gEMfitter: a highly parallel FFT-based 3D density fitting tool with GPU texture memory acceleration.

Area of Science:

• Structural biology
• Computational biology
• Biophysics

Background:

• Accurate atomic modeling of macromolecular assemblies from cryo-electron microscopy (cryo-EM) data is crucial.
• Fitting high-resolution protein structures into low-resolution cryo-EM density maps presents computational challenges.

Purpose of the Study:

• To develop and present gEMfitter, a novel, highly parallel program for fitting atomic structures into cryo-EM density maps.
• To leverage graphics processor unit (GPU) hardware for accelerated correlation searches in cryo-EM density fitting.

Main Methods:

• Developed gEMfitter, a highly parallel fast Fourier transform (FFT) based program.
• Utilized GPU texture memory for efficient rotation of 3D voxel grids, minimizing rotation costs.
• Implemented a parallel processing framework to scale speed-up with the number of CPUs or GPUs.

Main Results:

• Achieved up to a 26-fold speed-up compared to single-core CPU-based 3D correlation.
• Demonstrated linear increase in speed-up with the number of processing units (CPUs/GPUs).
• Obtained satisfactory fitting results for the GroEL-GroES complex using locally normalized cross-correlation with a Laplacian pre-filter, significantly outperforming existing methods like COLORES.

Conclusions:

• gEMfitter enables routine use of robust 3D correlation techniques for cryo-EM density fitting.
• The GPU-accelerated approach dramatically reduces computation time, making complex modeling feasible.
• This advancement facilitates more accurate and efficient atomic modeling of large biological assemblies.