Accelerating Exact Protein Structure Alignment with Graphics Processors

Among different structural alignment tools, DALIX is one capable of calculating an optimal structural alignment based on the DALI score in most cases. It outperforms DALI, one of the most popular structural alignment algorithms, on the alignment quality. However, the high time complexity of DALIX hinders its application to large protein or complex structure alignments. In this paper, we parallelize the major steps of DALIX on the GPU (Graphics Processing Units) to speed up its processing. Specifically, to better utilize the massive GPU thread parallelism, we design a two-level parallel algorithm for the dynamic programming, which is the most time-consuming component in the tool. We compact the decision table in the dynamic programming so that it can fit into the shared memory for inter-thread communication to further improve the performance. Results show that our GPU-DALIX achieves a speedup ranging from 5.5x to 20x, over the sequential version of DALIX on a set of real-world protein alignments. Especially, our GPU-DALIX provides significant performance improvement when the protein size is large or the structure is complex.