High‐performance transformation of protein structure representation from internal to Cartesian coordinates

Abstract

We present a highly parallel algorithm to convert internal coordinates of a polymeric molecule into Cartesian coordinates. Traditionally, converting the structures of polymers (e.g., proteins) from internal to Cartesian coordinates has been performed serially, due to an inherent linear dependency along the polymer chain. We show this dependency can be removed using a tree‐based concatenation of coordinate transforms between segments, and then parallelized efficiently on graphics processing units (GPUs). The conversion algorithm is applicable to protein engineering and fitting protein structures to experimental data, and we observe an order of magnitude speedup using parallel processing on a GPU compared to serial execution on a CPU.

MP-NeRF: A Massively Parallel Method for Accelerating Protein Structure Reconstruction from Internal Coordinates

“…Although the calculation can be parallelized across many polymers of similar length, this bottleneck is significant for applications that make intensive use of forward and reverse translation. Examples of such applications include the training of machine learning models 4 , protein structure refinement from NMR data 1 , analysis of protein structure changes 3 , and molecular dynamics simulations 5 .…”

“…However, it considered only the backbone, and the number of parallelized fragments was low (on the order of 1-10). A more recent implementation of a parallel NeRF algorithm 5 considered the sidechains as well, and performed a tree-based merge of the different fragments. However, it required specialized hardware such as CUDA-capable GPUs and there is friction when trying to adapt the implementation for different usecases.…”

“…With the development of more and better computational tools, some effort has been devoted in recent years to alleviating the reverse translation bottleneck 4,5 . These works have focused on the usage of high-performance, optimized code; the division of the backbone into different fragments (which are folded independently and later ensembled) 4 ; and tree ensembling algorithms 5 , among other strategies. These approaches, however, are often implemented specifically for Graphical Processing Units (GPUs), thus limiting usage to expensive and specialized hardware accelerators.…”

“…We introduce a massively-parallel (mp-NeRF) algorithm, which reorders the steps in the parallel implementations of the NeRF algorithm to unlock parallelization and optimize execution when possible. The algorithm leverages ideas from previous work 4,5 , and builds on top of them to provide an acceleration of over 400x-1200x, depending on the length of the protein being translated. It accomplishes this by decomposing operations into parallelizable units and using more efficient data structures that leverage the efficiency of matrix-based operations.…”

MP‐NeRF: A massively parallel method for accelerating protein structure reconstruction from internal coordinates

“…Examples of such applications include the training of machine learning models, 4 protein structure refinement from NMR data, 1 analysis of protein structure changes, 3 and molecular dynamics simulations. 5 With the development of more and better computational tools, some effort has been devoted in recent years to alleviating the reverse translation bottleneck. 4,5 These works have focused on the usage of high-performance, optimized code; the division of the backbone into different fragments (which are folded independently and later assembled) 4 ; and tree ensembling algorithms, 5 among other strategies.…”