Abstract
Genetic research utilizes a decomposition of trait variances and covariances into genetic and environmental parts. Our software package biMM is a computationally efficient implementation of a bivariate linear mixed model for settings where hundreds of traits have been measured on partially overlapping sets of individuals. Availability Implementation in R freely available at www.iki.fi/mpirinen .
Introduction
In genetic research, determining the genetic basis of traits and their correlations is essential. The Bivariate Linear Mixed Model (LMM) is the standard approach for partitioning variation into genetic and environmental components. However, LMMs suffer from $O(n^3)$ computational complexity relative to the number of individuals ($n$), making them impractical for large cohorts with hundreds of trait pairs when using standard software like GCTA or GEMMA.
Methodology
The biMM R package overcomes these computational barriers through three key strategies:
- Reusing Eigendecompositions: By identifying trait pairs measured on the same individuals, biMM reuses the expensive eigendecomposition of the Genetic Relationship Matrix (GRM), reducing complexity to $O(n^2)$.
- Smart Ordering: biMM treats trait sequencing as a "Shortest Hamiltonian Path" problem, using a greedy heuristic to find an order that maximizes sample overlap and minimizes new decompositions.
- Strategic Approximation: Users can set thresholds to impute or drop a small number of individuals to force sample overlap. This allows a single eigendecomposition to be applied to multiple trait pairs with minimal impact on results.
Performance and Accuracy
The authors evaluated biMM using 120 trait pairs from the Northern Finland Birth Cohort 1966 (NFBC1966), involving approximately 5,000 individuals.
Computational Speedup
The performance gains are dramatic. The approximate version of biMM completed the analysis in 0.05 hours, significantly faster than GEMMA (2.76 hours) and GCTA (21.39 hours).
Accuracy of Approximation
Despite the speed-enhancing approximations, the results for genetic variance ($V_G$) and genetic correlation ($\rho_G$) remain highly accurate. The estimates from the approximate method are nearly identical to the complete bivariate LMM results.
Conclusion
biMM provides a highly efficient solution for large-scale bivariate genetic analysis. By optimizing matrix operations and leveraging sample overlaps, it enables researchers to analyze hundreds of traits in minutes on standard hardware, maintaining the statistical rigor of the gold-standard LMM.
Summary updated