ADAMIXTURE: adaptive first-order optimization for biobank-scale genetic clustering.

Area of Science:

• Population genetics
• Computational biology
• Bioinformatics

Background:

• Estimating genetic clusters is crucial for demographic inference and adjusting for population structure in genetic association studies.
• The widely used ADMIXTURE software, based on the Expectation-Maximization (EM) algorithm, faces scalability challenges with large biobank-sized datasets.
• Existing acceleration methods for EM are computationally intensive, while faster EM-free methods may sacrifice accuracy.

Purpose of the Study:

• To develop a novel optimization framework, ADAMIXTURE, that enhances the computational efficiency and scalability of model-based population genetic clustering.
• To integrate the EM algorithm with Adaptive Moment Estimation (Adam) to overcome the limitations of existing methods.

Main Methods:

• Introduced ADAMIXTURE, a framework combining the EM algorithm with Adaptive Moment Estimation (Adam).
• Utilized first-order gradients with adaptive learning rates to approximate curvature, avoiding computationally expensive Hessian approximations.
• Developed a GPU implementation for ADAMIXTURE.

Main Results:

• ADAMIXTURE demonstrated superior convergence efficiency compared to second-order methods while retaining low computational complexity.
• Achieved substantial reductions in wall-clock runtime and improved scalability on simulated and large-scale empirical datasets.
• The GPU implementation processed half a million samples and variants in under 2 hours, representing a two-order-of-magnitude speedup.

Conclusions:

• ADAMIXTURE offers a computationally efficient and scalable solution for population genetic clustering.
• The method maintains or improves inference accuracy compared to state-of-the-art approaches.
• ADAMIXTURE significantly advances the analysis of large-scale genomic datasets.