Sparse Linear Algebra Accelerates Genotype Representation Graph Computation at Biobank Scale

Mikado reformulates genotype representation graph computation as sparse triangular solves, enabling principal components on the All of Us cohort in minutes and BOLT-LMM in under 19 minutes on one GPU.

• Why it matters: Biobank-scale analyses are limited by compute cost when hundreds of thousands to millions of samples and variants must be processed together, and GRG traversal has not used modern accelerators.
• What they did: Mikado uses a reverse topological ordering so the GRG adjacency matrix is strictly block-lower-triangular, turning genotype matrix-vector products into pipelined blocked sparse matrix-vector multiplies.
• The result: On All of Us, Mikado computes principal components in minutes where PLINK2 cannot finish one chromosome in 72 hours, and it runs genome-wide BOLT-LMM-inf in under 19 minutes on one GPU.

AI summary of the abstract · may contain errors · not medical advice · read the paper.