High performance graph analysis on parallel architectures

Abstract

Over the last decade pharmacology has been developing computational methods to enhance drug development and testing. A computational method called network pharmacology uses graph analysis tools to determine protein target sets that can lead on better targeted drugs for diseases as Cancer. One promising area of network-based pharmacology is the detection of protein groups that can produce better e ects if they are targeted together by drugs. However, the e cient prediction of such protein combinations is still a bottleneck in the area of computational biology. The computational burden of the algorithms used by such protein prediction strategies to characterise the importance of such proteins consists an additional challenge for the eld of network pharmacology. Such computationally expensive graph algorithms as the all pairs shortest path (APSP) computation can a ect the overall drug discovery process as needed network analysis results cannot be given on time. An ideal solution for these highly intensive computations could be the use of super-computing. However, graph algorithms have datadriven computation dictated by the structure of the graph and this can lead to low compute capacity utilisation with execution times dominated by memory latency. Therefore, this thesis seeks optimised solutions for the real-world graph problems of critical node detection and e ectiveness characterisation emerged from the collaboration with a pioneer company in the eld of network pharmacology as part of a Knowledge Transfer Partnership (KTP) / Secondment (KTS). In particular, we examine how genetic algorithms could bene t the prediction of protein complexes where their removal could produce a more e ective 'druggable' impact. Furthermore, we investigate how the problem of all pairs shortest path (APSP) computation can be bene ted by the use of emerging parallel hardware architectures as GPU- and FPGA- desktop-based accelerators. In particular, we address the problem of critical node detection with the development of a heuristic search method. It is based on a genetic algorithm that computes optimised node combinations where their removal causes greater impact than common impact analysis strategies. Furthermore, we design a general pattern for parallel network analysis on multi-core architectures that considers graph's embedded properties. It is a divide and conquer approach that decomposes a graph into smaller subgraphs based on its strongly connected components and computes the all pairs shortest paths concurrently on GPU. Furthermore, we use linear algebra to design an APSP approach based on the BFS algorithm. We use algebraic expressions to transform the problem of path computation to multiple independent matrix-vector multiplications that are executed concurrently on FPGA. Finally, we analyse how the optimised solutions of perturbation analysis and parallel graph processing provided in this thesis will impact the drug discovery process.