GWAS (Bellenguez 2022) | EBI GWAS Catalog GCST90027158, GRCh38-native |
755 MB |
| Corces C24 microglial ChromBPNet model + peaks | github.com/corceslab/variantapp |
~90 MB |
| 1000G EUR reference + baseline-LD v2.2 | Zenodo 10515792 |
~1.1 GB |
See DATA_PROVENANCE.md for exact URLs, accessions, versions, and licenses.
- AD GWAS: Bellenguez C, et al. Nat Genet 2022. GWAS Catalog GCST90027158 (PMID 35379992).
- Microglial ChromBPNet models + scATAC peaks: Corces MR, et al. Nat Genet 2020 (PMID 33106633); model weights from
corceslab/variantapp. - S-LDSC: Finucane HK, et al. Nat Genet 2015; Gazal S, et al. baseline-LD v2.2. Software:
bulik/ldsc(GPL-3.0), Python-3 port included. - Environmental-arm single-nucleus data: SEA-AD MTG (Allen Institute) and CEREBRI mouse TBI (GEO GSE269748).
This work was produced for the Built with Claude — Life Sciences hackathon (Research Track). All analysis was performed during the event on public data.
MIT (see LICENSE). Third-party data and models retain their own licenses as
documented in DATA_PROVENANCE.md.