RNA secondary structure prediction is one of the important research areas in modern bioinformatics and computational biology. PKNOTS is the most famous benchmark program and has been widely used to predict RNA secondary structure including pseudoknots. It adopts the standard 4D dynamic programming method and is the basis of many variants and improved algorithms. Unfortunately, the O(N6) computing requirements and complicated data dependency greatly limits the usefulness of PKNOTS package with the explosion in gene database size. In this paper, we present a fine-grained parallel PKNOTS algorithm and prototype system for accelerating RNA folding application on field programmable gate-array (FPGA) platform. We improved data locality by converting cycle nested relationship and reorganizing computing order of the elements in source code. We aggressively exploit data reuse, data dependency elimination and memory access scheduling strategies to minimize the need for loading data from external memory. To the best of our knowledge, our design is the first FPGA implementation for accelerating 4D dynamic programming problem for RNA folding application including pseudoknots. The experimental results show a factor of more than 11 × average speedup over the PKNOTS-1.05 software running on a PC platform with AMD Phenom 9650 Quad CPU for input RNA sequences. However, the power consumption of our FPGA accelerator is only about 50% of the general-purpose micro-processors.
Dynamic programming algorithms for RNA secondary structure prediction with pseudoknots
