Polygenic-Score-Based Disease Subtyping
Summary¶
Many complex diseases (type 2 diabetes, obesity, MASLD, schizophrenia, Alzheimer's disease) are genetically heterogeneous: the same clinical diagnosis can arise from distinct underlying mechanisms with different downstream complications. A recurring family of methods uses genome-wide association data — usually via some form of partitioned or component polygenic risk score (PRS) — to recover this heterogeneity as discrete or continuous genetic subtypes. The literature on this splits cleanly into three methodological strategies, distinguished by what gets clustered first.
The Three Approaches¶
| Approach | What is clustered | Representative methods |
|---|---|---|
| Mechanism-first | Variants/loci, by their shared multi-trait GWAS association pattern; a partitioned PRS is then computed per cluster | Udler et al. 2018 bNMF T2D clustering; Kim/Westerman/Udler 2023 high-throughput extension; Smith/Deutsch 2024 multi-ancestry extension; Chami et al. 2025 obesity subtyping; Jamialahmadi et al. 2024 MASLD partitioned PRS; PRSet; Genomic SEM; PDR |
| Individual-first | People, by their profile across several (often mechanism-first-derived) partitioned/component scores | Li et al. 2024 prediabetes subtyping; NetMoST; Gene-SGAN; DeGAs / dPRS |
| Subtype-prediction | Nothing new — reuses an already-recognized subtype's partitioned score(s) as a fixed predictor of clinical outcomes | DiCorpo et al. 2022 T2D partitioned-PRS outcome study |
These are not mutually exclusive: several papers in this list compose two approaches in sequence. Li et al. (2024), for example, take mechanism-first scores (from Udler et al.) as input and then apply an individual-first clustering step on top of them; DiCorpo et al. (2022) is the natural downstream validation of the same mechanism-first clusters, applied at much larger scale without deriving anything new.
Disease-Specific Applications¶
- Cardiovascular disease — see Polygenic Subtyping of Cardiovascular Disease for how these three families map onto CAD, MI, and heart failure, the one purpose-built CAD partition (Hu et al. 2025 Pleiotropy-Decomposed PRS), and proposed novel directions (bNMF-for-CAD with a dedicated lipoprotein(a) axis, individual-first pharmacogenomic biotyping).
Why the Distinction Matters¶
- Mechanism-first clusters are directly interpretable as biological pathways (beta-cell dysfunction, lipodystrophy, hepatic lipid retention, etc.) and generalize naturally to new cohorts once fixed, because the cluster loadings are defined on the variants, not on any particular sample's individuals. Their main limitation is that cluster resolution depends on the trait panel and GWAS available at derivation time.
- Individual-first clusters (or continuous component profiles) can reveal patient heterogeneity that a fixed set of mechanism-first scores doesn't fully capture — e.g. Li et al.'s cluster 6 prediabetes subtype, whose defining feature (very low proinsulin score without otherwise-extreme scores) only emerges from looking at combinations of scores across people. The tradeoff is that hard clustering (k-means, network-based, deep clustering) can be less stable and requires re-deriving cluster assignments in new cohorts, though DeGAs' continuous component-profiling avoids this specific issue.
- Subtype-prediction studies are where clinical utility is actually tested at scale (DiCorpo et al.'s 454,193 participants dwarf any single discovery cohort), but they cannot discover new heterogeneity — only confirm or refute the predictive value of subtypes already proposed elsewhere.
Citations¶
This page synthesizes material documented with full inline citations on its three child pages; see Mechanism-First Polygenic Risk Score Clustering, Individual-First Polygenic Risk Score Clustering, and Subtype-Prediction Polygenic Risk Scores for the primary sources and their exact supporting locations.
Synthesis: the three-approach taxonomy and the cross-approach relationships described in "Why the Distinction Matters" are this wiki's own organizing framework for the ingested literature, not a taxonomy stated as such in any single cited paper.