Genetic studies of Alzheimer's disease routinely end with a list of "enriched pathways", meaning groups of genes, such as immune signaling or membrane trafficking, that carry more risk signal than chance. These lists rarely lead to a mechanism, because they do not say what the pathway does or when it acts. This proposal turns such lists into falsifiable timing claims. Each pathway's inherited risk is summarized as a per-person score. That score is then tested against the order in which a patient's brain changes in longitudinal imaging from the Alzheimer's Disease Neuroimaging Initiative (ADNI): amyloid plaques appear first, tau tangles spread next, and tissue loss follows. A pathway that genuinely drives amyloid clearance should shift when amyloid turns abnormal and leave later steps alone. A pathway that merely correlates with risk should show no specific timing signature once it is compared against random gene sets of the same size.
Two things would be new. The first is a stage-resolved, calibrated test of pathway genetics on in vivo trajectories rather than on autopsy snapshots. The second is a built-in benchmark: pathways whose mechanism has already been established in the laboratory must land at the disease step the experiments predict. The deliverable is a short, ranked set of pathway-by-stage hypotheses. Each one names the cell type and the assay readout a wet-lab follow-up should use. This targets the gap between statistical association and mechanism more directly than adding one more enrichment table.
Core Claim
Most pathway enrichments reported for Alzheimer's imaging traits reflect diffuse polygenic signal rather than a specific biological role. The few that are real can be recognized because they act at a particular transition in the disease sequence, and they do so beyond size-matched random gene sets. The same stage-localization test, applied to pathways with known mechanisms, will recover the stage those mechanisms predict. Two examples are ApoE-dependent endolysosomal handling of amyloid-β, where Kanekiyo et al. traced impaired lysosomal trafficking under ApoE4, and microglial MS4A6A function, where a knockout study showed loss of neuroprotective microglia and impaired amyloid responses. Both should shift amyloid onset. A third, ApoE receptor signaling onto tau kinases (Freeman et al.), predicts a tau-stage effect. If these positive controls fail, the method is not trustworthy for novel pathways, and we will report that.
Why This Design, Given Prior Work
Three lines of work motivate the proposal. Pathway-partitioned polygenic scores have been built across thousands of Gene Ontology processes. Katsumata et al. linked "membrane" and "protein binding" scores to a latent amyloid/tau proteinopathy (corrected p = 1.7 × 10⁻⁶ and 0.00011). Aung et al. tied cell-cycle, apoptosis and angiogenesis scores to neuropathology with the APOE region excluded. Dutta et al. found network hub genes (VCP, IQCB1) and used the SEA-AD single-cell atlas to place them in microglia and neurons. All three work on end-stage autopsy cohorts, so they cannot tell an effect on when disease starts apart from an effect on how fast it moves. We reuse their pathway-score construction and cell-type lookup. We replace the autopsy outcome with in vivo timing.
Two ADNI studies are the closest neighbors, and both constrain the design. The longitudinal endophenotype GWAS fits straight-line mixed models, one biomarker at a time, and then runs MAGMA gene-set enrichment on the results. It reports immune and metabolic enrichment, but it treats each biomarker's slope as independent and assumes linear change. The multimodal progression-score GWAS aligns people along a sigmoidal amyloid-plus-hippocampus curve using the GRACE algorithm. It found LCORL (rs6850306, p = 1.03 × 10⁻⁸) where single-modality traits found nothing, but it compressed progression into one scalar and ran no pathway analysis. Our model keeps the sigmoidal alignment that made the second study work, and it keeps the biomarker-specific resolution the first study had.
Estarellas et al. supply the missing-data method. Their version of Subtype and Stage Inference, an unsupervised method that orders biomarker events and clusters people into progression subtypes, treats unobserved measurements as uniformly distributed. On synthetic data it stayed accurate with up to 40% missingness, where mean and nearest-neighbour imputation did not. Tau PET is missing in over 87% of ADNI cases (Burns et al.), so this likelihood-based handling is essential.
Kim et al.'s knowledge-driven binning is the main counterpoint on power. Gene burden tests on 750 ADNI genomes found nothing, while pathway bins tested with SKAT found FANCC. We take from this that ADNI alone is too small for discovery. We therefore estimate all variant weights outside ADNI and use ADNI only to test pre-specified pathway scores. This is a far lighter statistical burden than a genome-wide scan.
Research Program
We fit a latent-time disease progression model to ADNI participants who have genotypes and longitudinal amyloid PET, structural MRI, CSF and cognition. Tau PET enters where it is available, through the missing-data likelihood. Each biomarker $k$ follows a sigmoid in latent disease time, and each person $i$ has an individual time offset, as in GRACE. The genetic question enters through biomarker-specific onset shifts:
$$\tau_{ik} = \tau_k + \sum_p \gamma_{pk}\, s_{ip},$$
Here $\tau_{ik}$ is when biomarker $k$ turns abnormal for person $i$, $\tau_k$ is the population onset, $s_{ip}$ is person $i$'s standardized risk score for pathway $p$, and $\gamma_{pk}$ is how many years pathway $p$ moves biomarker $k$. A second coefficient on the curve's steepness separates "starts earlier" from "progresses faster".
A stage-localized mechanism predicts a characteristic pattern: $\gamma_{pk}$ is concentrated on one biomarker and the downstream events inherit only the shift propagated through the model. A diffuse polygenic effect instead moves all events together.
