neurodeg-microglia-convergence-r1

Extension of two prior OmiCairn lab-KB findings. (1) A Jonckheere–Terpstra test on the SEA-AD Microglia-PVM MTG atlas (CellxGene collection 1ca90a2d) found the LDAM module score (LGALS3, CD9, GPNMB, NPC2) significantly elevated in APOE4 carriers vs non-carriers (JT=993.0, p=0.006), while the DAM2 score (TREM2, LPL, CST7) showed no significant carrier shift; a follow-up tau regression was directionally positive but null (β_tau=+0.031, p=0.143, 95% CI [−0.007, +0.069]). (2) In AD prefrontal cortex (GSE157827, 21 donors) the most-disrupted ligand–receptor pairs were ECM–integrin interactions (LAMA2→ITGAV/ITGA6, COL4A2→ITGB8; microglia/endothelial→astrocyte; log2FC ≈ −2.0). Both findings were observed within Alzheimer's disease alone. The open AD/DLB/PDD prefrontal-cortex atlas GSE303823 (Brase/Harari-style cohort; processed object on Zenodo 11107597) independently reports a shared, myelin-enriched microglial state with increased MSR1-mediated phagocytosis across all three dementia subtypes.

This run asks whether the APOE4-linked lipid-droplet microglial program (LDAM) is a SHARED cross-disease feature of dementia or an AD-specific one, and whether the ECM–integrin disruption is, by contrast, AD-specific. Primary discovery cohort: GSE303823 (snRNA-seq PFC; NCI control n=4, AD n=4, DLB n=6, PDD n=6; ~20 donors, the processed rna_list.rds on Zenodo 11107597). Orthogonal AD-arm replication: SEA-AD on CellxGene (open; MTG + DLPFC; the Microglia-PVM object under collection 1ca90a2d, plus the full atlas). Held-out blind external validation: SEA-AD DLPFC (A9) — never used to define the microglial states or fit any model.

Hypotheses (pre-registered):

H1 (primary — shared vs AD-specific LDAM elevation across 3 diseases): The LDAM module score (LGALS3, CD9, GPNMB, NPC2) within the microglial compartment is elevated in EACH of AD, DLB and PDD relative to NCI controls in GSE303823, i.e. it is a SHARED dementia feature rather than AD-restricted. Decision rule: per-donor pseudobulk LDAM module score in microglia; for each disease vs NCI report Hedges' g effect size + p-value + 95% bootstrap CI, with Benjamini–Hochberg FDR across the 3 contrasts. H1 (shared) passes if all 3 diseases show a positive elevation with FDR<0.05 and 95% CI excluding 0. It is DECISIVELY classed AD-specific (the competing model) if only AD is significant and DLB/PDD CIs include 0. A partial pattern (e.g. AD+DLB but not PDD) is reported verbatim as partial sharing — NOT rounded up to "shared".

H2 (secondary — orthogonal method corroboration): The shared/AD-specific LDAM pattern in H1 is reproduced by an independent quantification: (a) compositional shift of an unsupervised LDAM-like microglial subcluster (Leiden, scored post-hoc by the LDAM markers) tested with a Dirichlet/scCODA-style per-donor proportion model, and (b) SCENIC/regulon or AUCell module activity for the same gene set. H2 corroborates H1 if the sign and the shared-vs-specific call agree across at least 2 of the 3 diseases under both orthogonal methods.

H3 (secondary — held-out blind external validation, AD arm only): A signature locked from GSE303823 (the per-gene LDAM weights defining the microglial program) is scored, WITHOUT refitting, on the held-out SEA-AD DLPFC (A9) atlas to test whether it separates high- vs low-pathology AD donors. Report AUC + 1000× bootstrap 95% CI. The lock file (gene list + weights + cluster definition) is frozen before touching SEA-AD DLPFC. A null AUC (CI spanning 0.5) is an informative, reportable outcome.

H4 (secondary — APOE4-specificity of the lipid-droplet sub-program): IF APOE genotype / APOE4-carrier status is present in GSE303823 and/or SEA-AD donors, the lipid-droplet sub-program (the LGALS3/GPNMB-high axis) is enriched specifically in APOE4 carriers, replicating the prior JT=993.0, p=0.006 SEA-AD finding in an independent cohort. Decision rule: carrier vs non-carrier Jonckheere–Terpstra (or Mann–Whitney) on the per-donor sub-program score, with effect size + p + 95% CI. COVARIATE DEPENDENCY: if APOE status is absent, H4 is UNTESTABLE — report that honestly and fall back to a genotype-agnostic description. Do NOT impute genotype.

H5 (secondary — NEGATIVE CONTROL: ECM–integrin disruption is AD-specific, not shared): The ECM–integrin ligand–receptor disruption observed in AD PFC (LAMA2→ITGAV/ITGA6, COL4A2→ITGB8; microglia/endothelial→ astrocyte) is AD-specific and is NOT shared by DLB or PDD. Decision rule: per-donor pseudobulk expression of the ECM ligands (LAMA2, COL4A2, and the integrin receptors) and/or a CellChat-style L-R disruption score, each disease vs NCI; report effect size + p + 95% CI. This is a pre-registered DISSOCIATION test: H5 passes (supporting the convergence story) if AD shows the disruption while DLB/PDD do NOT. If the ECM disruption is ALSO shared across diseases, that is an equally informative negative-control failure and must be reported as the headline finding, not buried.

H6 (secondary — honest-failure / sensitivity decomposition): Decompose the H1 LDAM effect into nuisance vs real-biology layers: (i) microglial-proportion differences across disease (does the score shift just track more microglia?), (ii) post-mortem-interval / age / sex confounds (covariate-adjusted re-fit), (iii) Leiden-resolution sweep and marker-set perturbation (swap to the expanded LDAM set adding APOE/CLU, per the prior sensitivity finding). State explicitly that resolution-instability or a proportion-driven artifact is itself the headline if it dominates.

A negative or partial result is fully informative. If the LDAM elevation turns out AD-specific (not shared), or the ECM negative control fails (disruption is shared after all), or H3 validation is null, THAT is the headline result. The analysis must NOT be retuned (re-clustering, marker-set cherry-picking) until H1 reads "shared".

The research question comes from the topic's current research brief. All attempts at a topic share it; earlier attempts may have started from an earlier wording.