BACKGROUND

Alzheimer's disease is biologically heterogeneous, and prioritizing therapeutic targets requires methods that resolve disease signals across patient subgroups. The TREAT-AD Target Risk Score (TRS) integrates genetics and multi-omics across AMP-AD cohorts to estimate genome-wide AD-associated effects and map them onto 19 AD biodomains1. Using the TRS pipeline1, we recomputed proteomic meta-analytic weights per subgroup across cohorts to characterize subgroup AD signatures. These signatures revealed broadly concordant AD-associated changes, but subgroup-specific differences at the pathway level remained unresolved. Resolving these differences requires direct subgroup comparison, which in meta-analytic settings is vulnerable to aggregation bias, as naive pooling can conflate cohort effects with subgroup effects2. To address this, we developed a principled framework that compares subgroups while controlling for cohort-level confounding. 

METHODS

Preprocessed protein expression matrices and sample metadata from AMP-AD were subset by user-specified criteria. Initially, for each subset, protein-level AD-associated effects (Hedges’ g3, AD vs. control) were estimated using random-effects meta-analysis of standardized mean differences across cohorts. These per-subgroup weights captured each subgroup's independent AD signal to compare to a full-cohort baseline. To compare subgroups, the pipeline computes Hedges' g per subgroup within each cohort first, then differences within cohort (Δg = g_A – g_B), canceling batch, platform, and population structure effects2. Δg is meta-analyzed across cohorts via REML random-effects models, yielding per-protein interaction z-scores (z = Δg/SE). A high |z| requires the subgroup difference to be both large and consistent across independent cohorts. Pathway discovery uses interaction GSEA4,5, a single gene set enrichment analysis ranked by interaction z-score, mapped to 19 AD biodomains. We applied this pipeline to AMP-AD proteomics across two axes: Sex (Male vs. Female) and APOE4 carrier status (ApoE4+ vs. Non-carrier). Cohorts lacking both subgroups within a comparison were excluded. For each significant GO term identified by interaction GSEA, subgroup dominance was assigned via a hierarchical majority vote: within each cohort, the subgroup with larger |g| was identified per protein; protein-level dominance was determined by majority across cohorts; term-level dominance was determined by majority across member proteins. This classification partitions enriched pathways into subgroup-dominant sets for visualization.

CONCLUSIONS