Protein-protein interaction network alignment using GPU

Abstract

The alignment of Protein-Protein interaction Networks is becoming an imperative phenomenon in Bio-informatics that leads to several vital results. These results can be used in numerous fields associated with Bio-informatics including the prediction/variation of evolutionary relationships finding cures for gene inflicted diseases (like cancer) and identifying probable therapies. However with the introduction of fast sequencing and other technologies that spawn large amounts of data for computing (since the proteins are very large in size and have many nodes and edges) limiting dynamics arise. These include performance scalability and time consumption. Recently CPU versions of the alignment procedures and computations have been introduced. However because of the large size of the proteins they are very time-consuming. Therefore in this thesis i propose a GPU version for performing the computations quickly and efficiently. This thesis is based on improving the efficiency of SPiNAL a polynomial time heuristic algorithm introduced by [1] that finds the similarities between pairs of PPi-Networks. in this thesis the sequential algorithm of SPiNAL is converted into a parallel algorithm using Heterogeneous Programming Library (HPL) that performs the computations in a massively parallel fashion on a single GPU with 448 thread processors a clock rate of 1.15 Giga Hertz and 6 Giga Bytes of DRAM. The modifications/enhancements to the algorithm result in a significant speedup as compared to the benchmark algorithms.

The alignment of Protein-Protein Interaction Networks is becoming an imperative phenomenon in Bio-Informatics that leads to several vital results. These results can be used in numerous fields associated with Bio-Informatics including the prediction/variation of evolutionary relationships, finding cures for gene inflicted diseases (like cancer) and identifying probable therapies. However, with the introduction of fast sequencing and other technologies that spawn large amounts of data for computing (since the proteins are very large in size and have many nodes and edges), limiting dynamics arise. These include performance, scalability and time consumption. Recently, CPU versions of the alignment procedures and computations have been introduced. However, because of the large size of the proteins, they are very time-consuming. Therefore, in this thesis, I propose a GPU version for performing the computations quickly and efficiently. This thesis is based on improving the efficiency of SPINAL, a polynomial time heuristic algorithm introduced by [1] that finds the similarities between pairs of PPI-Networks. In this thesis, the sequential algorithm of SPINAL is converted into a parallel algorithm using Heterogeneous Programming Library (HPL) that performs the computations in a massively parallel fashion on a single GPU with 448 thread processors, a clock rate of 1.15 Giga Hertz and 6 Giga Bytes of DRAM. The modifications/enhancements to the algorithm result in a significant speedup as compared to the benchmark algorithms.