Poster
166
(#166) Diffusion Maps: A Novel Approach to GWAS
Psych Congress 2025
Abstract: Objective: Genome-wide association studies (GWAS) underpin modern psychiatric genetics and biomarker discovery, yet their accuracy hinges on rigorous control of population structure. Principal components analysis (PCA), the current standard, captures only linear axes and can obscure fine-scale ancestry, inflating test statistics or masking true signals-especially in genetically diverse cohorts.
Methods: We applied diffusion maps (DM), a nonlinear manifold-learning technique, to a population-scale biobank (~270,000 UK Biobank participants). DM eigenvectors replaced top PCs in linear mixed-model GWAS of neuropsychiatric phenotypes (Alzheimer's disease, depression, schizophrenia) plus a geography-derived null phenotype (eastings coordinates). Performance was benchmarked against PCA using genomic-inflation metrics, LD score regression intercepts, QQ-plot behavior, recovery of canonical loci (e.g., APOE), and workflow runtime on multi-GPU hardware.
Results: Across multiple phenotypes, DM yielded visibly tighter QQ plots, lower inflation statistics, and intercepts trending closer to theoretical expectations. Canonical risk loci remained robust, while some biologically plausible association clusters-previously borderline under PCA-became clearer after DM adjustment. DM integrated efficiently into existing analytic pipelines, completing within standard GWAS runtimes without extensive parameter tuning.
Implications: By modeling non-linear genetic manifold structure, diffusion maps can reduce bias and uncover under-appreciated variation that informs psychiatric and Alzheimer's research. Early findings suggest DM offers a practical, geometry-aware improvement compared to PCA, with the potential to sharpen risk prediction, refine genetic architecture, and accelerate precision neuropsychiatry and therapeutic discovery in diverse populations.
Short Description: Diffusion maps, a nonlinear manifold-learning method, was tested as a replacement for principal components in GWAS of Alzheimer's disease, depression, and schizophrenia in ~270,000 UK Biobank participants. The approach produced cleaner QQ plots, lower genomic inflation, and clearer association clusters than PCA while preserving established loci such as APOE. The poster details methodology, performance benchmarks, practical guidance, and theoretical considerations for integrating diffusion maps into high-throughput psychiatric-genetics pipelines.
Name of Sponsoring Organization(s):
Methods: We applied diffusion maps (DM), a nonlinear manifold-learning technique, to a population-scale biobank (~270,000 UK Biobank participants). DM eigenvectors replaced top PCs in linear mixed-model GWAS of neuropsychiatric phenotypes (Alzheimer's disease, depression, schizophrenia) plus a geography-derived null phenotype (eastings coordinates). Performance was benchmarked against PCA using genomic-inflation metrics, LD score regression intercepts, QQ-plot behavior, recovery of canonical loci (e.g., APOE), and workflow runtime on multi-GPU hardware.
Results: Across multiple phenotypes, DM yielded visibly tighter QQ plots, lower inflation statistics, and intercepts trending closer to theoretical expectations. Canonical risk loci remained robust, while some biologically plausible association clusters-previously borderline under PCA-became clearer after DM adjustment. DM integrated efficiently into existing analytic pipelines, completing within standard GWAS runtimes without extensive parameter tuning.
Implications: By modeling non-linear genetic manifold structure, diffusion maps can reduce bias and uncover under-appreciated variation that informs psychiatric and Alzheimer's research. Early findings suggest DM offers a practical, geometry-aware improvement compared to PCA, with the potential to sharpen risk prediction, refine genetic architecture, and accelerate precision neuropsychiatry and therapeutic discovery in diverse populations.
Short Description: Diffusion maps, a nonlinear manifold-learning method, was tested as a replacement for principal components in GWAS of Alzheimer's disease, depression, and schizophrenia in ~270,000 UK Biobank participants. The approach produced cleaner QQ plots, lower genomic inflation, and clearer association clusters than PCA while preserving established loci such as APOE. The poster details methodology, performance benchmarks, practical guidance, and theoretical considerations for integrating diffusion maps into high-throughput psychiatric-genetics pipelines.
Name of Sponsoring Organization(s):
