Scalable computation of anisotropic vibrations for large macromolecular assemblies

Area of Science:

• Computational Biology
• Structural Biology
• Biophysics

Background:

• Normal Mode Analysis (NMA) is crucial for understanding macromolecular vibrations and dynamics.
• Analyzing large macromolecules (millions of atoms) is computationally intensive and challenging for existing NMA methods.
• Current NMA approaches are not optimized for parallel computing on Graphics Processing Units (GPUs).

Purpose of the Study:

• To develop a novel, efficient NMA method for large macromolecules.
• To leverage GPU computing for accelerated eigenproblem construction and diagonalization.
• To enable accurate calculation of low-frequency vibrational modes in megascale biological structures.

Main Methods:

• Implemented level-structure bandwidth-reducing algorithms for sparse matrix computations.
• Transformed NMA computations to a globally-sparse-yet-locally-dense format for GPU efficiency.
• Utilized and optimized Krylov-subspace eigensolvers with techniques like Chebyshev filtering and deflation for fast GPU-resident calculations.

Main Results:

• Achieved significant speedups (at least 250x) for calculating vibrational modes of large PDB structures (>2.4 million atoms).
• Enabled accurate computation of the first 1000 vibrational modes on GPUs.
• Demonstrated efficient batched tensor product execution on GPUs through optimized algorithms.

Conclusions:

• The developed GPU-accelerated NMA method overcomes computational bottlenecks for large macromolecules.
• This approach facilitates the study of dynamics in megascale biological assemblies.
• Provides a powerful tool for structural biology and computational biophysics research.