报 告 人：李启寨研究员（中国科学院）
Recent advances in high-throughput biotechnologies haveprovided an unprecedented opportunity for biomarker discovery, which, from astatistical point of view, can be cast as a variable selection problem. Thisproblem is challenging due to the high-dimensional and non-linear nature ofomics data and, in general, it suffers three difficulties: (i) an unknownfunctional form of the nonlinear system, (ii) variable selection consistency,and (iii) high-demanding computation. To circumvent the rst difficulty, weemploy a feed-forward neural network to approximate the unknown nonlinearfunction motivated by its universal approximation ability. To circumvent thesecond difficulty, we conduct structure selection for the neural network, whichinduces variable selection, by choosing appropriate prior distributions thatlead to the consistency of variable selection. To circumvent the thirddifficulty, we implement the population stochastic approximation Monte Carloalgorithm, a parallel adaptive Markov Chain Monte Carlo (MCMC) algorithm, onthe OpenMP platform which provides a linear speedup for the simulation. Thenumerical results indicate that the proposed method can execute very fast on amulticore computer and work very well for identi cation of relevant variables forgeneral high-dimensional nonlinear systems. The proposed method is successfullyapplied to selection of anticancer drug response genes for the drug sensitivitydata collected in the cancer cell line encyclopedia (CCLE) study.