Fig. 6: T1GRS reveals clinically relevant genetic subtypes of T1D.
From: Genetic association and machine learning improve the prediction of type 1 diabetes risk
a, ScanPy-derived UMAP of SHAP values from the discovery cohort. Unsupervised clustering resulted in four distinct genetic clusters, determined based on optimal silhouette and inertia scores (see the plot below; dashed line indicates optimal cluster number). T1D individuals from the validation cohort in AoU were projected onto these existing clusters, with the centroid of individuals in each cluster shown. The membership proportion of T1D patients for each cluster is shown on the right. b, Per-cluster log2 fold-change of mean SHAP values relative to other clusters for candidate MHC (top) and non-MHC (middle) loci. P values were calculated by a two-sided t test. Cluster IDs are shown in the bottom annotation. Enrichment of cell-type regulatory elements for high-feature-importance loci in each cluster. P values were calculated by a permutation test. c, Total SHAP values for T1D individuals in discovery cohorts, broken down by features assigned to the MHC locus or specific cell types, split by cluster. d, Cumulative incidence of T1D for each cluster by age. P values calculated by log-rank test: cluster 1 P = 1.13 × 10−19, cluster 2 P = 0.777, cluster 3 = 4.31 × 10−6 and cluster 4 P = 3.31 × 10−17. Dashed lines indicate the median age of onset for each cluster. e, OR of clinical complications for the discovery dataset and AoU validation cohort split by cluster. P values were calculated by a two-sided Cox proportional hazards test. Error bars represent the 95% confidence intervals. Avg., average; CI, confidence interval; panc., pancreas.
