Massively Parallel Mapping of Next Generation Sequence Reads Using GPUs

DNA sequence alignment problem can be broadly defined as the character-level comparison of DNA sequences obtained from one or more samples against a database of reference (i.e., consensus) genome sequence of the same or a similar species. High throughput sequencing (HTS) technologies were introduced in 2006 [6], and the latest iterations of HTS technologies are able to read the genome of a human individual in just three days for a cost of ∼ $1,000. However, they also present a computational problem since the analysis of the HTS data requires the comparison of >1 billion short (100 characters, or base pairs) “reads” against a very long (3 billion base pairs) reference genome. Since DNA molecules are composed of two opposing strands (i.e. two complementary strings), the number of required comparisons are doubled. Instead of local alignment of short vs long sequences, heuristics are applied to speed up the process. First, partial sequence matches, called “seeds”, are quickly found using either Burrows Wheeler Transform (BWT) [1] followed with Ferragina-Manzini Index (FM) [2], or a simple hash table [8]. Next, the candidate locations are verified using a dynamic programming alignment algorithm that calculates Levenshtein edit distance [3], which runs in quadratic time. Although these heuristics are substantially faster than local alignment, because of the repetitive nature of the human genome, they often require hundreds of verification runs per read, imposing a heavy computational burden. However, all of these billions of alignments are independent from each other, thus the read mapping problem presents itself as embarrassingly parallel. Our goal in this project is to develop and implement a GPGPUfriendly algorithm based on Levenshtein’s algorithm [3] that can compute millions of dynamic programming matrices concurrently. We implement our algorithms using the CUDA (Compute Unified Device Architecture) platform, and test them using the NVIDIA Tesla K20 GPGPU processors. In this work, we propose a massively parallel, fast, memory-aware Levenshtein edit distance algorithm model for graphics processing units, together with Ukkonen’s approximation algorithm [7] to prevent redundant calculations in matrices. Considering the memory limitations and very high number of available threads, our algorithm ensures maximum occupancy on GPGPUs. 2. Background