Pathway scores are built from external Alzheimer's GWAS summary statistics from the ADGC data used by Katsumata and Aung, with ADNI samples removed to avoid overlap. The APOE region is handled as its own score. Gene sets come from two sources. The first is curated Gene Ontology and KEGG sets. The second is cell-type-restricted versions of the same sets, keeping only genes expressed in the relevant SEA-AD cell type (for example, microglia or excitatory neurons). This addresses the context-blindness of standard enrichment.
Every pathway's localization statistic is compared against a null distribution from random gene sets matched on size, gene length and baseline association strength. This competitive test guards against the overlap and size bias that inflate classical enrichment. Pathways that pass are then annotated with cell type, stage, and a proposed assay readout: amyloid uptake and degradation, tau phosphorylation, or neuronal survival.
Baselines
The naive reference is a model with only APOE and a genome-wide polygenic score as onset modifiers. Any pathway result must add held-out trajectory likelihood beyond it. Otherwise the "mechanism" is just overall genetic load.
The strongest existing method is the pipeline of the longitudinal endophenotype GWAS: mixed-model slopes per biomarker followed by MAGMA enrichment. We will re-run it on the same participants, biomarkers and gene sets, and test its pathway calls for stage specificity under the same random-set null. That paper reports enriched immune and metabolic pathways but no predictive or calibration figure, so there is no reported result on Lacuna for held-out trajectory prediction or positive-control recovery by that pipeline. The comparison is therefore head-to-head on our metrics.
A second strong competitor is pathway scores regressed on a single GRACE progression score, following the DPS study. That study reported a single-locus hit with β = −0.07 ± 0.01. We win against these baselines if we recover more positive controls at the correct stage and hold false-positive localization at the nominal rate under random gene sets.
Three ablations isolate the key ingredients. We swap sigmoid latent time for linear slopes. We drop the cell-type restriction. We drop the matched random-set null in favour of plain Benjamini–Hochberg correction. All variants get identical data, fold splits and tuning budget.
Why not just…?
Why not just run pathway enrichment on the five SuStaIn subtypes? This is the most obvious move at the intersection, and we run it as a baseline. With 789 people split five ways, subtype-specific GWAS has almost no power, and assignment to subtypes is itself uncertain. Kim et al.'s null burden results at N = 750 suggest discovery inside ADNI will fail. Our design avoids discovery in ADNI and tests externally weighted scores against continuous timing parameters instead.
Why not just apply KGWAS to an ADNI progression trait? KGWAS finds 48.6% more true associations in N = 5,000 simulations by borrowing strength over a functional-genomics graph. That is a real advance for finding loci, but it outputs variants, not the disease step they act on. It is complementary rather than competing: we will add KGWAS-prioritized gene neighborhoods as an extra gene-set source and test them with the same localization statistic.
Why not just take top hits like LCORL or FANCC straight to the wet lab? Single hits come with no hypothesis about cell type or readout, so assays are chosen by guesswork. This proposal is meant to feed wet-lab work: its output specifies which assay to run. The positive-control benchmark tells us whether that specification can be trusted.
Why not stay with autopsy cohorts, which are larger? Autopsy scores measure terminal burden and cannot separate an earlier start from a faster progression. That separation is exactly what distinguishes a clearance mechanism from a vulnerability mechanism.
Decisive Checks
The first check is positive-control recovery. APOE-excluded endolysosomal and microglial pathway scores, together with the MS4A6A neighborhood, should shift amyloid onset. ApoE–kinase signaling genes should localize to tau. Failure here halts interpretation of novel pathways.
The second check is calibration. Across random matched gene sets, the rate of "localized" calls must match the nominal level. We expect standard MAGMA-style calls to exceed it.
The third check is held-out prediction. Pathway-shifted onsets must improve cross-validated log-likelihood of future biomarker values over the APOE-plus-global-score reference.
The fourth check is external consistency. Pathways localized to early amyloid should also predict MCI-to-AD conversion timing in NACC, the independent cohort where the DPS study replicated rs6850306 (HR = 0.593). This check is partial, because NACC lacks matched imaging.
We state one limitation plainly. ADNI is overwhelmingly of European ancestry, so any localization claims are restricted to that population until they are replicated elsewhere.
References
- Differential Regulation of Amyloid-β Endocytic Trafficking and Lysosomal Degradation by Apolipoprotein E Isoforms
- MS4A6A/Ms4a6d deficiency disrupts neuroprotective microglia functions and promotes inflammation in Alzheimer’s disease model
- Apolipoprotein E decreases tau kinases and phospho-tau levels in primary neurons
- Biological pathway‐based Alzheimer’s Disease polygenic risk score and proteinopathy
- Pathway‐specific polygenic risk score analysis to explore the inverse association between cancer and Alzheimer’s disease neuropathologic changes
- A systems‐biology approach to identify candidate genes associated with composite brain pathology scores in the Adult Changes of Thought (ACT) and the Religious Orders Study and Memory and Aging Project (ROSMAP) autopsy cohorts
- Pathway enrichment in genome‐wide analysis of longitudinal Alzheimer's disease biomarker endophenotypes
- Genetic study of multimodal imaging Alzheimer’s disease progression score implicates novel loci
- Multimodal subtypes identified in Alzheimer’s Disease Neuroimaging Initiative participants by missing-data-enabled subtype and stage inference
- Flexible Multimodal Neuroimaging Fusion for Alzheimer's Disease Progression Prediction
- Knowledge-driven binning approach for rare variant association analysis: application to neuroimaging biomarkers in Alzheimer's disease
- Small-cohort GWAS discovery with AI over massive functional genomics knowledge graph