Massive parallelization boosts big Bayesian multidimensional scaling.

Area of Science:

• Computational biology
• Statistical inference
• Big data analytics

Background:

• Bayesian multidimensional scaling (MDS) is valuable for analyzing complex scientific data, such as viral spread.
• Computational demands, scaling quadratically with observations, limit its application in big data scenarios.
• Massive parallelization techniques are needed to overcome these computational bottlenecks.

Purpose of the Study:

• To enhance the computational efficiency of Bayesian MDS for big data analysis.
• To apply accelerated Bayesian MDS to infer the global spread of seasonal influenza virus subtypes.
• To introduce MassiveMDS, an open-source library for large-scale Bayesian MDS.

Main Methods:

• Massive parallelization utilizing multi-core CPUs, instruction-level vectorization, and Graphics Processing Units (GPUs).
• Hamiltonian Monte Carlo for fitting the MDS model, achieving over 100-fold speedups.
• Incorporation of a phylogenetic extension to account for viral evolutionary history.

Main Results:

• GPUs accelerate Bayesian MDS calculations, enabling big data applications.
• Analysis of 5392 viral sequences and 14 million pairwise distances revealed influenza subtype H3N2 spreads most effectively.
• The phylogenetic extension adjusted for shared evolutionary history in spread analysis.

Conclusions:

• Optimized Bayesian MDS using GPUs significantly enhances computational speed for large datasets.
• Influenza H3N2 demonstrates the most effective global spread via air traffic.
• The MassiveMDS library facilitates large-scale Bayesian inference and analysis of complex scientific phenomena.