Engineering and Technology Degree Level: PhD Abstract ID# 879 ...
Graduate Category: Engineering and Technology Degree Level: PhD Abstract ID# 879
Parallel method for Accelerated Protein Coordinate Conversion Jaydeep P. Bardhan ([email protected])
Mahsa Bayati ([email protected])
Abstract
Miriam Leeser ([email protected])
Implementation
Proteins contain thousands to millions of atoms. Position represented using one of two methods: Cartesian coordinates Internal coordinates Different steps of protein modeling need one or the other representation Computational intensity and Large amount of simulation data Reasons for motion Thermal motion Atomic charges Influence of water and surrounding ions Forward Goal: rapid, efficient , scalable conversion inherent dependency in data Internal Cartesian structure Data Parallelism relative to neighboring atoms Task Parallelism Reverse Novel method for reverse translation
Background
Reverse -
Local Cartesian coordinate for each residue in protein back bone Merge each local coordinate into one intermediate Coordinate system Find Translation matrix in each merge Forms a reduction tree By proceeding to bottom of tree serial computation increases To prevent serial computation only calculate last residue Compute Cartesian Coordinate of atoms based on Translation matrixes
Results & Future Work Forward
Analysis
Cartesian Coordinate 𝐴𝑡𝑜𝑚𝑖 𝑡 = 𝑥𝑖 𝑡 , 𝑦𝑖 𝑡 , 𝑧𝑖 𝑡 Internal Coordinate - Bond distance
𝑟𝑖,𝑗 =
using CUDA-C with data streaming and overlapping computation , achieved 13-20x times speed up compared to
(𝑋𝑗 − 𝑋𝑖 )2 +(𝑦𝑗 − 𝑦𝑖 )2 +(𝑍𝑗 − 𝑍𝑖 )2
𝑐𝑜𝑠𝜃𝑖𝑗𝑘 =
CUDA-MATLAB: Convenient to use OpenMP: Easy to implement, acceptable speed up but limited potential for parallelization
𝑟𝑖𝑗 ∙𝑟𝑗𝑘 𝑟𝑗𝑖 ∙ 𝑟𝑗𝑘
-Angle
-Dihedral
𝑐𝑜𝑠 𝜑𝑖𝑗𝑘𝑙 = sin 𝜑𝑖𝑗𝑘𝑙 =
𝑟𝑖𝑗 ×𝑟𝑗𝑘
.
𝑟𝑗𝑘 ×𝑟𝑘𝑙
𝑟𝑖𝑗
𝑟𝑗𝑘 sin 𝜑𝑖𝑗𝑘 𝑟𝑗𝑘 𝑟𝑘𝑙 sin 𝜑𝑗𝑘𝑙 𝑟𝑘𝑙 ×𝑟𝑖𝑗 .𝑟𝑗𝑘 .𝑟𝑗𝑘
𝑟𝑗𝑘
𝑟𝑘𝑙 sin 𝜑𝑗𝑘𝑙 𝑟𝑖𝑗
Future Work
𝑟𝑗𝑘 sin 𝜑𝑖𝑗𝑘
Handle larger protein using multiple GPUs Internal to Cartesian conversion Side chains
Improper: all three atoms connected to a central atom
Implementation Forward Protein structure file (.Psf) List of atoms has bonds, angles, dihedrals, impropers, Molecular Dynamic Toolbox to parse protein structure file. Trajectory file (.dcd) binary format collection of atomic coordinates in different time frames
Reverse
Acknowledgment National Institute of General Medical Sciences supports the work of Jaydeep Bardhan (R21GM102642). The content is solely the responsibility of the authors.
Parallel method for Accelerated Protein Coordinate Conversion Jaydeep P. Bardhan ([email protected])
Mahsa Bayati ([email protected])
Abstract
Miriam Leeser ([email protected])
Implementation
Proteins contain thousands to millions of atoms. Position represented using one of two methods: Cartesian coordinates Internal coordinates Different steps of protein modeling need one or the other representation Computational intensity and Large amount of simulation data Reasons for motion Thermal motion Atomic charges Influence of water and surrounding ions Forward Goal: rapid, efficient , scalable conversion inherent dependency in data Internal Cartesian structure Data Parallelism relative to neighboring atoms Task Parallelism Reverse Novel method for reverse translation
Background
Reverse -
Local Cartesian coordinate for each residue in protein back bone Merge each local coordinate into one intermediate Coordinate system Find Translation matrix in each merge Forms a reduction tree By proceeding to bottom of tree serial computation increases To prevent serial computation only calculate last residue Compute Cartesian Coordinate of atoms based on Translation matrixes
Results & Future Work Forward
Analysis
Cartesian Coordinate 𝐴𝑡𝑜𝑚𝑖 𝑡 = 𝑥𝑖 𝑡 , 𝑦𝑖 𝑡 , 𝑧𝑖 𝑡 Internal Coordinate - Bond distance
𝑟𝑖,𝑗 =
using CUDA-C with data streaming and overlapping computation , achieved 13-20x times speed up compared to
(𝑋𝑗 − 𝑋𝑖 )2 +(𝑦𝑗 − 𝑦𝑖 )2 +(𝑍𝑗 − 𝑍𝑖 )2
𝑐𝑜𝑠𝜃𝑖𝑗𝑘 =
CUDA-MATLAB: Convenient to use OpenMP: Easy to implement, acceptable speed up but limited potential for parallelization
𝑟𝑖𝑗 ∙𝑟𝑗𝑘 𝑟𝑗𝑖 ∙ 𝑟𝑗𝑘
-Angle
-Dihedral
𝑐𝑜𝑠 𝜑𝑖𝑗𝑘𝑙 = sin 𝜑𝑖𝑗𝑘𝑙 =
𝑟𝑖𝑗 ×𝑟𝑗𝑘
.
𝑟𝑗𝑘 ×𝑟𝑘𝑙
𝑟𝑖𝑗
𝑟𝑗𝑘 sin 𝜑𝑖𝑗𝑘 𝑟𝑗𝑘 𝑟𝑘𝑙 sin 𝜑𝑗𝑘𝑙 𝑟𝑘𝑙 ×𝑟𝑖𝑗 .𝑟𝑗𝑘 .𝑟𝑗𝑘
𝑟𝑗𝑘
𝑟𝑘𝑙 sin 𝜑𝑗𝑘𝑙 𝑟𝑖𝑗
Future Work
𝑟𝑗𝑘 sin 𝜑𝑖𝑗𝑘
Handle larger protein using multiple GPUs Internal to Cartesian conversion Side chains
Improper: all three atoms connected to a central atom
Implementation Forward Protein structure file (.Psf) List of atoms has bonds, angles, dihedrals, impropers, Molecular Dynamic Toolbox to parse protein structure file. Trajectory file (.dcd) binary format collection of atomic coordinates in different time frames
Reverse
Acknowledgment National Institute of General Medical Sciences supports the work of Jaydeep Bardhan (R21GM102642). The content is solely the responsibility of the authors.
