Skip to content

Individual-First Polygenic Risk Score Clustering

Summary

Individual-first disease subtyping inverts the mechanism-first order of operations: rather than clustering variants into mechanistic groups and then scoring individuals, this approach starts from a set of already-defined partitioned or component genetic scores (often themselves produced by a mechanism-first analysis) and clusters people according to their profile across those scores. The resulting patient clusters can then be tested for differential clinical trajectories, imaging phenotypes, or treatment response.

Prediabetes Subtyping from Partitioned T2D Scores: Li et al. 2024

Li, Chen, Moon et al. took the five pathway-based partitioned PRS from Udler et al.'s T2D clustering (β-cell, low-proinsulin, obesity, lipodystrophy-like, liver-lipid) and used them as five input dimensions for k-means-based consensus clustering (1,000 k-means iterations with resampling, consensus-matrix hierarchical clustering, cluster number chosen by maximum average silhouette) of individuals with prediabetes.[1] In the HCHS/SOL cohort (3,677 prediabetic participants), six clusters emerged (sizes 13.2–20.0% of the sample); five of the six replicated in UK Biobank (16,284 participants; cluster 3 was not detected).[1]

Despite similar baseline glucose levels across clusters, T2D progression risk differed sharply over ~6-10 years of follow-up:

Cluster Profile HCHS/SOL RR (vs. cluster 1) UKBB HR (vs. cluster 1) Combined meta-analysis RR
Cluster 5 High lipodystrophy-like/liver-lipid/proinsulin scores, high HOMA-IR; low β-cell/obesity scores, BMI, HDL ("metabolically obese normal-weight"-like) 1.23 [0.96–1.57], P=0.096 1.35 [1.05–1.75], P=0.021 1.29 [1.08–1.53], P=0.005
Cluster 6 Distinctively very low proinsulin score; other four scores only slightly above average 1.39 [1.10–1.76], P=0.006 1.29 [1.00–1.67], P=0.049 1.34 [1.13–1.60], P<0.001

A healthy-lifestyle score interacted with cluster membership: in HCHS/SOL, lifestyle adherence reduced T2D risk overall (RR 0.73 [0.63–0.84]) but most strongly in cluster 5 (RR 0.65 [0.47–0.89]); in UK Biobank the pattern held (overall HR 0.80 [0.75–0.85]; cluster 5 HR 0.71 [0.62–0.81], the strongest benefit; cluster 1 HR 0.86 [0.75–1.00], the weakest). Critically, the elevated risk in cluster 5 relative to cluster 1 shrank from RR 1.77 [1.19–2.64] under an unhealthy lifestyle to RR 1.03 [0.91–1.18] under a healthy lifestyle — i.e. lifestyle intervention could largely neutralize the excess genetic risk carried by this cluster.[1]

Network-Based Clustering from Polygenic Haplotype Biomarkers: NetMoST

NetMoST builds a patient-similarity network directly from each individual's profile across many polygenic haplotype biomarkers (rather than a handful of pre-defined pathway scores) and applies network-based clustering. In 141 schizophrenia patients it recovered three biotypes — neurodevelopmental (36.9%), immune (28.4%), and calcium/neurotransmitter-transport (34.7%) — independently corroborated by distinct neuroimaging and cognitive profiles.[2]

Deep Clustering on Imaging + Genetic Profiles: Gene-SGAN

Gene-SGAN extends individual-first clustering to a weakly-supervised deep-learning setting, jointly clustering patients by their neuroimaging phenotype and genetic profile. It identified four Alzheimer's disease subtypes (ADNI, N=1,533) and five hypertension-related brain-change subtypes (UK Biobank, N=27,325), each subtype showing a distinct combination of neuroanatomical pattern and dominant genetic signature — see the tool page for the full subtype breakdown.[3]

Latent-Component Profiling: DeGAs / dPRS

DeGAs is a softer version of individual-first clustering: instead of assigning each person to a discrete cluster, it decomposes genome-wide SNP-trait associations into latent components via SVD and then profiles each individual's genetic risk as a weighted contribution from each component. This surfaces "outlier" individuals whose risk is driven almost entirely by one component (e.g. a fat-mass-only or fat-free-mass-only BMI profile, or one of five distinct myocardial-infarction risk-component profiles) rather than the typical multi-component blend — the same underlying question ("what is this person's profile across partitioned scores?") without forcing a hard cluster assignment.[4]

See Also

Citations

[1] Li, Y., Chen, G.-C., Moon, J.-Y., Arthur, R., Sotres-Alvarez, D., Daviglus, M. L., Pirzada, A., Mattei, J., Perreira, K. M., Rotter, J. I., Taylor, K. D., Chen, Y.-D. I., Wassertheil-Smoller, S., Wang, T., Rohan, T. E., Kaufman, J. D., Kaplan, R., & Qi, Q. (2024). Genetic Subtypes of Prediabetes, Healthy Lifestyle, and Risk of Type 2 Diabetes. Diabetes, 73(7), 1155–1165. DOI: 10.2337/db23-0699. Source: li2024-prediabetes-genetic-subtypes.md. Supports: all clustering, T2D-progression-risk, and lifestyle-interaction results above. Location: Full text — Methods and Results.

[2] Wei, X., Dong, S., Su, Z., Tang, L., Zhao, P., Pan, C., Wang, F., Tang, Y., Zhang, W., & Zhang, X. (2023). NetMoST: A network-based machine learning approach for subtyping schizophrenia using polygenic SNP allele biomarkers. arXiv preprint arXiv:2302.00104. Source: wei2023-netmost.md. Supports: NetMoST summary above; see NetMoST for full detail.

[3] Yang, Z., Wen, J., Abdulkadir, A., et al., & Davatzikos, C. (2024). Gene-SGAN: discovering disease subtypes with imaging and genetic signatures via multi-view weakly-supervised deep clustering. Nature Communications, 15, 354. DOI: 10.1038/s41467-023-44271-2. Source: yang2024-gene-sgan.md. Supports: Gene-SGAN summary above; see Gene-SGAN for full subtype breakdown.

[4] Aguirre, M., Tanigawa, Y., Venkataraman, G. R., Tibshirani, R., Hastie, T., & Rivas, M. A. (2021). Polygenic risk modeling with latent trait-related genetic components. European Journal of Human Genetics, 29, 1051–1061. DOI: 10.1038/s41431-021-00813-0. Source: aguirre2021-degas.md. Supports: DeGAs/dPRS summary above; see DeGAs / dPRS for full detail.