Parallel Hybrid Genetic Algorithm based Dynamic Load-Balancing for Mining of Molecular Fragments

Recent researches show a tremendous potential of applying in silico methods to the drug discovery processes via Virtual High-Throughput Screening (vHTS). The purpose of vHTS is to employ intensive mining techniques to test and analyze large libraries of chemicals or molecules to find regularities or patterns, such as molecular features that restrain a desired reaction, common substructures responsible for a protein binding etc. among specific classes of molecules in a relatively short period of time. The problem of such pattern mining is characterized by a highly irregular search space in which a reliable estimation of workloads is hard due to its dynamic nature, such as arrivals of new tasks at unpredictable intervals, breakdowns of processingunits, cancelations of tasks etc. Computational complexity of exploring such a search space for a large dataset renders sequential algorithms useless. This work presents a Parallel Hybrid Genetic Algorithm to address the problem of dynamic load-balancing to explore highly irregular search spaces. Basic Genetic Algorithm is hybridized based on local and problem-specific knowledge and reasoning. Two major design goals are considered: 1) optimize solution quality, 2) improve convergence speed. An efficient local search is integrated with Hybrid Genetic Algorithm to optimize the quality of solution produced by global search. Then Hybrid Genetic Algorithm is parallelized to improve the convergence speed. Experimental results on AIDS screening data set of the National Cancer Institute validate the purpose of the proposed algorithms. The performances of the proposed approaches are also compared to that of a Simple Genetic Algorithm solving the same problem